Update external/libyuv to r1837

Update external/libyuv to version r1837 (d53f1bee) from upstream.

This brings in some new functions that will be used by libavif.

Also update README.version with local modifications. The files/
subdirectory is a pristine copy of the upstream checkout except
for the local modifications that are listed in README.version.

Bug: b/241008246
Bug: b/228492909
Test: Builds. Media and Camera CTS tests pass.
Merged-In: I1bd041e475666e13fafd8c3792d7142a022d435a
Change-Id: I85df21181e582ff835aab1e9ef13ad3ac8421d30

189 files changed, 59346 insertions, 26803 deletions

diff --git a/METADATA b/METADATA
index d97975ca..bff062d8 100644
--- a/METADATA
+++ b/METADATA
@@ -1,3 +1,14 @@
+name: "libyuv"
+description:
+    "libyuv is an open source project that includes YUV scaling and conversion "
+    "functionality."
+
 third_party {
+  url {
+    type: GIT
+    value: "https://chromium.googlesource.com/libyuv/libyuv/"
+  }
+  version: "d53f1beecdd8d959f7a3f2e19bd0bd7e7227a233"
+  last_upgrade_date { year: 2022 month: 8 day: 5 }
   license_type: NOTICE
 }
diff --git a/README.version b/README.version
index 0e74ad15..5deb188e 100644
--- a/README.version
+++ b/README.version
@@ -1,3 +1,8 @@
-Version: r1732
+Version: r1837
 BugComponent: 42195
 Owner: lajos
+Local Modifications:
+ * Remove files/Android.mk (it messes with the android build system).
+ * Remove OWNERS files within files/ and all the subdirectories (except for
+   files/fuzz). Having these files breaks repo presubmit hooks since they
+   contain non @google.com email addresses.
diff --git a/UPDATING b/UPDATING
new file mode 100644
index 00000000..2679284c
--- /dev/null
+++ b/UPDATING
@@ -0,0 +1,36 @@
+To sync the libyuv checkout to an upstream revision, do the following:
+
+These commands are known to work from the external/libyuv directory of the
+Android tree's checkout.
+
+Step 1: Remove the files/ subdirectory.
+
+$ rm -rf files
+
+Step 2: Clone the libyuv repository from upstream.
+
+$ git clone https://chromium.googlesource.com/libyuv/libyuv files
+
+Step 3 (optional): Checkout a specific commit/tag.
+
+$ cd files
+$ git checkout <commit_or_tag>
+$ cd ..
+
+Step 4: Remove files that aren't necessary (Android.mk, .git and OWNERS).
+
+$ rm files/Android.mk
+$ rm -rf files/.git
+$ find files/ -name "OWNERS" | xargs rm
+
+Step 5: Update the version and last_upgrade_date fields in the METADATA file.
+
+Step 6: Update README.version with the version (can be found in
+        files/include/libyuv/version.h)
+
+Step 7: If any local modifications are being done, update README.version and
+        this file with updated instructions.
+
+Step 8: Ensure that libyuv builds and camera and media related CTS tests are
+        passing. If there are any linker errors about missing symbols, try
+	updating frameworks/av/media/libstagefright/export.lds.
diff --git a/files/.clang-format b/files/.clang-format
new file mode 100644
index 00000000..59d48705
--- /dev/null
+++ b/files/.clang-format
@@ -0,0 +1,6 @@
+# Defines the Chromium style for automatic reformatting.
+# http://clang.llvm.org/docs/ClangFormatStyleOptions.html
+BasedOnStyle: Chromium
+---
+Language: Java
+BasedOnStyle: Google
diff --git a/files/.gitignore b/files/.gitignore
index 711f09e0..20d679b7 100644
--- a/files/.gitignore
+++ b/files/.gitignore
@@ -1,86 +1,29 @@
 *.pyc
+.landmines
 pin-log.txt
 /base
 /build
 /buildtools
-/chromium/.gclient.tmp
-/chromium/.gclient.tmp_entries
-/chromium/.last_sync_chromium
-/chromium/src/
 /google_apis
 /links
 /links.db
+/ios
 /mojo
 /native_client
 /net
 /out
+/unit_test/out
+/source/out
 /sde-avx-sse-transition-out.txt
 /testing
-/third_party/android_platform
-/third_party/android_tools
-/third_party/appurify-python
-/third_party/asan
-/third_party/ashmem
-/third_party/binutils
-/third_party/BUILD.gn
-/third_party/catapult
-/third_party/drmemory
-/third_party/gflags/src
-/third_party/icu
-/third_party/ijar
-/third_party/instrumented_libraries
-/third_party/jsr-305
-/third_party/junit
-/third_party/libjpeg
-/third_party/libjpeg_turbo
-/third_party/libxml
-/third_party/llvm
-/third_party/llvm-build
-/third_party/lss
-/third_party/mockito
-/third_party/modp_b64
-/third_party/protobuf
-/third_party/requests
-/third_party/robolectric
-/third_party/WebKit
-/third_party/yasm
-/tools/android
-/tools/clang
-/tools/generate_library_loader
-/tools/gn
-/tools/grit
-/tools/gritsettings/README
-/tools/gritsettings/resource_ids
-/tools/gyp
-/tools/isolate_driver.py
-/tools/memory
-/tools/protoc_wrapper
-/tools/python
-/tools/sanitizer_options
-/tools/swarming_client
-/tools/tsan_suppressions
-/tools/valgrind
-/tools/valgrind-libyuv/libyuv_tests.bat
-/tools/valgrind-libyuv/libyuv_tests.py
-/tools/valgrind-libyuv/libyuv_tests.sh
-/tools/valgrind-libyuv/memcheck/OWNERS
-/tools/valgrind-libyuv/memcheck/PRESUBMIT.py
-/tools/valgrind-libyuv/memcheck/suppressions.txt
-/tools/valgrind-libyuv/memcheck/suppressions_mac.txt
-/tools/valgrind-libyuv/memcheck/suppressions_win32.txt
-/tools/valgrind-libyuv/tsan/OWNERS
-/tools/valgrind-libyuv/tsan/PRESUBMIT.py
-/tools/valgrind-libyuv/tsan/suppressions.txt
-/tools/valgrind-libyuv/tsan/suppressions_mac.txt
-/tools/valgrind-libyuv/tsan/suppressions_win32.txt
-/tools/vim
-/tools/win
+/third_party
+/tools
 
 # Files generated by CMake build
 cmake_install.cmake
 CMakeCache.txt
 CMakeFiles/
-convert
+yuvconvert
 libgtest.a
 libyuv.a
 libyuv_unittest
diff --git a/files/.gn b/files/.gn
index 63dad32d..a765caa5 100644
--- a/files/.gn
+++ b/files/.gn
@@ -6,9 +6,15 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
+import("//build/dotfile_settings.gni")
+
 # The location of the build configuration file.
 buildconfig = "//build/config/BUILDCONFIG.gn"
 
+# The python interpreter to use by default. On Windows, this will look
+# for python3.exe and python3.bat.
+script_executable = "python3"
+
 # The secondary source root is a parallel directory tree where
 # GN build files are placed when they can not be placed directly
 # in the source tree, e.g. for third party source trees.
@@ -23,24 +29,12 @@ check_targets = [ "//libyuv/*" ]
 # These are the list of GN files that run exec_script. This whitelist exists
 # to force additional review for new uses of exec_script, which is strongly
 # discouraged except for gypi_to_gn calls.
-exec_script_whitelist = [
-  "//build/config/BUILD.gn",
-  "//build/config/android/BUILD.gn",
-  "//build/config/android/config.gni",
-  "//build/config/android/internal_rules.gni",
-  "//build/config/android/rules.gni",
-  "//build/config/compiler/BUILD.gn",
-  "//build/config/gcc/gcc_version.gni",
-  "//build/config/ios/ios_sdk.gni",
-  "//build/config/linux/BUILD.gn",
-  "//build/config/linux/pkg_config.gni",
-  "//build/config/mac/mac_sdk.gni",
-  "//build/config/posix/BUILD.gn",
-  "//build/config/sysroot.gni",
-  "//build/config/win/visual_studio_version.gni",
-  "//build/gn_helpers.py",
-  "//build/gypi_to_gn.py",
-  "//build/toolchain/gcc_toolchain.gni",
-  "//build/toolchain/mac/BUILD.gn",
-  "//build/toolchain/win/BUILD.gn",
-]
+exec_script_whitelist = build_dotfile_settings.exec_script_whitelist +
+                        [ "//build_overrides/build.gni" ]
+
+default_args = {
+  mac_sdk_min = "10.12"
+
+  # https://bugs.chromium.org/p/libyuv/issues/detail?id=826
+  ios_deployment_target = "10.0"
+}
diff --git a/files/.vpython b/files/.vpython
new file mode 100644
index 00000000..4a64fd21
--- /dev/null
+++ b/files/.vpython
@@ -0,0 +1,52 @@
+# This is a vpython "spec" file.
+#
+# It describes patterns for python wheel dependencies of the python scripts in
+# the chromium repo, particularly for dependencies that have compiled components
+# (since pure-python dependencies can be easily vendored into third_party).
+#
+# When vpython is invoked, it finds this file and builds a python VirtualEnv,
+# containing all of the dependencies described in this file, fetching them from
+# CIPD (the "Chrome Infrastructure Package Deployer" service). Unlike `pip`,
+# this never requires the end-user machine to have a working python extension
+# compilation environment. All of these packages are built using:
+#   https://chromium.googlesource.com/infra/infra/+/master/infra/tools/dockerbuild/
+#
+# All python scripts in the repo share this same spec, to avoid dependency
+# fragmentation.
+#
+# If you have depot_tools installed in your $PATH, you can invoke python scripts
+# in this repo by running them as you normally would run them, except
+# substituting `vpython` instead of `python` on the command line, e.g.:
+#   vpython path/to/script.py some --arguments
+#
+# Read more about `vpython` and how to modify this file here:
+#   https://chromium.googlesource.com/infra/infra/+/master/doc/users/vpython.md
+
+python_version: "2.7"
+
+# Used by:
+#   third_party/catapult
+wheel: <
+  name: "infra/python/wheels/psutil/${platform}_${py_python}_${py_abi}"
+  version: "version:5.2.2"
+>
+
+# Used by:
+#   third_party/catapult
+wheel: <
+  name: "infra/python/wheels/pypiwin32/${vpython_platform}"
+  version: "version:219"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+>
+
+# Used by:
+#   build/android
+wheel: <
+  name: "infra/python/wheels/requests-py2_py3"
+  version: "version:2.13.0"
+>
diff --git a/files/.vpython3 b/files/.vpython3
new file mode 100644
index 00000000..0a9aa38b
--- /dev/null
+++ b/files/.vpython3
@@ -0,0 +1,405 @@
+# This is a vpython "spec" file.
+#
+# It describes patterns for python wheel dependencies of the python scripts in
+# the chromium repo, particularly for dependencies that have compiled components
+# (since pure-python dependencies can be easily vendored into third_party).
+#
+# When vpython is invoked, it finds this file and builds a python VirtualEnv,
+# containing all of the dependencies described in this file, fetching them from
+# CIPD (the "Chrome Infrastructure Package Deployer" service). Unlike `pip`,
+# this never requires the end-user machine to have a working python extension
+# compilation environment. All of these packages are built using:
+#   https://chromium.googlesource.com/infra/infra/+/main/infra/tools/dockerbuild/
+#
+# All python scripts in the repo share this same spec, to avoid dependency
+# fragmentation.
+#
+# If you have depot_tools installed in your $PATH, you can invoke python scripts
+# in this repo by running them as you normally would run them, except
+# substituting `vpython` instead of `python` on the command line, e.g.:
+#   vpython path/to/script.py some --arguments
+#
+# Read more about `vpython` and how to modify this file here:
+#   https://chromium.googlesource.com/infra/infra/+/main/doc/users/vpython.md
+
+python_version: "3.8"
+
+# The default set of platforms vpython checks does not yet include mac-arm64.
+# Setting `verify_pep425_tag` to the list of platforms we explicitly must support
+# allows us to ensure that vpython specs stay mac-arm64-friendly
+verify_pep425_tag: [
+    {python: "cp38", abi: "cp38", platform: "manylinux1_x86_64"},
+    {python: "cp38", abi: "cp38", platform: "linux_arm64"},
+
+    {python: "cp38", abi: "cp38", platform: "macosx_10_10_intel"},
+    {python: "cp38", abi: "cp38", platform: "macosx_11_0_arm64"},
+
+    {python: "cp38", abi: "cp38", platform: "win32"},
+    {python: "cp38", abi: "cp38", platform: "win_amd64"}
+]
+
+# Used by:
+#   build/android/pylib/local/emulator/avd.py
+#   components/policy/test_support/policy_testserver.py
+wheel: <
+  name: "infra/python/wheels/protobuf-py2_py3"
+  version: "version:3.15.8"
+>
+
+# TODO(https://crbug.com/898348): Add in necessary wheels as Python3 versions
+# become available.
+wheel: <
+  name: "infra/python/wheels/six-py2_py3"
+  version: "version:1.15.0"
+>
+
+# Common utilities.
+# Use the same versions specified by //third_party/catapult/.vpython3 so that
+# Chromium tests using Telemetry function properly.
+wheel: <
+  name: "infra/python/wheels/numpy/${vpython_platform}"
+  version: "version:1.20.3"
+  # A newer version of numpy is required on ARM64, but it breaks older OS versions.
+  not_match_tag <
+    platform: "macosx_11_0_arm64"
+  >
+>
+wheel: <
+  name: "infra/python/wheels/numpy/mac-arm64_cp38_cp38"
+  version: "version:1.21.1"
+  match_tag <
+    platform: "macosx_11_0_arm64"
+  >
+>
+wheel: <
+  name: "infra/python/wheels/psutil/${vpython_platform}"
+  version: "version:5.8.0.chromium.2"
+>
+wheel: <
+  name: "infra/python/wheels/requests-py2_py3"
+  version: "version:2.26.0"
+>
+
+# Used by various python unit tests.
+wheel: <
+  name: "infra/python/wheels/mock-py2_py3"
+  version: "version:2.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/parameterized-py2_py3"
+  version: "version:0.7.1"
+>
+wheel: <
+  name: "infra/python/wheels/pbr-py2_py3"
+  version: "version:3.0.0"
+>
+
+wheel: <
+  name: "infra/python/wheels/pyfakefs-py2_py3"
+  version: "version:3.7.2"
+>
+
+# Used by:
+#   build/chromeos/test_runner.py
+wheel: <
+  name: "infra/python/wheels/jsonlines-py2_py3"
+  version: "version:1.2.0"
+>
+wheel: <
+  name: "infra/python/wheels/python-dateutil-py2_py3"
+  version: "version:2.7.3"
+>
+
+# Used by WPT importer
+wheel: <
+  name: "infra/python/wheels/charset_normalizer-py3"
+  version: "version:2.0.4"
+>
+wheel: <
+  name: "infra/python/wheels/pyasn1-py2_py3"
+  version: "version:0.4.5"
+>
+wheel: <
+  name: "infra/python/wheels/pyasn1_modules-py2_py3"
+  version: "version:0.2.4"
+>
+wheel: <
+  name: "infra/python/wheels/rsa-py2_py3"
+  version: "version:3.4.2"
+>
+wheel: <
+  name: "infra/python/wheels/cachetools-py2_py3"
+  version: "version:2.0.1"
+>
+wheel: <
+  name: "infra/python/wheels/uritemplate-py2_py3"
+  version: "version:3.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-auth-py2_py3"
+  version: "version:1.25.0"
+>
+wheel: <
+  name: "infra/python/wheels/googleapis-common-protos-py2_py3"
+  version: "version:1.52.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-api-core-py2_py3"
+  version: "version:1.25.1"
+>
+wheel: <
+  name: "infra/python/wheels/google-auth-httplib2-py2_py3"
+  version: "version:0.1.0"
+>
+wheel: <
+  name: "infra/python/wheels/google-api-python-client-py3"
+  version: "version:2.2.0"
+>
+wheel: <
+  name: "infra/python/wheels/oauth2client-py2_py3"
+  version: "version:3.0.0"
+>
+
+# Used by Web Platform Tests (WPT) codebase in
+# //third_party/blink/web_tests/external/wpt/tools/
+wheel: <
+  name: "infra/python/wheels/html5lib-py2_py3"
+  version: "version:1.0.1"
+>
+wheel: <
+  name: "infra/python/wheels/mozdebug-py2_py3"
+  version: "version:0.2"
+>
+wheel: <
+  name: "infra/python/wheels/mozinfo-py2_py3"
+  version: "version:1.2.2"
+>
+wheel: <
+  name: "infra/python/wheels/mozlog-py2_py3"
+  version: "version:7.1.0"
+>
+wheel: <
+  name: "infra/python/wheels/mozprocess-py2_py3"
+  version: "version:1.2.1"
+>
+wheel: <
+  name: "infra/python/wheels/urllib3-py2_py3"
+  version: "version:1.24.3"
+>
+wheel: <
+  name: "infra/python/wheels/blessings-py2_py3"
+  version: "version:1.7"
+>
+wheel: <
+  name: "infra/python/wheels/mozfile-py2_py3"
+  version: "version:2.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/mozterm-py2_py3"
+  version: "version:1.0.0"
+>
+wheel: <
+  name: "infra/python/wheels/webencodings-py2_py3"
+  version: "version:0.5.1"
+>
+wheel: <
+  name: "infra/python/wheels/certifi-py2_py3"
+  version: "version:2020.11.8"
+>
+wheel: <
+  name: "infra/python/wheels/chardet-py2_py3"
+  version: "version:3.0.4"
+>
+wheel: <
+  name: "infra/python/wheels/idna-py2_py3"
+  version: "version:2.8"
+>
+wheel: <
+  name: "infra/python/wheels/distro-py2_py3"
+  version: "version:1.4.0"
+>
+wheel: <
+  name: "infra/python/wheels/pillow/linux-amd64_cp38_cp38"
+  version: "version:8.1.2"
+>
+wheel: <
+  name: "infra/python/wheels/aioquic/${vpython_platform}"
+  version: "version:0.9.15"
+>
+wheel: <
+  name: "infra/python/wheels/pylsqpack/${vpython_platform}"
+  version: "version:0.3.12"
+>
+wheel: <
+  name: "infra/python/wheels/cryptography/${vpython_platform}"
+  version: "version:3.3.1.chromium.1"
+>
+wheel: <
+  name: "infra/python/wheels/cffi/${vpython_platform}"
+  version: "version:1.14.5"
+>
+wheel: <
+  name: "infra/python/wheels/pycparser-py2_py3"
+  version: "version:2.19"
+>
+
+# Used by:
+#   chrome/test/chromedriver/test/run_webdriver_tests.py
+wheel: <
+  name: "infra/python/wheels/iniconfig-py3"
+  version: "version:1.1.1"
+>
+
+wheel: <
+  name: "infra/python/wheels/packaging-py2_py3"
+  version: "version:16.8"
+>
+
+wheel: <
+  name: "infra/python/wheels/pyparsing-py2_py3"
+  version: "version:2.4.7"
+>
+
+wheel: <
+  name: "infra/python/wheels/toml-py3"
+  version: "version:0.10.1"
+>
+
+wheel <
+  name: "infra/python/wheels/pytest-py3"
+  version: "version:6.2.2"
+>
+
+wheel <
+  name: "infra/python/wheels/pytest-asyncio-py3"
+  version: "version:0.14.0"
+>
+
+wheel <
+  name: "infra/python/wheels/attrs-py2_py3"
+  version: "version:20.3.0"
+>
+
+wheel <
+  name: "infra/python/wheels/six-py2_py3"
+  version: "version:1.15.0"
+>
+
+wheel <
+  name: "infra/python/wheels/more-itertools-py2_py3"
+  version: "version:4.1.0"
+>
+
+wheel <
+  name: "infra/python/wheels/pluggy-py3"
+  version: "version:0.13.1"
+>
+
+wheel <
+  name: "infra/python/wheels/py-py2_py3"
+  version: "version:1.10.0"
+>
+
+wheel <
+  name: "infra/python/wheels/funcsigs-py2_py3"
+  version: "version:1.0.2"
+>
+
+wheel: <
+  name: "infra/python/wheels/atomicwrites-py2_py3"
+  version: "version:1.3.0"
+>
+
+wheel: <
+  name: "infra/python/wheels/colorama-py2_py3"
+  version: "version:0.4.1"
+>
+
+# Used by:
+#   testing/buildbot/generate_buildbot_json_coveragetest.py
+wheel: <
+  name: "infra/python/wheels/coverage/${vpython_platform}"
+  version: "version:5.5.chromium.2"
+>
+
+# Used by:
+#   //content/test/gpu
+wheel: <
+  name: "infra/python/wheels/pathos/${vpython_platform}"
+  version: "version:0.2.7.chromium.4"
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "manylinux1_i686"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_mips64"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_armv6l"
+  >
+  not_match_tag <
+    abi: "cp27mu"
+    platform: "linux_armv7l"
+  >
+>
+
+# Used by:
+#   //tools/infra/find_bad_builds.py
+wheel: <
+  name: "infra/python/wheels/pytz-py2_py3"
+  version: "version:2018.4"
+>
+
+# Used by:
+#   //third_party/blink/tools/blinkpy/web_tests/port/server_process.py
+wheel: <
+  name: "infra/python/wheels/pywin32/${vpython_platform}"
+  version: "version:300"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+>
+
+# Used by:
+#   //content/test/gpu/gpu_tests/color_profile_manager_mac.py
+wheel: <
+  name: "infra/python/wheels/pyobjc/${vpython_platform}"
+  version: "version:7.3.chromium.1"
+  match_tag: <
+    platform: "macosx_10_10_intel"
+  >
+>
+
+# Used by:
+#   tools/perf/core/results_dashboard.py
+wheel: <
+  name: "infra/python/wheels/httplib2-py3"
+  version: "version:0.19.1"
+>
+
+# Used by:
+#   tools/perf/flakiness_cli
+wheel: <
+  name: "infra/python/wheels/pandas/${vpython_platform}"
+  version: "version:1.3.2.chromium.1"
+  match_tag: <
+    platform: "win32"
+  >
+  match_tag: <
+    platform: "win_amd64"
+  >
+  match_tag: <
+    platform: "manylinux1_i686"
+  >
+  match_tag: <
+    platform: "manylinux1_x86_64"
+  >
+  match_tag: <
+    platform: "macosx_10_6_intel"
+  >
+>
diff --git a/files/AUTHORS b/files/AUTHORS
index 9686ac13..28c08956 100644
--- a/files/AUTHORS
+++ b/files/AUTHORS
@@ -2,3 +2,5 @@
 # Name or Organization <email address>
 
 Google Inc.
+
+Ivan Pavlotskiy <ivan.pavlotskiy@lgepartner.com>
diff --git a/files/Android.bp b/files/Android.bp
index 20b8c234..36156287 100644
--- a/files/Android.bp
+++ b/files/Android.bp
@@ -17,7 +17,6 @@ license {
     ],
     license_text: [
         "LICENSE",
-        "LICENSE_THIRD_PARTY",
         "PATENTS",
     ],
 }
@@ -35,31 +34,31 @@ cc_library {
         "source/compare.cc",
         "source/compare_common.cc",
         "source/compare_gcc.cc",
+        "source/compare_msa.cc",
         "source/compare_neon.cc",
         "source/compare_neon64.cc",
-        "source/compare_mmi.cc",
-        "source/compare_msa.cc",
         "source/convert.cc",
         "source/convert_argb.cc",
         "source/convert_from.cc",
         "source/convert_from_argb.cc",
+        "source/convert_jpeg.cc",
         "source/convert_to_argb.cc",
         "source/convert_to_i420.cc",
         "source/cpu_id.cc",
+        "source/mjpeg_decoder.cc",
+        "source/mjpeg_validate.cc",
         "source/planar_functions.cc",
         "source/rotate.cc",
         "source/rotate_any.cc",
         "source/rotate_argb.cc",
         "source/rotate_common.cc",
         "source/rotate_gcc.cc",
-        "source/rotate_mmi.cc",
         "source/rotate_msa.cc",
         "source/rotate_neon.cc",
         "source/rotate_neon64.cc",
         "source/row_any.cc",
         "source/row_common.cc",
         "source/row_gcc.cc",
-        "source/row_mmi.cc",
         "source/row_msa.cc",
         "source/row_neon.cc",
         "source/row_neon64.cc",
@@ -68,14 +67,12 @@ cc_library {
         "source/scale_argb.cc",
         "source/scale_common.cc",
         "source/scale_gcc.cc",
-        "source/scale_mmi.cc",
         "source/scale_msa.cc",
         "source/scale_neon.cc",
         "source/scale_neon64.cc",
+        "source/scale_rgb.cc",
+        "source/scale_uv.cc",
         "source/video_common.cc",
-        "source/convert_jpeg.cc",
-        "source/mjpeg_decoder.cc",
-        "source/mjpeg_validate.cc",
     ],
 
     cflags: [
@@ -122,7 +119,6 @@ cc_test {
     shared_libs: ["libjpeg"],
     cflags: ["-Wall", "-Werror"],
     srcs: [
-        "unit_test/unit_test.cc",
         "unit_test/basictypes_test.cc",
         "unit_test/color_test.cc",
         "unit_test/compare_test.cc",
@@ -134,7 +130,10 @@ cc_test {
         "unit_test/rotate_argb_test.cc",
         "unit_test/rotate_test.cc",
         "unit_test/scale_argb_test.cc",
+        "unit_test/scale_rgb_test.cc",
         "unit_test/scale_test.cc",
+        "unit_test/scale_uv_test.cc",
+        "unit_test/unit_test.cc",
         "unit_test/video_common_test.cc",
     ],
 }
@@ -158,6 +157,15 @@ cc_test {
 }
 
 cc_test {
+    name: "i444tonv12_eg",
+    gtest: false,
+    srcs: [
+        "util/i444tonv12_eg.cc",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
     name: "psnr",
     gtest: false,
     srcs: [
@@ -169,6 +177,15 @@ cc_test {
 }
 
 cc_test {
+    name: "yuvconstants",
+    gtest: false,
+    srcs: [
+        "util/yuvconstants.c",
+    ],
+    static_libs: ["libyuv"],
+}
+
+cc_test {
     name: "yuvconvert",
     gtest: false,
     srcs: [
diff --git a/files/BUILD.gn b/files/BUILD.gn
index 8904fd6c..a72ff065 100644
--- a/files/BUILD.gn
+++ b/files/BUILD.gn
@@ -6,12 +6,12 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-import("libyuv.gni")
 import("//testing/test.gni")
+import("libyuv.gni")
 
 declare_args() {
-  # Set to false to disable building with gflags.
-  libyuv_use_gflags = true
+  # Set to false to disable building with absl flags.
+  libyuv_use_absl_flags = true
 
   # When building a shared library using a target in WebRTC or
   # Chromium projects that depends on libyuv, setting this flag
@@ -27,20 +27,24 @@ config("libyuv_config") {
   if (is_android && current_cpu != "arm64") {
     ldflags = [ "-Wl,--dynamic-linker,/system/bin/linker" ]
   }
+
+  if (!libyuv_use_neon) {
+    defines = [ "LIBYUV_DISABLE_NEON" ]
+  }
 }
 
 # This target is built when no specific target is specified on the command line.
 group("default") {
   testonly = true
-  deps = [
-    ":libyuv",
-  ]
+  deps = [ ":libyuv" ]
   if (libyuv_include_tests) {
     deps += [
       ":compare",
       ":cpuid",
+      ":i444tonv12_eg",
       ":libyuv_unittest",
       ":psnr",
+      ":yuvconstants",
       ":yuvconvert",
     ]
   }
@@ -52,13 +56,9 @@ group("libyuv") {
 
   if (is_win && target_cpu == "x64") {
     # Compile with clang in order to get inline assembly
-    public_deps = [
-      ":libyuv_internal(//build/toolchain/win:win_clang_x64)",
-    ]
+    public_deps = [ ":libyuv_internal(//build/toolchain/win:win_clang_x64)" ]
   } else {
-    public_deps = [
-      ":libyuv_internal",
-    ]
+    public_deps = [ ":libyuv_internal" ]
   }
 
   if (libyuv_use_neon) {
@@ -69,11 +69,7 @@ group("libyuv") {
     deps += [ ":libyuv_msa" ]
   }
 
-  if (libyuv_use_mmi) {
-    deps += [ ":libyuv_mmi" ]
-  }
-
-  if (!is_ios) {
+  if (!is_ios && !libyuv_disable_jpeg) {
     # Make sure that clients of libyuv link with libjpeg. This can't go in
     # libyuv_internal because in Windows x64 builds that will generate a clang
     # build of libjpeg, and we don't want two copies.
@@ -102,7 +98,9 @@ static_library("libyuv_internal") {
     "include/libyuv/row.h",
     "include/libyuv/scale.h",
     "include/libyuv/scale_argb.h",
+    "include/libyuv/scale_rgb.h",
     "include/libyuv/scale_row.h",
+    "include/libyuv/scale_uv.h",
     "include/libyuv/version.h",
     "include/libyuv/video_common.h",
 
@@ -137,6 +135,8 @@ static_library("libyuv_internal") {
     "source/scale_argb.cc",
     "source/scale_common.cc",
     "source/scale_gcc.cc",
+    "source/scale_rgb.cc",
+    "source/scale_uv.cc",
     "source/scale_win.cc",
     "source/video_common.cc",
   ]
@@ -150,7 +150,7 @@ static_library("libyuv_internal") {
     configs += [ "//build/config/gcc:symbol_visibility_default" ]
   }
 
-  if (!is_ios) {
+  if (!is_ios && !libyuv_disable_jpeg) {
     defines += [ "HAVE_JPEG" ]
 
     # Needed to pull in libjpeg headers. Can't add //third_party:jpeg to deps
@@ -192,9 +192,7 @@ if (libyuv_use_neon) {
       "source/scale_neon64.cc",
     ]
 
-    deps = [
-      ":libyuv_internal",
-    ]
+    deps = [ ":libyuv_internal" ]
 
     public_configs = [ ":libyuv_config" ]
 
@@ -225,27 +223,7 @@ if (libyuv_use_msa) {
       "source/scale_msa.cc",
     ]
 
-    deps = [
-      ":libyuv_internal",
-    ]
-
-    public_configs = [ ":libyuv_config" ]
-  }
-}
-
-if (libyuv_use_mmi) {
-  static_library("libyuv_mmi") {
-    sources = [
-      # MMI Source Files
-      "source/compare_mmi.cc",
-      "source/rotate_mmi.cc",
-      "source/row_mmi.cc",
-      "source/scale_mmi.cc",
-    ]
-
-    deps = [
-      ":libyuv_internal",
-    ]
+    deps = [ ":libyuv_internal" ]
 
     public_configs = [ ":libyuv_config" ]
   }
@@ -275,8 +253,6 @@ if (libyuv_include_tests) {
     testonly = true
 
     sources = [
-      # sources
-      # headers
       "unit_test/basictypes_test.cc",
       "unit_test/color_test.cc",
       "unit_test/compare_test.cc",
@@ -288,7 +264,9 @@ if (libyuv_include_tests) {
       "unit_test/rotate_argb_test.cc",
       "unit_test/rotate_test.cc",
       "unit_test/scale_argb_test.cc",
+      "unit_test/scale_rgb_test.cc",
       "unit_test/scale_test.cc",
+      "unit_test/scale_uv_test.cc",
       "unit_test/unit_test.cc",
       "unit_test/unit_test.h",
       "unit_test/video_common_test.cc",
@@ -300,19 +278,20 @@ if (libyuv_include_tests) {
     ]
 
     defines = []
-    if (libyuv_use_gflags) {
-      defines += [ "LIBYUV_USE_GFLAGS" ]
-      deps += [ "//third_party/gflags" ]
+    if (libyuv_use_absl_flags) {
+      defines += [ "LIBYUV_USE_ABSL_FLAGS" ]
+      deps += [
+        "//third_party/abseil-cpp/absl/flags:flag",
+        "//third_party/abseil-cpp/absl/flags:parse",
+      ]
     }
 
     configs += [ ":libyuv_unittest_warnings_config" ]
 
-    public_deps = [
-      "//testing/gtest",
-    ]
+    public_deps = [ "//testing/gtest" ]
     public_configs = [ ":libyuv_unittest_config" ]
 
-    if (is_linux) {
+    if (is_linux || is_chromeos) {
       cflags = [ "-fexceptions" ]
     }
     if (is_ios) {
@@ -349,10 +328,8 @@ if (libyuv_include_tests) {
       # sources
       "util/compare.cc",
     ]
-    deps = [
-      ":libyuv",
-    ]
-    if (is_linux) {
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
       cflags = [ "-fexceptions" ]
     }
   }
@@ -362,10 +339,19 @@ if (libyuv_include_tests) {
       # sources
       "util/yuvconvert.cc",
     ]
-    deps = [
-      ":libyuv",
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
+      cflags = [ "-fexceptions" ]
+    }
+  }
+
+  executable("yuvconstants") {
+    sources = [
+      # sources
+      "util/yuvconstants.c",
     ]
-    if (is_linux) {
+    deps = [ ":libyuv" ]
+    if (is_linux || is_chromeos) {
       cflags = [ "-fexceptions" ]
     }
   }
@@ -377,22 +363,26 @@ if (libyuv_include_tests) {
       "util/psnr_main.cc",
       "util/ssim.cc",
     ]
-    deps = [
-      ":libyuv",
-    ]
+    deps = [ ":libyuv" ]
 
     if (!is_ios && !libyuv_disable_jpeg) {
       defines = [ "HAVE_JPEG" ]
     }
   }
 
+  executable("i444tonv12_eg") {
+    sources = [
+      # sources
+      "util/i444tonv12_eg.cc",
+    ]
+    deps = [ ":libyuv" ]
+  }
+
   executable("cpuid") {
     sources = [
       # sources
       "util/cpuid.c",
     ]
-    deps = [
-      ":libyuv",
-    ]
+    deps = [ ":libyuv" ]
   }
 }
diff --git a/files/CMakeLists.txt b/files/CMakeLists.txt
index 188a26b7..d190507b 100644
--- a/files/CMakeLists.txt
+++ b/files/CMakeLists.txt
@@ -22,6 +22,10 @@ LIST ( SORT			ly_unittest_sources )
 
 INCLUDE_DIRECTORIES( BEFORE ${ly_inc_dir} )
 
+if(MSVC)
+  ADD_DEFINITIONS ( -D_CRT_SECURE_NO_WARNINGS )
+endif()
+
 # this creates the static library (.a)
 ADD_LIBRARY				( ${ly_lib_static} STATIC ${ly_source_files} )
 
@@ -29,6 +33,9 @@ ADD_LIBRARY				( ${ly_lib_static} STATIC ${ly_source_files} )
 ADD_LIBRARY				( ${ly_lib_shared} SHARED ${ly_source_files} )
 SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES OUTPUT_NAME "${ly_lib_name}" )
 SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES PREFIX "lib" )
+if(WIN32)
+  SET_TARGET_PROPERTIES	( ${ly_lib_shared} PROPERTIES IMPORT_PREFIX "lib" )
+endif()
 
 # this creates the conversion tool
 ADD_EXECUTABLE			( yuvconvert ${ly_base_dir}/util/yuvconvert.cc )
@@ -71,12 +78,6 @@ if(TEST)
   if(NACL AND NACL_LIBC STREQUAL "newlib")
     target_link_libraries(libyuv_unittest glibc-compat)
   endif()
-
-  find_library(GFLAGS_LIBRARY gflags)
-  if(NOT GFLAGS_LIBRARY STREQUAL "GFLAGS_LIBRARY-NOTFOUND")
-    target_link_libraries(libyuv_unittest gflags)
-    add_definitions(-DLIBYUV_USE_GFLAGS)
-  endif()
 endif()
 
 
diff --git a/files/DEPS b/files/DEPS
index c5f81b86..3cf2dbe0 100644
--- a/files/DEPS
+++ b/files/DEPS
@@ -1,90 +1,92 @@
+gclient_gn_args_file = 'src/build/config/gclient_args.gni'
+gclient_gn_args = [
+  'generate_location_tags',
+]
+
 vars = {
   'chromium_git': 'https://chromium.googlesource.com',
-  'chromium_revision': '4476bd69d1c8e4e1cde8633d3b33c992f7d3a6d0',
-  'swarming_revision': '0e3e1c4dc4e79f25a5b58fcbc135dc93183c0c54',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling lss
-  # and whatever else without interference from each other.
-  'lss_revision': 'e6527b0cd469e3ff5764785dadcb39bf7d787154',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling catapult
-  # and whatever else without interference from each other.
-  'catapult_revision': 'a24a725f7834c16b3628bfb63f349b3480bf9592',
-  # the commit queue can handle CLs rolling android_sdk_build-tools_version
-  # and whatever else without interference from each other.
-  'android_sdk_build-tools_version': 'DLK621q5_Bga5EsOr7cp6bHWWxFKx6UHLu_Ix_m3AckC',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_emulator_version
-  # and whatever else without interference from each other.
-  'android_sdk_emulator_version': 'ki7EDQRAiZAUYlnTWR1XmI6cJTk65fJ-DNZUU1zrtS8C',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_extras_version
-  # and whatever else without interference from each other.
-  'android_sdk_extras_version': 'iIwhhDox5E-mHgwUhCz8JACWQCpUjdqt5KTY9VLugKQC',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_patcher_version
-  # and whatever else without interference from each other.
-  'android_sdk_patcher_version': 'I6FNMhrXlpB-E1lOhMlvld7xt9lBVNOO83KIluXDyA0C',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_platform-tools_version
-  # and whatever else without interference from each other.
-  'android_sdk_platform-tools_version': '4Y2Cb2LGzoc-qt-oIUIlhySotJaKeE3ELFedSVe6Uk8C',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_platforms_version
-  # and whatever else without interference from each other.
-  'android_sdk_platforms_version': 'Kg2t9p0YnQk8bldUv4VA3o156uPXLUfIFAmVZ-Gm5ewC',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_sources_version
-  # and whatever else without interference from each other.
-  'android_sdk_sources_version': 'K9uEn3JvNELEVjjVK_GQD3ZQD3rqAnJSxCWxjmUmRkgC',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_tools_version
-  # and whatever else without interference from each other.
-  'android_sdk_tools_version': 'wYcRQC2WHsw2dKWs4EA7fw9Qsyzu1ds1_fRjKmGxe5QC',
-  # Three lines of non-changing comments so that
-  # the commit queue can handle CLs rolling android_sdk_tools-lint_version
-  # and whatever else without interference from each other.
-  'android_sdk_tools-lint_version': '89hXqZYzCum3delB5RV7J_QyWkaRodqdtQS0s3LMh3wC',
+  'chromium_revision': '829c6df33dce1085a61d8fd44209fc84bbf9a6a7',
+  'gn_version': 'git_revision:6f13aaac55a977e1948910942675c69f2b4f7a94',
+
+  # Keep the Chromium default of generating location tags.
+  'generate_location_tags': True,
 }
 
 deps = {
   'src/build':
-    Var('chromium_git') + '/chromium/src/build' + '@' + '669e41d6f18842ed5740449662a71b715dc607c6',
+    Var('chromium_git') + '/chromium/src/build' + '@' + 'dcea3443035f48d58193788e0bc56daca4e5db33',
   'src/buildtools':
-    Var('chromium_git') + '/chromium/buildtools.git' + '@' + '0e1cbc4eab6861b0c84bf2ed9a3c4b7aa2063819',
+    Var('chromium_git') + '/chromium/src/buildtools' + '@' + '075dd7e22837a69189003e4fa84499acf63188cf',
   'src/testing':
-    Var('chromium_git') + '/chromium/src/testing' + '@' + 'b1c6aeebeabcc177a83ff0a33dc6c3ab03d4aa94',
+    Var('chromium_git') + '/chromium/src/testing' + '@' + 'f4e42be13265ec304b0f3085eee2b15f30f44077',
   'src/third_party':
-    Var('chromium_git') + '/chromium/src/third_party' + '@' + 'be3e0fc18f2e9ea14d0e9369e539eae5986335fd',
+    Var('chromium_git') + '/chromium/src/third_party' + '@' + '42c249feeb71bc0cd184849f0509aefef599343d',
+
+  'src/buildtools/linux64': {
+    'packages': [
+      {
+        'package': 'gn/gn/linux-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_linux',
+  },
+  'src/buildtools/mac': {
+    'packages': [
+      {
+        'package': 'gn/gn/mac-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_mac',
+  },
+  'src/buildtools/win': {
+    'packages': [
+      {
+        'package': 'gn/gn/windows-amd64',
+        'version': Var('gn_version'),
+      }
+    ],
+    'dep_type': 'cipd',
+    'condition': 'checkout_win',
+  },
+
+  'src/buildtools/clang_format/script':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/clang/tools/clang-format.git' + '@' + '99876cacf78329e5f99c244dbe42ccd1654517a0',
+  'src/buildtools/third_party/libc++/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libcxx.git' + '@' + '79a2e924d96e2fc1e4b937c42efd08898fa472d7',
+  'src/buildtools/third_party/libc++abi/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libcxxabi.git' + '@' + '665b74f7d1b3bb295cd6ba7d8fcec1acd3d2ac84',
+  'src/buildtools/third_party/libunwind/trunk':
+    Var('chromium_git') + '/external/github.com/llvm/llvm-project/libunwind.git' + '@' + 'f51a154281bdfe746c46c07cd4fb05be97f9441d',
+
   'src/third_party/catapult':
-    Var('chromium_git') + '/catapult.git' + '@' + Var('catapult_revision'),
+    Var('chromium_git') + '/catapult.git' + '@' + '75423c310eb303d28978be892fcf7b9c2c824909',
   'src/third_party/colorama/src':
     Var('chromium_git') + '/external/colorama.git' + '@' + '799604a1041e9b3bc5d2789ecbd7e8db2e18e6b8',
+  'src/third_party/depot_tools':
+    Var('chromium_git') + '/chromium/tools/depot_tools.git' + '@' + '2ffa1bde797a8127c0f72908d0bd74051fd65d0d',
   'src/third_party/freetype/src':
-    Var('chromium_git') + '/chromium/src/third_party/freetype2.git' + '@' + 'd01e28f41f8810c8ea422b854f8722659589fa99',
+    Var('chromium_git') + '/chromium/src/third_party/freetype2.git' + '@' + 'cff026d41599945498044d2f4dcc0e610ffb6929',
   'src/third_party/googletest/src':
-    Var('chromium_git') + '/external/github.com/google/googletest.git' + '@' + '879ac092fde0a19e1b3a61b2546b2a422b1528bc',
+    Var('chromium_git') + '/external/github.com/google/googletest.git' + '@' + 'e2f3978937c0244508135f126e2617a7734a68be',
   'src/third_party/harfbuzz-ng/src':
-    Var('chromium_git') + '/external/github.com/harfbuzz/harfbuzz.git' + '@' + '26c5b54fb09fb45e02c9c4618bcea4958c698953',
+    Var('chromium_git') + '/external/github.com/harfbuzz/harfbuzz.git' + '@' + '64b29dbd5994a511acee69cb9b45ad650ef88359',
   'src/third_party/libjpeg_turbo':
-    Var('chromium_git') + '/chromium/deps/libjpeg_turbo.git' + '@' + '61a2bbaa9aec89cb2c882d87ace6aba9aee49bb9',
-  'src/third_party/yasm/source/patched-yasm':
-    Var('chromium_git') + '/chromium/deps/yasm/patched-yasm.git' + '@' + '720b70524a4424b15fc57e82263568c8ba0496ad',
+    Var('chromium_git') + '/chromium/deps/libjpeg_turbo.git' + '@' + '02959c3ee17abacfd1339ec22ea93301292ffd56',
+  'src/third_party/nasm':
+    Var('chromium_git') + '/chromium/deps/nasm.git' + '@' + '9215e8e1d0fe474ffd3e16c1a07a0f97089e6224',
   'src/tools':
-    Var('chromium_git') + '/chromium/src/tools' + '@' + '419541c8352b3b75a99c9a5a7c0d1e7b92f3fcf7',
-  'src/tools/swarming_client':
-    Var('chromium_git') + '/infra/luci/client-py.git' + '@' +  Var('swarming_revision'),
+    Var('chromium_git') + '/chromium/src/tools' + '@' + '198dc879529652b39ba6e223bcc0bcad5f1facd6',
 
   # libyuv-only dependencies (not present in Chromium).
-  'src/third_party/gflags':
-    Var('chromium_git') + '/external/webrtc/deps/third_party/gflags' + '@' + '892576179b45861b53e04a112996a738309cf364',
-  'src/third_party/gflags/src':
-    Var('chromium_git') + '/external/github.com/gflags/gflags' + '@' + '03bebcb065c83beff83d50ae025a55a4bf94dfca',
   'src/third_party/gtest-parallel':
     Var('chromium_git') + '/external/webrtc/deps/third_party/gtest-parallel' + '@' + '1dad0e9f6d82ff994130b529d7d814b40eb32b0e',
 
   'src/third_party/lss': {
-    'url': Var('chromium_git') + '/linux-syscall-support.git' + '@' + Var('lss_revision'),
+    'url': Var('chromium_git') + '/linux-syscall-support.git' + '@' + '92a65a8f5d705d1928874420c8d0d15bde8c89e5',
     'condition': 'checkout_android or checkout_linux',
   },
 
@@ -93,25 +95,27 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/accessibility-test-framework',
-              'version': 'version:2.1-cr0',
+              'version': 'b5ec1e56e58e56bc1a0c77d43111c37f9b512c8a',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
   'src/third_party/auto/src': {
-    'url': Var('chromium_git') + '/external/github.com/google/auto.git' + '@' + '8a81a858ae7b78a1aef71ac3905fade0bbd64e82',
+    'url': Var('chromium_git') + '/external/github.com/google/auto.git' + '@' + 'fe67d853d6356943dc79541c892ab6d3e6a7b61a',
     'condition': 'checkout_android',
   },
+  'src/third_party/boringssl/src':
+    'https://boringssl.googlesource.com/boringssl.git' + '@' + '3a667d10e94186fd503966f5638e134fe9fb4080',
   'src/base': {
-    'url': Var('chromium_git') + '/chromium/src/base' + '@' + '162a5d66ad148f26bbbe6b6ecaf5c1bafa2173e6',
+    'url': Var('chromium_git') + '/chromium/src/base' + '@' + 'e9e639622449a893a1b5e32781d072cec08ead72',
     'condition': 'checkout_android',
   },
   'src/third_party/bazel': {
       'packages': [
           {
               'package': 'chromium/third_party/bazel',
-              'version': 'version:0.10.0',
+              'version': 'VjMsf48QUWw8n7XtJP2AuSjIGmbQeYdWdwyxVvIRLmAC',
           },
       ],
       'condition': 'checkout_android',
@@ -121,77 +125,89 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/bouncycastle',
-              'version': 'version:1.46-cr0',
+              'version': 'c078e87552ba26e776566fdaf0f22cd8712743d0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
   'src/third_party/android_ndk': {
-    'url': Var('chromium_git') + '/android_ndk.git' + '@' + '4e2cea441bfd43f0863d14f57b1e1844260b9884',
+    'url': Var('chromium_git') + '/android_ndk.git' + '@' + '401019bf85744311b26c88ced255cd53401af8b7',
+    'condition': 'checkout_android',
+  },
+  'src/third_party/androidx': {
+    'packages': [
+      {
+          'package': 'chromium/third_party/androidx',
+          'version': '6d8ij5pzYh29WWjPbdbAWFBJSA1nUgkWf2p6wCVZKIsC',
+      },
+    ],
     'condition': 'checkout_android',
+    'dep_type': 'cipd',
   },
   'src/third_party/android_support_test_runner': {
       'packages': [
           {
               'package': 'chromium/third_party/android_support_test_runner',
-              'version': 'version:0.5-cr0',
+              'version': '96d4bf848cd210fdcbca6bcc8c1b4b39cbd93141',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-  'src/third_party/android_tools': {
-    'url': Var('chromium_git') + '/android_tools.git' + '@' + 'e958d6ea74442d4e0849bb8a018d215a0e78981d',
-    'condition': 'checkout_android',
-  },
   'src/third_party/android_sdk/public': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_sdk/public/build-tools',
-              'version': Var('android_sdk_build-tools_version'),
+              'package': 'chromium/third_party/android_sdk/public/build-tools/31.0.0',
+              'version': 'tRoD45SCi7UleQqSV7MrMQO1_e5P8ysphkCcj6z_cCQC',
           },
           {
               'package': 'chromium/third_party/android_sdk/public/emulator',
-              'version': Var('android_sdk_emulator_version'),
+              'version': 'gMHhUuoQRKfxr-MBn3fNNXZtkAVXtOwMwT7kfx8jkIgC',
           },
           {
               'package': 'chromium/third_party/android_sdk/public/extras',
-              'version': Var('android_sdk_extras_version'),
+              'version': 'ppQ4TnqDvBHQ3lXx5KPq97egzF5X2FFyOrVHkGmiTMQC',
           },
           {
               'package': 'chromium/third_party/android_sdk/public/patcher',
-              'version': Var('android_sdk_patcher_version'),
+              'version': 'I6FNMhrXlpB-E1lOhMlvld7xt9lBVNOO83KIluXDyA0C',
           },
           {
               'package': 'chromium/third_party/android_sdk/public/platform-tools',
-              'version': Var('android_sdk_platform-tools_version'),
+              'version': 'g7n_-r6yJd_SGRklujGB1wEt8iyr77FZTUJVS9w6O34C',
           },
           {
-              'package': 'chromium/third_party/android_sdk/public/platforms',
-              'version': Var('android_sdk_platforms_version'),
+              'package': 'chromium/third_party/android_sdk/public/platforms/android-31',
+              'version': 'lL3IGexKjYlwjO_1Ga-xwxgwbE_w-lmi2Zi1uOlWUIAC',
           },
           {
-              'package': 'chromium/third_party/android_sdk/public/sources',
-              'version': Var('android_sdk_sources_version'),
+              'package': 'chromium/third_party/android_sdk/public/sources/android-31',
+              'version': '_a_BcnANjPYw5mSKlNHa7GFY8yc1kdqj2rmQgac7yUcC',
           },
           {
-              'package': 'chromium/third_party/android_sdk/public/tools',
-              'version': Var('android_sdk_tools_version'),
-          },
-          {
-              'package': 'chromium/third_party/android_sdk/public/tools-lint',
-              'version': Var('android_sdk_tools-lint_version'),
+              'package': 'chromium/third_party/android_sdk/public/cmdline-tools',
+              'version': 'Ez2NWws2SJYCF6qw2O-mSCqK6424l3ZdSTpppLyVR_cC',
           },
       ],
-      'condition': 'checkout_android_native_support',
+      'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
+  'src/tools/clang/dsymutil': {
+    'packages': [
+      {
+        'package': 'chromium/llvm-build-tools/dsymutil',
+        'version': 'OWlhXkmj18li3yhJk59Kmjbc5KdgLh56TwCd1qBdzlIC',
+      }
+    ],
+    'condition': 'checkout_mac',
+    'dep_type': 'cipd',
+  },
   'src/third_party/android_build_tools/aapt2': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_tools_aapt2',
-              'version': 'version:3.2.0-alpha18-4804415-cr0',
+              'package': 'chromium/third_party/android_build_tools/aapt2',
+              'version': 'version:3.6.0-alpha03-5516695-cr0',
           },
       ],
       'condition': 'checkout_android',
@@ -201,14 +217,14 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/byte_buddy',
-              'version': 'version:1.4.17-cr0',
+              'version': 'c9b53316603fc2d997c899c7ca1707f809b918cd',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
   'src/third_party/ced/src': {
-    'url': Var('chromium_git') + '/external/github.com/google/compact_enc_det.git' + '@' + '94c367a1fe3a13207f4b22604fcfd1d9f9ddf6d9',
+    'url': Var('chromium_git') + '/external/github.com/google/compact_enc_det.git' + '@' + 'ba412eaaacd3186085babcd901679a48863c7dd5',
     'condition': 'checkout_android',
   },
   'src/third_party/errorprone/lib': {
@@ -223,7 +239,7 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/gson',
-              'version': 'version:2.8.0-cr0',
+              'version': '681931c9778045903a0ed59856ce2dd8dd7bf7ca',
           },
       ],
       'condition': 'checkout_android',
@@ -233,7 +249,7 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/guava',
-              'version': 'version:23.0-cr0',
+              'version': 'a6fba501f3a0de88b9be1daa2052632de5b96a46',
           },
       ],
       'condition': 'checkout_android',
@@ -243,20 +259,21 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/hamcrest',
-              'version': 'version:1.3-cr0',
+              'version': '37eccfc658fe79695d6abb6dd497463c4372032f',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
+
   'src/third_party/icu': {
-    'url': Var('chromium_git') + '/chromium/deps/icu.git' + '@' + 'd65301491c513d49163ad29c853eb85c02c8d5b4',
+    'url': Var('chromium_git') + '/chromium/deps/icu.git' + '@' + 'bf66d373ae781a3498f2babe7b61d933dd774b82',
   },
   'src/third_party/icu4j': {
       'packages': [
           {
               'package': 'chromium/third_party/icu4j',
-              'version': 'version:53.1-cr0',
+              'version': 'e87e5bed2b4935913ee26a3ebd0b723ee2344354',
           },
       ],
       'condition': 'checkout_android',
@@ -266,7 +283,21 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/intellij',
-              'version': 'version:12.0-cr0',
+              'version': '77c2721b024b36ee073402c08e6d8428c0295336',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/jdk': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/jdk',
+              'version': 'PfRSnxe8Od6WU4zBXomq-zsgcJgWmm3z4gMQNB-r2QcC',
+          },
+          {
+              'package': 'chromium/third_party/jdk/extras',
+              'version': 'fkhuOQ3r-zKtWEdKplpo6k0vKkjl-LY_rJTmtzFCQN4C',
           },
       ],
       'condition': 'checkout_android',
@@ -280,6 +311,10 @@ deps = {
     'url': Var('chromium_git') + '/external/junit.git' + '@' + '64155f8a9babcfcf4263cf4d08253a1556e75481',
     'condition': 'checkout_android',
   },
+  'src/third_party/libunwindstack': {
+      'url': Var('chromium_git') + '/chromium/src/third_party/libunwindstack.git' + '@' + '6868358481bb1e5e20d155c1084dc436c88b5e6b',
+      'condition': 'checkout_android',
+  },
   'src/third_party/mockito/src': {
     'url': Var('chromium_git') + '/external/mockito/mockito.git' + '@' + '04a2a289a4222f80ad20717c25144981210d2eac',
     'condition': 'checkout_android',
@@ -288,7 +323,7 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/objenesis',
-              'version': 'version:2.4-cr0',
+              'version': '9e367f55e5a65781ee77bfcbaa88fb82b30e75c0',
           },
       ],
       'condition': 'checkout_android',
@@ -298,7 +333,7 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/ow2_asm',
-              'version': 'version:5.0.1-cr0',
+              'version': 'NNAhdJzMdnutUVqfSJm5v0tVazA9l3Dd6CRwH6N4Q5kC',
           },
       ],
       'condition': 'checkout_android',
@@ -308,7 +343,7 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/r8',
-              'version': 'version:1.0.30',
+              'version': 'Nu_mvQJe34CotIXadFlA3w732CJ9EvQGuVs4udcZedAC',
           },
       ],
       'condition': 'checkout_android',
@@ -318,537 +353,1765 @@ deps = {
       'packages': [
           {
               'package': 'chromium/third_party/proguard',
-              'version': '3bd778c422ea5496de2ef25c007a517dbb5ce5ca',
+              'version': 'Fd91BJFVlmiO6c46YMTsdy7n2f5Sk2hVVGlzPLvqZPsC',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
   'src/third_party/requests/src': {
-    'url': Var('chromium_git') + '/external/github.com/kennethreitz/requests.git' + '@' + 'f172b30356d821d180fa4ecfa3e71c7274a32de4',
+    'url': Var('chromium_git') + '/external/github.com/kennethreitz/requests.git' + '@' + 'refs/tags/v2.23.0',
     'condition': 'checkout_android',
   },
   'src/third_party/robolectric': {
       'packages': [
           {
               'package': 'chromium/third_party/robolectric',
-              'version': 'version:3.5.1',
+              'version': 'iC6RDM5EH3GEAzR-1shW_Mg0FeeNE5shq1okkFfuuNQC',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-  'src/third_party/robolectric/robolectric': {
-    'url': Var('chromium_git') + '/external/robolectric.git' + '@' + '7e067f1112e1502caa742f7be72d37b5678d3403',
-    'condition': 'checkout_android',
-  },
   'src/third_party/sqlite4java': {
       'packages': [
           {
               'package': 'chromium/third_party/sqlite4java',
-              'version': 'version:0.282-cr0',
+              'version': '889660698187baa7c8b0d79f7bf58563125fbd66',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-  'src/third_party/ub-uiautomator/lib': {
-    'url': Var('chromium_git') + '/chromium/third_party/ub-uiautomator.git' + '@' + '00270549ce3161ae72ceb24712618ea28b4f9434',
-    'condition': 'checkout_android',
-  },
-  'src/third_party/xstream': {
+  'src/third_party/turbine': {
       'packages': [
           {
-              'package': 'chromium/third_party/xstream',
-              'version': 'version:1.4.8-cr0',
+              'package': 'chromium/third_party/turbine',
+              'version': 'Om6yIEXgJxuqghErK29h9RcMH6VaymMbxwScwXmcN6EC',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
+  'src/third_party/ub-uiautomator/lib': {
+    'url': Var('chromium_git') + '/chromium/third_party/ub-uiautomator.git' + '@' + '00270549ce3161ae72ceb24712618ea28b4f9434',
+    'condition': 'checkout_android',
+  },
 
   # iOS deps:
   'src/ios': {
-    'url': Var('chromium_git') + '/chromium/src/ios' + '@' + '44be3c093cf2db7ab4cf1997d6a1a07722f1f391',
+    'url': Var('chromium_git') + '/chromium/src/ios' + '@' + '81826d980c159f949c2c7901f4dbec9a09788964',
     'condition': 'checkout_ios'
   },
 
-  # Win deps:
-  # Dependencies used by libjpeg-turbo
-  'src/third_party/yasm/binaries': {
-    'url': Var('chromium_git') + '/chromium/deps/yasm/binaries.git' + '@' + '52f9b3f4b0aa06da24ef8b123058bb61ee468881',
-    'condition': 'checkout_win',
-  },
-
+  # Everything coming after this is automatically updated by the auto-roller.
   # === ANDROID_DEPS Generated Code Start ===
-  # Generated by //tools/android/roll/android_deps/fetch_all.sh
+
   'src/third_party/android_deps/libs/android_arch_core_common': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/android_arch_core_common',
-              'version': 'version:1.0.0-cr0',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/android_arch_core_runtime': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_core_runtime',
+              'version': 'version:2@1.1.1.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/android_arch_lifecycle_common': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_common',
-              'version': 'version:1.0.0-cr0',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/android_arch_lifecycle_common_java8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_common_java8',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/android_arch_lifecycle_livedata': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_livedata',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/android_arch_lifecycle_livedata_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_livedata_core',
+              'version': 'version:2@1.1.1.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/android_arch_lifecycle_runtime': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_runtime',
-              'version': 'version:1.0.0-cr0',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/android_arch_lifecycle_viewmodel': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/android_arch_lifecycle_viewmodel',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/backport_util_concurrent_backport_util_concurrent': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/backport_util_concurrent_backport_util_concurrent',
+              'version': 'version:2@3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/classworlds_classworlds': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/classworlds_classworlds',
+              'version': 'version:2@1.1-alpha-2.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_animated_vector_drawable': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_animated_vector_drawable',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_appcompat_v7': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_appcompat_v7',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_asynclayoutinflater': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_asynclayoutinflater',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_cardview_v7': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_cardview_v7',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_design': {
+  'src/third_party/android_deps/libs/com_android_support_collections': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_design',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_collections',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_gridlayout_v7': {
+  'src/third_party/android_deps/libs/com_android_support_coordinatorlayout': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_gridlayout_v7',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_coordinatorlayout',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_leanback_v17': {
+  'src/third_party/android_deps/libs/com_android_support_cursoradapter': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_leanback_v17',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_cursoradapter',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_mediarouter_v7': {
+  'src/third_party/android_deps/libs/com_android_support_customview': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_mediarouter_v7',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_customview',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_multidex': {
+  'src/third_party/android_deps/libs/com_android_support_design': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_multidex',
-              'version': 'version:1.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_design',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_palette_v7': {
+  'src/third_party/android_deps/libs/com_android_support_documentfile': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_palette_v7',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_documentfile',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_preference_leanback_v17': {
+  'src/third_party/android_deps/libs/com_android_support_drawerlayout': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_preference_leanback_v17',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_drawerlayout',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_preference_v14': {
+  'src/third_party/android_deps/libs/com_android_support_interpolator': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_preference_v14',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_interpolator',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_preference_v7': {
+  'src/third_party/android_deps/libs/com_android_support_loader': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_preference_v7',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_loader',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_localbroadcastmanager': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_localbroadcastmanager',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_multidex': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_multidex',
+              'version': 'version:2@1.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_print': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_print',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_recyclerview_v7': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_recyclerview_v7',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_slidingpanelayout': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_slidingpanelayout',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_annotations': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_annotations',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_compat': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_compat',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_core_ui': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_core_ui',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_core_utils': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_core_utils',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_fragment': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_fragment',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_support_media_compat': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_support_media_compat',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_support_v13': {
+  'src/third_party/android_deps/libs/com_android_support_support_v4': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_v13',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_v4',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_support_v4': {
+  'src/third_party/android_deps/libs/com_android_support_support_vector_drawable': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_v4',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_vector_drawable',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
-  'src/third_party/android_deps/libs/com_android_support_support_vector_drawable': {
+  'src/third_party/android_deps/libs/com_android_support_swiperefreshlayout': {
       'packages': [
           {
-              'package': 'chromium/third_party/android_deps/libs/com_android_support_support_vector_drawable',
-              'version': 'version:27.0.0-cr0',
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_swiperefreshlayout',
+              'version': 'version:2@28.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_android_support_transition': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_android_support_transition',
-              'version': 'version:27.0.0-cr0',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_versionedparcelable': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_versionedparcelable',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_support_viewpager': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_support_viewpager',
+              'version': 'version:2@28.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_tools_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_common',
+              'version': 'version:2@30.0.0-alpha10.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs_configuration': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_desugar_jdk_libs_configuration',
+              'version': 'version:2@1.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_tools_layoutlib_layoutlib_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_layoutlib_layoutlib_api',
+              'version': 'version:2@30.0.0-alpha10.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_android_tools_sdk_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_android_tools_sdk_common',
+              'version': 'version:2@30.0.0-alpha10.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_github_ben_manes_caffeine_caffeine': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_github_ben_manes_caffeine_caffeine',
+              'version': 'version:2@2.8.8.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_github_kevinstern_software_and_algorithms': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_github_kevinstern_software_and_algorithms',
+              'version': 'version:2@1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_datatransport_transport_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_datatransport_transport_api',
+              'version': 'version:2@2.2.1.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth_api_phone': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth_api_phone',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.5.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_auth_base': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_auth_base',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_base': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_base',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.5.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_basement': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_basement',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.5.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_cast': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cast',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_cast_framework': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cast_framework',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_clearcut': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_clearcut',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_cloud_messaging': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_cloud_messaging',
+              'version': 'version:2@16.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_fido': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_fido',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@19.0.0-beta.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_flags': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_flags',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_gcm': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_gcm',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_iid': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_iid',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_instantapps': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_instantapps',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_location': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_location',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_phenotype': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_phenotype',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_places_placereport': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_places_placereport',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_gms_play_services_stats': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_stats',
+              'version': 'version:2@17.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_tasks': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_tasks',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@17.2.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_vision': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_vision',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@18.0.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_gms_play_services_vision_common': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_gms_play_services_vision_common',
-              'version': 'version:12.0.1-cr0',
+              'version': 'version:2@18.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_android_material_material': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_android_material_material',
+              'version': 'version:2@1.6.0-alpha01.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_google_android_play_core': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_google_android_play_core',
-              'version': 'version:1.3.0-cr0',
+              'version': 'version:2@1.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_auto_auto_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_auto_common',
+              'version': 'version:2@1.1.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_auto_service_auto_service': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_service_auto_service',
+              'version': 'version:2@1.0-rc6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_auto_service_auto_service_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_service_auto_service_annotations',
+              'version': 'version:2@1.0-rc6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_auto_value_auto_value_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_auto_value_auto_value_annotations',
+              'version': 'version:2@1.7.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_code_findbugs_jformatstring': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_code_findbugs_jformatstring',
+              'version': 'version:2@3.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_code_findbugs_jsr305': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_code_findbugs_jsr305',
+              'version': 'version:2@3.0.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_code_gson_gson': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_code_gson_gson',
+              'version': 'version:2@2.8.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_dagger_dagger': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger',
+              'version': 'version:2@2.30.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_compiler': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_compiler',
+              'version': 'version:2@2.30.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_producers': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_producers',
+              'version': 'version:2@2.30.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_dagger_dagger_spi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_dagger_dagger_spi',
+              'version': 'version:2@2.30.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_annotation': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_annotation',
+              'version': 'version:2@2.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_annotations',
+              'version': 'version:2@2.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_check_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_check_api',
+              'version': 'version:2@2.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_core': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_core',
+              'version': 'version:2@2.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_error_prone_type_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_error_prone_type_annotations',
+              'version': 'version:2@2.10.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_javac': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_javac',
+              'version': 'version:2@9+181-r4173-1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_errorprone_javac_shaded': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_errorprone_javac_shaded',
+              'version': 'version:2@9-dev-r4023-3.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_annotations',
+              'version': 'version:2@16.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_common',
+              'version': 'version:2@19.5.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_components': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_components',
+              'version': 'version:2@16.1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_encoders': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_encoders',
+              'version': 'version:2@16.1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_encoders_json': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_encoders_json',
+              'version': 'version:2@17.1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_iid': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_iid',
+              'version': 'version:2@21.0.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_iid_interop': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_iid_interop',
+              'version': 'version:2@17.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_installations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_installations',
+              'version': 'version:2@16.3.5.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_installations_interop': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_installations_interop',
+              'version': 'version:2@16.0.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_measurement_connector': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_measurement_connector',
+              'version': 'version:2@18.0.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_firebase_firebase_messaging': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_firebase_firebase_messaging',
+              'version': 'version:2@21.0.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_flatbuffers_flatbuffers_java': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_flatbuffers_flatbuffers_java',
+              'version': 'version:2@2.0.3.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_googlejavaformat_google_java_format': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_googlejavaformat_google_java_format',
+              'version': 'version:2@1.5.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_guava_failureaccess': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_failureaccess',
+              'version': 'version:2@1.0.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_guava_guava': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_guava',
+              'version': 'version:2@31.0-jre.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_guava_guava_android': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_guava_android',
+              'version': 'version:2@31.0-android.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_guava_listenablefuture': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_guava_listenablefuture',
+              'version': 'version:2@1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_j2objc_j2objc_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_j2objc_j2objc_annotations',
+              'version': 'version:2@1.3.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_protobuf_protobuf_java': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_protobuf_protobuf_java',
+              'version': 'version:2@3.4.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_google_protobuf_protobuf_javalite': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_google_protobuf_protobuf_javalite',
+              'version': 'version:2@3.13.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_googlecode_java_diff_utils_diffutils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_googlecode_java_diff_utils_diffutils',
+              'version': 'version:2@1.3.0.cr0',
           },
       ],
       'condition': 'checkout_android',
       'dep_type': 'cipd',
   },
-
   'src/third_party/android_deps/libs/com_squareup_javapoet': {
       'packages': [
           {
               'package': 'chromium/third_party/android_deps/libs/com_squareup_javapoet',
-              'version': 'version:1.11.0-cr0',
+              'version': 'version:2@1.13.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/com_squareup_javawriter': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/com_squareup_javawriter',
+              'version': 'version:2@2.1.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/io_github_java_diff_utils_java_diff_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/io_github_java_diff_utils_java_diff_utils',
+              'version': 'version:2@4.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/javax_annotation_javax_annotation_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_annotation_javax_annotation_api',
+              'version': 'version:2@1.3.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/javax_annotation_jsr250_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_annotation_jsr250_api',
+              'version': 'version:2@1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/javax_inject_javax_inject': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/javax_inject_javax_inject',
+              'version': 'version:2@1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/nekohtml_nekohtml': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/nekohtml_nekohtml',
+              'version': 'version:2@1.9.6.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/nekohtml_xercesminimal': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/nekohtml_xercesminimal',
+              'version': 'version:2@1.9.6.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/net_ltgt_gradle_incap_incap': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_ltgt_gradle_incap_incap',
+              'version': 'version:2@0.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/net_sf_kxml_kxml2': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/net_sf_kxml_kxml2',
+              'version': 'version:2@2.3.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_ant_ant': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_ant_ant',
+              'version': 'version:2@1.8.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_ant_ant_launcher': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_ant_ant_launcher',
+              'version': 'version:2@1.8.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_ant_tasks': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_ant_tasks',
+              'version': 'version:2@2.1.3.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_artifact': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_artifact',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_artifact_manager': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_artifact_manager',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_error_diagnostics': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_error_diagnostics',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_model': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_model',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_plugin_registry': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_plugin_registry',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_profile': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_profile',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_project': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_project',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_repository_metadata': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_repository_metadata',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_maven_settings': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_maven_settings',
+              'version': 'version:2@2.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_wagon_wagon_file': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_wagon_wagon_file',
+              'version': 'version:2@1.0-beta-6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_wagon_wagon_http_lightweight': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_wagon_wagon_http_lightweight',
+              'version': 'version:2@1.0-beta-6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_wagon_wagon_http_shared': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_wagon_wagon_http_shared',
+              'version': 'version:2@1.0-beta-6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_apache_maven_wagon_wagon_provider_api': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_apache_maven_wagon_wagon_provider_api',
+              'version': 'version:2@1.0-beta-6.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ccil_cowan_tagsoup_tagsoup': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ccil_cowan_tagsoup_tagsoup',
+              'version': 'version:2@1.2.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_checkerframework_checker_compat_qual': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_checker_compat_qual',
+              'version': 'version:2@2.5.5.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_checkerframework_checker_qual': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_checker_qual',
+              'version': 'version:2@3.12.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_checkerframework_dataflow_errorprone': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_checkerframework_dataflow_errorprone',
+              'version': 'version:2@3.15.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_codehaus_mojo_animal_sniffer_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_codehaus_mojo_animal_sniffer_annotations',
+              'version': 'version:2@1.17.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_codehaus_plexus_plexus_container_default': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_codehaus_plexus_plexus_container_default',
+              'version': 'version:2@1.0-alpha-9-stable-1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_codehaus_plexus_plexus_interpolation': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_codehaus_plexus_plexus_interpolation',
+              'version': 'version:2@1.11.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_codehaus_plexus_plexus_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_codehaus_plexus_plexus_utils',
+              'version': 'version:2@1.5.15.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_eclipse_jgit_org_eclipse_jgit': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_eclipse_jgit_org_eclipse_jgit',
+              'version': 'version:2@4.4.1.201607150455-r.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_annotations',
+              'version': 'version:2@13.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib',
+              'version': 'version:2@1.6.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_common': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_common',
+              'version': 'version:2@1.6.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk7': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk7',
+              'version': 'version:2@1.5.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk8': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlin_kotlin_stdlib_jdk8',
+              'version': 'version:2@1.5.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_android': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_android',
+              'version': 'version:2@1.5.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_core_jvm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_coroutines_core_jvm',
+              'version': 'version:2@1.5.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_metadata_jvm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_jetbrains_kotlinx_kotlinx_metadata_jvm',
+              'version': 'version:2@0.1.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ow2_asm_asm': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm',
+              'version': 'version:2@7.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_analysis': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_analysis',
+              'version': 'version:2@7.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_commons': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_commons',
+              'version': 'version:2@7.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_tree': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_tree',
+              'version': 'version:2@7.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_ow2_asm_asm_util': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_ow2_asm_asm_util',
+              'version': 'version:2@7.0.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_pcollections_pcollections': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_pcollections_pcollections',
+              'version': 'version:2@2.1.2.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_annotations': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_annotations',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_junit': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_junit',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_pluginapi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_pluginapi',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_plugins_maven_dependency_resolver': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_plugins_maven_dependency_resolver',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_resources': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_resources',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_robolectric': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_robolectric',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_sandbox': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_sandbox',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_shadowapi': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadowapi',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_shadows_framework': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadows_framework',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_shadows_playservices': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_shadows_playservices',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_utils': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_utils',
+              'version': 'version:2@4.3.1.cr0',
+          },
+      ],
+      'condition': 'checkout_android',
+      'dep_type': 'cipd',
+  },
+  'src/third_party/android_deps/libs/org_robolectric_utils_reflector': {
+      'packages': [
+          {
+              'package': 'chromium/third_party/android_deps/libs/org_robolectric_utils_reflector',
+              'version': 'version:2@4.3.1.cr0',
           },
       ],
       'condition': 'checkout_android',
@@ -858,16 +2121,13 @@ deps = {
   # === ANDROID_DEPS Generated Code End ===
 }
 
-# Define rules for which include paths are allowed in our source.
-include_rules = [ '+gflags' ]
-
 pre_deps_hooks = [
   {
     # Remove any symlinks from before 177567c518b121731e507e9b9c4049c4dc96e4c8.
     # TODO(kjellander): Remove this in March 2017.
     'name': 'cleanup_links',
     'pattern': '.',
-    'action': ['python', 'src/cleanup_links.py'],
+    'action': ['python3', 'src/cleanup_links.py'],
   },
 ]
 
@@ -879,7 +2139,7 @@ hooks = [
     'name': 'landmines',
     'pattern': '.',
     'action': [
-        'python',
+        'python3',
         'src/build/landmines.py',
         '--landmine-scripts',
         'src/tools_libyuv/get_landmines.py',
@@ -892,108 +2152,88 @@ hooks = [
     'name': 'sysroot_arm',
     'pattern': '.',
     'condition': 'checkout_linux and checkout_arm',
-    'action': ['python', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
                '--arch=arm'],
   },
   {
     'name': 'sysroot_arm64',
     'pattern': '.',
     'condition': 'checkout_linux and checkout_arm64',
-    'action': ['python', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
                '--arch=arm64'],
   },
   {
     'name': 'sysroot_x86',
     'pattern': '.',
     'condition': 'checkout_linux and (checkout_x86 or checkout_x64)',
-    'action': ['python', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
                '--arch=x86'],
   },
   {
     'name': 'sysroot_mips',
     'pattern': '.',
     'condition': 'checkout_linux and checkout_mips',
-    'action': ['python', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
                '--arch=mips'],
   },
   {
     'name': 'sysroot_x64',
     'pattern': '.',
     'condition': 'checkout_linux and checkout_x64',
-    'action': ['python', 'src/build/linux/sysroot_scripts/install-sysroot.py',
+    'action': ['python3', 'src/build/linux/sysroot_scripts/install-sysroot.py',
                '--arch=x64'],
   },
   {
     # Update the Windows toolchain if necessary.
     'name': 'win_toolchain',
     'pattern': '.',
-    'action': ['python', 'src/build/vs_toolchain.py', 'update'],
+    'action': ['python3', 'src/build/vs_toolchain.py', 'update'],
   },
   {
     # Update the Mac toolchain if necessary.
     'name': 'mac_toolchain',
     'pattern': '.',
-    'action': ['python', 'src/build/mac_toolchain.py'],
+    'action': ['python3', 'src/build/mac_toolchain.py'],
+    'condition': 'checkout_mac',
   },
-  # Pull binutils for linux, enabled debug fission for faster linking /
-  # debugging when used with clang on Ubuntu Precise.
-  # https://code.google.com/p/chromium/issues/detail?id=352046
   {
-    'name': 'binutils',
-    'pattern': 'src/third_party/binutils',
-    'action': [
-        'python',
-        'src/third_party/binutils/download.py',
-    ],
+    'name': 'msan_chained_origins',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                "--no_resume",
+                "--no_auth",
+                "--bucket", "chromium-instrumented-libraries",
+                "-s", "src/third_party/instrumented_libraries/binaries/msan-chained-origins.tgz.sha1",
+              ],
+  },
+  {
+    'name': 'msan_no_origins',
+    'pattern': '.',
+    'condition': 'checkout_instrumented_libraries',
+    'action': [ 'python3',
+                'src/third_party/depot_tools/download_from_google_storage.py',
+                "--no_resume",
+                "--no_auth",
+                "--bucket", "chromium-instrumented-libraries",
+                "-s", "src/third_party/instrumented_libraries/binaries/msan-no-origins.tgz.sha1",
+              ],
   },
   {
     # Pull clang if needed or requested via GYP_DEFINES.
     # Note: On Win, this should run after win_toolchain, as it may use it.
     'name': 'clang',
     'pattern': '.',
-    'action': ['python', 'src/tools/clang/scripts/update.py'],
+    'action': ['python3', 'src/tools/clang/scripts/update.py'],
   },
   {
     # Update LASTCHANGE.
     'name': 'lastchange',
     'pattern': '.',
-    'action': ['python', 'src/build/util/lastchange.py',
+    'action': ['python3', 'src/build/util/lastchange.py',
                '-o', 'src/build/util/LASTCHANGE'],
   },
-  # Pull GN binaries.
-  {
-    'name': 'gn_win',
-    'pattern': '.',
-    'action': [ 'download_from_google_storage',
-                '--no_resume',
-                '--platform=win32',
-                '--no_auth',
-                '--bucket', 'chromium-gn',
-                '-s', 'src/buildtools/win/gn.exe.sha1',
-    ],
-  },
-  {
-    'name': 'gn_mac',
-    'pattern': '.',
-    'action': [ 'download_from_google_storage',
-                '--no_resume',
-                '--platform=darwin',
-                '--no_auth',
-                '--bucket', 'chromium-gn',
-                '-s', 'src/buildtools/mac/gn.sha1',
-    ],
-  },
-  {
-    'name': 'gn_linux64',
-    'pattern': '.',
-    'action': [ 'download_from_google_storage',
-                '--no_resume',
-                '--platform=linux*',
-                '--no_auth',
-                '--bucket', 'chromium-gn',
-                '-s', 'src/buildtools/linux64/gn.sha1',
-    ],
-  },
   # Pull clang-format binaries using checked-in hashes.
   {
     'name': 'clang_format_win',
@@ -1020,6 +2260,7 @@ hooks = [
   {
     'name': 'clang_format_linux',
     'pattern': '.',
+    'condition': 'host_os == "linux"',
     'action': [ 'download_from_google_storage',
                 '--no_resume',
                 '--platform=linux*',
@@ -1074,23 +2315,16 @@ hooks = [
                '--root', 'src',
     ],
   },
-  # Android dependencies. Many are downloaded using Google Storage these days.
-  # They're copied from https://cs.chromium.org/chromium/src/DEPS for all
-  # such dependencies we share with Chromium.
   {
-    # This downloads SDK extras and puts them in the
-    # third_party/android_sdk/public/extras directory.
-    'name': 'sdkextras',
-    'condition': 'checkout_android',
+    'name': 'Generate component metadata for tests',
     'pattern': '.',
-    'action': ['vpython',
-               'src/build/android/play_services/update.py',
-               'download'
+    'action': [
+      'vpython3',
+      'src/testing/generate_location_tags.py',
+      '--out',
+      'src/testing/location_tags.json',
     ],
   },
 ]
 
-recursedeps = [
-  # buildtools provides clang_format, libc++, and libc++abi.
-  'src/buildtools',
-]
+recursedeps = []
diff --git a/files/DIR_METADATA b/files/DIR_METADATA
new file mode 100644
index 00000000..8bc04f15
--- /dev/null
+++ b/files/DIR_METADATA
@@ -0,0 +1,3 @@
+monorail {
+  component: "Internals>Images>Codecs"
+}
diff --git a/files/LICENSE_THIRD_PARTY b/files/LICENSE_THIRD_PARTY
deleted file mode 100644
index a71591e7..00000000
--- a/files/LICENSE_THIRD_PARTY
+++ /dev/null
@@ -1,8 +0,0 @@
-This source tree contains third party source code which is governed by third
-party licenses. This file contains references to files which are under other
-licenses than the one provided in the LICENSE file in the root of the source
-tree.
-
-Files governed by third party licenses:
-source/x86inc.asm
-
diff --git a/files/PRESUBMIT.py b/files/PRESUBMIT.py
index 2cf1542f..d3901caf 100755..100644
--- a/files/PRESUBMIT.py
+++ b/files/PRESUBMIT.py
@@ -6,50 +6,30 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-import os
-
-
-def _RunPythonTests(input_api, output_api):
-  def join(*args):
-    return input_api.os_path.join(input_api.PresubmitLocalPath(), *args)
-
-  test_directories = [
-      root for root, _, files in os.walk(join('tools_libyuv'))
-      if any(f.endswith('_test.py') for f in files)
-  ]
-
-  tests = []
-  for directory in test_directories:
-    tests.extend(
-      input_api.canned_checks.GetUnitTestsInDirectory(
-          input_api,
-          output_api,
-          directory,
-          whitelist=[r'.+_test\.py$']))
-  return input_api.RunTests(tests, parallel=True)
-
+# Runs PRESUBMIT.py in py3 mode by git cl presubmit.
+USE_PYTHON3 = True
 
 def _CommonChecks(input_api, output_api):
   """Checks common to both upload and commit."""
   results = []
   results.extend(input_api.canned_checks.RunPylint(input_api, output_api,
-      black_list=(r'^base[\\\/].*\.py$',
-                  r'^build[\\\/].*\.py$',
-                  r'^buildtools[\\\/].*\.py$',
-                  r'^ios[\\\/].*\.py$',
-                  r'^out.*[\\\/].*\.py$',
-                  r'^testing[\\\/].*\.py$',
-                  r'^third_party[\\\/].*\.py$',
-                  r'^tools[\\\/].*\.py$',
-                  # TODO(kjellander): should arguably be checked.
-                  r'^tools_libyuv[\\\/]valgrind[\\\/].*\.py$',
-                  r'^xcodebuild.*[\\\/].*\.py$',),
+      files_to_skip=(r'^base[\\\/].*\.py$',
+                     r'^build[\\\/].*\.py$',
+                     r'^buildtools[\\\/].*\.py$',
+                     r'^ios[\\\/].*\.py$',
+                     r'^out.*[\\\/].*\.py$',
+                     r'^testing[\\\/].*\.py$',
+                     r'^third_party[\\\/].*\.py$',
+                     r'^tools[\\\/].*\.py$',
+                     # TODO(kjellander): should arguably be checked.
+                     r'^tools_libyuv[\\\/]valgrind[\\\/].*\.py$',
+                     r'^xcodebuild.*[\\\/].*\.py$',),
       disabled_warnings=['F0401',  # Failed to import x
                          'E0611',  # No package y in x
                          'W0232',  # Class has no __init__ method
                         ],
-      pylintrc='pylintrc'))
-  results.extend(_RunPythonTests(input_api, output_api))
+      pylintrc='pylintrc',
+      version='2.7'))
   return results
 
 
diff --git a/files/README.chromium b/files/README.chromium
index bddc2023..3f68e21e 100644
--- a/files/README.chromium
+++ b/files/README.chromium
@@ -1,8 +1,9 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1732
+Version: 1837
 License: BSD
 License File: LICENSE
 
 Description:
 libyuv is an open source project that includes YUV conversion and scaling functionality.
+
diff --git a/files/all.gyp b/files/all.gyp
deleted file mode 100644
index 88a74842..00000000
--- a/files/all.gyp
+++ /dev/null
@@ -1,21 +0,0 @@
-# Copyright 2013 The LibYuv Project Authors. All rights reserved.

-#

-# Use of this source code is governed by a BSD-style license

-# that can be found in the LICENSE file in the root of the source

-# tree. An additional intellectual property rights grant can be found

-# in the file PATENTS. All contributing project authors may

-# be found in the AUTHORS file in the root of the source tree.

-

-# all.gyp and All target are for benefit of android gyp build.

-{

-  'targets': [

-    {

-      'target_name': 'All',

-      'type': 'none',

-      'dependencies': [

-        'libyuv.gyp:*',

-        'libyuv_test.gyp:*',

-      ],

-    },

-  ],

-}

diff --git a/files/build_overrides/build.gni b/files/build_overrides/build.gni
index 6d8319b9..c8490313 100644
--- a/files/build_overrides/build.gni
+++ b/files/build_overrides/build.gni
@@ -6,9 +6,6 @@
 # in the file PATENTS. All contributing project authors may
 # be found in the AUTHORS file in the root of the source tree.
 
-# Some non-Chromium builds don't use Chromium's third_party/binutils.
-linux_use_bundled_binutils_override = true
-
 # Variable that can be used to support multiple build scenarios, like having
 # Chromium specific targets in a client project's GN file etc.
 build_with_chromium = false
@@ -44,3 +41,20 @@ if (host_os == "mac") {
              "hermetic toolchain if the minimum OS version is not met.")
   use_system_xcode = _result == 0
 }
+
+declare_args() {
+  # Tracing support requires //third_party/perfetto.
+  enable_base_tracing = false
+  use_perfetto_client_library = false
+
+  # Limits the defined //third_party/android_deps targets to only "buildCompile"
+  # and "buildCompileNoDeps" targets. This is useful for third-party
+  # repositories which do not use JUnit tests. For instance,
+  # limit_android_deps == true removes "gn gen" requirement for
+  # //third_party/robolectric .
+  limit_android_deps = false
+
+  # Allows googletest to pretty-print various absl types.
+  # Defined here rather than in gtest.gni to match chromium.
+  gtest_enable_absl_printers = true
+}
diff --git a/files/chromium/.gclient b/files/chromium/.gclient
deleted file mode 100644
index c1a86ecf..00000000
--- a/files/chromium/.gclient
+++ /dev/null
@@ -1,20 +0,0 @@
-solutions = [{
-  'name': 'src',
-  'url': 'https://chromium.googlesource.com/chromium/src.git',
-  'deps_file': '.DEPS.git',
-  'managed': False,
-  'custom_deps': {
-    # Skip syncing some large dependencies Libyuv will never need.
-    'src/third_party/cld_2/src': None,
-    'src/third_party/ffmpeg': None,
-    'src/third_party/hunspell_dictionaries': None,
-    'src/third_party/liblouis/src': None,
-    'src/third_party/pdfium': None,
-    'src/third_party/skia': None,
-    'src/third_party/trace-viewer': None,
-    'src/third_party/webrtc': None,
-  },
-  'safesync_url': ''
-}]
-
-cache_dir = None
diff --git a/files/chromium/README b/files/chromium/README
deleted file mode 100644
index 127f4b52..00000000
--- a/files/chromium/README
+++ /dev/null
@@ -1,5 +0,0 @@
-This .gclient file is used to do download a copy of Chromium.
-Libyuv uses the Chromium build toolchain and a number of shared
-dependencies by creating symlinks to folders in this checkout,
-using the ../setup_links.py script.
-
diff --git a/files/cleanup_links.py b/files/cleanup_links.py
index ba290789..7d1eba9b 100755
--- a/files/cleanup_links.py
+++ b/files/cleanup_links.py
@@ -1,4 +1,5 @@
-#!/usr/bin/env python
+#!/usr/bin/env vpython3
+
 # Copyright 2017 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -18,8 +19,8 @@ landing that change, this script cleans up any old symlinks, avoiding annoying
 manual cleanup needed in order to complete gclient sync.
 """
 
+import argparse
 import logging
-import optparse
 import os
 import shelve
 import subprocess
@@ -32,14 +33,14 @@ LINKS_DB = 'links'
 # Version management to make future upgrades/downgrades easier to support.
 SCHEMA_VERSION = 1
 
-class WebRTCLinkSetup(object):
+class WebRTCLinkSetup():
   def __init__(self, links_db, dry_run=False):
     self._dry_run = dry_run
     self._links_db = links_db
 
   def CleanupLinks(self):
     logging.debug('CleanupLinks')
-    for source, link_path  in self._links_db.iteritems():
+    for source, link_path  in self._links_db.tems():
       if source == 'SCHEMA_VERSION':
         continue
       if os.path.islink(link_path) or sys.platform.startswith('win'):
@@ -71,15 +72,15 @@ def _initialize_database(filename):
 
 
 def main():
-  parser = optparse.OptionParser()
-  parser.add_option('-d', '--dry-run', action='store_true', default=False,
-                    help='Print what would be done, but don\'t perform any '
-                         'operations. This will automatically set logging to '
-                         'verbose.')
-  parser.add_option('-v', '--verbose', action='store_const',
-                    const=logging.DEBUG, default=logging.INFO,
-                    help='Print verbose output for debugging.')
-  options, _ = parser.parse_args()
+  p = argparse.ArgumentParser()
+  p.add_argument('-d', '--dry-run', action='store_true', default=False,
+                 help='Print what would be done, but don\'t perform any '
+                      'operations. This will automatically set logging to '
+                      'verbose.')
+  p.add_argument('-v', '--verbose', action='store_const',
+                 const=logging.DEBUG, default=logging.INFO,
+                 help='Print verbose output for debugging.')
+  options = p.parse_args()
 
   if options.dry_run:
     options.verbose = logging.DEBUG
diff --git a/files/codereview.settings b/files/codereview.settings
index 00ba1d37..b226fae5 100644
--- a/files/codereview.settings
+++ b/files/codereview.settings
@@ -1,6 +1,5 @@
-# This file is used by git cl to get repository specific information.
+# This file is used by `git cl` to get repository specific information.
 CODE_REVIEW_SERVER: codereview.chromium.org
 GERRIT_HOST: True
 PROJECT: libyuv
-TRY_ON_UPLOAD: False
 VIEW_VC: https://chromium.googlesource.com/libyuv/libyuv/+/
diff --git a/files/docs/deprecated_builds.md b/files/docs/deprecated_builds.md
index 29e0bf9b..ba42966c 100644
--- a/files/docs/deprecated_builds.md
+++ b/files/docs/deprecated_builds.md
@@ -239,6 +239,7 @@ If you get a compile error for atlthunk.lib on Windows, read http://www.chromium
     ninja -C out/Debug libyuv_unittest
     ninja -C out/Debug compare
     ninja -C out/Debug yuvconvert
+    ninja -C out/Debug yuvconstants
     ninja -C out/Debug psnr
     ninja -C out/Debug cpuid
 
diff --git a/files/docs/environment_variables.md b/files/docs/environment_variables.md
index cd8159ad..dd5d59fb 100644
--- a/files/docs/environment_variables.md
+++ b/files/docs/environment_variables.md
@@ -22,6 +22,7 @@ By default the cpu is detected and the most advanced form of SIMD is used.  But
     LIBYUV_DISABLE_F16C
     LIBYUV_DISABLE_AVX512BW
     LIBYUV_DISABLE_AVX512VL
+    LIBYUV_DISABLE_AVX512VNNI
     LIBYUV_DISABLE_AVX512VBMI
     LIBYUV_DISABLE_AVX512VBMI2
     LIBYUV_DISABLE_AVX512VBITALG
@@ -34,7 +35,10 @@ By default the cpu is detected and the most advanced form of SIMD is used.  But
 
 ## MIPS CPUs
     LIBYUV_DISABLE_MSA
-    LIBYUV_DISABLE_MMI
+
+## LOONGARCH CPUs
+    LIBYUV_DISABLE_LSX
+    LIBYUV_DISABLE_LASX
 
 # Test Width/Height/Repeat
 
diff --git a/files/docs/formats.md b/files/docs/formats.md
index 97e8ce05..12ea9465 100644
--- a/files/docs/formats.md
+++ b/files/docs/formats.md
@@ -4,7 +4,9 @@ Formats (FOURCC) supported by libyuv are detailed here.
 
 # Core Formats
 
-There are 2 core formats supported by libyuv - I420 and ARGB.  All YUV formats can be converted to/from I420.  All RGB formats can be converted to/from ARGB.
+There are 2 core formats supported by libyuv - I420 and ARGB.
+  All YUV formats can be converted to/from I420.
+  All RGB formats can be converted to/from ARGB.
 
 Filtering functions such as scaling and planar functions work on I420 and/or ARGB.
 
@@ -36,7 +38,7 @@ This is how OSX formats map to libyuv
 
 The following is extracted from video_common.h as a complete list of formats supported by libyuv.
     enum FourCC {
-      // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+      // 10 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
       FOURCC_I420 = FOURCC('I', '4', '2', '0'),
       FOURCC_I422 = FOURCC('I', '4', '2', '2'),
       FOURCC_I444 = FOURCC('I', '4', '4', '4'),
@@ -46,16 +48,20 @@ The following is extracted from video_common.h as a complete list of formats sup
       FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
       FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
       FOURCC_H010 = FOURCC('H', '0', '1', '0'),  // unofficial fourcc. 10 bit lsb
+      FOURCC_U010 = FOURCC('U', '0', '1', '0'),  // bt.2020, unofficial fourcc.
+                                                 // 10 bit lsb
 
       // 1 Secondary YUV format: row biplanar.
-      FOURCC_M420 = FOURCC('M', '4', '2', '0'),
+      FOURCC_M420 = FOURCC('M', '4', '2', '0'),  // deprecated.
 
-      // 11 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc
+      // 13 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc, 2 64 bpp
       FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
       FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
       FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
       FOURCC_AR30 = FOURCC('A', 'R', '3', '0'),  // 10 bit per channel. 2101010.
       FOURCC_AB30 = FOURCC('A', 'B', '3', '0'),  // ABGR version of 10 bit
+      FOURCC_AR64 = FOURCC('A', 'R', '6', '4'),  // 16 bit per channel.
+      FOURCC_AB64 = FOURCC('A', 'B', '6', '4'),  // ABGR version of 16 bit
       FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
       FOURCC_RAW = FOURCC('r', 'a', 'w', ' '),
       FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
@@ -66,7 +72,7 @@ The following is extracted from video_common.h as a complete list of formats sup
       // 1 Primary Compressed YUV format.
       FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
 
-      // 8 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+      // 11 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
       FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
       FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
       FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
@@ -75,6 +81,9 @@ The following is extracted from video_common.h as a complete list of formats sup
       FOURCC_J400 = FOURCC('J', '4', '0', '0'),  // unofficial fourcc
       FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // unofficial fourcc
       FOURCC_H422 = FOURCC('H', '4', '2', '2'),  // unofficial fourcc
+      FOURCC_U420 = FOURCC('U', '4', '2', '0'),  // bt.2020, unofficial fourcc
+      FOURCC_U422 = FOURCC('U', '4', '2', '2'),  // bt.2020, unofficial fourcc
+      FOURCC_U444 = FOURCC('U', '4', '4', '4'),  // bt.2020, unofficial fourcc
 
       // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
       FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
@@ -104,6 +113,27 @@ The following is extracted from video_common.h as a complete list of formats sup
         I444, NV24 and NV42 are full width, full height
         I400 and J400 have no chroma channel.
 
+# Color space
+      The YUV formats start with a letter to specify the color space. e.g. I420
+        I = BT.601 limited range
+        J = BT.601 full range     (J = JPEG that uses this)
+        H = BT.709 limited range  (H for HD)
+        F = BT.709 full range     (F for Full range)
+        U = BT.2020 limited range (U for UHD)
+        V = BT.2020 full range
+        For YUV to RGB conversions, a matrix can be passed.  See also convert_argh.h
+
+# HDR formats
+      Planar formats with 10 or 12 bits use the following fourcc:
+        I010, I012, P010, P012 are half width, half height
+        I210, I212, P210, P212 are half width, full height
+        I410, I412, P410, P412 are full width, full height
+      where
+        I is the color space (see above) and 3 planes: Y, U and V.
+        P is a biplanar format, similar to NV12 but 16 bits, with the valid bits in the high bits.  There is a Y plane and a UV plane.
+        0, 2 or 4 is the last digit of subsampling: 4:2:0, 4:2:2, or 4:4:4
+        10 or 12 is the bits per channel.  The bits are in the low bits of a 16 bit channel.
+
 # The ARGB FOURCC
 
 There are 4 ARGB layouts - ARGB, BGRA, ABGR and RGBA.  ARGB is most common by far, used for screen formats, and windows webcam drivers.
@@ -152,6 +182,13 @@ The 2 bit alpha has 4 values.  Here are the comparable 8 bit alpha values.
 The 10 bit RGB values range from 0 to 1023.
 XR30 is the same as AR30 but with no alpha channel.
 
+# AB64 and AR64
+
+AB64 is similar to ABGR, with 16 bit (2 bytes) per channel. Each channel stores an unsigned short.
+In memory R is the lowest and A is the highest.
+Each channel has value ranges from 0 to 65535.
+AR64 is similar to ARGB.
+
 # NV12 and NV21
 
 NV12 is a biplanar format with a full sized Y plane followed by a single
@@ -161,3 +198,11 @@ The 12 in NV12 refers to 12 bits per pixel.  NV12 has a half width and half
 height chroma channel, and therefore is a 420 subsampling.
 NV16 is 16 bits per pixel, with half width and full height.  aka 422.
 NV24 is 24 bits per pixel with full sized chroma channel. aka 444.
+Most NV12 functions allow the destination Y pointer to be NULL.
+
+# YUY2 and UYVY
+
+YUY2 is a packed YUV format with half width, full height.
+
+YUY2 is YUYV in memory
+UYVY is UYVY in memory
diff --git a/files/docs/getting_started.md b/files/docs/getting_started.md
index 4426b606..15b19ab2 100644
--- a/files/docs/getting_started.md
+++ b/files/docs/getting_started.md
@@ -165,6 +165,7 @@ Running test with C code:
     ninja -C out/Debug libyuv_unittest
     ninja -C out/Debug compare
     ninja -C out/Debug yuvconvert
+    ninja -C out/Debug yuvconstants
     ninja -C out/Debug psnr
     ninja -C out/Debug cpuid
 
@@ -179,8 +180,8 @@ Running test with C code:
 
 mips
 
-   gn gen out/Release "--args=is_debug=false target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" mips_use_mmi=true is_component_build=false use_sysroot=false use_gold=false"
-   gn gen out/Debug "--args=is_debug=true target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" mips_use_mmi=true is_component_build=false use_sysroot=false use_gold=false"
+   gn gen out/Release "--args=is_debug=false target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" is_component_build=false use_sysroot=false use_gold=false"
+   gn gen out/Debug "--args=is_debug=true target_os=\"linux\" target_cpu=\"mips64el\" mips_arch_variant=\"loongson3\" is_component_build=false use_sysroot=false use_gold=false"
    ninja -v -C out/Debug libyuv_unittest
    ninja -v -C out/Release libyuv_unittest
 
@@ -190,7 +191,7 @@ mips
 
     make V=1 -f linux.mk
     make V=1 -f linux.mk clean
-    make V=1 -f linux.mk CXX=clang++
+    make V=1 -f linux.mk CXX=clang++ CC=clang
 
 ## Building the library with cmake
 
diff --git a/files/docs/rotation.md b/files/docs/rotation.md
index fb84fce5..a08430fd 100644
--- a/files/docs/rotation.md
+++ b/files/docs/rotation.md
@@ -100,4 +100,8 @@ Inverting can be achieved with almost any libyuv function by passing a negative
 
 I420Mirror and ARGBMirror can also be used to rotate by 180 degrees by passing a negative height.
 
+# Cropping - Vertical Flip
 
+When cropping from a subsampled format like NV21, the method of setting the start pointers wont work for odd crop start y on the UV plane.
+If the height after cropping will be odd, invert the source - point to the last row, negate the strides, and pass negative height, which
+will re-invert the image as the conversion outputs.
diff --git a/files/download_vs_toolchain.py b/files/download_vs_toolchain.py
index 4b345789..6bc086d6 100644
--- a/files/download_vs_toolchain.py
+++ b/files/download_vs_toolchain.py
@@ -1,5 +1,5 @@
-#!/usr/bin/env python
-#
+#!/usr/bin/env vpython3
+
 # Copyright 2014 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -22,7 +22,7 @@ sys.path.insert(0, os.path.join(checkout_root, 'build'))
 sys.path.insert(0, os.path.join(checkout_root, 'tools', 'find_depot_tools'))
 
 
-import vs_toolchain
+import vs_toolchain  # pylint: disable=wrong-import-position
 
 
 if __name__ == '__main__':
diff --git a/files/gyp_libyuv b/files/gyp_libyuv
deleted file mode 100755
index 445b924f..00000000
--- a/files/gyp_libyuv
+++ /dev/null
@@ -1,101 +0,0 @@
-#!/usr/bin/env python
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This script is used to run GYP for libyuv. It contains selected parts of the
-# main function from the src/build/gyp_chromium file.
-
-import glob
-import os
-import shlex
-import sys
-
-checkout_root = os.path.dirname(os.path.realpath(__file__))
-
-sys.path.insert(0, os.path.join(checkout_root, 'build'))
-import gyp_chromium
-import gyp_helper
-import vs_toolchain
-
-sys.path.insert(0, os.path.join(checkout_root, 'tools', 'gyp', 'pylib'))
-import gyp
-
-def GetSupplementalFiles():
-  """Returns a list of the supplemental files that are included in all GYP
-  sources."""
-  # Can't use the one in gyp_chromium since the directory location of the root
-  # is different.
-  return glob.glob(os.path.join(checkout_root, '*', 'supplement.gypi'))
-
-
-if __name__ == '__main__':
-  args = sys.argv[1:]
-
-  if int(os.environ.get('GYP_CHROMIUM_NO_ACTION', 0)):
-    print 'Skipping gyp_libyuv due to GYP_CHROMIUM_NO_ACTION env var.'
-    sys.exit(0)
-
-  # This could give false positives since it doesn't actually do real option
-  # parsing.  Oh well.
-  gyp_file_specified = False
-  for arg in args:
-    if arg.endswith('.gyp'):
-      gyp_file_specified = True
-      break
-
-  # If we didn't get a file, assume 'all.gyp' in the root of the checkout.
-  if not gyp_file_specified:
-    # Because of a bug in gyp, simply adding the abspath to all.gyp doesn't
-    # work, but chdir'ing and adding the relative path does. Spooky :/
-    os.chdir(checkout_root)
-    args.append('all.gyp')
-
-  # There shouldn't be a circular dependency relationship between .gyp files,
-  args.append('--no-circular-check')
-
-  # Default to ninja unless GYP_GENERATORS is set.
-  if not os.environ.get('GYP_GENERATORS'):
-    os.environ['GYP_GENERATORS'] = 'ninja'
-
-  vs2013_runtime_dll_dirs = None
-  if int(os.environ.get('DEPOT_TOOLS_WIN_TOOLCHAIN', '1')):
-    vs2013_runtime_dll_dirs = vs_toolchain.SetEnvironmentAndGetRuntimeDllDirs()
-
-  # Enforce gyp syntax checking. This adds about 20% execution time.
-  args.append('--check')
-
-  supplemental_includes = gyp_chromium.GetSupplementalFiles()
-  gyp_vars_dict = gyp_chromium.GetGypVars(supplemental_includes)
-
-  # Automatically turn on crosscompile support for platforms that need it.
-  if all(('ninja' in os.environ.get('GYP_GENERATORS', ''),
-          gyp_vars_dict.get('OS') in ['android', 'ios'],
-          'GYP_CROSSCOMPILE' not in os.environ)):
-    os.environ['GYP_CROSSCOMPILE'] = '1'
-
-  args.extend(['-I' + i for i in
-               gyp_chromium.additional_include_files(supplemental_includes,
-                                                     args)])
-
-  # Set the gyp depth variable to the root of the checkout.
-  args.append('--depth=' + os.path.relpath(checkout_root))
-
-  print 'Updating projects from gyp files...'
-  sys.stdout.flush()
-
-  # Off we go...
-  gyp_rc = gyp.main(args)
-
-  if vs2013_runtime_dll_dirs:
-    x64_runtime, x86_runtime = vs2013_runtime_dll_dirs
-    vs_toolchain.CopyVsRuntimeDlls(
-        os.path.join(checkout_root, gyp_chromium.GetOutputDirectory()),
-        (x86_runtime, x64_runtime))
-
-  sys.exit(gyp_rc)
diff --git a/files/gyp_libyuv.py b/files/gyp_libyuv.py
deleted file mode 100644
index bb32ec39..00000000
--- a/files/gyp_libyuv.py
+++ /dev/null
@@ -1,28 +0,0 @@
-#!/usr/bin/env python
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-
-# This script is a modified copy of the src/build/gyp_chromium.py file.
-# It is needed for parallel processing.
-
-# This file is (possibly, depending on python version) imported by
-# gyp_libyuv when GYP_PARALLEL=1 and it creates sub-processes
-# through the multiprocessing library.
-
-# Importing in Python 2.6 (fixed in 2.7) on Windows doesn't search for
-# imports that don't end in .py (and aren't directories with an
-# __init__.py). This wrapper makes "import gyp_libyuv" work with
-# those old versions and makes it possible to execute gyp_libyuv.py
-# directly on Windows where the extension is useful.
-
-import os
-
-path = os.path.abspath(os.path.split(__file__)[0])
-execfile(os.path.join(path, 'gyp_libyuv'))
diff --git a/files/include/libyuv.h b/files/include/libyuv.h
index aeffd5ef..a06e1233 100644
--- a/files/include/libyuv.h
+++ b/files/include/libyuv.h
@@ -26,6 +26,7 @@
 #include "libyuv/scale.h"
 #include "libyuv/scale_argb.h"
 #include "libyuv/scale_row.h"
+#include "libyuv/scale_uv.h"
 #include "libyuv/version.h"
 #include "libyuv/video_common.h"
 
diff --git a/files/include/libyuv/compare_row.h b/files/include/libyuv/compare_row.h
index e95b9d93..d8e82d72 100644
--- a/files/include/libyuv/compare_row.h
+++ b/files/include/libyuv/compare_row.h
@@ -55,20 +55,20 @@ extern "C" {
 
 // The following are available for Visual C and clangcl 32 bit:
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
+    !defined(__clang__) &&                                                   \
     (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
 #define HAS_HASHDJB2_AVX2
 #define HAS_SUMSQUAREERROR_AVX2
 #endif
 
-// The following are available for GCC and clangcl 64 bit:
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+// The following are available for GCC and clangcl:
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 #define HAS_HAMMINGDISTANCE_SSSE3
 #endif
 
-// The following are available for GCC and clangcl 64 bit:
+// The following are available for GCC and clangcl:
 #if !defined(LIBYUV_DISABLE_X86) && defined(CLANG_HAS_AVX2) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+    (defined(__x86_64__) || defined(__i386__))
 #define HAS_HAMMINGDISTANCE_AVX2
 #endif
 
@@ -84,11 +84,6 @@ extern "C" {
 #define HAS_SUMSQUAREERROR_MSA
 #endif
 
-#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
-#define HAS_HAMMINGDISTANCE_MMI
-#define HAS_SUMSQUAREERROR_MMI
-#endif
-
 uint32_t HammingDistance_C(const uint8_t* src_a,
                            const uint8_t* src_b,
                            int count);
@@ -107,9 +102,6 @@ uint32_t HammingDistance_NEON(const uint8_t* src_a,
 uint32_t HammingDistance_MSA(const uint8_t* src_a,
                              const uint8_t* src_b,
                              int count);
-uint32_t HammingDistance_MMI(const uint8_t* src_a,
-                             const uint8_t* src_b,
-                             int count);
 uint32_t SumSquareError_C(const uint8_t* src_a,
                           const uint8_t* src_b,
                           int count);
@@ -125,9 +117,6 @@ uint32_t SumSquareError_NEON(const uint8_t* src_a,
 uint32_t SumSquareError_MSA(const uint8_t* src_a,
                             const uint8_t* src_b,
                             int count);
-uint32_t SumSquareError_MMI(const uint8_t* src_a,
-                            const uint8_t* src_b,
-                            int count);
 
 uint32_t HashDjb2_C(const uint8_t* src, int count, uint32_t seed);
 uint32_t HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed);
diff --git a/files/include/libyuv/convert.h b/files/include/libyuv/convert.h
index f571142f..46d37159 100644
--- a/files/include/libyuv/convert.h
+++ b/files/include/libyuv/convert.h
@@ -42,6 +42,21 @@ int I444ToI420(const uint8_t* src_y,
                int width,
                int height);
 
+// Convert I444 to NV12.
+LIBYUV_API
+int I444ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
 // Convert I444 to NV21.
 LIBYUV_API
 int I444ToNV21(const uint8_t* src_y,
@@ -74,6 +89,68 @@ int I422ToI420(const uint8_t* src_y,
                int width,
                int height);
 
+// Convert I422 to I444.
+LIBYUV_API
+int I422ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I422 to I210.
+LIBYUV_API
+int I422ToI210(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert MM21 to NV12.
+LIBYUV_API
+int MM21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert MM21 to I420.
+LIBYUV_API
+int MM21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
 // Convert I422 to NV21.
 LIBYUV_API
 int I422ToNV21(const uint8_t* src_y,
@@ -107,6 +184,23 @@ int I420Copy(const uint8_t* src_y,
              int width,
              int height);
 
+// Convert I420 to I444.
+LIBYUV_API
+int I420ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
 // Copy I010 to I010
 #define I010ToI010 I010Copy
 #define H010ToH010 I010Copy
@@ -144,6 +238,246 @@ int I010ToI420(const uint16_t* src_y,
                int width,
                int height);
 
+#define H210ToH420 I210ToI420
+LIBYUV_API
+int I210ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H210ToH422 I210ToI422
+LIBYUV_API
+int I210ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H410ToH444 I410ToI444
+LIBYUV_API
+int I410ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H012ToH420 I012ToI420
+LIBYUV_API
+int I012ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H212ToH422 I212ToI422
+LIBYUV_API
+int I212ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define H412ToH444 I412ToI444
+LIBYUV_API
+int I412ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define I412ToI012 I410ToI010
+#define H410ToH010 I410ToI010
+#define H412ToH012 I410ToI010
+LIBYUV_API
+int I410ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+#define I212ToI012 I210ToI010
+#define H210ToH010 I210ToI010
+#define H212ToH012 I210ToI010
+LIBYUV_API
+int I210ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I010 to I410
+LIBYUV_API
+int I010ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I012 to I412
+#define I012ToI412 I010ToI410
+
+// Convert I210 to I410
+LIBYUV_API
+int I210ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
+// Convert I212 to I412
+#define I212ToI412 I210ToI410
+
+// Convert I010 to P010
+LIBYUV_API
+int I010ToP010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I210 to P210
+LIBYUV_API
+int I210ToP210(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I012 to P012
+LIBYUV_API
+int I012ToP012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert I212 to P212
+LIBYUV_API
+int I212ToP212(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
 // Convert I400 (grey) to I420.
 LIBYUV_API
 int I400ToI420(const uint8_t* src_y,
@@ -200,6 +534,70 @@ int NV21ToI420(const uint8_t* src_y,
                int width,
                int height);
 
+// Convert NV12 to NV24.
+LIBYUV_API
+int NV12ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert NV16 to NV24.
+LIBYUV_API
+int NV16ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P010 to P410.
+LIBYUV_API
+int P010ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P012 to P412.
+#define P012ToP412 P010ToP410
+
+// Convert P016 to P416.
+#define P016ToP416 P010ToP410
+
+// Convert P210 to P410.
+LIBYUV_API
+int P210ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Convert P212 to P412.
+#define P212ToP412 P210ToP410
+
+// Convert P216 to P416.
+#define P216ToP416 P210ToP410
+
 // Convert YUY2 to I420.
 LIBYUV_API
 int YUY2ToI420(const uint8_t* src_yuy2,
@@ -248,19 +646,6 @@ int AYUVToNV21(const uint8_t* src_ayuv,
                int width,
                int height);
 
-// Convert M420 to I420.
-LIBYUV_API
-int M420ToI420(const uint8_t* src_m420,
-               int src_stride_m420,
-               uint8_t* dst_y,
-               int dst_stride_y,
-               uint8_t* dst_u,
-               int dst_stride_u,
-               uint8_t* dst_v,
-               int dst_stride_v,
-               int width,
-               int height);
-
 // Convert Android420 to I420.
 LIBYUV_API
 int Android420ToI420(const uint8_t* src_y,
@@ -370,6 +755,19 @@ int RAWToI420(const uint8_t* src_raw,
               int width,
               int height);
 
+// RGB big endian (rgb in memory) to J420.
+LIBYUV_API
+int RAWToJ420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height);
+
 // RGB16 (RGBP fourcc) little endian to I420.
 LIBYUV_API
 int RGB565ToI420(const uint8_t* src_rgb565,
@@ -418,7 +816,15 @@ int RGB24ToJ400(const uint8_t* src_rgb24,
                 int width,
                 int height);
 
-#ifdef HAVE_JPEG
+// RGB big endian (rgb in memory) to J400.
+LIBYUV_API
+int RAWToJ400(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_yj,
+              int dst_stride_yj,
+              int width,
+              int height);
+
 // src_width/height provided by capture.
 // dst_width/height for clipping determine final size.
 LIBYUV_API
@@ -448,13 +854,25 @@ int MJPGToNV21(const uint8_t* sample,
                int dst_width,
                int dst_height);
 
+// JPEG to NV12
+LIBYUV_API
+int MJPGToNV12(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height);
+
 // Query size of MJPG in pixels.
 LIBYUV_API
 int MJPGSize(const uint8_t* sample,
              size_t sample_size,
              int* width,
              int* height);
-#endif
 
 // Convert camera sample to I420 with cropping, rotation and vertical flip.
 // "src_size" is needed to parse MJPG.
diff --git a/files/include/libyuv/convert_argb.h b/files/include/libyuv/convert_argb.h
index e8ed1f59..f66d20ce 100644
--- a/files/include/libyuv/convert_argb.h
+++ b/files/include/libyuv/convert_argb.h
@@ -14,17 +14,72 @@
 #include "libyuv/basic_types.h"
 
 #include "libyuv/rotate.h"  // For enum RotationMode.
-
-// TODO(fbarchard): This set of functions should exactly match convert.h
-// TODO(fbarchard): Add tests. Create random content of right size and convert
-// with C vs Opt and or to I420 and compare.
-// TODO(fbarchard): Some of these functions lack parameter setting.
+#include "libyuv/scale.h"   // For enum FilterMode.
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+// Conversion matrix for YUV to RGB
+LIBYUV_API extern const struct YuvConstants kYuvI601Constants;   // BT.601
+LIBYUV_API extern const struct YuvConstants kYuvJPEGConstants;   // BT.601 full
+LIBYUV_API extern const struct YuvConstants kYuvH709Constants;   // BT.709
+LIBYUV_API extern const struct YuvConstants kYuvF709Constants;   // BT.709 full
+LIBYUV_API extern const struct YuvConstants kYuv2020Constants;   // BT.2020
+LIBYUV_API extern const struct YuvConstants kYuvV2020Constants;  // BT.2020 full
+
+// Conversion matrix for YVU to BGR
+LIBYUV_API extern const struct YuvConstants kYvuI601Constants;   // BT.601
+LIBYUV_API extern const struct YuvConstants kYvuJPEGConstants;   // BT.601 full
+LIBYUV_API extern const struct YuvConstants kYvuH709Constants;   // BT.709
+LIBYUV_API extern const struct YuvConstants kYvuF709Constants;   // BT.709 full
+LIBYUV_API extern const struct YuvConstants kYvu2020Constants;   // BT.2020
+LIBYUV_API extern const struct YuvConstants kYvuV2020Constants;  // BT.2020 full
+
+// Macros for end swapped destination Matrix conversions.
+// Swap UV and pass mirrored kYvuJPEGConstants matrix.
+// TODO(fbarchard): Add macro for each Matrix function.
+#define kYuvI601ConstantsVU kYvuI601Constants
+#define kYuvJPEGConstantsVU kYvuJPEGConstants
+#define kYuvH709ConstantsVU kYvuH709Constants
+#define kYuvF709ConstantsVU kYvuF709Constants
+#define kYuv2020ConstantsVU kYvu2020Constants
+#define kYuvV2020ConstantsVU kYvuV2020Constants
+
+#define NV12ToABGRMatrix(a, b, c, d, e, f, g, h, i) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV21ToABGRMatrix(a, b, c, d, e, f, g, h, i) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV12ToRAWMatrix(a, b, c, d, e, f, g, h, i) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, g##VU, h, i)
+#define NV21ToRAWMatrix(a, b, c, d, e, f, g, h, i) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, g##VU, h, i)
+#define I010ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I010ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I210ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I210ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I410ToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I410ToARGBMatrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I010ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I010ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I210ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I210ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I410ToAB30Matrix(a, b, c, d, e, f, g, h, i, j, k) \
+  I410ToAR30Matrix(a, b, e, f, c, d, g, h, i##VU, j, k)
+#define I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I420AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I422AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I444AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I010AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I210AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+#define I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, k, l, m, n) \
+  I410AlphaToARGBMatrix(a, b, e, f, c, d, g, h, i, j, k##VU, l, m, n)
+
 // Alias.
 #define ARGBToARGB ARGBCopy
 
@@ -50,7 +105,7 @@ int I420ToARGB(const uint8_t* src_y,
                int width,
                int height);
 
-// Duplicate prototype for function in convert_from.h for remoting.
+// Convert I420 to ABGR.
 LIBYUV_API
 int I420ToABGR(const uint8_t* src_y,
                int src_stride_y,
@@ -63,19 +118,292 @@ int I420ToABGR(const uint8_t* src_y,
                int width,
                int height);
 
-// Convert I010 to ARGB.
+// Convert J420 to ARGB.
 LIBYUV_API
-int I010ToARGB(const uint16_t* src_y,
+int J420ToARGB(const uint8_t* src_y,
                int src_stride_y,
-               const uint16_t* src_u,
+               const uint8_t* src_u,
                int src_stride_u,
-               const uint16_t* src_v,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert J420 to ABGR.
+LIBYUV_API
+int J420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H420 to ARGB.
+LIBYUV_API
+int H420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H420 to ABGR.
+LIBYUV_API
+int H420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U420 to ARGB.
+LIBYUV_API
+int U420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U420 to ABGR.
+LIBYUV_API
+int U420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert J422 to ABGR.
+LIBYUV_API
+int J422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H422 to ARGB.
+LIBYUV_API
+int H422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert H422 to ABGR.
+LIBYUV_API
+int H422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U422 to ARGB.
+LIBYUV_API
+int U422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U422 to ABGR.
+LIBYUV_API
+int U422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert I444 to ABGR.
+LIBYUV_API
+int I444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert J444 to ARGB.
+LIBYUV_API
+int J444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert J444 to ABGR.
+LIBYUV_API
+int J444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H444 to ARGB.
+LIBYUV_API
+int H444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height);
 
+// Convert H444 to ABGR.
+LIBYUV_API
+int H444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U444 to ARGB.
+LIBYUV_API
+int U444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U444 to ABGR.
+LIBYUV_API
+int U444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
 // Convert I010 to ARGB.
 LIBYUV_API
 int I010ToARGB(const uint16_t* src_y,
@@ -128,52 +456,104 @@ int H010ToABGR(const uint16_t* src_y,
                int width,
                int height);
 
-// Convert I422 to ARGB.
+// Convert U010 to ARGB.
 LIBYUV_API
-int I422ToARGB(const uint8_t* src_y,
+int U010ToARGB(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height);
 
-// Convert I444 to ARGB.
+// Convert U010 to ABGR.
 LIBYUV_API
-int I444ToARGB(const uint8_t* src_y,
+int U010ToABGR(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I210 to ARGB.
+LIBYUV_API
+int I210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height);
 
-// Convert J444 to ARGB.
+// Convert I210 to ABGR.
 LIBYUV_API
-int J444ToARGB(const uint8_t* src_y,
+int I210ToABGR(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert H210 to ARGB.
+LIBYUV_API
+int H210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height);
 
-// Convert I444 to ABGR.
+// Convert H210 to ABGR.
 LIBYUV_API
-int I444ToABGR(const uint8_t* src_y,
+int H210ToABGR(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert U210 to ARGB.
+LIBYUV_API
+int U210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert U210 to ABGR.
+LIBYUV_API
+int U210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_abgr,
                int dst_stride_abgr,
@@ -212,6 +592,70 @@ int I420AlphaToABGR(const uint8_t* src_y,
                     int height,
                     int attenuate);
 
+// Convert I422 with Alpha to preattenuated ARGB.
+LIBYUV_API
+int I422AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I422 with Alpha to preattenuated ABGR.
+LIBYUV_API
+int I422AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I444 with Alpha to preattenuated ARGB.
+LIBYUV_API
+int I444AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate);
+
+// Convert I444 with Alpha to preattenuated ABGR.
+LIBYUV_API
+int I444AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate);
+
 // Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
 LIBYUV_API
 int I400ToARGB(const uint8_t* src_y,
@@ -332,15 +776,6 @@ int NV21ToRAW(const uint8_t* src_y,
               int width,
               int height);
 
-// Convert M420 to ARGB.
-LIBYUV_API
-int M420ToARGB(const uint8_t* src_m420,
-               int src_stride_m420,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
-               int width,
-               int height);
-
 // Convert YUY2 to ARGB.
 LIBYUV_API
 int YUY2ToARGB(const uint8_t* src_yuy2,
@@ -359,126 +794,113 @@ int UYVYToARGB(const uint8_t* src_uyvy,
                int width,
                int height);
 
-// Convert J420 to ARGB.
-LIBYUV_API
-int J420ToARGB(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
-               int width,
-               int height);
-
-// Convert J422 to ARGB.
+// Convert I010 to AR30.
 LIBYUV_API
-int J422ToARGB(const uint8_t* src_y,
+int I010ToAR30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
                int width,
                int height);
 
-// Convert J420 to ABGR.
+// Convert H010 to AR30.
 LIBYUV_API
-int J420ToABGR(const uint8_t* src_y,
+int H010ToAR30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_abgr,
-               int dst_stride_abgr,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
                int width,
                int height);
 
-// Convert J422 to ABGR.
+// Convert I010 to AB30.
 LIBYUV_API
-int J422ToABGR(const uint8_t* src_y,
+int I010ToAB30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_abgr,
-               int dst_stride_abgr,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
                int width,
                int height);
 
-// Convert H420 to ARGB.
+// Convert H010 to AB30.
 LIBYUV_API
-int H420ToARGB(const uint8_t* src_y,
+int H010ToAB30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
                int width,
                int height);
 
-// Convert H422 to ARGB.
+// Convert U010 to AR30.
 LIBYUV_API
-int H422ToARGB(const uint8_t* src_y,
+int U010ToAR30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
                int width,
                int height);
 
-// Convert H420 to ABGR.
+// Convert U010 to AB30.
 LIBYUV_API
-int H420ToABGR(const uint8_t* src_y,
+int U010ToAB30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_abgr,
-               int dst_stride_abgr,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
                int width,
                int height);
 
-// Convert H422 to ABGR.
+// Convert I210 to AR30.
 LIBYUV_API
-int H422ToABGR(const uint8_t* src_y,
+int I210ToAR30(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_abgr,
-               int dst_stride_abgr,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
                int width,
                int height);
 
-// Convert H010 to ARGB.
+// Convert I210 to AB30.
 LIBYUV_API
-int H010ToARGB(const uint16_t* src_y,
+int I210ToAB30(const uint16_t* src_y,
                int src_stride_y,
                const uint16_t* src_u,
                int src_stride_u,
                const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
                int width,
                int height);
 
-// Convert I010 to AR30.
+// Convert H210 to AR30.
 LIBYUV_API
-int I010ToAR30(const uint16_t* src_y,
+int H210ToAR30(const uint16_t* src_y,
                int src_stride_y,
                const uint16_t* src_u,
                int src_stride_u,
@@ -489,35 +911,35 @@ int I010ToAR30(const uint16_t* src_y,
                int width,
                int height);
 
-// Convert H010 to AR30.
+// Convert H210 to AB30.
 LIBYUV_API
-int H010ToAR30(const uint16_t* src_y,
+int H210ToAB30(const uint16_t* src_y,
                int src_stride_y,
                const uint16_t* src_u,
                int src_stride_u,
                const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_ar30,
-               int dst_stride_ar30,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
                int width,
                int height);
 
-// Convert I010 to AB30.
+// Convert U210 to AR30.
 LIBYUV_API
-int I010ToAB30(const uint16_t* src_y,
+int U210ToAR30(const uint16_t* src_y,
                int src_stride_y,
                const uint16_t* src_u,
                int src_stride_u,
                const uint16_t* src_v,
                int src_stride_v,
-               uint8_t* dst_ab30,
-               int dst_stride_ab30,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
                int width,
                int height);
 
-// Convert H010 to AB30.
+// Convert U210 to AB30.
 LIBYUV_API
-int H010ToAB30(const uint16_t* src_y,
+int U210ToAB30(const uint16_t* src_y,
                int src_stride_y,
                const uint16_t* src_u,
                int src_stride_u,
@@ -576,6 +998,15 @@ int RAWToARGB(const uint8_t* src_raw,
               int width,
               int height);
 
+// RGB big endian (rgb in memory) to RGBA.
+LIBYUV_API
+int RAWToRGBA(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_rgba,
+              int dst_stride_rgba,
+              int width,
+              int height);
+
 // RGB16 (RGBP fourcc) little endian to ARGB.
 LIBYUV_API
 int RGB565ToARGB(const uint8_t* src_rgb565,
@@ -635,7 +1066,42 @@ int AR30ToAB30(const uint8_t* src_ar30,
                int width,
                int height);
 
-#ifdef HAVE_JPEG
+// Convert AR64 to ARGB.
+LIBYUV_API
+int AR64ToARGB(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert AB64 to ABGR.
+#define AB64ToABGR AR64ToARGB
+
+// Convert AB64 to ARGB.
+LIBYUV_API
+int AB64ToARGB(const uint16_t* src_ab64,
+               int src_stride_ab64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+// Convert AR64 to ABGR.
+#define AR64ToABGR AB64ToARGB
+
+// Convert AR64 To AB64.
+LIBYUV_API
+int AR64ToAB64(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height);
+
+// Convert AB64 To AR64.
+#define AB64ToAR64 AR64ToAB64
+
 // src_width/height provided by capture
 // dst_width/height for clipping determine final size.
 LIBYUV_API
@@ -647,7 +1113,6 @@ int MJPGToARGB(const uint8_t* sample,
                int src_height,
                int dst_width,
                int dst_height);
-#endif
 
 // Convert Android420 to ARGB.
 LIBYUV_API
@@ -677,6 +1142,1009 @@ int Android420ToABGR(const uint8_t* src_y,
                      int width,
                      int height);
 
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int H420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int H420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int J420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
+
+LIBYUV_API
+int J420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int J420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int H420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+LIBYUV_API
+int I422ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height);
+
+// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+
+LIBYUV_API
+int I420ToRGB565Dither(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height);
+
+// Convert I420 to AR30.
+LIBYUV_API
+int I420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert I420 to AB30.
+LIBYUV_API
+int I420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert H420 to AR30.
+LIBYUV_API
+int H420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height);
+
+// Convert H420 to AB30.
+LIBYUV_API
+int H420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height);
+
+// Convert I420 to ARGB with matrix.
+LIBYUV_API
+int I420ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I422 to ARGB with matrix.
+LIBYUV_API
+int I422ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I444 to ARGB with matrix.
+LIBYUV_API
+int I444ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 420 YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 420 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 444 YUV to ARGB with matrix.
+LIBYUV_API
+int I410ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// multiply 12 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I012ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 12 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I012ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 422 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert 10 bit 444 YUV to ARGB with matrix.
+LIBYUV_API
+int I410ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P010 to ARGB with matrix.
+LIBYUV_API
+int P010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P210 to ARGB with matrix.
+LIBYUV_API
+int P210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P010 to AR30 with matrix.
+LIBYUV_API
+int P010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert P210 to AR30 with matrix.
+LIBYUV_API
+int P210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// P012 and P010 use most significant bits so the conversion is the same.
+// Convert P012 to ARGB with matrix.
+#define P012ToARGBMatrix P010ToARGBMatrix
+// Convert P012 to AR30 with matrix.
+#define P012ToAR30Matrix P010ToAR30Matrix
+// Convert P212 to ARGB with matrix.
+#define P212ToARGBMatrix P210ToARGBMatrix
+// Convert P212 to AR30 with matrix.
+#define P212ToAR30Matrix P210ToAR30Matrix
+
+// Convert P016 to ARGB with matrix.
+#define P016ToARGBMatrix P010ToARGBMatrix
+// Convert P016 to AR30 with matrix.
+#define P016ToAR30Matrix P010ToAR30Matrix
+// Convert P216 to ARGB with matrix.
+#define P216ToARGBMatrix P210ToARGBMatrix
+// Convert P216 to AR30 with matrix.
+#define P216ToAR30Matrix P210ToAR30Matrix
+
+// Convert I420 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I420AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I422 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I422AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I444 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I444AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I010 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I010AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I210 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I210AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert I410 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I410AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate);
+
+// Convert NV12 to ARGB with matrix.
+LIBYUV_API
+int NV12ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert NV21 to ARGB with matrix.
+LIBYUV_API
+int NV21ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_vu,
+                     int src_stride_vu,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert NV12 to RGB565 with matrix.
+LIBYUV_API
+int NV12ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_uv,
+                       int src_stride_uv,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height);
+
+// Convert NV12 to RGB24 with matrix.
+LIBYUV_API
+int NV12ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_uv,
+                      int src_stride_uv,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert NV21 to RGB24 with matrix.
+LIBYUV_API
+int NV21ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_vu,
+                      int src_stride_vu,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert Android420 to ARGB with matrix.
+LIBYUV_API
+int Android420ToARGBMatrix(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height);
+
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I422ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to RGBA with matrix.
+LIBYUV_API
+int I420ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to RGB24 with matrix.
+LIBYUV_API
+int I420ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height);
+
+// Convert I420 to RGB565 with specified color matrix.
+LIBYUV_API
+int I420ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height);
+
+// Convert I420 to AR30 with matrix.
+LIBYUV_API
+int I420ToAR30Matrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height);
+
+// Convert I420 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I420ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I422 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I422ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I010 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int I010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I210 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int I210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I010 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I210 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert I420 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I420AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I422 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I422AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I010 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I010AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert I210 with Alpha to attenuated ARGB with matrix and UV filter mode.
+LIBYUV_API
+int I210AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter);
+
+// Convert P010 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int P010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P210 to ARGB with matrix and UV filter mode.
+LIBYUV_API
+int P210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P010 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int P010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
+// Convert P210 to AR30 with matrix and UV filter mode.
+LIBYUV_API
+int P210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter);
+
 // Convert camera sample to ARGB with cropping, rotation and vertical flip.
 // "sample_size" is needed to parse MJPG.
 // "dst_stride_argb" number of bytes in a row of the dst_argb plane.
diff --git a/files/include/libyuv/convert_from.h b/files/include/libyuv/convert_from.h
index 861418d0..32f42a63 100644
--- a/files/include/libyuv/convert_from.h
+++ b/files/include/libyuv/convert_from.h
@@ -23,6 +23,7 @@ extern "C" {
 
 // Convert 8 bit YUV to 10 bit.
 #define H420ToH010 I420ToI010
+LIBYUV_API
 int I420ToI010(const uint8_t* src_y,
                int src_stride_y,
                const uint8_t* src_u,
@@ -38,6 +39,24 @@ int I420ToI010(const uint8_t* src_y,
                int width,
                int height);
 
+// Convert 8 bit YUV to 12 bit.
+#define H420ToH012 I420ToI012
+LIBYUV_API
+int I420ToI012(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height);
+
 LIBYUV_API
 int I420ToI422(const uint8_t* src_y,
                int src_stride_y,
@@ -131,6 +150,10 @@ int I420ToUYVY(const uint8_t* src_y,
                int width,
                int height);
 
+// The following are from convert_argb.h
+// DEPRECATED: The prototypes will be removed in future.  Use convert_argb.h
+
+// Convert I420 to ARGB.
 LIBYUV_API
 int I420ToARGB(const uint8_t* src_y,
                int src_stride_y,
@@ -143,18 +166,7 @@ int I420ToARGB(const uint8_t* src_y,
                int width,
                int height);
 
-LIBYUV_API
-int I420ToBGRA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_bgra,
-               int dst_stride_bgra,
-               int width,
-               int height);
-
+// Convert I420 to ABGR.
 LIBYUV_API
 int I420ToABGR(const uint8_t* src_y,
                int src_stride_y,
@@ -167,181 +179,6 @@ int I420ToABGR(const uint8_t* src_y,
                int width,
                int height);
 
-LIBYUV_API
-int I420ToRGBA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_rgba,
-               int dst_stride_rgba,
-               int width,
-               int height);
-
-LIBYUV_API
-int I420ToRGB24(const uint8_t* src_y,
-                int src_stride_y,
-                const uint8_t* src_u,
-                int src_stride_u,
-                const uint8_t* src_v,
-                int src_stride_v,
-                uint8_t* dst_rgb24,
-                int dst_stride_rgb24,
-                int width,
-                int height);
-
-LIBYUV_API
-int I420ToRAW(const uint8_t* src_y,
-              int src_stride_y,
-              const uint8_t* src_u,
-              int src_stride_u,
-              const uint8_t* src_v,
-              int src_stride_v,
-              uint8_t* dst_raw,
-              int dst_stride_raw,
-              int width,
-              int height);
-
-LIBYUV_API
-int H420ToRGB24(const uint8_t* src_y,
-                int src_stride_y,
-                const uint8_t* src_u,
-                int src_stride_u,
-                const uint8_t* src_v,
-                int src_stride_v,
-                uint8_t* dst_rgb24,
-                int dst_stride_rgb24,
-                int width,
-                int height);
-
-LIBYUV_API
-int H420ToRAW(const uint8_t* src_y,
-              int src_stride_y,
-              const uint8_t* src_u,
-              int src_stride_u,
-              const uint8_t* src_v,
-              int src_stride_v,
-              uint8_t* dst_raw,
-              int dst_stride_raw,
-              int width,
-              int height);
-
-LIBYUV_API
-int I420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height);
-
-LIBYUV_API
-int J420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height);
-
-LIBYUV_API
-int H420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height);
-
-LIBYUV_API
-int I422ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height);
-
-// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
-// Values in dither matrix from 0 to 7 recommended.
-// The order of the dither matrix is first byte is upper left.
-
-LIBYUV_API
-int I420ToRGB565Dither(const uint8_t* src_y,
-                       int src_stride_y,
-                       const uint8_t* src_u,
-                       int src_stride_u,
-                       const uint8_t* src_v,
-                       int src_stride_v,
-                       uint8_t* dst_rgb565,
-                       int dst_stride_rgb565,
-                       const uint8_t* dither4x4,
-                       int width,
-                       int height);
-
-LIBYUV_API
-int I420ToARGB1555(const uint8_t* src_y,
-                   int src_stride_y,
-                   const uint8_t* src_u,
-                   int src_stride_u,
-                   const uint8_t* src_v,
-                   int src_stride_v,
-                   uint8_t* dst_argb1555,
-                   int dst_stride_argb1555,
-                   int width,
-                   int height);
-
-LIBYUV_API
-int I420ToARGB4444(const uint8_t* src_y,
-                   int src_stride_y,
-                   const uint8_t* src_u,
-                   int src_stride_u,
-                   const uint8_t* src_v,
-                   int src_stride_v,
-                   uint8_t* dst_argb4444,
-                   int dst_stride_argb4444,
-                   int width,
-                   int height);
-
-// Convert I420 to AR30.
-LIBYUV_API
-int I420ToAR30(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_ar30,
-               int dst_stride_ar30,
-               int width,
-               int height);
-
-// Convert H420 to AR30.
-LIBYUV_API
-int H420ToAR30(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_ar30,
-               int dst_stride_ar30,
-               int width,
-               int height);
-
 // Convert I420 to specified format.
 // "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
 //    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
diff --git a/files/include/libyuv/convert_from_argb.h b/files/include/libyuv/convert_from_argb.h
index cbbef6fe..2a488838 100644
--- a/files/include/libyuv/convert_from_argb.h
+++ b/files/include/libyuv/convert_from_argb.h
@@ -77,6 +77,10 @@ int ARGBToAR30(const uint8_t* src_argb,
                int width,
                int height);
 
+// Aliases
+#define ABGRToRGB24 ARGBToRAW
+#define ABGRToRAW ARGBToRGB24
+
 // Convert ARGB To RGB24.
 LIBYUV_API
 int ARGBToRGB24(const uint8_t* src_argb,
@@ -149,6 +153,30 @@ int ARGBToI444(const uint8_t* src_argb,
                int width,
                int height);
 
+// Convert ARGB to AR64.
+LIBYUV_API
+int ARGBToAR64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ar64,
+               int dst_stride_ar64,
+               int width,
+               int height);
+
+// Convert ABGR to AB64.
+#define ABGRToAB64 ARGBToAR64
+
+// Convert ARGB to AB64.
+LIBYUV_API
+int ARGBToAB64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height);
+
+// Convert ABGR to AR64.
+#define ABGRToAR64 ARGBToAB64
+
 // Convert ARGB To I422.
 LIBYUV_API
 int ARGBToI422(const uint8_t* src_argb,
@@ -210,6 +238,15 @@ int ARGBToJ400(const uint8_t* src_argb,
                int width,
                int height);
 
+// Convert RGBA to J400. (JPeg full range).
+LIBYUV_API
+int RGBAToJ400(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height);
+
 // Convert ARGB to I400.
 LIBYUV_API
 int ARGBToI400(const uint8_t* src_argb,
@@ -250,25 +287,25 @@ int ARGBToNV21(const uint8_t* src_argb,
                int width,
                int height);
 
-// Convert ARGB To NV21.
+// Convert ABGR To NV12.
 LIBYUV_API
-int ARGBToNV21(const uint8_t* src_argb,
-               int src_stride_argb,
+int ABGRToNV12(const uint8_t* src_abgr,
+               int src_stride_abgr,
                uint8_t* dst_y,
                int dst_stride_y,
-               uint8_t* dst_vu,
-               int dst_stride_vu,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
                int width,
                int height);
 
-// Convert ABGR To NV12.
+// Convert ABGR To NV21.
 LIBYUV_API
-int ABGRToNV12(const uint8_t* src_abgr,
+int ABGRToNV21(const uint8_t* src_abgr,
                int src_stride_abgr,
                uint8_t* dst_y,
                int dst_stride_y,
-               uint8_t* dst_uv,
-               int dst_stride_uv,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
                int width,
                int height);
 
@@ -290,6 +327,17 @@ int ARGBToUYVY(const uint8_t* src_argb,
                int width,
                int height);
 
+// RAW to JNV21 full range NV21
+LIBYUV_API
+int RAWToJNV21(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/include/libyuv/cpu_id.h b/files/include/libyuv/cpu_id.h
index b01cd25c..fb90c6c7 100644
--- a/files/include/libyuv/cpu_id.h
+++ b/files/include/libyuv/cpu_id.h
@@ -40,15 +40,20 @@ static const int kCpuHasF16C = 0x2000;
 static const int kCpuHasGFNI = 0x4000;
 static const int kCpuHasAVX512BW = 0x8000;
 static const int kCpuHasAVX512VL = 0x10000;
-static const int kCpuHasAVX512VBMI = 0x20000;
-static const int kCpuHasAVX512VBMI2 = 0x40000;
-static const int kCpuHasAVX512VBITALG = 0x80000;
-static const int kCpuHasAVX512VPOPCNTDQ = 0x100000;
+static const int kCpuHasAVX512VNNI = 0x20000;
+static const int kCpuHasAVX512VBMI = 0x40000;
+static const int kCpuHasAVX512VBMI2 = 0x80000;
+static const int kCpuHasAVX512VBITALG = 0x100000;
+static const int kCpuHasAVX512VPOPCNTDQ = 0x200000;
 
 // These flags are only valid on MIPS processors.
-static const int kCpuHasMIPS = 0x200000;
-static const int kCpuHasMSA = 0x400000;
-static const int kCpuHasMMI = 0x800000;
+static const int kCpuHasMIPS = 0x400000;
+static const int kCpuHasMSA = 0x800000;
+
+// These flags are only valid on LOONGARCH processors.
+static const int kCpuHasLOONGARCH = 0x2000000;
+static const int kCpuHasLSX = 0x4000000;
+static const int kCpuHasLASX = 0x8000000;
 
 // Optional init function. TestCpuFlag does an auto-init.
 // Returns cpu_info flags.
@@ -71,6 +76,8 @@ static __inline int TestCpuFlag(int test_flag) {
 // Internal function for parsing /proc/cpuinfo.
 LIBYUV_API
 int ArmCpuCaps(const char* cpuinfo_name);
+LIBYUV_API
+int MipsCpuCaps(const char* cpuinfo_name);
 
 // For testing, allow CPU flags to be disabled.
 // ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
diff --git a/files/include/libyuv/loongson_intrinsics.h b/files/include/libyuv/loongson_intrinsics.h
new file mode 100644
index 00000000..1d613def
--- /dev/null
+++ b/files/include/libyuv/loongson_intrinsics.h
@@ -0,0 +1,1949 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H
+#define INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H
+
+/*
+ * Copyright (c) 2022 Loongson Technology Corporation Limited
+ * All rights reserved.
+ * Contributed by Shiyou Yin <yinshiyou-hf@loongson.cn>
+ *                Xiwei Gu   <guxiwei-hf@loongson.cn>
+ *                Lu Wang    <wanglu@loongson.cn>
+ *
+ * This file is a header file for loongarch builtin extension.
+ *
+ */
+
+#ifndef LOONGSON_INTRINSICS_H
+#define LOONGSON_INTRINSICS_H
+
+/**
+ * MAJOR version: Macro usage changes.
+ * MINOR version: Add new functions, or bug fixes.
+ * MICRO version: Comment changes or implementation changes.
+ */
+#define LSOM_VERSION_MAJOR 1
+#define LSOM_VERSION_MINOR 1
+#define LSOM_VERSION_MICRO 0
+
+#define DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1) \
+  {                                               \
+    _OUT0 = _INS(_IN0);                           \
+    _OUT1 = _INS(_IN1);                           \
+  }
+
+#define DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1) \
+  {                                                           \
+    _OUT0 = _INS(_IN0, _IN1);                                 \
+    _OUT1 = _INS(_IN2, _IN3);                                 \
+  }
+
+#define DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1) \
+  {                                                                       \
+    _OUT0 = _INS(_IN0, _IN1, _IN2);                                       \
+    _OUT1 = _INS(_IN3, _IN4, _IN5);                                       \
+  }
+
+#define DUP4_ARG1(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1, _OUT2, _OUT3) \
+  {                                                                         \
+    DUP2_ARG1(_INS, _IN0, _IN1, _OUT0, _OUT1);                              \
+    DUP2_ARG1(_INS, _IN2, _IN3, _OUT2, _OUT3);                              \
+  }
+
+#define DUP4_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, _OUT0, \
+                  _OUT1, _OUT2, _OUT3)                                         \
+  {                                                                            \
+    DUP2_ARG2(_INS, _IN0, _IN1, _IN2, _IN3, _OUT0, _OUT1);                     \
+    DUP2_ARG2(_INS, _IN4, _IN5, _IN6, _IN7, _OUT2, _OUT3);                     \
+  }
+
+#define DUP4_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _IN6, _IN7, _IN8, \
+                  _IN9, _IN10, _IN11, _OUT0, _OUT1, _OUT2, _OUT3)             \
+  {                                                                           \
+    DUP2_ARG3(_INS, _IN0, _IN1, _IN2, _IN3, _IN4, _IN5, _OUT0, _OUT1);        \
+    DUP2_ARG3(_INS, _IN6, _IN7, _IN8, _IN9, _IN10, _IN11, _OUT2, _OUT3);      \
+  }
+
+#ifdef __loongarch_sx
+#include <lsxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               Then the results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_b(__m128i in_c,
+                                        __m128i in_h,
+                                        __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_b(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               unsigned byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               The results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_bu(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_bu(__m128i in_c,
+                                         __m128i in_h,
+                                         __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_bu(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               The results plus to signed half-word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_bu_b(in_c, in_h, in_l)
+ *        in_c : 1,1,1,1, 1,1,1,1
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : -1,-2,-3,-4, -5,-6,-7,-8, 1,2,3,4, 5,6,7,8
+ *         out : -4,-24,-60,-112, 6,26,62,114
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_h_bu_b(__m128i in_c,
+                                           __m128i in_h,
+                                           __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_h_bu_b(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of half-word vector elements
+ * Arguments   : Inputs  - in_c, in_h, in_l
+ *               Outputs - out
+ *               Return Type - __m128i
+ * Details     : Signed half-word elements from in_h are multiplied by
+ *               signed half-word elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ *               Then the results plus to signed word elements from in_c.
+ * Example     : out = __lsx_vdp2add_h_b(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1
+ *         out : 23,40,41,26
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2add_w_h(__m128i in_c,
+                                        __m128i in_h,
+                                        __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmaddwev_w_h(in_c, in_h, in_l);
+  out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_b(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_b(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_b(in_h, in_l);
+  out = __lsx_vmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               unsigned byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_bu(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_bu(in_h, in_l);
+  out = __lsx_vmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_h_bu_b(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,-1
+ *         out : 22,38,38,22, 22,38,38,6
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_h_bu_b(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_h_bu_b(in_h, in_l);
+  out = __lsx_vmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs  - in_h, in_l
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied by
+ *               signed byte elements from in_l, and then added adjacent to
+ *               each other to get results with the twice size of input.
+ * Example     : out = __lsx_vdp2_w_h(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22
+ * =============================================================================
+ */
+static inline __m128i __lsx_vdp2_w_h(__m128i in_h, __m128i in_l) {
+  __m128i out;
+
+  out = __lsx_vmulwev_w_h(in_h, in_l);
+  out = __lsx_vmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ *               out = ((_in) < (min)) ? (min) : (((_in) > (max)) ? (max) :
+ *               (_in))
+ * Arguments   : Inputs  - _in  (input vector)
+ *                       - min  (min threshold)
+ *                       - max  (max threshold)
+ *               Outputs - out  (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : out = __lsx_vclip_h(_in)
+ *         _in : -8,2,280,249, -8,255,280,249
+ *         min : 1,1,1,1, 1,1,1,1
+ *         max : 9,9,9,9, 9,9,9,9
+ *         out : 1,2,9,9, 1,9,9,9
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip_h(__m128i _in, __m128i min, __m128i max) {
+  __m128i out;
+
+  out = __lsx_vmax_h(min, _in);
+  out = __lsx_vmin_h(max, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments   : Inputs  - _in
+ *               Outputs - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from _in are clamped between 0 and 255.
+ * Example     : out = __lsx_vclip255_h(_in)
+ *         _in : -8,255,280,249, -8,255,280,249
+ *         out : 0,255,255,249, 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_h(__m128i _in) {
+  __m128i out;
+
+  out = __lsx_vmaxi_h(_in, 0);
+  out = __lsx_vsat_hu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Set each element of vector between 0 and 255
+ * Arguments   : Inputs  - _in
+ *               Outputs - out
+ *               Return Type - word
+ * Details     : Signed byte elements from _in are clamped between 0 and 255.
+ * Example     : out = __lsx_vclip255_w(_in)
+ *         _in : -8,255,280,249
+ *         out : 0,255,255,249
+ * =============================================================================
+ */
+static inline __m128i __lsx_vclip255_w(__m128i _in) {
+  __m128i out;
+
+  out = __lsx_vmaxi_w(_in, 0);
+  out = __lsx_vsat_wu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Swap two variables
+ * Arguments   : Inputs  - _in0, _in1
+ *               Outputs - _in0, _in1 (in-place)
+ * Details     : Swapping of two input variables using xor
+ * Example     : LSX_SWAP(_in0, _in1)
+ *        _in0 : 1,2,3,4
+ *        _in1 : 5,6,7,8
+ *   _in0(out) : 5,6,7,8
+ *   _in1(out) : 1,2,3,4
+ * =============================================================================
+ */
+#define LSX_SWAP(_in0, _in1)         \
+  {                                  \
+    _in0 = __lsx_vxor_v(_in0, _in1); \
+    _in1 = __lsx_vxor_v(_in0, _in1); \
+    _in0 = __lsx_vxor_v(_in0, _in1); \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with word elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3
+ *               Outputs - out0, out1, out2, out3
+ * Details     :
+ * Example     :
+ *               1, 2, 3, 4            1, 5, 9,13
+ *               5, 6, 7, 8    to      2, 6,10,14
+ *               9,10,11,12  =====>    3, 7,11,15
+ *              13,14,15,16            4, 8,12,16
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE4x4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    __m128i _t0, _t1, _t2, _t3;                                                \
+                                                                               \
+    _t0 = __lsx_vilvl_w(_in1, _in0);                                           \
+    _t1 = __lsx_vilvh_w(_in1, _in0);                                           \
+    _t2 = __lsx_vilvl_w(_in3, _in2);                                           \
+    _t3 = __lsx_vilvh_w(_in3, _in2);                                           \
+    _out0 = __lsx_vilvl_d(_t2, _t0);                                           \
+    _out1 = __lsx_vilvh_d(_t2, _t0);                                           \
+    _out2 = __lsx_vilvl_d(_t3, _t1);                                           \
+    _out3 = __lsx_vilvh_d(_t3, _t1);                                           \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with byte elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *               _out7
+ * Details     : The rows of the matrix become columns, and the columns
+ *               become rows.
+ * Example     : LSX_TRANSPOSE8x8_B
+ *        _in0 : 00,01,02,03,04,05,06,07, 00,00,00,00,00,00,00,00
+ *        _in1 : 10,11,12,13,14,15,16,17, 00,00,00,00,00,00,00,00
+ *        _in2 : 20,21,22,23,24,25,26,27, 00,00,00,00,00,00,00,00
+ *        _in3 : 30,31,32,33,34,35,36,37, 00,00,00,00,00,00,00,00
+ *        _in4 : 40,41,42,43,44,45,46,47, 00,00,00,00,00,00,00,00
+ *        _in5 : 50,51,52,53,54,55,56,57, 00,00,00,00,00,00,00,00
+ *        _in6 : 60,61,62,63,64,65,66,67, 00,00,00,00,00,00,00,00
+ *        _in7 : 70,71,72,73,74,75,76,77, 00,00,00,00,00,00,00,00
+ *
+ *      _ out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ *      _ out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ *      _ out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ *      _ out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ *      _ out4 : 04,14,24,34,44,54,64,74, 00,00,00,00,00,00,00,00
+ *      _ out5 : 05,15,25,35,45,55,65,75, 00,00,00,00,00,00,00,00
+ *      _ out6 : 06,16,26,36,46,56,66,76, 00,00,00,00,00,00,00,00
+ *      _ out7 : 07,17,27,37,47,57,67,77, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    __m128i zero = {0};                                                     \
+    __m128i shuf8 = {0x0F0E0D0C0B0A0908, 0x1716151413121110};               \
+    __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                         \
+                                                                            \
+    _t0 = __lsx_vilvl_b(_in2, _in0);                                        \
+    _t1 = __lsx_vilvl_b(_in3, _in1);                                        \
+    _t2 = __lsx_vilvl_b(_in6, _in4);                                        \
+    _t3 = __lsx_vilvl_b(_in7, _in5);                                        \
+    _t4 = __lsx_vilvl_b(_t1, _t0);                                          \
+    _t5 = __lsx_vilvh_b(_t1, _t0);                                          \
+    _t6 = __lsx_vilvl_b(_t3, _t2);                                          \
+    _t7 = __lsx_vilvh_b(_t3, _t2);                                          \
+    _out0 = __lsx_vilvl_w(_t6, _t4);                                        \
+    _out2 = __lsx_vilvh_w(_t6, _t4);                                        \
+    _out4 = __lsx_vilvl_w(_t7, _t5);                                        \
+    _out6 = __lsx_vilvh_w(_t7, _t5);                                        \
+    _out1 = __lsx_vshuf_b(zero, _out0, shuf8);                              \
+    _out3 = __lsx_vshuf_b(zero, _out2, shuf8);                              \
+    _out5 = __lsx_vshuf_b(zero, _out4, shuf8);                              \
+    _out7 = __lsx_vshuf_b(zero, _out6, shuf8);                              \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with half-word elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+ *               Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details     :
+ * Example     :
+ *              00,01,02,03,04,05,06,07           00,10,20,30,40,50,60,70
+ *              10,11,12,13,14,15,16,17           01,11,21,31,41,51,61,71
+ *              20,21,22,23,24,25,26,27           02,12,22,32,42,52,62,72
+ *              30,31,32,33,34,35,36,37    to     03,13,23,33,43,53,63,73
+ *              40,41,42,43,44,45,46,47  ======>  04,14,24,34,44,54,64,74
+ *              50,51,52,53,54,55,56,57           05,15,25,35,45,55,65,75
+ *              60,61,62,63,64,65,66,67           06,16,26,36,46,56,66,76
+ *              70,71,72,73,74,75,76,77           07,17,27,37,47,57,67,77
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    __m128i _s0, _s1, _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;               \
+                                                                            \
+    _s0 = __lsx_vilvl_h(_in6, _in4);                                        \
+    _s1 = __lsx_vilvl_h(_in7, _in5);                                        \
+    _t0 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t1 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvh_h(_in6, _in4);                                        \
+    _s1 = __lsx_vilvh_h(_in7, _in5);                                        \
+    _t2 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t3 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvl_h(_in2, _in0);                                        \
+    _s1 = __lsx_vilvl_h(_in3, _in1);                                        \
+    _t4 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t5 = __lsx_vilvh_h(_s1, _s0);                                          \
+    _s0 = __lsx_vilvh_h(_in2, _in0);                                        \
+    _s1 = __lsx_vilvh_h(_in3, _in1);                                        \
+    _t6 = __lsx_vilvl_h(_s1, _s0);                                          \
+    _t7 = __lsx_vilvh_h(_s1, _s0);                                          \
+                                                                            \
+    _out0 = __lsx_vpickev_d(_t0, _t4);                                      \
+    _out2 = __lsx_vpickev_d(_t1, _t5);                                      \
+    _out4 = __lsx_vpickev_d(_t2, _t6);                                      \
+    _out6 = __lsx_vpickev_d(_t3, _t7);                                      \
+    _out1 = __lsx_vpickod_d(_t0, _t4);                                      \
+    _out3 = __lsx_vpickod_d(_t1, _t5);                                      \
+    _out5 = __lsx_vpickod_d(_t2, _t6);                                      \
+    _out7 = __lsx_vpickod_d(_t3, _t7);                                      \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x4 byte block into 4x8
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3      (input 8x4 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3  (output 4x8 byte block)
+ *               Return Type - as per RTYPE
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LSX_TRANSPOSE8x4_B
+ *        _in0 : 00,01,02,03,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in1 : 10,11,12,13,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in2 : 20,21,22,23,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in3 : 30,31,32,33,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in4 : 40,41,42,43,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in5 : 50,51,52,53,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in6 : 60,61,62,63,00,00,00,00, 00,00,00,00,00,00,00,00
+ *        _in7 : 70,71,72,73,00,00,00,00, 00,00,00,00,00,00,00,00
+ *
+ *       _out0 : 00,10,20,30,40,50,60,70, 00,00,00,00,00,00,00,00
+ *       _out1 : 01,11,21,31,41,51,61,71, 00,00,00,00,00,00,00,00
+ *       _out2 : 02,12,22,32,42,52,62,72, 00,00,00,00,00,00,00,00
+ *       _out3 : 03,13,23,33,43,53,63,73, 00,00,00,00,00,00,00,00
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE8x4_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7, \
+                           _out0, _out1, _out2, _out3)                     \
+  {                                                                        \
+    __m128i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                            \
+                                                                           \
+    _tmp0_m = __lsx_vpackev_w(_in4, _in0);                                 \
+    _tmp1_m = __lsx_vpackev_w(_in5, _in1);                                 \
+    _tmp2_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp0_m = __lsx_vpackev_w(_in6, _in2);                                 \
+    _tmp1_m = __lsx_vpackev_w(_in7, _in3);                                 \
+                                                                           \
+    _tmp3_m = __lsx_vilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp0_m = __lsx_vilvl_h(_tmp3_m, _tmp2_m);                             \
+    _tmp1_m = __lsx_vilvh_h(_tmp3_m, _tmp2_m);                             \
+                                                                           \
+    _out0 = __lsx_vilvl_w(_tmp1_m, _tmp0_m);                               \
+    _out2 = __lsx_vilvh_w(_tmp1_m, _tmp0_m);                               \
+    _out1 = __lsx_vilvh_d(_out2, _out0);                                   \
+    _out3 = __lsx_vilvh_d(_out0, _out2);                                   \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 16x8 block with byte elements in vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7, in8
+ *                         in9, in10, in11, in12, in13, in14, in15
+ *               Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+ * Details     :
+ * Example     :
+ *              000,001,002,003,004,005,006,007
+ *              008,009,010,011,012,013,014,015
+ *              016,017,018,019,020,021,022,023
+ *              024,025,026,027,028,029,030,031
+ *              032,033,034,035,036,037,038,039
+ *              040,041,042,043,044,045,046,047        000,008,...,112,120
+ *              048,049,050,051,052,053,054,055        001,009,...,113,121
+ *              056,057,058,059,060,061,062,063   to   002,010,...,114,122
+ *              064,068,066,067,068,069,070,071 =====> 003,011,...,115,123
+ *              072,073,074,075,076,077,078,079        004,012,...,116,124
+ *              080,081,082,083,084,085,086,087        005,013,...,117,125
+ *              088,089,090,091,092,093,094,095        006,014,...,118,126
+ *              096,097,098,099,100,101,102,103        007,015,...,119,127
+ *              104,105,106,107,108,109,110,111
+ *              112,113,114,115,116,117,118,119
+ *              120,121,122,123,124,125,126,127
+ * =============================================================================
+ */
+#define LSX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                            _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                            _out6, _out7)                                    \
+  {                                                                          \
+    __m128i _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5, _tmp6, _tmp7;          \
+    __m128i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                          \
+    DUP4_ARG2(__lsx_vilvl_b, _in2, _in0, _in3, _in1, _in6, _in4, _in7, _in5, \
+              _tmp0, _tmp1, _tmp2, _tmp3);                                   \
+    DUP4_ARG2(__lsx_vilvl_b, _in10, _in8, _in11, _in9, _in14, _in12, _in15,  \
+              _in13, _tmp4, _tmp5, _tmp6, _tmp7);                            \
+    DUP2_ARG2(__lsx_vilvl_b, _tmp1, _tmp0, _tmp3, _tmp2, _t0, _t2);          \
+    DUP2_ARG2(__lsx_vilvh_b, _tmp1, _tmp0, _tmp3, _tmp2, _t1, _t3);          \
+    DUP2_ARG2(__lsx_vilvl_b, _tmp5, _tmp4, _tmp7, _tmp6, _t4, _t6);          \
+    DUP2_ARG2(__lsx_vilvh_b, _tmp5, _tmp4, _tmp7, _tmp6, _t5, _t7);          \
+    DUP2_ARG2(__lsx_vilvl_w, _t2, _t0, _t3, _t1, _tmp0, _tmp4);              \
+    DUP2_ARG2(__lsx_vilvh_w, _t2, _t0, _t3, _t1, _tmp2, _tmp6);              \
+    DUP2_ARG2(__lsx_vilvl_w, _t6, _t4, _t7, _t5, _tmp1, _tmp5);              \
+    DUP2_ARG2(__lsx_vilvh_w, _t6, _t4, _t7, _t5, _tmp3, _tmp7);              \
+    DUP2_ARG2(__lsx_vilvl_d, _tmp1, _tmp0, _tmp3, _tmp2, _out0, _out2);      \
+    DUP2_ARG2(__lsx_vilvh_d, _tmp1, _tmp0, _tmp3, _tmp2, _out1, _out3);      \
+    DUP2_ARG2(__lsx_vilvl_d, _tmp5, _tmp4, _tmp7, _tmp6, _out4, _out6);      \
+    DUP2_ARG2(__lsx_vilvh_d, _tmp5, _tmp4, _tmp7, _tmp6, _out5, _out7);      \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments   : Inputs  - in0, in1, in2, in3
+ *               Outputs - out0, out1, out2, out3
+ * Details     : Butterfly operation
+ * Example     :
+ *               out0 = in0 + in3;
+ *               out1 = in1 + in2;
+ *               out2 = in1 - in2;
+ *               out3 = in0 - in3;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_b(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_b(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_b(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_b(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_h(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_h(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_h(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_h(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_w(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_w(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_w(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_w(_in0, _in3);                                         \
+  }
+#define LSX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                           \
+    _out0 = __lsx_vadd_d(_in0, _in3);                                         \
+    _out1 = __lsx_vadd_d(_in1, _in2);                                         \
+    _out2 = __lsx_vsub_d(_in1, _in2);                                         \
+    _out3 = __lsx_vsub_d(_in0, _in3);                                         \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, ~
+ *               Outputs - _out0, _out1, _out2, _out3, ~
+ * Details     : Butterfly operation
+ * Example     :
+ *              _out0 = _in0 + _in7;
+ *              _out1 = _in1 + _in6;
+ *              _out2 = _in2 + _in5;
+ *              _out3 = _in3 + _in4;
+ *              _out4 = _in3 - _in4;
+ *              _out5 = _in2 - _in5;
+ *              _out6 = _in1 - _in6;
+ *              _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LSX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_b(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_b(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_b(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_b(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_b(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_b(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_b(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_b(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_h(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_h(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_h(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_h(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_h(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_h(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_h(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_h(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_w(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_w(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_w(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_w(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_w(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_w(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_w(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_w(_in0, _in7);                                      \
+  }
+
+#define LSX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                          _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                          _out7)                                           \
+  {                                                                        \
+    _out0 = __lsx_vadd_d(_in0, _in7);                                      \
+    _out1 = __lsx_vadd_d(_in1, _in6);                                      \
+    _out2 = __lsx_vadd_d(_in2, _in5);                                      \
+    _out3 = __lsx_vadd_d(_in3, _in4);                                      \
+    _out4 = __lsx_vsub_d(_in3, _in4);                                      \
+    _out5 = __lsx_vsub_d(_in2, _in5);                                      \
+    _out6 = __lsx_vsub_d(_in1, _in6);                                      \
+    _out7 = __lsx_vsub_d(_in0, _in7);                                      \
+  }
+
+#endif  // LSX
+
+#ifdef __loongarch_asx
+#include <lasxintrin.h>
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_bu(in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Signed byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplication results of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_b(in_h, in_l);
+  out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector.
+ * Example     : out = __lasx_xvdp2_w_h(in_h, in_l)
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1
+ *         out : 22,38,38,22, 22,38,38,22
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of word vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed double
+ * Details     : Signed word elements from in_h are multiplied with
+ *               signed word elements from in_l producing a result
+ *               twice the size of input i.e. signed double-word.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the out vector.
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_d_w(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_d_w(in_h, in_l);
+  out = __lasx_xvmaddwod_d_w(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. unsigned word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the out vector
+ * Example     : See out = __lasx_xvdp2_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2_w_hu_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_hu_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Signed byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_b(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_b(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_bu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_bu(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product & addition of byte vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               signed byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Then this multiplied results of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_h_bu_b(__m256i in_c,
+                                             __m256i in_h,
+                                             __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_h_bu_b(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu_b(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - per RTYPE
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 1,2,3,4
+ *        in_h : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8,
+ *        in_l : 8,7,6,5, 4,3,2,1, 8,7,6,5, 4,3,2,1,
+ *         out : 23,40,41,26, 23,40,41,26
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_h(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               unsigned halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector.
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_hu(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Unsigned halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added to the in_c vector
+ * Example     : See out = __lasx_xvdp2add_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2add_w_hu_h(__m256i in_c,
+                                             __m256i in_h,
+                                             __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmaddwev_w_hu_h(in_c, in_h, in_l);
+  out = __lasx_xvmaddwod_w_hu_h(out, in_h, in_l);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Unsigned Dot Product and Subtract
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed halfword
+ * Details     : Unsigned byte elements from in_h are multiplied with
+ *               unsigned byte elements from in_l producing a result
+ *               twice the size of input i.e. signed halfword.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added together and subtracted from double width elements
+ *               in_c vector.
+ * Example     : See out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_h_bu(__m256i in_c,
+                                           __m256i in_h,
+                                           __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_h_bu(in_h, in_l);
+  out = __lasx_xvmaddwod_h_bu(out, in_h, in_l);
+  out = __lasx_xvsub_h(in_c, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Vector Signed Dot Product and Subtract
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               Signed halfword elements from in_l producing a result
+ *               twice the size of input i.e. signed word.
+ *               Multiplication result of adjacent odd-even elements
+ *               are added together and subtracted from double width elements
+ *               in_c vector.
+ * Example     : out = __lasx_xvdp2sub_w_h(in_c, in_h, in_l)
+ *        in_c : 0,0,0,0, 0,0,0,0
+ *        in_h : 3,1,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1
+ *        in_l : 2,1,1,0, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ *         out : -7,-3,0,0, 0,-1,0,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp2sub_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  out = __lasx_xvsub_w(in_c, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Dot product of halfword vector elements
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ *               Return Type - signed word
+ * Details     : Signed halfword elements from in_h are multiplied with
+ *               signed halfword elements from in_l producing a result
+ *               four times the size of input i.e. signed doubleword.
+ *               Then this multiplication results of four adjacent elements
+ *               are added together and stored to the out vector.
+ * Example     : out = __lasx_xvdp4_d_h(in_h, in_l)
+ *        in_h :  3,1,3,0, 0,0,0,1, 0,0,1,-1, 0,0,0,1
+ *        in_l : -2,1,1,0, 1,0,0,0, 0,0,1, 0, 1,0,0,1
+ *         out : -2,0,1,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvdp4_d_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvmulwev_w_h(in_h, in_l);
+  out = __lasx_xvmaddwod_w_h(out, in_h, in_l);
+  out = __lasx_xvhaddw_d_w(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               higher half of the two-fold sign extension (signed byte
+ *               to signed halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwh_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwh_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvh_b(in_h, in_l);
+  out = __lasx_xvhaddw_h_b(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The high half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               higher half of the two-fold sign extension (signed halfword
+ *               to signed word) and stored to the out vector.
+ * Example     : out = __lasx_xvaddwh_w_h(in_h, in_l)
+ *        in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ *         out : 1,0,0,-1, 1,0,0, 2
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwh_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvh_h(in_h, in_l);
+  out = __lasx_xvhaddw_w_h(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               lower half of the two-fold sign extension (signed byte
+ *               to signed halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_b(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_b(in_h, in_l);
+  out = __lasx_xvhaddw_h_b(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are added after the
+ *               lower half of the two-fold sign extension (signed halfword
+ *               to signed word) and stored to the out vector.
+ * Example     : out = __lasx_xvaddwl_w_h(in_h, in_l)
+ *        in_h : 3, 0,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 1,0,1, 0, 1,0,0,1
+ *         out : 5,-1,4,2, 1,0,2,-1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_h(in_h, in_l);
+  out = __lasx_xvhaddw_w_h(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The out vector and the out vector are added after the
+ *               lower half of the two-fold zero extension (unsigned byte
+ *               to unsigned halfword) and stored to the out vector.
+ * Example     : See out = __lasx_xvaddwl_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddwl_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvilvl_b(in_h, in_l);
+  out = __lasx_xvhaddw_hu_bu(out, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_l vector after double zero extension (unsigned byte to
+ *               signed halfword)，added to the in_h vector.
+ * Example     : See out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_h_h_bu(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvsllwil_hu_bu(in_l, 0);
+  out = __lasx_xvadd_h(in_h, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are expanded and
+ *               added after being doubled.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_l vector after double sign extension (signed halfword to
+ *               signed word), added to the in_h vector.
+ * Example     : out = __lasx_xvaddw_w_w_h(in_h, in_l)
+ *        in_h : 0, 1,0,0, -1,0,0,1,
+ *        in_l : 2,-1,1,2,  1,0,0,0, 0,0,1,0, 1,0,0,1,
+ *         out : 2, 0,1,2, -1,0,1,1,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvaddw_w_w_h(__m256i in_h, __m256i in_l) {
+  __m256i out;
+
+  out = __lasx_xvsllwil_w_h(in_l, 0);
+  out = __lasx_xvadd_w(in_h, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ *               of the lower half of the vector.
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed halfword
+ *               to signed word), and the result is added to the vector in_c,
+ *               then stored to the out vector.
+ * Example     : out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ *        in_c : 1,2,3,4, 5,6,7,8
+ *        in_h : 1,2,3,4, 1,2,3,4, 5,6,7,8, 5,6,7,8
+ *        in_l : 200, 300, 400, 500,  2000, 3000, 4000, 5000,
+ *              -200,-300,-400,-500, -2000,-3000,-4000,-5000
+ *         out : 201, 602,1203,2004, -995, -1794,-2793,-3992
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwl_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+  tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+  tmp0 = __lasx_xvmul_w(tmp0, tmp1);
+  out = __lasx_xvadd_w(tmp0, in_c);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication and addition calculation after expansion
+ *               of the higher half of the vector.
+ * Arguments   : Inputs - in_c, in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the higher half of the two-fold sign extension (signed
+ *               halfword to signed word), and the result is added to
+ *               the vector in_c, then stored to the out vector.
+ * Example     : See out = __lasx_xvmaddwl_w_h(in_c, in_h, in_l)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmaddwh_w_h(__m256i in_c,
+                                          __m256i in_h,
+                                          __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvilvh_h(in_h, in_h);
+  tmp1 = __lasx_xvilvh_h(in_l, in_l);
+  tmp0 = __lasx_xvmulwev_w_h(tmp0, tmp1);
+  out = __lasx_xvadd_w(tmp0, in_c);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ *               half of the vector.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed
+ *               halfword to signed word), then stored to the out vector.
+ * Example     : out = __lasx_xvmulwl_w_h(in_h, in_l)
+ *        in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ *         out : 6,1,3,0, 0,0,1,0
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwl_w_h(__m256i in_h, __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvsllwil_w_h(in_h, 0);
+  tmp1 = __lasx_xvsllwil_w_h(in_l, 0);
+  out = __lasx_xvmul_w(tmp0, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Multiplication calculation after expansion of the lower
+ *               half of the vector.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector and the in_l vector are multiplied after
+ *               the lower half of the two-fold sign extension (signed
+ *               halfword to signed word), then stored to the out vector.
+ * Example     : out = __lasx_xvmulwh_w_h(in_h, in_l)
+ *        in_h : 3,-1,3,0, 0,0,0,-1, 0,0,1,-1, 0,0,0,1
+ *        in_l : 2,-1,1,2, 1,0,0, 0, 0,0,1, 0, 1,0,0,1
+ *         out : 0,0,0,0, 0,0,0,1
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvmulwh_w_h(__m256i in_h, __m256i in_l) {
+  __m256i tmp0, tmp1, out;
+
+  tmp0 = __lasx_xvilvh_h(in_h, in_h);
+  tmp1 = __lasx_xvilvh_h(in_l, in_l);
+  out = __lasx_xvmulwev_w_h(tmp0, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : The low half of the vector elements are added to the high half
+ *               after being doubled, then saturated.
+ * Arguments   : Inputs - in_h, in_l
+ *               Output - out
+ * Details     : The in_h vector adds the in_l vector after the lower half of
+ *               the two-fold zero extension (unsigned byte to unsigned
+ *               halfword) and then saturated. The results are stored to the out
+ *               vector.
+ * Example     : out = __lasx_xvsaddw_hu_hu_bu(in_h, in_l)
+ *        in_h : 2,65532,1,2, 1,0,0,0, 0,0,1,0, 1,0,0,1
+ *        in_l : 3,6,3,0, 0,0,0,1, 0,0,1,1, 0,0,0,1, 3,18,3,0, 0,0,0,1, 0,0,1,1,
+ *               0,0,0,1
+ *        out  : 5,65535,4,2, 1,0,0,1, 3,18,4,0, 1,0,0,2,
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsaddw_hu_hu_bu(__m256i in_h, __m256i in_l) {
+  __m256i tmp1, out;
+  __m256i zero = {0};
+
+  tmp1 = __lasx_xvilvl_b(zero, in_l);
+  out = __lasx_xvsadd_hu(in_h, tmp1);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all halfword elements of input vector between min & max
+ *               out = ((in) < (min)) ? (min) : (((in) > (max)) ? (max) : (in))
+ * Arguments   : Inputs  - in    (input vector)
+ *                       - min   (min threshold)
+ *                       - max   (max threshold)
+ *               Outputs - in    (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : out = __lasx_xvclip_h(in, min, max)
+ *          in : -8,2,280,249, -8,255,280,249, 4,4,4,4, 5,5,5,5
+ *         min : 1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1
+ *         max : 9,9,9,9, 9,9,9,9, 9,9,9,9, 9,9,9,9
+ *         out : 1,2,9,9, 1,9,9,9, 4,4,4,4, 5,5,5,5
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip_h(__m256i in, __m256i min, __m256i max) {
+  __m256i out;
+
+  out = __lasx_xvmax_h(min, in);
+  out = __lasx_xvmin_h(max, out);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed halfword elements of input vector
+ *               between 0 & 255
+ * Arguments   : Inputs  - in   (input vector)
+ *               Outputs - out  (output vector with clipped elements)
+ *               Return Type - signed halfword
+ * Example     : See out = __lasx_xvclip255_w(in)
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_h(__m256i in) {
+  __m256i out;
+
+  out = __lasx_xvmaxi_h(in, 0);
+  out = __lasx_xvsat_hu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Clip all signed word elements of input vector
+ *               between 0 & 255
+ * Arguments   : Inputs - in   (input vector)
+ *               Output - out  (output vector with clipped elements)
+ *               Return Type - signed word
+ * Example     : out = __lasx_xvclip255_w(in)
+ *          in : -8,255,280,249, -8,255,280,249
+ *         out :  0,255,255,249,  0,255,255,249
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvclip255_w(__m256i in) {
+  __m256i out;
+
+  out = __lasx_xvmaxi_w(in, 0);
+  out = __lasx_xvsat_wu(out, 7);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ *               elements in output vector. If 'idx < 8' use xvsplati_l_*,
+ *               if 'idx >= 8' use xvsplati_h_*.
+ * Arguments   : Inputs - in, idx
+ *               Output - out
+ * Details     : Idx element value from in vector is replicated to all
+ *               elements in out vector.
+ *               Valid index range for halfword operation is 0-7
+ * Example     : out = __lasx_xvsplati_l_h(in, idx)
+ *          in : 20,10,11,12, 13,14,15,16, 0,0,2,0, 0,0,0,0
+ *         idx : 0x02
+ *         out : 11,11,11,11, 11,11,11,11, 11,11,11,11, 11,11,11,11
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_l_h(__m256i in, int idx) {
+  __m256i out;
+
+  out = __lasx_xvpermi_q(in, in, 0x02);
+  out = __lasx_xvreplve_h(out, idx);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Indexed halfword element values are replicated to all
+ *               elements in output vector. If 'idx < 8' use xvsplati_l_*,
+ *               if 'idx >= 8' use xvsplati_h_*.
+ * Arguments   : Inputs - in, idx
+ *               Output - out
+ * Details     : Idx element value from in vector is replicated to all
+ *               elements in out vector.
+ *               Valid index range for halfword operation is 0-7
+ * Example     : out = __lasx_xvsplati_h_h(in, idx)
+ *          in : 20,10,11,12, 13,14,15,16, 0,2,0,0, 0,0,0,0
+ *         idx : 0x09
+ *         out : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ * =============================================================================
+ */
+static inline __m256i __lasx_xvsplati_h_h(__m256i in, int idx) {
+  __m256i out;
+
+  out = __lasx_xvpermi_q(in, in, 0x13);
+  out = __lasx_xvreplve_h(out, idx);
+  return out;
+}
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with double-word elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ * Example     : LASX_TRANSPOSE4x4_D
+ *        _in0 : 1,2,3,4
+ *        _in1 : 1,2,3,4
+ *        _in2 : 1,2,3,4
+ *        _in3 : 1,2,3,4
+ *
+ *       _out0 : 1,1,1,1
+ *       _out1 : 2,2,2,2
+ *       _out2 : 3,3,3,3
+ *       _out3 : 4,4,4,4
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, \
+                            _out3)                                       \
+  {                                                                      \
+    __m256i _tmp0, _tmp1, _tmp2, _tmp3;                                  \
+    _tmp0 = __lasx_xvilvl_d(_in1, _in0);                                 \
+    _tmp1 = __lasx_xvilvh_d(_in1, _in0);                                 \
+    _tmp2 = __lasx_xvilvl_d(_in3, _in2);                                 \
+    _tmp3 = __lasx_xvilvh_d(_in3, _in2);                                 \
+    _out0 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x20);                        \
+    _out2 = __lasx_xvpermi_q(_tmp2, _tmp0, 0x31);                        \
+    _out1 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x20);                        \
+    _out3 = __lasx_xvpermi_q(_tmp3, _tmp1, 0x31);                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with word elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *               _out7
+ * Example     : LASX_TRANSPOSE8x8_W
+ *        _in0 : 1,2,3,4,5,6,7,8
+ *        _in1 : 2,2,3,4,5,6,7,8
+ *        _in2 : 3,2,3,4,5,6,7,8
+ *        _in3 : 4,2,3,4,5,6,7,8
+ *        _in4 : 5,2,3,4,5,6,7,8
+ *        _in5 : 6,2,3,4,5,6,7,8
+ *        _in6 : 7,2,3,4,5,6,7,8
+ *        _in7 : 8,2,3,4,5,6,7,8
+ *
+ *       _out0 : 1,2,3,4,5,6,7,8
+ *       _out1 : 2,2,2,2,2,2,2,2
+ *       _out2 : 3,3,3,3,3,3,3,3
+ *       _out3 : 4,4,4,4,4,4,4,4
+ *       _out4 : 5,5,5,5,5,5,5,5
+ *       _out5 : 6,6,6,6,6,6,6,6
+ *       _out6 : 7,7,7,7,7,7,7,7
+ *       _out7 : 8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _s0_m, _s1_m;                                                    \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+                                                                             \
+    _s0_m = __lasx_xvilvl_w(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvl_w(_in3, _in1);                                     \
+    _tmp0_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp1_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_w(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvh_w(_in3, _in1);                                     \
+    _tmp2_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp3_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvl_w(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvl_w(_in7, _in5);                                     \
+    _tmp4_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp5_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_w(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvh_w(_in7, _in5);                                     \
+    _tmp6_m = __lasx_xvilvl_w(_s1_m, _s0_m);                                 \
+    _tmp7_m = __lasx_xvilvh_w(_s1_m, _s0_m);                                 \
+    _out0 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x20);                        \
+    _out1 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x20);                        \
+    _out2 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x20);                        \
+    _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x20);                        \
+    _out4 = __lasx_xvpermi_q(_tmp4_m, _tmp0_m, 0x31);                        \
+    _out5 = __lasx_xvpermi_q(_tmp5_m, _tmp1_m, 0x31);                        \
+    _out6 = __lasx_xvpermi_q(_tmp6_m, _tmp2_m, 0x31);                        \
+    _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp3_m, 0x31);                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ *                         _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ *                         (input 16x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x16 byte block)
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : See LASX_TRANSPOSE16x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                             _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                             _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                             _out6, _out7)                                    \
+  {                                                                           \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                               \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                               \
+                                                                              \
+    _tmp0_m = __lasx_xvilvl_b(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvl_b(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvl_b(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvl_b(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvl_b(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvl_b(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvl_b(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvl_b(_in15, _in13);                                  \
+    _out0 = __lasx_xvilvl_b(_tmp1_m, _tmp0_m);                                \
+    _out1 = __lasx_xvilvh_b(_tmp1_m, _tmp0_m);                                \
+    _out2 = __lasx_xvilvl_b(_tmp3_m, _tmp2_m);                                \
+    _out3 = __lasx_xvilvh_b(_tmp3_m, _tmp2_m);                                \
+    _out4 = __lasx_xvilvl_b(_tmp5_m, _tmp4_m);                                \
+    _out5 = __lasx_xvilvh_b(_tmp5_m, _tmp4_m);                                \
+    _out6 = __lasx_xvilvl_b(_tmp7_m, _tmp6_m);                                \
+    _out7 = __lasx_xvilvh_b(_tmp7_m, _tmp6_m);                                \
+    _tmp0_m = __lasx_xvilvl_w(_out2, _out0);                                  \
+    _tmp2_m = __lasx_xvilvh_w(_out2, _out0);                                  \
+    _tmp4_m = __lasx_xvilvl_w(_out3, _out1);                                  \
+    _tmp6_m = __lasx_xvilvh_w(_out3, _out1);                                  \
+    _tmp1_m = __lasx_xvilvl_w(_out6, _out4);                                  \
+    _tmp3_m = __lasx_xvilvh_w(_out6, _out4);                                  \
+    _tmp5_m = __lasx_xvilvl_w(_out7, _out5);                                  \
+    _tmp7_m = __lasx_xvilvh_w(_out7, _out5);                                  \
+    _out0 = __lasx_xvilvl_d(_tmp1_m, _tmp0_m);                                \
+    _out1 = __lasx_xvilvh_d(_tmp1_m, _tmp0_m);                                \
+    _out2 = __lasx_xvilvl_d(_tmp3_m, _tmp2_m);                                \
+    _out3 = __lasx_xvilvh_d(_tmp3_m, _tmp2_m);                                \
+    _out4 = __lasx_xvilvl_d(_tmp5_m, _tmp4_m);                                \
+    _out5 = __lasx_xvilvh_d(_tmp5_m, _tmp4_m);                                \
+    _out6 = __lasx_xvilvl_d(_tmp7_m, _tmp6_m);                                \
+    _out7 = __lasx_xvilvh_d(_tmp7_m, _tmp6_m);                                \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 16x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,
+ *                         _in8, _in9, _in10, _in11, _in12, _in13, _in14, _in15
+ *                         (input 16x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x16 byte block)
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LASX_TRANSPOSE16x8_H
+ *        _in0 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in1 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in2 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in3 : 4,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in4 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in5 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in6 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in7 : 8,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in8 : 9,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *        _in9 : 1,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in10 : 0,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in11 : 2,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in12 : 3,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in13 : 7,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in14 : 5,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *       _in15 : 6,2,3,4,5,6,7,8,0,0,0,0,0,0,0,0
+ *
+ *       _out0 : 1,2,3,4,5,6,7,8,9,1,0,2,3,7,5,6
+ *       _out1 : 2,2,2,2,2,2,2,2,2,2,2,2,2,2,2,2
+ *       _out2 : 3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3
+ *       _out3 : 4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4
+ *       _out4 : 5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+ *       _out5 : 6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6
+ *       _out6 : 7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
+ *       _out7 : 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE16x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                             _in8, _in9, _in10, _in11, _in12, _in13, _in14,   \
+                             _in15, _out0, _out1, _out2, _out3, _out4, _out5, \
+                             _out6, _out7)                                    \
+  {                                                                           \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                               \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                               \
+    __m256i _t0, _t1, _t2, _t3, _t4, _t5, _t6, _t7;                           \
+                                                                              \
+    _tmp0_m = __lasx_xvilvl_h(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvl_h(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvl_h(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvl_h(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvl_h(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvl_h(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvl_h(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvl_h(_in15, _in13);                                  \
+    _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m);                                  \
+    _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m);                                  \
+    _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m);                                  \
+    _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m);                                  \
+    _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m);                                  \
+    _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m);                                  \
+    _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m);                                  \
+    _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m);                                  \
+    _tmp0_m = __lasx_xvilvl_d(_t2, _t0);                                      \
+    _tmp2_m = __lasx_xvilvh_d(_t2, _t0);                                      \
+    _tmp4_m = __lasx_xvilvl_d(_t3, _t1);                                      \
+    _tmp6_m = __lasx_xvilvh_d(_t3, _t1);                                      \
+    _tmp1_m = __lasx_xvilvl_d(_t6, _t4);                                      \
+    _tmp3_m = __lasx_xvilvh_d(_t6, _t4);                                      \
+    _tmp5_m = __lasx_xvilvl_d(_t7, _t5);                                      \
+    _tmp7_m = __lasx_xvilvh_d(_t7, _t5);                                      \
+    _out0 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20);                         \
+    _out1 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20);                         \
+    _out2 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20);                         \
+    _out3 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20);                         \
+                                                                              \
+    _tmp0_m = __lasx_xvilvh_h(_in2, _in0);                                    \
+    _tmp1_m = __lasx_xvilvh_h(_in3, _in1);                                    \
+    _tmp2_m = __lasx_xvilvh_h(_in6, _in4);                                    \
+    _tmp3_m = __lasx_xvilvh_h(_in7, _in5);                                    \
+    _tmp4_m = __lasx_xvilvh_h(_in10, _in8);                                   \
+    _tmp5_m = __lasx_xvilvh_h(_in11, _in9);                                   \
+    _tmp6_m = __lasx_xvilvh_h(_in14, _in12);                                  \
+    _tmp7_m = __lasx_xvilvh_h(_in15, _in13);                                  \
+    _t0 = __lasx_xvilvl_h(_tmp1_m, _tmp0_m);                                  \
+    _t1 = __lasx_xvilvh_h(_tmp1_m, _tmp0_m);                                  \
+    _t2 = __lasx_xvilvl_h(_tmp3_m, _tmp2_m);                                  \
+    _t3 = __lasx_xvilvh_h(_tmp3_m, _tmp2_m);                                  \
+    _t4 = __lasx_xvilvl_h(_tmp5_m, _tmp4_m);                                  \
+    _t5 = __lasx_xvilvh_h(_tmp5_m, _tmp4_m);                                  \
+    _t6 = __lasx_xvilvl_h(_tmp7_m, _tmp6_m);                                  \
+    _t7 = __lasx_xvilvh_h(_tmp7_m, _tmp6_m);                                  \
+    _tmp0_m = __lasx_xvilvl_d(_t2, _t0);                                      \
+    _tmp2_m = __lasx_xvilvh_d(_t2, _t0);                                      \
+    _tmp4_m = __lasx_xvilvl_d(_t3, _t1);                                      \
+    _tmp6_m = __lasx_xvilvh_d(_t3, _t1);                                      \
+    _tmp1_m = __lasx_xvilvl_d(_t6, _t4);                                      \
+    _tmp3_m = __lasx_xvilvh_d(_t6, _t4);                                      \
+    _tmp5_m = __lasx_xvilvl_d(_t7, _t5);                                      \
+    _tmp7_m = __lasx_xvilvh_d(_t7, _t5);                                      \
+    _out4 = __lasx_xvpermi_q(_tmp1_m, _tmp0_m, 0x20);                         \
+    _out5 = __lasx_xvpermi_q(_tmp3_m, _tmp2_m, 0x20);                         \
+    _out6 = __lasx_xvpermi_q(_tmp5_m, _tmp4_m, 0x20);                         \
+    _out7 = __lasx_xvpermi_q(_tmp7_m, _tmp6_m, 0x20);                         \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 4x4 block with halfword elements in vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ *               Return Type - signed halfword
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE4x4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, \
+                            _out3)                                       \
+  {                                                                      \
+    __m256i _s0_m, _s1_m;                                                \
+                                                                         \
+    _s0_m = __lasx_xvilvl_h(_in1, _in0);                                 \
+    _s1_m = __lasx_xvilvl_h(_in3, _in2);                                 \
+    _out0 = __lasx_xvilvl_w(_s1_m, _s0_m);                               \
+    _out2 = __lasx_xvilvh_w(_s1_m, _s0_m);                               \
+    _out1 = __lasx_xvilvh_d(_out0, _out0);                               \
+    _out3 = __lasx_xvilvh_d(_out2, _out2);                               \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose input 8x8 byte block
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7
+ *                         (input 8x8 byte block)
+ *               Outputs - _out0, _out1, _out2, _out3, _out4, _out5, _out6,
+ *                         _out7 (output 8x8 byte block)
+ * Example     : See LASX_TRANSPOSE8x8_H
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+    _tmp0_m = __lasx_xvilvl_b(_in2, _in0);                                   \
+    _tmp1_m = __lasx_xvilvl_b(_in3, _in1);                                   \
+    _tmp2_m = __lasx_xvilvl_b(_in6, _in4);                                   \
+    _tmp3_m = __lasx_xvilvl_b(_in7, _in5);                                   \
+    _tmp4_m = __lasx_xvilvl_b(_tmp1_m, _tmp0_m);                             \
+    _tmp5_m = __lasx_xvilvh_b(_tmp1_m, _tmp0_m);                             \
+    _tmp6_m = __lasx_xvilvl_b(_tmp3_m, _tmp2_m);                             \
+    _tmp7_m = __lasx_xvilvh_b(_tmp3_m, _tmp2_m);                             \
+    _out0 = __lasx_xvilvl_w(_tmp6_m, _tmp4_m);                               \
+    _out2 = __lasx_xvilvh_w(_tmp6_m, _tmp4_m);                               \
+    _out4 = __lasx_xvilvl_w(_tmp7_m, _tmp5_m);                               \
+    _out6 = __lasx_xvilvh_w(_tmp7_m, _tmp5_m);                               \
+    _out1 = __lasx_xvbsrl_v(_out0, 8);                                       \
+    _out3 = __lasx_xvbsrl_v(_out2, 8);                                       \
+    _out5 = __lasx_xvbsrl_v(_out4, 8);                                       \
+    _out7 = __lasx_xvbsrl_v(_out6, 8);                                       \
+  }
+
+/*
+ * =============================================================================
+ * Description : Transpose 8x8 block with halfword elements in vectors.
+ * Arguments   : Inputs  - _in0, _in1, ~
+ *               Outputs - _out0, _out1, ~
+ * Details     : The rows of the matrix become columns, and the columns become
+ *               rows.
+ * Example     : LASX_TRANSPOSE8x8_H
+ *        _in0 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in1 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ *        _in2 : 8,2,3,4, 5,6,7,8, 8,2,3,4, 5,6,7,8
+ *        _in3 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in4 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ *        _in5 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in6 : 1,2,3,4, 5,6,7,8, 1,2,3,4, 5,6,7,8
+ *        _in7 : 9,2,3,4, 5,6,7,8, 9,2,3,4, 5,6,7,8
+ *
+ *       _out0 : 1,8,8,1, 9,1,1,9, 1,8,8,1, 9,1,1,9
+ *       _out1 : 2,2,2,2, 2,2,2,2, 2,2,2,2, 2,2,2,2
+ *       _out2 : 3,3,3,3, 3,3,3,3, 3,3,3,3, 3,3,3,3
+ *       _out3 : 4,4,4,4, 4,4,4,4, 4,4,4,4, 4,4,4,4
+ *       _out4 : 5,5,5,5, 5,5,5,5, 5,5,5,5, 5,5,5,5
+ *       _out5 : 6,6,6,6, 6,6,6,6, 6,6,6,6, 6,6,6,6
+ *       _out6 : 7,7,7,7, 7,7,7,7, 7,7,7,7, 7,7,7,7
+ *       _out7 : 8,8,8,8, 8,8,8,8, 8,8,8,8, 8,8,8,8
+ * =============================================================================
+ */
+#define LASX_TRANSPOSE8x8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                            _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                            _out7)                                           \
+  {                                                                          \
+    __m256i _s0_m, _s1_m;                                                    \
+    __m256i _tmp0_m, _tmp1_m, _tmp2_m, _tmp3_m;                              \
+    __m256i _tmp4_m, _tmp5_m, _tmp6_m, _tmp7_m;                              \
+                                                                             \
+    _s0_m = __lasx_xvilvl_h(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvl_h(_in7, _in5);                                     \
+    _tmp0_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp1_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_h(_in6, _in4);                                     \
+    _s1_m = __lasx_xvilvh_h(_in7, _in5);                                     \
+    _tmp2_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp3_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+                                                                             \
+    _s0_m = __lasx_xvilvl_h(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvl_h(_in3, _in1);                                     \
+    _tmp4_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp5_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+    _s0_m = __lasx_xvilvh_h(_in2, _in0);                                     \
+    _s1_m = __lasx_xvilvh_h(_in3, _in1);                                     \
+    _tmp6_m = __lasx_xvilvl_h(_s1_m, _s0_m);                                 \
+    _tmp7_m = __lasx_xvilvh_h(_s1_m, _s0_m);                                 \
+                                                                             \
+    _out0 = __lasx_xvpickev_d(_tmp0_m, _tmp4_m);                             \
+    _out2 = __lasx_xvpickev_d(_tmp1_m, _tmp5_m);                             \
+    _out4 = __lasx_xvpickev_d(_tmp2_m, _tmp6_m);                             \
+    _out6 = __lasx_xvpickev_d(_tmp3_m, _tmp7_m);                             \
+    _out1 = __lasx_xvpickod_d(_tmp0_m, _tmp4_m);                             \
+    _out3 = __lasx_xvpickod_d(_tmp1_m, _tmp5_m);                             \
+    _out5 = __lasx_xvpickod_d(_tmp2_m, _tmp6_m);                             \
+    _out7 = __lasx_xvpickod_d(_tmp3_m, _tmp7_m);                             \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 4 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3
+ *               Outputs - _out0, _out1, _out2, _out3
+ * Details     : Butterfly operation
+ * Example     : LASX_BUTTERFLY_4
+ *               _out0 = _in0 + _in3;
+ *               _out1 = _in1 + _in2;
+ *               _out2 = _in1 - _in2;
+ *               _out3 = _in0 - _in3;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_4_B(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_b(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_b(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_b(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_b(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_H(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_h(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_h(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_h(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_h(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_W(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_w(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_w(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_w(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_w(_in0, _in3);                                        \
+  }
+#define LASX_BUTTERFLY_4_D(_in0, _in1, _in2, _in3, _out0, _out1, _out2, _out3) \
+  {                                                                            \
+    _out0 = __lasx_xvadd_d(_in0, _in3);                                        \
+    _out1 = __lasx_xvadd_d(_in1, _in2);                                        \
+    _out2 = __lasx_xvsub_d(_in1, _in2);                                        \
+    _out3 = __lasx_xvsub_d(_in0, _in3);                                        \
+  }
+
+/*
+ * =============================================================================
+ * Description : Butterfly of 8 input vectors
+ * Arguments   : Inputs  - _in0, _in1, _in2, _in3, ~
+ *               Outputs - _out0, _out1, _out2, _out3, ~
+ * Details     : Butterfly operation
+ * Example     : LASX_BUTTERFLY_8
+ *               _out0 = _in0 + _in7;
+ *               _out1 = _in1 + _in6;
+ *               _out2 = _in2 + _in5;
+ *               _out3 = _in3 + _in4;
+ *               _out4 = _in3 - _in4;
+ *               _out5 = _in2 - _in5;
+ *               _out6 = _in1 - _in6;
+ *               _out7 = _in0 - _in7;
+ * =============================================================================
+ */
+#define LASX_BUTTERFLY_8_B(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_b(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_b(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_b(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_b(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_b(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_b(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_b(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_b(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_H(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_h(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_h(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_h(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_h(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_h(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_h(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_h(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_h(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_W(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_w(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_w(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_w(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_w(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_w(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_w(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_w(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_w(_in0, _in7);                                     \
+  }
+
+#define LASX_BUTTERFLY_8_D(_in0, _in1, _in2, _in3, _in4, _in5, _in6, _in7,  \
+                           _out0, _out1, _out2, _out3, _out4, _out5, _out6, \
+                           _out7)                                           \
+  {                                                                         \
+    _out0 = __lasx_xvadd_d(_in0, _in7);                                     \
+    _out1 = __lasx_xvadd_d(_in1, _in6);                                     \
+    _out2 = __lasx_xvadd_d(_in2, _in5);                                     \
+    _out3 = __lasx_xvadd_d(_in3, _in4);                                     \
+    _out4 = __lasx_xvsub_d(_in3, _in4);                                     \
+    _out5 = __lasx_xvsub_d(_in2, _in5);                                     \
+    _out6 = __lasx_xvsub_d(_in1, _in6);                                     \
+    _out7 = __lasx_xvsub_d(_in0, _in7);                                     \
+  }
+
+#endif  // LASX
+
+/*
+ * =============================================================================
+ * Description : Print out elements in vector.
+ * Arguments   : Inputs  - RTYPE, _element_num, _in0, _enter
+ *               Outputs -
+ * Details     : Print out '_element_num' elements in 'RTYPE' vector '_in0', if
+ *               '_enter' is TRUE, prefix "\nVP:" will be added first.
+ * Example     : VECT_PRINT(v4i32,4,in0,1); // in0: 1,2,3,4
+ *               VP:1,2,3,4,
+ * =============================================================================
+ */
+#define VECT_PRINT(RTYPE, element_num, in0, enter) \
+  {                                                \
+    RTYPE _tmp0 = (RTYPE)in0;                      \
+    int _i = 0;                                    \
+    if (enter)                                     \
+      printf("\nVP:");                             \
+    for (_i = 0; _i < element_num; _i++)           \
+      printf("%d,", _tmp0[_i]);                    \
+  }
+
+#endif /* LOONGSON_INTRINSICS_H */
+#endif /* INCLUDE_LIBYUV_LOONGSON_INTRINSICS_H */
diff --git a/files/include/libyuv/macros_msa.h b/files/include/libyuv/macros_msa.h
index 29997ce1..b9a44fcc 100644
--- a/files/include/libyuv/macros_msa.h
+++ b/files/include/libyuv/macros_msa.h
@@ -81,25 +81,35 @@
   })
 #endif  // !(__mips == 64)
 #else   // !(__mips_isa_rev >= 6)
-#define LW(psrc)                                       \
-  ({                                                   \
-    const uint8_t* psrc_lw_m = (const uint8_t*)(psrc); \
-    uint32_t val_m;                                    \
-    asm volatile("ulw  %[val_m],  %[psrc_lw_m]  \n"    \
-                 : [val_m] "=r"(val_m)                 \
-                 : [psrc_lw_m] "m"(*psrc_lw_m));       \
-    val_m;                                             \
+#define LW(psrc)                                \
+  ({                                            \
+    uint8_t* psrc_lw_m = (uint8_t*)(psrc);      \
+    uint32_t val_lw_m;                          \
+                                                \
+    __asm__ volatile(                           \
+        "lwr %[val_lw_m], 0(%[psrc_lw_m]) \n\t" \
+        "lwl %[val_lw_m], 3(%[psrc_lw_m]) \n\t" \
+                                                \
+        : [val_lw_m] "=&r"(val_lw_m)            \
+        : [psrc_lw_m] "r"(psrc_lw_m));          \
+                                                \
+    val_lw_m;                                   \
   })
 
 #if (__mips == 64)
-#define LD(psrc)                                       \
-  ({                                                   \
-    const uint8_t* psrc_ld_m = (const uint8_t*)(psrc); \
-    uint64_t val_m = 0;                                \
-    asm volatile("uld  %[val_m],  %[psrc_ld_m]  \n"    \
-                 : [val_m] "=r"(val_m)                 \
-                 : [psrc_ld_m] "m"(*psrc_ld_m));       \
-    val_m;                                             \
+#define LD(psrc)                                \
+  ({                                            \
+    uint8_t* psrc_ld_m = (uint8_t*)(psrc);      \
+    uint64_t val_ld_m = 0;                      \
+                                                \
+    __asm__ volatile(                           \
+        "ldr %[val_ld_m], 0(%[psrc_ld_m]) \n\t" \
+        "ldl %[val_ld_m], 7(%[psrc_ld_m]) \n\t" \
+                                                \
+        : [val_ld_m] "=&r"(val_ld_m)            \
+        : [psrc_ld_m] "r"(psrc_ld_m));          \
+                                                \
+    val_ld_m;                                   \
   })
 #else  // !(__mips == 64)
 #define LD(psrc)                                                         \
@@ -140,6 +150,9 @@
 #define LD_B(RTYPE, psrc) *((RTYPE*)(psrc)) /* NOLINT */
 #define LD_UB(...) LD_B(const v16u8, __VA_ARGS__)
 
+#define LD_H(RTYPE, psrc) *((RTYPE*)(psrc)) /* NOLINT */
+#define LD_UH(...) LD_H(const v8u16, __VA_ARGS__)
+
 #define ST_B(RTYPE, in, pdst) *((RTYPE*)(pdst)) = (in) /* NOLINT */
 #define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
 
diff --git a/files/include/libyuv/planar_functions.h b/files/include/libyuv/planar_functions.h
index f6f5b3ed..1ef2256b 100644
--- a/files/include/libyuv/planar_functions.h
+++ b/files/include/libyuv/planar_functions.h
@@ -83,6 +83,28 @@ void SetPlane(uint8_t* dst_y,
               int height,
               uint32_t value);
 
+// Convert a plane of tiles of 16 x H to linear.
+LIBYUV_API
+void DetilePlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height,
+                 int tile_height);
+
+// Convert a UV plane of tiles of 16 x H into linear U and V planes.
+LIBYUV_API
+void DetileSplitUVPlane(const uint8_t* src_uv,
+                        int src_stride_uv,
+                        uint8_t* dst_u,
+                        int dst_stride_u,
+                        uint8_t* dst_v,
+                        int dst_stride_v,
+                        int width,
+                        int height,
+                        int tile_height);
+
 // Split interleaved UV plane into separate U and V planes.
 LIBYUV_API
 void SplitUVPlane(const uint8_t* src_uv,
@@ -105,6 +127,72 @@ void MergeUVPlane(const uint8_t* src_u,
                   int width,
                   int height);
 
+// Split interleaved msb UV plane into separate lsb U and V planes.
+LIBYUV_API
+void SplitUVPlane_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_u,
+                     int dst_stride_u,
+                     uint16_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     int depth);
+
+// Merge separate lsb U and V planes into one interleaved msb UV plane.
+LIBYUV_API
+void MergeUVPlane_16(const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height,
+                     int depth);
+
+// Convert lsb plane to msb plane
+LIBYUV_API
+void ConvertToMSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth);
+
+// Convert msb plane to lsb plane
+LIBYUV_API
+void ConvertToLSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth);
+
+// Scale U and V to half width and height and merge into interleaved UV plane.
+// width and height are source size, allowing odd sizes.
+// Use for converting I444 or I422 to NV12.
+LIBYUV_API
+void HalfMergeUVPlane(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height);
+
+// Swap U and V channels in interleaved UV plane.
+LIBYUV_API
+void SwapUVPlane(const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_vu,
+                 int dst_stride_vu,
+                 int width,
+                 int height);
+
 // Split interleaved RGB plane into separate R, G and B planes.
 LIBYUV_API
 void SplitRGBPlane(const uint8_t* src_rgb,
@@ -131,6 +219,92 @@ void MergeRGBPlane(const uint8_t* src_r,
                    int width,
                    int height);
 
+// Split interleaved ARGB plane into separate R, G, B and A planes.
+// dst_a can be NULL to discard alpha plane.
+LIBYUV_API
+void SplitARGBPlane(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_r,
+                    int dst_stride_r,
+                    uint8_t* dst_g,
+                    int dst_stride_g,
+                    uint8_t* dst_b,
+                    int dst_stride_b,
+                    uint8_t* dst_a,
+                    int dst_stride_a,
+                    int width,
+                    int height);
+
+// Merge separate R, G, B and A planes into one interleaved ARGB plane.
+// src_a can be NULL to fill opaque value to alpha.
+LIBYUV_API
+void MergeARGBPlane(const uint8_t* src_r,
+                    int src_stride_r,
+                    const uint8_t* src_g,
+                    int src_stride_g,
+                    const uint8_t* src_b,
+                    int src_stride_b,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height);
+
+// Merge separate 'depth' bit R, G and B planes stored in lsb
+// into one interleaved XR30 plane.
+// depth should in range [10, 16]
+LIBYUV_API
+void MergeXR30Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    uint8_t* dst_ar30,
+                    int dst_stride_ar30,
+                    int width,
+                    int height,
+                    int depth);
+
+// Merge separate 'depth' bit R, G, B and A planes stored in lsb
+// into one interleaved AR64 plane.
+// src_a can be NULL to fill opaque value to alpha.
+// depth should in range [1, 16]
+LIBYUV_API
+void MergeAR64Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    const uint16_t* src_a,
+                    int src_stride_a,
+                    uint16_t* dst_ar64,
+                    int dst_stride_ar64,
+                    int width,
+                    int height,
+                    int depth);
+
+// Merge separate 'depth' bit R, G, B and A planes stored in lsb
+// into one interleaved ARGB plane.
+// src_a can be NULL to fill opaque value to alpha.
+// depth should in range [8, 16]
+LIBYUV_API
+void MergeARGB16To8Plane(const uint16_t* src_r,
+                         int src_stride_r,
+                         const uint16_t* src_g,
+                         int src_stride_g,
+                         const uint16_t* src_b,
+                         int src_stride_b,
+                         const uint16_t* src_a,
+                         int src_stride_a,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height,
+                         int depth);
+
 // Copy I400.  Supports inverting.
 LIBYUV_API
 int I400ToI400(const uint8_t* src_y,
@@ -178,6 +352,48 @@ int I444Copy(const uint8_t* src_y,
              int width,
              int height);
 
+// Copy I210 to I210.
+#define I210ToI210 I210Copy
+LIBYUV_API
+int I210Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height);
+
+// Copy NV12. Supports inverting.
+int NV12Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_uv,
+             int src_stride_uv,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_uv,
+             int dst_stride_uv,
+             int width,
+             int height);
+
+// Copy NV21. Supports inverting.
+int NV21Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_vu,
+             int src_stride_vu,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_vu,
+             int dst_stride_vu,
+             int width,
+             int height);
+
 // Convert YUY2 to I422.
 LIBYUV_API
 int YUY2ToI422(const uint8_t* src_yuy2,
@@ -245,6 +461,14 @@ int YUY2ToY(const uint8_t* src_yuy2,
             int width,
             int height);
 
+LIBYUV_API
+int UYVYToY(const uint8_t* src_uyvy,
+            int src_stride_uyvy,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height);
+
 // Convert I420 to I400. (calls CopyPlane ignoring u/v).
 LIBYUV_API
 int I420ToI400(const uint8_t* src_y,
@@ -293,6 +517,22 @@ int I400Mirror(const uint8_t* src_y,
                int height);
 
 // Alias
+#define NV12ToNV12Mirror NV12Mirror
+
+// NV12 mirror.
+LIBYUV_API
+int NV12Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height);
+
+// Alias
 #define ARGBToARGBMirror ARGBMirror
 
 // ARGB mirror.
@@ -304,56 +544,35 @@ int ARGBMirror(const uint8_t* src_argb,
                int width,
                int height);
 
-// Convert NV12 to RGB565.
-LIBYUV_API
-int NV12ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_uv,
-                 int src_stride_uv,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height);
+// Alias
+#define RGB24ToRGB24Mirror RGB24Mirror
 
-// I422ToARGB is in convert_argb.h
-// Convert I422 to BGRA.
+// RGB24 mirror.
 LIBYUV_API
-int I422ToBGRA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_bgra,
-               int dst_stride_bgra,
-               int width,
-               int height);
+int RGB24Mirror(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height);
 
-// Convert I422 to ABGR.
+// Mirror a plane of data.
 LIBYUV_API
-int I422ToABGR(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_abgr,
-               int dst_stride_abgr,
-               int width,
-               int height);
+void MirrorPlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height);
 
-// Convert I422 to RGBA.
+// Mirror a plane of UV data.
 LIBYUV_API
-int I422ToRGBA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_rgba,
-               int dst_stride_rgba,
-               int width,
-               int height);
+void MirrorUVPlane(const uint8_t* src_uv,
+                   int src_stride_uv,
+                   uint8_t* dst_uv,
+                   int dst_stride_uv,
+                   int width,
+                   int height);
 
 // Alias
 #define RGB24ToRAW RAWToRGB24
@@ -734,6 +953,19 @@ int ARGBBlur(const uint8_t* src_argb,
              int height,
              int radius);
 
+// Gaussian 5x5 blur a float plane.
+// Coefficients of 1, 4, 6, 4, 1.
+// Each destination pixel is a blur of the 5x5
+// pixels from the source.
+// Source edges are clamped.
+LIBYUV_API
+int GaussPlane_F32(const float* src,
+                   int src_stride,
+                   float* dst,
+                   int dst_stride,
+                   int width,
+                   int height);
+
 // Multiply ARGB image by ARGB value.
 LIBYUV_API
 int ARGBShade(const uint8_t* src_argb,
@@ -759,6 +991,21 @@ int InterpolatePlane(const uint8_t* src0,
                      int height,
                      int interpolation);
 
+// Interpolate between two images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src0
+// and 255 means 1% src0 and 99% src1.
+LIBYUV_API
+int InterpolatePlane_16(const uint16_t* src0,
+                        int src_stride0,  // measured in 16 bit pixels
+                        const uint16_t* src1,
+                        int src_stride1,
+                        uint16_t* dst,
+                        int dst_stride,
+                        int width,
+                        int height,
+                        int interpolation);
+
 // Interpolate between two ARGB images using specified amount of interpolation
 // Internally calls InterpolatePlane with width * 4 (bpp).
 LIBYUV_API
@@ -815,7 +1062,7 @@ void ARGBAffineRow_SSE2(const uint8_t* src_argb,
                         int width);
 
 // Shuffle ARGB channel order.  e.g. BGRA to ARGB.
-// shuffler is 16 bytes and must be aligned.
+// shuffler is 16 bytes.
 LIBYUV_API
 int ARGBShuffle(const uint8_t* src_bgra,
                 int src_stride_bgra,
@@ -825,6 +1072,17 @@ int ARGBShuffle(const uint8_t* src_bgra,
                 int width,
                 int height);
 
+// Shuffle AR64 channel order.  e.g. AR64 to AB64.
+// shuffler is 16 bytes.
+LIBYUV_API
+int AR64Shuffle(const uint16_t* src_ar64,
+                int src_stride_ar64,
+                uint16_t* dst_ar64,
+                int dst_stride_ar64,
+                const uint8_t* shuffler,
+                int width,
+                int height);
+
 // Sobel ARGB effect with planar output.
 LIBYUV_API
 int ARGBSobelToPlane(const uint8_t* src_argb,
diff --git a/files/include/libyuv/rotate.h b/files/include/libyuv/rotate.h
index c64e0216..684ed5e6 100644
--- a/files/include/libyuv/rotate.h
+++ b/files/include/libyuv/rotate.h
@@ -49,6 +49,24 @@ int I420Rotate(const uint8_t* src_y,
                int height,
                enum RotationMode mode);
 
+// Rotate I422 frame.
+LIBYUV_API
+int I422Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
 // Rotate I444 frame.
 LIBYUV_API
 int I444Rotate(const uint8_t* src_y,
@@ -83,6 +101,26 @@ int NV12ToI420Rotate(const uint8_t* src_y,
                      int height,
                      enum RotationMode mode);
 
+// Convert Android420 to I420 with rotation.
+// "rotation" can be 0, 90, 180 or 270.
+LIBYUV_API
+int Android420ToI420Rotate(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_y,
+                           int dst_stride_y,
+                           uint8_t* dst_u,
+                           int dst_stride_u,
+                           uint8_t* dst_v,
+                           int dst_stride_v,
+                           int width,
+                           int height,
+                           enum RotationMode rotation);
+
 // Rotate a plane by 0, 90, 180, or 270.
 LIBYUV_API
 int RotatePlane(const uint8_t* src,
@@ -118,39 +156,51 @@ void RotatePlane270(const uint8_t* src,
                     int width,
                     int height);
 
-LIBYUV_API
-void RotateUV90(const uint8_t* src,
-                int src_stride,
-                uint8_t* dst_a,
-                int dst_stride_a,
-                uint8_t* dst_b,
-                int dst_stride_b,
-                int width,
-                int height);
-
 // Rotations for when U and V are interleaved.
-// These functions take one input pointer and
+// These functions take one UV input pointer and
 // split the data into two buffers while
-// rotating them. Deprecated.
+// rotating them.
+// width and height expected to be half size for NV12.
+LIBYUV_API
+int SplitRotateUV(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height,
+                  enum RotationMode mode);
+
+LIBYUV_API
+void SplitRotateUV90(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     uint8_t* dst_b,
+                     int dst_stride_b,
+                     int width,
+                     int height);
+
 LIBYUV_API
-void RotateUV180(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height);
+void SplitRotateUV180(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 LIBYUV_API
-void RotateUV270(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height);
+void SplitRotateUV270(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 // The 90 and 270 functions are based on transposes.
 // Doing a transpose with reversing the read/write
@@ -165,14 +215,14 @@ void TransposePlane(const uint8_t* src,
                     int height);
 
 LIBYUV_API
-void TransposeUV(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height);
+void SplitTransposeUV(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/files/include/libyuv/rotate_row.h b/files/include/libyuv/rotate_row.h
index 022293ee..aa8528a9 100644
--- a/files/include/libyuv/rotate_row.h
+++ b/files/include/libyuv/rotate_row.h
@@ -32,8 +32,9 @@ extern "C" {
 #define LIBYUV_DISABLE_X86
 #endif
 #endif
-// The following are available for Visual C and clangcl 32 bit:
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+// The following are available for Visual C 32 bit:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
+    !defined(__clang__)
 #define HAS_TRANSPOSEWX8_SSSE3
 #define HAS_TRANSPOSEUVWX8_SSE2
 #endif
@@ -60,9 +61,9 @@ extern "C" {
 #define HAS_TRANSPOSEUVWX16_MSA
 #endif
 
-#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
-#define HAS_TRANSPOSEWX8_MMI
-#define HAS_TRANSPOSEUVWX8_MMI
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_TRANSPOSEWX16_LSX
+#define HAS_TRANSPOSEUVWX16_LSX
 #endif
 
 void TransposeWxH_C(const uint8_t* src,
@@ -92,11 +93,6 @@ void TransposeWx8_SSSE3(const uint8_t* src,
                         uint8_t* dst,
                         int dst_stride,
                         int width);
-void TransposeWx8_MMI(const uint8_t* src,
-                      int src_stride,
-                      uint8_t* dst,
-                      int dst_stride,
-                      int width);
 void TransposeWx8_Fast_SSSE3(const uint8_t* src,
                              int src_stride,
                              uint8_t* dst,
@@ -107,6 +103,11 @@ void TransposeWx16_MSA(const uint8_t* src,
                        uint8_t* dst,
                        int dst_stride,
                        int width);
+void TransposeWx16_LSX(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width);
 
 void TransposeWx8_Any_NEON(const uint8_t* src,
                            int src_stride,
@@ -118,11 +119,6 @@ void TransposeWx8_Any_SSSE3(const uint8_t* src,
                             uint8_t* dst,
                             int dst_stride,
                             int width);
-void TransposeWx8_Any_MMI(const uint8_t* src,
-                          int src_stride,
-                          uint8_t* dst,
-                          int dst_stride,
-                          int width);
 void TransposeWx8_Fast_Any_SSSE3(const uint8_t* src,
                                  int src_stride,
                                  uint8_t* dst,
@@ -133,6 +129,11 @@ void TransposeWx16_Any_MSA(const uint8_t* src,
                            uint8_t* dst,
                            int dst_stride,
                            int width);
+void TransposeWx16_Any_LSX(const uint8_t* src,
+                           int src_stride,
+                           uint8_t* dst,
+                           int dst_stride,
+                           int width);
 
 void TransposeUVWxH_C(const uint8_t* src,
                       int src_stride,
@@ -171,13 +172,6 @@ void TransposeUVWx8_NEON(const uint8_t* src,
                          uint8_t* dst_b,
                          int dst_stride_b,
                          int width);
-void TransposeUVWx8_MMI(const uint8_t* src,
-                        int src_stride,
-                        uint8_t* dst_a,
-                        int dst_stride_a,
-                        uint8_t* dst_b,
-                        int dst_stride_b,
-                        int width);
 void TransposeUVWx16_MSA(const uint8_t* src,
                          int src_stride,
                          uint8_t* dst_a,
@@ -185,6 +179,13 @@ void TransposeUVWx16_MSA(const uint8_t* src,
                          uint8_t* dst_b,
                          int dst_stride_b,
                          int width);
+void TransposeUVWx16_LSX(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width);
 
 void TransposeUVWx8_Any_SSE2(const uint8_t* src,
                              int src_stride,
@@ -200,13 +201,6 @@ void TransposeUVWx8_Any_NEON(const uint8_t* src,
                              uint8_t* dst_b,
                              int dst_stride_b,
                              int width);
-void TransposeUVWx8_Any_MMI(const uint8_t* src,
-                            int src_stride,
-                            uint8_t* dst_a,
-                            int dst_stride_a,
-                            uint8_t* dst_b,
-                            int dst_stride_b,
-                            int width);
 void TransposeUVWx16_Any_MSA(const uint8_t* src,
                              int src_stride,
                              uint8_t* dst_a,
@@ -214,6 +208,13 @@ void TransposeUVWx16_Any_MSA(const uint8_t* src,
                              uint8_t* dst_b,
                              int dst_stride_b,
                              int width);
+void TransposeUVWx16_Any_LSX(const uint8_t* src,
+                             int src_stride,
+                             uint8_t* dst_a,
+                             int dst_stride_a,
+                             uint8_t* dst_b,
+                             int dst_stride_b,
+                             int width);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/files/include/libyuv/row.h b/files/include/libyuv/row.h
index 9bb48850..1a1cf4b6 100644
--- a/files/include/libyuv/row.h
+++ b/files/include/libyuv/row.h
@@ -74,8 +74,10 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && \
     (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
 // Conversions:
-#define HAS_ABGRTOUVROW_SSSE3
 #define HAS_ABGRTOYROW_SSSE3
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ABGRTOUVROW_SSSE3
+#endif
 #define HAS_ARGB1555TOARGBROW_SSE2
 #define HAS_ARGB4444TOARGBROW_SSE2
 #define HAS_ARGBEXTRACTALPHAROW_SSE2
@@ -87,18 +89,19 @@ extern "C" {
 #define HAS_ARGBTORGB24ROW_SSSE3
 #define HAS_ARGBTORGB565DITHERROW_SSE2
 #define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#if !defined(LIBYUV_BIT_EXACT)
 #define HAS_ARGBTOUV444ROW_SSSE3
 #define HAS_ARGBTOUVJROW_SSSE3
 #define HAS_ARGBTOUVROW_SSSE3
-#define HAS_ARGBTOYJROW_SSSE3
-#define HAS_ARGBTOYROW_SSSE3
 #define HAS_BGRATOUVROW_SSSE3
-#define HAS_BGRATOYROW_SSSE3
+#endif
 #define HAS_COPYROW_ERMS
 #define HAS_COPYROW_SSE2
 #define HAS_H422TOARGBROW_SSSE3
 #define HAS_HALFFLOATROW_SSE2
-#define HAS_I400TOARGBROW_SSE2
 #define HAS_I422TOARGB1555ROW_SSSE3
 #define HAS_I422TOARGB4444ROW_SSSE3
 #define HAS_I422TOARGBROW_SSSE3
@@ -108,11 +111,12 @@ extern "C" {
 #define HAS_I422TOUYVYROW_SSE2
 #define HAS_I422TOYUY2ROW_SSE2
 #define HAS_I444TOARGBROW_SSSE3
+#define HAS_INTERPOLATEROW_SSSE3
 #define HAS_J400TOARGBROW_SSE2
 #define HAS_J422TOARGBROW_SSSE3
 #define HAS_MERGEUVROW_SSE2
 #define HAS_MIRRORROW_SSSE3
-#define HAS_MIRRORUVROW_SSSE3
+#define HAS_MIRRORSPLITUVROW_SSSE3
 #define HAS_NV12TOARGBROW_SSSE3
 #define HAS_NV12TORGB24ROW_SSSE3
 #define HAS_NV12TORGB565ROW_SSSE3
@@ -124,8 +128,12 @@ extern "C" {
 #define HAS_RGB24TOARGBROW_SSSE3
 #define HAS_RGB24TOYROW_SSSE3
 #define HAS_RGB565TOARGBROW_SSE2
-#define HAS_RGBATOUVROW_SSSE3
 #define HAS_RGBATOYROW_SSSE3
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_RGB24TOYJROW_SSSE3
+#define HAS_RAWTOYJROW_SSSE3
+#define HAS_RGBATOUVROW_SSSE3
+#endif
 #define HAS_SETROW_ERMS
 #define HAS_SETROW_X86
 #define HAS_SPLITUVROW_SSE2
@@ -141,7 +149,9 @@ extern "C" {
 // Effects:
 #define HAS_ARGBADDROW_SSE2
 #define HAS_ARGBAFFINEROW_SSE2
+#if !defined(LIBYUV_BIT_EXACT)
 #define HAS_ARGBATTENUATEROW_SSSE3
+#endif
 #define HAS_ARGBBLENDROW_SSSE3
 #define HAS_ARGBCOLORMATRIXROW_SSSE3
 #define HAS_ARGBCOLORTABLEROW_X86
@@ -160,7 +170,6 @@ extern "C" {
 #define HAS_BLENDPLANEROW_SSSE3
 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
-#define HAS_INTERPOLATEROW_SSSE3
 #define HAS_RGBCOLORTABLEROW_X86
 #define HAS_SOBELROW_SSE2
 #define HAS_SOBELTOPLANEROW_SSE2
@@ -175,6 +184,7 @@ extern "C" {
 // TODO(fbarchard): fix build error on android_full_debug=1
 // https://code.google.com/p/libyuv/issues/detail?id=517
 #define HAS_I422ALPHATOARGBROW_SSSE3
+#define HAS_I444ALPHATOARGBROW_SSSE3
 #endif
 #endif
 
@@ -190,15 +200,18 @@ extern "C" {
 #define HAS_ARGBPOLYNOMIALROW_AVX2
 #define HAS_ARGBSHUFFLEROW_AVX2
 #define HAS_ARGBTORGB565DITHERROW_AVX2
-#define HAS_ARGBTOUVJROW_AVX2
-#define HAS_ARGBTOUVROW_AVX2
 #define HAS_ARGBTOYJROW_AVX2
+#define HAS_RAWTOYJROW_AVX2
+#define HAS_RGB24TOYJROW_AVX2
 #define HAS_ARGBTOYROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ARGBTOUVJROW_AVX2
+#define HAS_ARGBTOUVROW_AVX2
+#endif
 #define HAS_COPYROW_AVX
 #define HAS_H422TOARGBROW_AVX2
 #define HAS_HALFFLOATROW_AVX2
 //  #define HAS_HALFFLOATROW_F16C  // Enable to test halffloat cast
-#define HAS_I400TOARGBROW_AVX2
 #define HAS_I422TOARGB1555ROW_AVX2
 #define HAS_I422TOARGB4444ROW_AVX2
 #define HAS_I422TOARGBROW_AVX2
@@ -227,7 +240,9 @@ extern "C" {
 
 // Effects:
 #define HAS_ARGBADDROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
 #define HAS_ARGBATTENUATEROW_AVX2
+#endif
 #define HAS_ARGBMULTIPLYROW_AVX2
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBUNATTENUATEROW_AVX2
@@ -238,13 +253,14 @@ extern "C" {
 // TODO(fbarchard): fix build error on android_full_debug=1
 // https://code.google.com/p/libyuv/issues/detail?id=517
 #define HAS_I422ALPHATOARGBROW_AVX2
+#define HAS_I444ALPHATOARGBROW_AVX2
 #endif
 #endif
 
-// The following are available for AVX2 Visual C and clangcl 32 bit:
+// The following are available for AVX2 Visual C 32 bit:
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER) && \
-    (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2))
+    !defined(__clang__) && defined(VISUALC_HAS_AVX2)
 #define HAS_ARGB1555TOARGBROW_AVX2
 #define HAS_ARGB4444TOARGBROW_AVX2
 #define HAS_ARGBTOARGB1555ROW_AVX2
@@ -257,62 +273,148 @@ extern "C" {
 // The following are also available on x64 Visual C.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && defined(_M_X64) && \
     (!defined(__clang__) || defined(__SSSE3__))
+#define HAS_I444ALPHATOARGBROW_SSSE3
+#define HAS_I444TOARGBROW_SSSE3
 #define HAS_I422ALPHATOARGBROW_SSSE3
 #define HAS_I422TOARGBROW_SSSE3
 #endif
 
 // The following are available for gcc/clang x86 platforms:
 // TODO(fbarchard): Port to Visual C
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 #define HAS_ABGRTOAR30ROW_SSSE3
 #define HAS_ARGBTOAR30ROW_SSSE3
+#define HAS_ARGBTOAR64ROW_SSSE3
+#define HAS_ARGBTOAB64ROW_SSSE3
+#define HAS_AR64TOARGBROW_SSSE3
+#define HAS_AB64TOARGBROW_SSSE3
 #define HAS_CONVERT16TO8ROW_SSSE3
 #define HAS_CONVERT8TO16ROW_SSE2
-// I210 is for H010.  2 = 422.  I for 601 vs H for 709.
+#define HAS_DETILEROW_SSE2
+#define HAS_DETILESPLITUVROW_SSSE3
+#define HAS_HALFMERGEUVROW_SSSE3
 #define HAS_I210TOAR30ROW_SSSE3
 #define HAS_I210TOARGBROW_SSSE3
+#define HAS_I212TOAR30ROW_SSSE3
+#define HAS_I212TOARGBROW_SSSE3
+#define HAS_I400TOARGBROW_SSE2
 #define HAS_I422TOAR30ROW_SSSE3
+#define HAS_I410TOAR30ROW_SSSE3
+#define HAS_I410TOARGBROW_SSSE3
+#define HAS_MERGEARGBROW_SSE2
+#define HAS_MERGEXRGBROW_SSE2
 #define HAS_MERGERGBROW_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV21TOYUV24ROW_SSSE3
+#define HAS_P210TOAR30ROW_SSSE3
+#define HAS_P210TOARGBROW_SSSE3
+#define HAS_P410TOAR30ROW_SSSE3
+#define HAS_P410TOARGBROW_SSSE3
+#define HAS_RAWTORGBAROW_SSSE3
+#define HAS_RGB24MIRRORROW_SSSE3
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_RGBATOYJROW_SSSE3
+#endif
+#define HAS_SPLITARGBROW_SSE2
+#define HAS_SPLITARGBROW_SSSE3
+#define HAS_SPLITXRGBROW_SSE2
+#define HAS_SPLITXRGBROW_SSSE3
 #define HAS_SPLITRGBROW_SSSE3
+#define HAS_SWAPUVROW_SSSE3
+
+#if defined(__x86_64__) || !defined(__pic__)
+// TODO(fbarchard): fix build error on android_full_debug=1
+// https://code.google.com/p/libyuv/issues/detail?id=517
+#define HAS_I210ALPHATOARGBROW_SSSE3
+#define HAS_I410ALPHATOARGBROW_SSSE3
+#endif
 #endif
 
 // The following are available for AVX2 gcc/clang x86 platforms:
 // TODO(fbarchard): Port to Visual C
-#if !defined(LIBYUV_DISABLE_X86) &&                                       \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER))) && \
+#if !defined(LIBYUV_DISABLE_X86) &&               \
+    (defined(__x86_64__) || defined(__i386__)) && \
     (defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
 #define HAS_ABGRTOAR30ROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_ABGRTOUVROW_AVX2
+#define HAS_ABGRTOYROW_AVX2
+#endif
 #define HAS_ARGBTOAR30ROW_AVX2
 #define HAS_ARGBTORAWROW_AVX2
 #define HAS_ARGBTORGB24ROW_AVX2
+#define HAS_ARGBTOAR64ROW_AVX2
+#define HAS_ARGBTOAB64ROW_AVX2
+#define HAS_AR64TOARGBROW_AVX2
+#define HAS_AB64TOARGBROW_AVX2
 #define HAS_CONVERT16TO8ROW_AVX2
+#define HAS_INTERPOLATEROW_16TO8_AVX2
 #define HAS_CONVERT8TO16ROW_AVX2
+#define HAS_DIVIDEROW_16_AVX2
+#define HAS_HALFMERGEUVROW_AVX2
+#define HAS_MERGEAR64ROW_AVX2
+#define HAS_MERGEARGB16TO8ROW_AVX2
+#define HAS_MERGEARGBROW_AVX2
+#define HAS_MERGEXR30ROW_AVX2
+#define HAS_MERGEXR64ROW_AVX2
+#define HAS_MERGEXRGB16TO8ROW_AVX2
+#define HAS_MERGEXRGBROW_AVX2
+#define HAS_NV21TOYUV24ROW_AVX2
 #define HAS_I210TOAR30ROW_AVX2
 #define HAS_I210TOARGBROW_AVX2
+#define HAS_I212TOAR30ROW_AVX2
+#define HAS_I212TOARGBROW_AVX2
+#define HAS_I400TOARGBROW_AVX2
+#define HAS_I410TOAR30ROW_AVX2
+#define HAS_I410TOARGBROW_AVX2
+#define HAS_P210TOAR30ROW_AVX2
+#define HAS_P210TOARGBROW_AVX2
+#define HAS_P410TOAR30ROW_AVX2
+#define HAS_P410TOARGBROW_AVX2
 #define HAS_I422TOAR30ROW_AVX2
 #define HAS_I422TOUYVYROW_AVX2
 #define HAS_I422TOYUY2ROW_AVX2
 #define HAS_MERGEUVROW_16_AVX2
+#define HAS_MIRRORUVROW_AVX2
 #define HAS_MULTIPLYROW_16_AVX2
-// TODO(fbarchard): Fix AVX2 version of YUV24
-// #define HAS_NV21TOYUV24ROW_AVX2
+#if !defined(LIBYUV_BIT_EXACT)
+#define HAS_RGBATOYJROW_AVX2
+#endif
+#define HAS_SPLITARGBROW_AVX2
+#define HAS_SPLITXRGBROW_AVX2
+#define HAS_SPLITUVROW_16_AVX2
+#define HAS_SWAPUVROW_AVX2
+
+#if defined(__x86_64__) || !defined(__pic__)
+// TODO(fbarchard): fix build error on android_full_debug=1
+// https://code.google.com/p/libyuv/issues/detail?id=517
+#define HAS_I210ALPHATOARGBROW_AVX2
+#define HAS_I410ALPHATOARGBROW_AVX2
+#endif
 #endif
 
 // The following are available for AVX512 clang x86 platforms:
 // TODO(fbarchard): Port to GCC and Visual C
 // TODO(fbarchard): re-enable HAS_ARGBTORGB24ROW_AVX512VBMI. Issue libyuv:789
-#if !defined(LIBYUV_DISABLE_X86) &&                                       \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER))) && \
-    (defined(CLANG_HAS_AVX512))
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || defined(__i386__)) && (defined(CLANG_HAS_AVX512))
 #define HAS_ARGBTORGB24ROW_AVX512VBMI
 #endif
 
+// The following are available for AVX512 clang x64 platforms:
+// TODO(fbarchard): Port to x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(__x86_64__) && \
+    (defined(CLANG_HAS_AVX512))
+#define HAS_I422TOARGBROW_AVX512BW
+#endif
+
 // The following are available on Neon platforms:
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_AB64TOARGBROW_NEON
 #define HAS_ABGRTOUVROW_NEON
 #define HAS_ABGRTOYROW_NEON
+#define HAS_AR64TOARGBROW_NEON
 #define HAS_ARGB1555TOARGBROW_NEON
 #define HAS_ARGB1555TOUVROW_NEON
 #define HAS_ARGB1555TOYROW_NEON
@@ -321,6 +423,8 @@ extern "C" {
 #define HAS_ARGB4444TOYROW_NEON
 #define HAS_ARGBEXTRACTALPHAROW_NEON
 #define HAS_ARGBSETROW_NEON
+#define HAS_ARGBTOAB64ROW_NEON
+#define HAS_ARGBTOAR64ROW_NEON
 #define HAS_ARGBTOARGB1555ROW_NEON
 #define HAS_ARGBTOARGB4444ROW_NEON
 #define HAS_ARGBTORAWROW_NEON
@@ -338,8 +442,13 @@ extern "C" {
 #define HAS_BGRATOUVROW_NEON
 #define HAS_BGRATOYROW_NEON
 #define HAS_BYTETOFLOATROW_NEON
+#define HAS_CONVERT16TO8ROW_NEON
 #define HAS_COPYROW_NEON
+#define HAS_DETILEROW_NEON
+#define HAS_DETILESPLITUVROW_NEON
+#define HAS_DIVIDEROW_16_NEON
 #define HAS_HALFFLOATROW_NEON
+#define HAS_HALFMERGEUVROW_NEON
 #define HAS_I400TOARGBROW_NEON
 #define HAS_I422ALPHATOARGBROW_NEON
 #define HAS_I422TOARGB1555ROW_NEON
@@ -350,11 +459,24 @@ extern "C" {
 #define HAS_I422TORGBAROW_NEON
 #define HAS_I422TOUYVYROW_NEON
 #define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444ALPHATOARGBROW_NEON
 #define HAS_I444TOARGBROW_NEON
+#define HAS_INTERPOLATEROW_16_NEON
+#define HAS_INTERPOLATEROW_NEON
 #define HAS_J400TOARGBROW_NEON
+#define HAS_MERGEAR64ROW_NEON
+#define HAS_MERGEARGB16TO8ROW_NEON
+#define HAS_MERGEARGBROW_NEON
+#define HAS_MERGEUVROW_16_NEON
 #define HAS_MERGEUVROW_NEON
+#define HAS_MERGEXR30ROW_NEON
+#define HAS_MERGEXR64ROW_NEON
+#define HAS_MERGEXRGB16TO8ROW_NEON
+#define HAS_MERGEXRGBROW_NEON
 #define HAS_MIRRORROW_NEON
+#define HAS_MIRRORSPLITUVROW_NEON
 #define HAS_MIRRORUVROW_NEON
+#define HAS_MULTIPLYROW_16_NEON
 #define HAS_NV12TOARGBROW_NEON
 #define HAS_NV12TORGB24ROW_NEON
 #define HAS_NV12TORGB565ROW_NEON
@@ -363,20 +485,29 @@ extern "C" {
 #define HAS_NV21TOYUV24ROW_NEON
 #define HAS_RAWTOARGBROW_NEON
 #define HAS_RAWTORGB24ROW_NEON
+#define HAS_RAWTORGBAROW_NEON
+#define HAS_RAWTOUVJROW_NEON
 #define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYJROW_NEON
 #define HAS_RAWTOYROW_NEON
 #define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVJROW_NEON
 #define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYJROW_NEON
 #define HAS_RGB24TOYROW_NEON
 #define HAS_RGB565TOARGBROW_NEON
 #define HAS_RGB565TOUVROW_NEON
 #define HAS_RGB565TOYROW_NEON
 #define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYJROW_NEON
 #define HAS_RGBATOYROW_NEON
 #define HAS_SETROW_NEON
+#define HAS_SPLITARGBROW_NEON
 #define HAS_SPLITRGBROW_NEON
+#define HAS_SPLITUVROW_16_NEON
 #define HAS_SPLITUVROW_NEON
-#define HAS_UVToVUROW_NEON
+#define HAS_SPLITXRGBROW_NEON
+#define HAS_SWAPUVROW_NEON
 #define HAS_UYVYTOARGBROW_NEON
 #define HAS_UYVYTOUV422ROW_NEON
 #define HAS_UYVYTOUVROW_NEON
@@ -393,13 +524,13 @@ extern "C" {
 #define HAS_ARGBCOLORMATRIXROW_NEON
 #define HAS_ARGBGRAYROW_NEON
 #define HAS_ARGBMIRRORROW_NEON
+#define HAS_RGB24MIRRORROW_NEON
 #define HAS_ARGBMULTIPLYROW_NEON
 #define HAS_ARGBQUANTIZEROW_NEON
 #define HAS_ARGBSEPIAROW_NEON
 #define HAS_ARGBSHADEROW_NEON
 #define HAS_ARGBSHUFFLEROW_NEON
 #define HAS_ARGBSUBTRACTROW_NEON
-#define HAS_INTERPOLATEROW_NEON
 #define HAS_SOBELROW_NEON
 #define HAS_SOBELTOPLANEROW_NEON
 #define HAS_SOBELXROW_NEON
@@ -409,8 +540,10 @@ extern "C" {
 
 // The following are available on AArch64 platforms:
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
-#define HAS_FLOATDIVTOBYTEROW_NEON
+#define HAS_INTERPOLATEROW_16TO8_NEON
 #define HAS_SCALESUMSAMPLES_NEON
+#define HAS_GAUSSROW_F32_NEON
+#define HAS_GAUSSCOL_F32_NEON
 #endif
 #if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
 #define HAS_ABGRTOUVROW_MSA
@@ -448,21 +581,27 @@ extern "C" {
 #define HAS_BGRATOYROW_MSA
 #define HAS_HALFFLOATROW_MSA
 #define HAS_I400TOARGBROW_MSA
-#define HAS_I422ALPHATOARGBROW_MSA
-#define HAS_I422TOARGBROW_MSA
-#define HAS_I422TORGB24ROW_MSA
-#define HAS_I422TORGBAROW_MSA
 #define HAS_I422TOUYVYROW_MSA
 #define HAS_I422TOYUY2ROW_MSA
+#define HAS_I422TOARGBROW_MSA
+#define HAS_I422TORGBAROW_MSA
+#define HAS_I422ALPHATOARGBROW_MSA
+#define HAS_I422TORGB24ROW_MSA
+#define HAS_I422TORGB565ROW_MSA
+#define HAS_I422TOARGB4444ROW_MSA
+#define HAS_I422TOARGB1555ROW_MSA
+#define HAS_NV12TOARGBROW_MSA
+#define HAS_NV12TORGB565ROW_MSA
+#define HAS_NV21TOARGBROW_MSA
+#define HAS_YUY2TOARGBROW_MSA
+#define HAS_UYVYTOARGBROW_MSA
 #define HAS_I444TOARGBROW_MSA
 #define HAS_INTERPOLATEROW_MSA
 #define HAS_J400TOARGBROW_MSA
 #define HAS_MERGEUVROW_MSA
 #define HAS_MIRRORROW_MSA
 #define HAS_MIRRORUVROW_MSA
-#define HAS_NV12TOARGBROW_MSA
-#define HAS_NV12TORGB565ROW_MSA
-#define HAS_NV21TOARGBROW_MSA
+#define HAS_MIRRORSPLITUVROW_MSA
 #define HAS_RAWTOARGBROW_MSA
 #define HAS_RAWTORGB24ROW_MSA
 #define HAS_RAWTOUVROW_MSA
@@ -482,88 +621,113 @@ extern "C" {
 #define HAS_SOBELXYROW_MSA
 #define HAS_SOBELYROW_MSA
 #define HAS_SPLITUVROW_MSA
-#define HAS_UYVYTOARGBROW_MSA
 #define HAS_UYVYTOUVROW_MSA
 #define HAS_UYVYTOYROW_MSA
-#define HAS_YUY2TOARGBROW_MSA
 #define HAS_YUY2TOUV422ROW_MSA
 #define HAS_YUY2TOUVROW_MSA
 #define HAS_YUY2TOYROW_MSA
 #endif
 
-#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
-#define HAS_ABGRTOUVROW_MMI
-#define HAS_ABGRTOYROW_MMI
-#define HAS_ARGB1555TOARGBROW_MMI
-#define HAS_ARGB1555TOUVROW_MMI
-#define HAS_ARGB1555TOYROW_MMI
-#define HAS_ARGB4444TOARGBROW_MMI
-#define HAS_ARGB4444TOUVROW_MMI
-#define HAS_ARGB4444TOYROW_MMI
-#define HAS_ARGBADDROW_MMI
-#define HAS_ARGBATTENUATEROW_MMI
-#define HAS_ARGBBLENDROW_MMI
-#define HAS_ARGBCOLORMATRIXROW_MMI
-#define HAS_ARGBCOPYALPHAROW_MMI
-#define HAS_ARGBCOPYYTOALPHAROW_MMI
-#define HAS_ARGBEXTRACTALPHAROW_MMI
-#define HAS_ARGBGRAYROW_MMI
-#define HAS_ARGBMIRRORROW_MMI
-#define HAS_ARGBMULTIPLYROW_MMI
-#define HAS_ARGBSEPIAROW_MMI
-#define HAS_ARGBSHADEROW_MMI
-#define HAS_ARGBSHUFFLEROW_MMI
-#define HAS_ARGBSUBTRACTROW_MMI
-#define HAS_ARGBTOARGB1555ROW_MMI
-#define HAS_ARGBTOARGB4444ROW_MMI
-#define HAS_ARGBTORAWROW_MMI
-#define HAS_ARGBTORGB24ROW_MMI
-#define HAS_ARGBTORGB565DITHERROW_MMI
-#define HAS_ARGBTORGB565ROW_MMI
-#define HAS_ARGBTOUV444ROW_MMI
-#define HAS_ARGBTOUVJROW_MMI
-#define HAS_ARGBTOUVROW_MMI
-#define HAS_ARGBTOYJROW_MMI
-#define HAS_ARGBTOYROW_MMI
-#define HAS_BGRATOUVROW_MMI
-#define HAS_BGRATOYROW_MMI
-#define HAS_BLENDPLANEROW_MMI
-#define HAS_COMPUTECUMULATIVESUMROW_MMI
-#define HAS_CUMULATIVESUMTOAVERAGEROW_MMI
-#define HAS_HALFFLOATROW_MMI
-#define HAS_I400TOARGBROW_MMI
-#define HAS_I422TOUYVYROW_MMI
-#define HAS_I422TOYUY2ROW_MMI
-#define HAS_INTERPOLATEROW_MMI
-#define HAS_J400TOARGBROW_MMI
-#define HAS_MERGERGBROW_MMI
-#define HAS_MERGEUVROW_MMI
-#define HAS_MIRRORROW_MMI
-#define HAS_MIRRORUVROW_MMI
-#define HAS_RAWTOARGBROW_MMI
-#define HAS_RAWTORGB24ROW_MMI
-#define HAS_RAWTOUVROW_MMI
-#define HAS_RAWTOYROW_MMI
-#define HAS_RGB24TOARGBROW_MMI
-#define HAS_RGB24TOUVROW_MMI
-#define HAS_RGB24TOYROW_MMI
-#define HAS_RGB565TOARGBROW_MMI
-#define HAS_RGB565TOUVROW_MMI
-#define HAS_RGB565TOYROW_MMI
-#define HAS_RGBATOUVROW_MMI
-#define HAS_RGBATOYROW_MMI
-#define HAS_SOBELROW_MMI
-#define HAS_SOBELTOPLANEROW_MMI
-#define HAS_SOBELXROW_MMI
-#define HAS_SOBELXYROW_MMI
-#define HAS_SOBELYROW_MMI
-#define HAS_SPLITRGBROW_MMI
-#define HAS_SPLITUVROW_MMI
-#define HAS_UYVYTOUVROW_MMI
-#define HAS_UYVYTOYROW_MMI
-#define HAS_YUY2TOUV422ROW_MMI
-#define HAS_YUY2TOUVROW_MMI
-#define HAS_YUY2TOYROW_MMI
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_ARGB4444TOARGBROW_LSX
+#define HAS_ARGB1555TOARGBROW_LSX
+#define HAS_RGB565TOARGBROW_LSX
+#define HAS_RGB24TOARGBROW_LSX
+#define HAS_RAWTOARGBROW_LSX
+#define HAS_ARGB1555TOYROW_LSX
+#define HAS_ARGB1555TOUVROW_LSX
+#define HAS_RGB565TOYROW_LSX
+#define HAS_RGB565TOUVROW_LSX
+#define HAS_RGB24TOYROW_LSX
+#define HAS_RGB24TOUVROW_LSX
+#define HAS_RAWTOYROW_LSX
+#define HAS_RAWTOUVROW_LSX
+#define HAS_NV12TOARGBROW_LSX
+#define HAS_NV12TORGB565ROW_LSX
+#define HAS_NV21TOARGBROW_LSX
+#define HAS_SOBELROW_LSX
+#define HAS_SOBELTOPLANEROW_LSX
+#define HAS_SOBELXYROW_LSX
+#define HAS_ARGBTOYJROW_LSX
+#define HAS_BGRATOYROW_LSX
+#define HAS_BGRATOUVROW_LSX
+#define HAS_ABGRTOYROW_LSX
+#define HAS_ABGRTOUVROW_LSX
+#define HAS_RGBATOYROW_LSX
+#define HAS_RGBATOUVROW_LSX
+#define HAS_ARGBTOUVJROW_LSX
+#define HAS_I444TOARGBROW_LSX
+#define HAS_I400TOARGBROW_LSX
+#define HAS_J400TOARGBROW_LSX
+#define HAS_YUY2TOARGBROW_LSX
+#define HAS_UYVYTOARGBROW_LSX
+#define HAS_INTERPOLATEROW_LSX
+#define HAS_ARGBSETROW_LSX
+#define HAS_RAWTORGB24ROW_LSX
+#define HAS_MERGEUVROW_LSX
+#define HAS_ARGBEXTRACTALPHAROW_LSX
+#define HAS_ARGBBLENDROW_LSX
+#define HAS_ARGBQUANTIZEROW_LSX
+#define HAS_ARGBCOLORMATRIXROW_LSX
+#define HAS_SPLITUVROW_LSX
+#define HAS_SETROW_LSX
+#define HAS_MIRRORSPLITUVROW_LSX
+#endif
+
+#if !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
+#define HAS_I422TOARGBROW_LASX
+#define HAS_I422TORGBAROW_LASX
+#define HAS_I422ALPHATOARGBROW_LASX
+#define HAS_I422TOYUY2ROW_LASX
+#define HAS_I422TOUYVYROW_LASX
+#define HAS_MIRRORROW_LASX
+#define HAS_MIRRORUVROW_LASX
+#define HAS_ARGBMIRRORROW_LASX
+#define HAS_I422TORGB24ROW_LASX
+#define HAS_I422TORGB565ROW_LASX
+#define HAS_I422TOARGB4444ROW_LASX
+#define HAS_I422TOARGB1555ROW_LASX
+#define HAS_YUY2TOUVROW_LASX
+#define HAS_YUY2TOYROW_LASX
+#define HAS_YUY2TOUV422ROW_LASX
+#define HAS_UYVYTOYROW_LASX
+#define HAS_UYVYTOUVROW_LASX
+#define HAS_UYVYTOUV422ROW_LASX
+#define HAS_ARGBTOYROW_LASX
+#define HAS_ARGBTOUVROW_LASX
+#define HAS_ARGBTORGB24ROW_LASX
+#define HAS_ARGBTORAWROW_LASX
+#define HAS_ARGBTORGB565ROW_LASX
+#define HAS_ARGBTOARGB1555ROW_LASX
+#define HAS_ARGBTOARGB4444ROW_LASX
+#define HAS_ARGBTOUV444ROW_LASX
+#define HAS_ARGBMULTIPLYROW_LASX
+#define HAS_ARGBADDROW_LASX
+#define HAS_ARGBSUBTRACTROW_LASX
+#define HAS_ARGBATTENUATEROW_LASX
+#define HAS_ARGBTORGB565DITHERROW_LASX
+#define HAS_ARGBSHUFFLEROW_LASX
+#define HAS_ARGBSHADEROW_LASX
+#define HAS_ARGBGRAYROW_LASX
+#define HAS_ARGBSEPIAROW_LASX
+#define HAS_ARGB4444TOARGBROW_LASX
+#define HAS_ARGB1555TOARGBROW_LASX
+#define HAS_RGB565TOARGBROW_LASX
+#define HAS_RGB24TOARGBROW_LASX
+#define HAS_RAWTOARGBROW_LASX
+#define HAS_ARGB1555TOYROW_LASX
+#define HAS_ARGB1555TOUVROW_LASX
+#define HAS_RGB565TOYROW_LASX
+#define HAS_RGB565TOUVROW_LASX
+#define HAS_RGB24TOYROW_LASX
+#define HAS_RGB24TOUVROW_LASX
+#define HAS_RAWTOYROW_LASX
+#define HAS_RAWTOUVROW_LASX
+#define HAS_NV12TOARGBROW_LASX
+#define HAS_NV12TORGB565ROW_LASX
+#define HAS_NV21TOARGBROW_LASX
+#define HAS_ARGBTOYJROW_LASX
+#define HAS_ARGBTOUVJROW_LASX
 #endif
 
 #if defined(_MSC_VER) && !defined(__CLR_VER) && !defined(__clang__)
@@ -572,8 +736,10 @@ extern "C" {
 #else
 #define SIMD_ALIGNED(var) __declspec(align(16)) var
 #endif
+#define LIBYUV_NOINLINE __declspec(noinline)
 typedef __declspec(align(16)) int16_t vec16[8];
 typedef __declspec(align(16)) int32_t vec32[4];
+typedef __declspec(align(16)) float vecf32[4];
 typedef __declspec(align(16)) int8_t vec8[16];
 typedef __declspec(align(16)) uint16_t uvec16[8];
 typedef __declspec(align(16)) uint32_t uvec32[4];
@@ -591,8 +757,10 @@ typedef __declspec(align(32)) uint8_t ulvec8[32];
 #else
 #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
 #endif
+#define LIBYUV_NOINLINE __attribute__((noinline))
 typedef int16_t __attribute__((vector_size(16))) vec16;
 typedef int32_t __attribute__((vector_size(16))) vec32;
+typedef float __attribute__((vector_size(16))) vecf32;
 typedef int8_t __attribute__((vector_size(16))) vec8;
 typedef uint16_t __attribute__((vector_size(16))) uvec16;
 typedef uint32_t __attribute__((vector_size(16))) uvec32;
@@ -605,8 +773,10 @@ typedef uint32_t __attribute__((vector_size(32))) ulvec32;
 typedef uint8_t __attribute__((vector_size(32))) ulvec8;
 #else
 #define SIMD_ALIGNED(var) var
+#define LIBYUV_NOINLINE
 typedef int16_t vec16[8];
 typedef int32_t vec32[4];
+typedef float vecf32[4];
 typedef int8_t vec8[16];
 typedef uint16_t uvec16[8];
 typedef uint32_t uvec32[4];
@@ -619,55 +789,30 @@ typedef uint32_t ulvec32[8];
 typedef uint8_t ulvec8[32];
 #endif
 
-#if defined(__aarch64__)
-// This struct is for Arm64 color conversion.
-struct YuvConstants {
-  uvec16 kUVToRB;
-  uvec16 kUVToRB2;
-  uvec16 kUVToG;
-  uvec16 kUVToG2;
-  vec16 kUVBiasBGR;
-  vec32 kYToRgb;
-};
-#elif defined(__arm__)
-// This struct is for ArmV7 color conversion.
+#if defined(__aarch64__) || defined(__arm__)
+// This struct is for ARM color conversion.
 struct YuvConstants {
-  uvec8 kUVToRB;
-  uvec8 kUVToG;
-  vec16 kUVBiasBGR;
-  vec32 kYToRgb;
+  uvec8 kUVCoeff;
+  vec16 kRGBCoeffBias;
 };
 #else
 // This struct is for Intel color conversion.
 struct YuvConstants {
-  int8_t kUVToB[32];
-  int8_t kUVToG[32];
-  int8_t kUVToR[32];
-  int16_t kUVBiasB[16];
-  int16_t kUVBiasG[16];
-  int16_t kUVBiasR[16];
+  uint8_t kUVToB[32];
+  uint8_t kUVToG[32];
+  uint8_t kUVToR[32];
   int16_t kYToRgb[16];
+  int16_t kYBiasToRgb[16];
 };
 
 // Offsets into YuvConstants structure
 #define KUVTOB 0
 #define KUVTOG 32
 #define KUVTOR 64
-#define KUVBIASB 96
-#define KUVBIASG 128
-#define KUVBIASR 160
-#define KYTORGB 192
-#endif
-
-// Conversion matrix for YUV to RGB
-extern const struct YuvConstants SIMD_ALIGNED(kYuvI601Constants);  // BT.601
-extern const struct YuvConstants SIMD_ALIGNED(kYuvJPEGConstants);  // JPeg
-extern const struct YuvConstants SIMD_ALIGNED(kYuvH709Constants);  // BT.709
+#define KYTORGB 96
+#define KYBIASTORGB 128
 
-// Conversion matrix for YVU to BGR
-extern const struct YuvConstants SIMD_ALIGNED(kYvuI601Constants);  // BT.601
-extern const struct YuvConstants SIMD_ALIGNED(kYvuJPEGConstants);  // JPeg
-extern const struct YuvConstants SIMD_ALIGNED(kYvuH709Constants);  // BT.709
+#endif
 
 #define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a)-1)))
 
@@ -755,6 +900,13 @@ void I422ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
 void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
                              const uint8_t* src_u,
                              const uint8_t* src_v,
@@ -762,12 +914,6 @@ void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
                              uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void I422ToARGBRow_NEON(const uint8_t* src_y,
-                        const uint8_t* src_u,
-                        const uint8_t* src_v,
-                        uint8_t* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width);
 void I422ToRGBARow_NEON(const uint8_t* src_y,
                         const uint8_t* src_u,
                         const uint8_t* src_v,
@@ -841,6 +987,12 @@ void I444ToARGBRow_MSA(const uint8_t* src_y,
                        uint8_t* dst_argb,
                        const struct YuvConstants* yuvconstants,
                        int width);
+void I444ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
 
 void I422ToARGBRow_MSA(const uint8_t* src_y,
                        const uint8_t* src_u,
@@ -848,12 +1000,24 @@ void I422ToARGBRow_MSA(const uint8_t* src_y,
                        uint8_t* dst_argb,
                        const struct YuvConstants* yuvconstants,
                        int width);
+void I422ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
 void I422ToRGBARow_MSA(const uint8_t* src_y,
                        const uint8_t* src_u,
                        const uint8_t* src_v,
                        uint8_t* dst_argb,
                        const struct YuvConstants* yuvconstants,
                        int width);
+void I422ToRGBARow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
 void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
                             const uint8_t* src_u,
                             const uint8_t* src_v,
@@ -861,30 +1025,61 @@ void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
                             uint8_t* dst_argb,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I422AlphaToARGBRow_LASX(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
 void I422ToRGB24Row_MSA(const uint8_t* src_y,
                         const uint8_t* src_u,
                         const uint8_t* src_v,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
+void I422ToRGB24Row_LASX(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
 void I422ToRGB565Row_MSA(const uint8_t* src_y,
                          const uint8_t* src_u,
                          const uint8_t* src_v,
                          uint8_t* dst_rgb565,
                          const struct YuvConstants* yuvconstants,
                          int width);
+void I422ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
 void I422ToARGB4444Row_MSA(const uint8_t* src_y,
                            const uint8_t* src_u,
                            const uint8_t* src_v,
                            uint8_t* dst_argb4444,
                            const struct YuvConstants* yuvconstants,
                            int width);
+void I422ToARGB4444Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422ToARGB1555Row_MSA(const uint8_t* src_y,
                            const uint8_t* src_u,
                            const uint8_t* src_v,
                            uint8_t* dst_argb1555,
                            const struct YuvConstants* yuvconstants,
                            int width);
+void I422ToARGB1555Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void NV12ToARGBRow_MSA(const uint8_t* src_y,
                        const uint8_t* src_uv,
                        uint8_t* dst_argb,
@@ -909,23 +1104,73 @@ void UYVYToARGBRow_MSA(const uint8_t* src_uyvy,
                        const struct YuvConstants* yuvconstants,
                        int width);
 
+void NV12ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void NV12ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void NV12ToRGB565Row_LSX(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void NV12ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void NV21ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void NV21ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void YUY2ToARGBRow_LSX(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void UYVYToARGBRow_LSX(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+
 void ARGBToYRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void ARGBToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void ABGRToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void ARGBToYJRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void ARGBToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYJRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGBAToYJRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void RGBAToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width);
 void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width);
 void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width);
 void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width);
 void RGB24ToYRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
 void RAWToYRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width);
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width);
 void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width);
 void ARGBToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
 void ARGBToYJRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ARGBToYJRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYJRow_LSX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void ARGBToYJRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width);
 void ARGBToUV444Row_NEON(const uint8_t* src_argb,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
@@ -939,20 +1184,20 @@ void ARGBToUV444Row_MSA(const uint8_t* src_argb,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
-void ARGBToUVRow_MSA(const uint8_t* src_argb0,
-                     int src_stride_argb,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width);
-void ARGBToUV444Row_MMI(const uint8_t* src_argb,
-                        uint8_t* dst_u,
-                        uint8_t* dst_v,
-                        int width);
-void ARGBToUVRow_MMI(const uint8_t* src_argb0,
+void ARGBToUVRow_MSA(const uint8_t* src_argb,
                      int src_stride_argb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
+void ARGBToUVRow_LASX(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUV444Row_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
 void ARGBToUVJRow_NEON(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
@@ -983,6 +1228,16 @@ void RAWToUVRow_NEON(const uint8_t* src_raw,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
 void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
                         int src_stride_rgb565,
                         uint8_t* dst_u,
@@ -998,32 +1253,32 @@ void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
-void ARGBToUVJRow_MSA(const uint8_t* src_rgb0,
+void ARGBToUVJRow_MSA(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width);
-void BGRAToUVRow_MSA(const uint8_t* src_rgb0,
+void BGRAToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void ABGRToUVRow_MSA(const uint8_t* src_rgb0,
+void ABGRToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void RGBAToUVRow_MSA(const uint8_t* src_rgb0,
+void RGBAToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void RGB24ToUVRow_MSA(const uint8_t* src_rgb0,
+void RGB24ToUVRow_MSA(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width);
-void RAWToUVRow_MSA(const uint8_t* src_rgb0,
+void RAWToUVRow_MSA(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
@@ -1038,56 +1293,78 @@ void ARGB1555ToUVRow_MSA(const uint8_t* src_argb1555,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
-                      int src_stride_rgb,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width);
-void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
-                     int src_stride_rgb,
+void BGRAToUVRow_LSX(const uint8_t* src_bgra,
+                     int src_stride_bgra,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
-                     int src_stride_rgb,
+void ABGRToUVRow_LSX(const uint8_t* src_abgr,
+                     int src_stride_abgr,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
-                     int src_stride_rgb,
+void RGBAToUVRow_LSX(const uint8_t* src_rgba,
+                     int src_stride_rgba,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
-                      int src_stride_rgb,
+void ARGBToUVJRow_LSX(const uint8_t* src_argb,
+                      int src_stride_argb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width);
-void RAWToUVRow_MMI(const uint8_t* src_rgb0,
-                    int src_stride_rgb,
-                    uint8_t* dst_u,
-                    uint8_t* dst_v,
-                    int width);
-void RGB565ToUVRow_MMI(const uint8_t* src_rgb565,
-                       int src_stride_rgb565,
+void ARGBToUVJRow_LASX(const uint8_t* src_argb,
+                       int src_stride_argb,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
-void ARGB1555ToUVRow_MMI(const uint8_t* src_argb1555,
+void ARGB1555ToUVRow_LSX(const uint8_t* src_argb1555,
                          int src_stride_argb1555,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void ARGB4444ToUVRow_MMI(const uint8_t* src_argb4444,
-                         int src_stride_argb4444,
-                         uint8_t* dst_u,
-                         uint8_t* dst_v,
-                         int width);
+void ARGB1555ToUVRow_LASX(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB565ToUVRow_LSX(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RGB565ToUVRow_LASX(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RGB24ToUVRow_LSX(const uint8_t* src_rgb24,
+                      int src_stride_rgb24,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width);
+void RAWToUVRow_LSX(const uint8_t* src_raw,
+                    int src_stride_raw,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width);
+void RAWToUVRow_LASX(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
 void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width);
 void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width);
 void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width);
 void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width);
 void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width);
 void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
 void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
                          uint8_t* dst_y,
@@ -1095,46 +1372,65 @@ void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
 void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
                          uint8_t* dst_y,
                          int width);
-void BGRAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ABGRToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGBAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGB24ToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RAWToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+void BGRAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void ABGRToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGBAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RGB24ToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
+void RAWToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width);
 void RGB565ToYRow_MSA(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
 void ARGB1555ToYRow_MSA(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
-void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGB565ToYRow_MMI(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
-void ARGB1555ToYRow_MMI(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
-void ARGB4444ToYRow_MMI(const uint8_t* src_argb4444, uint8_t* dst_y, int width);
-
-void ARGBToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ARGBToYJRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void BGRAToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void ABGRToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGBAToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RGB24ToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
-void RAWToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width);
+
+void BGRAToYRow_LSX(const uint8_t* src_bgra, uint8_t* dst_y, int width);
+void ABGRToYRow_LSX(const uint8_t* src_abgr, uint8_t* dst_y, int width);
+void RGBAToYRow_LSX(const uint8_t* src_rgba, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_LSX(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
+void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width);
+void RGB565ToYRow_LSX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void RGB565ToYRow_LASX(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
+void RGB24ToYRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
+void RAWToYRow_LSX(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width);
+
+void ARGBToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void ARGBToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGBAToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void BGRAToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void ABGRToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGBAToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGB24ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RGB24ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RAWToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
+void RAWToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width);
 void RGB565ToYRow_C(const uint8_t* src_rgb565, uint8_t* dst_y, int width);
 void ARGB1555ToYRow_C(const uint8_t* src_argb1555, uint8_t* dst_y, int width);
 void ARGB4444ToYRow_C(const uint8_t* src_argb4444, uint8_t* dst_y, int width);
 void ARGBToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void BGRAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ABGRToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGBAToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RGB24ToYRow_Any_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_y, int width);
-void RAWToYRow_Any_SSSE3(const uint8_t* src_raw, uint8_t* dst_y, int width);
+void RGB24ToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void RAWToYRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGBToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void BGRAToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ABGRToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGBAToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGB24ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RAWToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYJRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGB565ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGB1555ToYRow_Any_NEON(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
@@ -1153,88 +1449,104 @@ void RGB565ToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void ARGB1555ToYRow_Any_MSA(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
-void BGRAToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ABGRToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RGBAToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ARGBToYJRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ARGBToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RGB24ToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RAWToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RGB565ToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ARGB1555ToYRow_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_ptr,
-                            int width);
-void ARGB4444ToYRow_Any_MMI(const uint8_t* src_ptr,
+
+void BGRAToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ABGRToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGBAToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB565ToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_LSX(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
 
-void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
+void RGB565ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RGB24ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYJRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGB1555ToYRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+
+void ARGBToUVRow_AVX2(const uint8_t* src_argb,
                       int src_stride_argb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width);
-void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
+void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
+void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
-void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
+void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
-void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
+void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
                         int src_stride_argb,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
-void BGRAToUVRow_SSSE3(const uint8_t* src_bgra0,
+void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
                        int src_stride_bgra,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
-void ABGRToUVRow_SSSE3(const uint8_t* src_abgr0,
+void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
                        int src_stride_abgr,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
-void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
+void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
                        int src_stride_rgba,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width);
 void ARGBToUVRow_Any_AVX2(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ABGRToUVRow_Any_AVX2(const uint8_t* src_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
 void ARGBToUVJRow_Any_AVX2(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void ARGBToUVRow_Any_SSSE3(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void ARGBToUVJRow_Any_SSSE3(const uint8_t* src_ptr,
-                            int src_stride_ptr,
+                            int src_stride,
                             uint8_t* dst_u,
                             uint8_t* dst_v,
                             int width);
 void BGRAToUVRow_Any_SSSE3(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void ABGRToUVRow_Any_SSSE3(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void RGBAToUVRow_Any_SSSE3(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
@@ -1243,7 +1555,7 @@ void ARGBToUV444Row_Any_NEON(const uint8_t* src_ptr,
                              uint8_t* dst_v,
                              int width);
 void ARGBToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -1256,57 +1568,67 @@ void ARGBToUVRow_Any_MSA(const uint8_t* src_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void ARGBToUV444Row_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_u,
-                            uint8_t* dst_v,
-                            int width);
-void ARGBToUVRow_Any_MMI(const uint8_t* src_ptr,
-                         int src_stride_ptr,
-                         uint8_t* dst_u,
-                         uint8_t* dst_v,
-                         int width);
+void ARGBToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void ARGBToUV444Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
 void ARGBToUVJRow_Any_NEON(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void BGRAToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
 void ABGRToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
 void RGBAToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
 void RGB24ToUVRow_Any_NEON(const uint8_t* src_ptr,
-                           int src_stride_ptr,
+                           int src_stride,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
 void RAWToUVRow_Any_NEON(const uint8_t* src_ptr,
-                         int src_stride_ptr,
+                         int src_stride,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
+void RGB24ToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                            int src_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void RAWToUVJRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
 void RGB565ToUVRow_Any_NEON(const uint8_t* src_ptr,
-                            int src_stride_ptr,
+                            int src_stride,
                             uint8_t* dst_u,
                             uint8_t* dst_v,
                             int width);
 void ARGB1555ToUVRow_Any_NEON(const uint8_t* src_ptr,
-                              int src_stride_ptr,
+                              int src_stride,
                               uint8_t* dst_u,
                               uint8_t* dst_v,
                               int width);
 void ARGB4444ToUVRow_Any_NEON(const uint8_t* src_ptr,
-                              int src_stride_ptr,
+                              int src_stride,
                               uint8_t* dst_u,
                               uint8_t* dst_v,
                               int width);
@@ -1350,96 +1672,126 @@ void ARGB1555ToUVRow_Any_MSA(const uint8_t* src_ptr,
                              uint8_t* dst_u,
                              uint8_t* dst_v,
                              int width);
-void ARGBToUVJRow_Any_MMI(const uint8_t* src_ptr,
-                          int src_stride_ptr,
-                          uint8_t* dst_u,
-                          uint8_t* dst_v,
-                          int width);
-void BGRAToUVRow_Any_MMI(const uint8_t* src_ptr,
+void ABGRToUVRow_Any_LSX(const uint8_t* src_ptr,
                          int src_stride_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void ABGRToUVRow_Any_MMI(const uint8_t* src_ptr,
+void BGRAToUVRow_Any_LSX(const uint8_t* src_ptr,
                          int src_stride_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void RGBAToUVRow_Any_MMI(const uint8_t* src_ptr,
+void RGBAToUVRow_Any_LSX(const uint8_t* src_ptr,
                          int src_stride_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void RGB24ToUVRow_Any_MMI(const uint8_t* src_ptr,
+void ARGBToUVJRow_Any_LSX(const uint8_t* src_ptr,
                           int src_stride_ptr,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
-void RAWToUVRow_Any_MMI(const uint8_t* src_ptr,
-                        int src_stride_ptr,
-                        uint8_t* dst_u,
-                        uint8_t* dst_v,
-                        int width);
-void RGB565ToUVRow_Any_MMI(const uint8_t* src_ptr,
+void ARGBToUVJRow_Any_LASX(const uint8_t* src_ptr,
                            int src_stride_ptr,
                            uint8_t* dst_u,
                            uint8_t* dst_v,
                            int width);
-void ARGB1555ToUVRow_Any_MMI(const uint8_t* src_ptr,
-                             int src_stride_ptr,
-                             uint8_t* dst_u,
-                             uint8_t* dst_v,
-                             int width);
-void ARGB4444ToUVRow_Any_MMI(const uint8_t* src_ptr,
+void ARGB1555ToUVRow_Any_LSX(const uint8_t* src_ptr,
                              int src_stride_ptr,
                              uint8_t* dst_u,
                              uint8_t* dst_v,
                              int width);
-void ARGBToUVRow_C(const uint8_t* src_rgb0,
+void ARGB1555ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                              int src_stride_ptr,
+                              uint8_t* dst_u,
+                              uint8_t* dst_v,
+                              int width);
+void RGB565ToUVRow_Any_LSX(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RGB565ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                            int src_stride_ptr,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void RGB24ToUVRow_Any_LSX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void RGB24ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                           int src_stride_ptr,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void RAWToUVRow_Any_LSX(const uint8_t* src_ptr,
+                        int src_stride_ptr,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void RAWToUVRow_Any_LASX(const uint8_t* src_ptr,
+                         int src_stride_ptr,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
+void ARGBToUVRow_C(const uint8_t* src_rgb,
                    int src_stride_rgb,
                    uint8_t* dst_u,
                    uint8_t* dst_v,
                    int width);
-void ARGBToUVJRow_C(const uint8_t* src_rgb0,
+void ARGBToUVJRow_C(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width);
-void ARGBToUVRow_C(const uint8_t* src_rgb0,
+void ARGBToUVRow_C(const uint8_t* src_rgb,
                    int src_stride_rgb,
                    uint8_t* dst_u,
                    uint8_t* dst_v,
                    int width);
-void ARGBToUVJRow_C(const uint8_t* src_rgb0,
+void ARGBToUVJRow_C(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width);
-void BGRAToUVRow_C(const uint8_t* src_rgb0,
+void BGRAToUVRow_C(const uint8_t* src_rgb,
                    int src_stride_rgb,
                    uint8_t* dst_u,
                    uint8_t* dst_v,
                    int width);
-void ABGRToUVRow_C(const uint8_t* src_rgb0,
+void ABGRToUVRow_C(const uint8_t* src_rgb,
                    int src_stride_rgb,
                    uint8_t* dst_u,
                    uint8_t* dst_v,
                    int width);
-void RGBAToUVRow_C(const uint8_t* src_rgb0,
+void RGBAToUVRow_C(const uint8_t* src_rgb,
                    int src_stride_rgb,
                    uint8_t* dst_u,
                    uint8_t* dst_v,
                    int width);
-void RGB24ToUVRow_C(const uint8_t* src_rgb0,
+void RGB24ToUVRow_C(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width);
-void RAWToUVRow_C(const uint8_t* src_rgb0,
+void RAWToUVRow_C(const uint8_t* src_rgb,
                   int src_stride_rgb,
                   uint8_t* dst_u,
                   uint8_t* dst_v,
                   int width);
+void RGB24ToUVJRow_C(const uint8_t* src_rgb,
+                     int src_stride_rgb,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width);
+void RAWToUVJRow_C(const uint8_t* src_rgb,
+                   int src_stride_rgb,
+                   uint8_t* dst_u,
+                   uint8_t* dst_v,
+                   int width);
 void RGB565ToUVRow_C(const uint8_t* src_rgb565,
                      int src_stride_rgb565,
                      uint8_t* dst_u,
@@ -1474,41 +1826,52 @@ void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_SSSE3(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width);
-void MirrorRow_MMI(const uint8_t* src, uint8_t* dst, int width);
+void MirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void MirrorRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void MirrorRow_Any_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void MirrorRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void MirrorRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void MirrorRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_MSA(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_LASX(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width);
+void MirrorUVRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void MirrorUVRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 
-void MirrorUVRow_SSSE3(const uint8_t* src,
-                       uint8_t* dst_u,
-                       uint8_t* dst_v,
-                       int width);
-void MirrorUVRow_NEON(const uint8_t* src_uv,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width);
-void MirrorUVRow_MSA(const uint8_t* src_uv,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width);
-void MirrorUVRow_MMI(const uint8_t* src_uv,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width);
-void MirrorUVRow_C(const uint8_t* src_uv,
-                   uint8_t* dst_u,
-                   uint8_t* dst_v,
-                   int width);
+void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void MirrorSplitUVRow_MSA(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void MirrorSplitUVRow_LSX(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
+void MirrorSplitUVRow_C(const uint8_t* src_uv,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
 
 void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
-void ARGBMirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width);
 void ARGBMirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width);
-void ARGBMirrorRow_MMI(const uint8_t* src, uint8_t* dst, int width);
+void ARGBMirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width);
 void ARGBMirrorRow_C(const uint8_t* src, uint8_t* dst, int width);
 void ARGBMirrorRow_Any_AVX2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
@@ -1520,7 +1883,23 @@ void ARGBMirrorRow_Any_NEON(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
 void ARGBMirrorRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ARGBMirrorRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBMirrorRow_Any_LASX(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+
+void RGB24MirrorRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_rgb24,
+                          int width);
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width);
+void RGB24MirrorRow_C(const uint8_t* src_rgb24, uint8_t* dst_rgb24, int width);
+void RGB24MirrorRow_Any_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void RGB24MirrorRow_Any_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
 
 void SplitUVRow_C(const uint8_t* src_uv,
                   uint8_t* dst_u,
@@ -1542,7 +1921,7 @@ void SplitUVRow_MSA(const uint8_t* src_uv,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width);
-void SplitUVRow_MMI(const uint8_t* src_uv,
+void SplitUVRow_LSX(const uint8_t* src_uv,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width);
@@ -1562,11 +1941,56 @@ void SplitUVRow_Any_MSA(const uint8_t* src_ptr,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
-void SplitUVRow_Any_MMI(const uint8_t* src_ptr,
+void SplitUVRow_Any_LSX(const uint8_t* src_ptr,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
+void DetileRow_C(const uint8_t* src,
+                 ptrdiff_t src_tile_stride,
+                 uint8_t* dst,
+                 int width);
 
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width);
+void DetileRow_Any_NEON(const uint8_t* src,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst,
+                        int width);
+void DetileRow_SSE2(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width);
+void DetileRow_Any_SSE2(const uint8_t* src,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst,
+                        int width);
+void DetileSplitUVRow_C(const uint8_t* src_uv,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width);
+void DetileSplitUVRow_SSSE3(const uint8_t* src_uv,
+                            ptrdiff_t src_tile_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width);
+void DetileSplitUVRow_Any_SSSE3(const uint8_t* src_uv,
+                                ptrdiff_t src_tile_stride,
+                                uint8_t* dst_u,
+                                uint8_t* dst_v,
+                                int width);
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width);
+void DetileSplitUVRow_Any_NEON(const uint8_t* src_uv,
+                               ptrdiff_t src_tile_stride,
+                               uint8_t* dst_u,
+                               uint8_t* dst_v,
+                               int width);
 void MergeUVRow_C(const uint8_t* src_u,
                   const uint8_t* src_v,
                   uint8_t* dst_uv,
@@ -1587,7 +2011,7 @@ void MergeUVRow_MSA(const uint8_t* src_u,
                     const uint8_t* src_v,
                     uint8_t* dst_uv,
                     int width);
-void MergeUVRow_MMI(const uint8_t* src_u,
+void MergeUVRow_LSX(const uint8_t* src_u,
                     const uint8_t* src_v,
                     uint8_t* dst_uv,
                     int width);
@@ -1607,11 +2031,39 @@ void MergeUVRow_Any_MSA(const uint8_t* y_buf,
                         const uint8_t* uv_buf,
                         uint8_t* dst_ptr,
                         int width);
-void MergeUVRow_Any_MMI(const uint8_t* y_buf,
+void MergeUVRow_Any_LSX(const uint8_t* y_buf,
                         const uint8_t* uv_buf,
                         uint8_t* dst_ptr,
                         int width);
 
+void HalfMergeUVRow_C(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int width);
+
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width);
+
+void HalfMergeUVRow_SSSE3(const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          uint8_t* dst_uv,
+                          int width);
+
+void HalfMergeUVRow_AVX2(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width);
+
 void SplitRGBRow_C(const uint8_t* src_rgb,
                    uint8_t* dst_r,
                    uint8_t* dst_g,
@@ -1627,11 +2079,6 @@ void SplitRGBRow_NEON(const uint8_t* src_rgb,
                       uint8_t* dst_g,
                       uint8_t* dst_b,
                       int width);
-void SplitRGBRow_MMI(const uint8_t* src_rgb,
-                     uint8_t* dst_r,
-                     uint8_t* dst_g,
-                     uint8_t* dst_b,
-                     int width);
 void SplitRGBRow_Any_SSSE3(const uint8_t* src_ptr,
                            uint8_t* dst_r,
                            uint8_t* dst_g,
@@ -1642,11 +2089,6 @@ void SplitRGBRow_Any_NEON(const uint8_t* src_ptr,
                           uint8_t* dst_g,
                           uint8_t* dst_b,
                           int width);
-void SplitRGBRow_Any_MMI(const uint8_t* src_ptr,
-                         uint8_t* dst_r,
-                         uint8_t* dst_g,
-                         uint8_t* dst_b,
-                         int width);
 
 void MergeRGBRow_C(const uint8_t* src_r,
                    const uint8_t* src_g,
@@ -1663,11 +2105,6 @@ void MergeRGBRow_NEON(const uint8_t* src_r,
                       const uint8_t* src_b,
                       uint8_t* dst_rgb,
                       int width);
-void MergeRGBRow_MMI(const uint8_t* src_r,
-                     const uint8_t* src_g,
-                     const uint8_t* src_b,
-                     uint8_t* dst_rgb,
-                     int width);
 void MergeRGBRow_Any_SSSE3(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
@@ -1678,31 +2115,449 @@ void MergeRGBRow_Any_NEON(const uint8_t* src_r,
                           const uint8_t* src_b,
                           uint8_t* dst_rgb,
                           int width);
-void MergeRGBRow_Any_MMI(const uint8_t* src_r,
-                         const uint8_t* src_g,
-                         const uint8_t* src_b,
-                         uint8_t* dst_rgb,
+void MergeARGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    const uint8_t* src_a,
+                    uint8_t* dst_argb,
+                    int width);
+void MergeARGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeARGBRow_Any_SSE2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeARGBRow_Any_AVX2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeARGBRow_Any_NEON(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           const uint8_t* a_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void SplitARGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    uint8_t* dst_a,
+                    int width);
+void SplitARGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        uint8_t* dst_a,
+                        int width);
+void SplitARGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width);
+void SplitARGBRow_Any_SSE2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void SplitARGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_r,
+                            uint8_t* dst_g,
+                            uint8_t* dst_b,
+                            uint8_t* dst_a,
+                            int width);
+void SplitARGBRow_Any_AVX2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void SplitARGBRow_Any_NEON(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           uint8_t* dst_a,
+                           int width);
+void MergeXRGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    uint8_t* dst_argb,
+                    int width);
+void MergeXRGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width);
+void MergeXRGBRow_Any_SSE2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeXRGBRow_Any_AVX2(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void MergeXRGBRow_Any_NEON(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           int width);
+void SplitXRGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    int width);
+void SplitXRGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        int width);
+void SplitXRGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width);
+void SplitXRGBRow_Any_SSE2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+void SplitXRGBRow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_r,
+                            uint8_t* dst_g,
+                            uint8_t* dst_b,
+                            int width);
+void SplitXRGBRow_Any_AVX2(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+void SplitXRGBRow_Any_NEON(const uint8_t* src_ptr,
+                           uint8_t* dst_r,
+                           uint8_t* dst_g,
+                           uint8_t* dst_b,
+                           int width);
+
+void MergeXR30Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint8_t* dst_ar30,
+                    int depth,
+                    int width);
+void MergeAR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    const uint16_t* src_a,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width);
+void MergeARGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         const uint16_t* src_a,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width);
+void MergeXR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width);
+void MergeXRGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         uint8_t* dst_argb,
+                         int depth,
                          int width);
+void MergeXR30Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width);
+void MergeAR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeARGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeXRGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width);
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width);
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width);
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width);
+void MergeXR30Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint8_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeAR64Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           const uint16_t* a_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXR64Row_Any_AVX2(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeARGB16To8Row_Any_AVX2(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                const uint16_t* a_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXRGB16To8Row_Any_AVX2(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXR30Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint8_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXR30Row_10_Any_NEON(const uint16_t* r_buf,
+                              const uint16_t* g_buf,
+                              const uint16_t* b_buf,
+                              uint8_t* dst_ptr,
+                              int depth,
+                              int width);
+void MergeAR64Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           const uint16_t* a_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeARGB16To8Row_Any_NEON(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                const uint16_t* a_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
+void MergeXR64Row_Any_NEON(const uint16_t* r_buf,
+                           const uint16_t* g_buf,
+                           const uint16_t* b_buf,
+                           uint16_t* dst_ptr,
+                           int depth,
+                           int width);
+void MergeXRGB16To8Row_Any_NEON(const uint16_t* r_buf,
+                                const uint16_t* g_buf,
+                                const uint16_t* b_buf,
+                                uint8_t* dst_ptr,
+                                int depth,
+                                int width);
 
 void MergeUVRow_16_C(const uint16_t* src_u,
                      const uint16_t* src_v,
                      uint16_t* dst_uv,
-                     int scale, /* 64 for 10 bit */
+                     int depth,
                      int width);
 void MergeUVRow_16_AVX2(const uint16_t* src_u,
                         const uint16_t* src_v,
                         uint16_t* dst_uv,
-                        int scale,
+                        int depth,
+                        int width);
+void MergeUVRow_16_Any_AVX2(const uint16_t* src_u,
+                            const uint16_t* src_v,
+                            uint16_t* dst_uv,
+                            int depth,
+                            int width);
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
                         int width);
+void MergeUVRow_16_Any_NEON(const uint16_t* src_u,
+                            const uint16_t* src_v,
+                            uint16_t* dst_uv,
+                            int depth,
+                            int width);
+
+void SplitUVRow_16_C(const uint16_t* src_uv,
+                     uint16_t* dst_u,
+                     uint16_t* dst_v,
+                     int depth,
+                     int width);
+void SplitUVRow_16_AVX2(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width);
+void SplitUVRow_16_Any_AVX2(const uint16_t* src_uv,
+                            uint16_t* dst_u,
+                            uint16_t* dst_v,
+                            int depth,
+                            int width);
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width);
+void SplitUVRow_16_Any_NEON(const uint16_t* src_uv,
+                            uint16_t* dst_u,
+                            uint16_t* dst_v,
+                            int depth,
+                            int width);
 
-void MultiplyRow_16_AVX2(const uint16_t* src_y,
-                         uint16_t* dst_y,
-                         int scale,
-                         int width);
 void MultiplyRow_16_C(const uint16_t* src_y,
                       uint16_t* dst_y,
                       int scale,
                       int width);
+void MultiplyRow_16_AVX2(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width);
+void MultiplyRow_16_Any_AVX2(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int scale,
+                             int width);
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width);
+void MultiplyRow_16_Any_NEON(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int scale,
+                             int width);
+
+void DivideRow_16_C(const uint16_t* src_y,
+                    uint16_t* dst_y,
+                    int scale,
+                    int width);
+void DivideRow_16_AVX2(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width);
+void DivideRow_16_Any_AVX2(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           int scale,
+                           int width);
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width);
+void DivideRow_16_Any_NEON(const uint16_t* src_ptr,
+                           uint16_t* dst_ptr,
+                           int scale,
+                           int width);
 
 void Convert8To16Row_C(const uint8_t* src_y,
                        uint16_t* dst_y,
@@ -1745,6 +2600,14 @@ void Convert16To8Row_Any_AVX2(const uint16_t* src_ptr,
                               uint8_t* dst_ptr,
                               int scale,
                               int width);
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width);
+void Convert16To8Row_Any_NEON(const uint16_t* src_y,
+                              uint8_t* dst_y,
+                              int scale,
+                              int width);
 
 void CopyRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void CopyRow_AVX(const uint8_t* src, uint8_t* dst, int width);
@@ -1761,16 +2624,12 @@ void CopyRow_16_C(const uint16_t* src, uint16_t* dst, int count);
 void ARGBCopyAlphaRow_C(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
-void ARGBCopyAlphaRow_MMI(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyAlphaRow_Any_SSE2(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
 void ARGBCopyAlphaRow_Any_AVX2(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
-void ARGBCopyAlphaRow_Any_MMI(const uint8_t* src_ptr,
-                              uint8_t* dst_ptr,
-                              int width);
 
 void ARGBExtractAlphaRow_C(const uint8_t* src_argb, uint8_t* dst_a, int width);
 void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb,
@@ -1785,7 +2644,7 @@ void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
 void ARGBExtractAlphaRow_MSA(const uint8_t* src_argb,
                              uint8_t* dst_a,
                              int width);
-void ARGBExtractAlphaRow_MMI(const uint8_t* src_argb,
+void ARGBExtractAlphaRow_LSX(const uint8_t* src_argb,
                              uint8_t* dst_a,
                              int width);
 void ARGBExtractAlphaRow_Any_SSE2(const uint8_t* src_ptr,
@@ -1800,31 +2659,29 @@ void ARGBExtractAlphaRow_Any_NEON(const uint8_t* src_ptr,
 void ARGBExtractAlphaRow_Any_MSA(const uint8_t* src_ptr,
                                  uint8_t* dst_ptr,
                                  int width);
-void ARGBExtractAlphaRow_Any_MMI(const uint8_t* src_ptr,
+void ARGBExtractAlphaRow_Any_LSX(const uint8_t* src_ptr,
                                  uint8_t* dst_ptr,
                                  int width);
 
 void ARGBCopyYToAlphaRow_C(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
-void ARGBCopyYToAlphaRow_MMI(const uint8_t* src, uint8_t* dst, int width);
 void ARGBCopyYToAlphaRow_Any_SSE2(const uint8_t* src_ptr,
                                   uint8_t* dst_ptr,
                                   int width);
 void ARGBCopyYToAlphaRow_Any_AVX2(const uint8_t* src_ptr,
                                   uint8_t* dst_ptr,
                                   int width);
-void ARGBCopyYToAlphaRow_Any_MMI(const uint8_t* src_ptr,
-                                 uint8_t* dst_ptr,
-                                 int width);
 
 void SetRow_C(uint8_t* dst, uint8_t v8, int width);
 void SetRow_MSA(uint8_t* dst, uint8_t v8, int width);
 void SetRow_X86(uint8_t* dst, uint8_t v8, int width);
 void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width);
 void SetRow_NEON(uint8_t* dst, uint8_t v8, int width);
+void SetRow_LSX(uint8_t* dst, uint8_t v8, int width);
 void SetRow_Any_X86(uint8_t* dst_ptr, uint8_t v32, int width);
 void SetRow_Any_NEON(uint8_t* dst_ptr, uint8_t v32, int width);
+void SetRow_Any_LSX(uint8_t* dst_ptr, uint8_t v32, int width);
 
 void ARGBSetRow_C(uint8_t* dst_argb, uint32_t v32, int width);
 void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width);
@@ -1832,6 +2689,8 @@ void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width);
 void ARGBSetRow_Any_NEON(uint8_t* dst_ptr, uint32_t v32, int width);
 void ARGBSetRow_MSA(uint8_t* dst_argb, uint32_t v32, int width);
 void ARGBSetRow_Any_MSA(uint8_t* dst_ptr, uint32_t v32, int width);
+void ARGBSetRow_LSX(uint8_t* dst_argb, uint32_t v32, int width);
+void ARGBSetRow_Any_LSX(uint8_t* dst_ptr, uint32_t v32, int width);
 
 // ARGBShufflers for BGRAToARGB etc.
 void ARGBShuffleRow_C(const uint8_t* src_argb,
@@ -1854,10 +2713,10 @@ void ARGBShuffleRow_MSA(const uint8_t* src_argb,
                         uint8_t* dst_argb,
                         const uint8_t* shuffler,
                         int width);
-void ARGBShuffleRow_MMI(const uint8_t* src_argb,
-                        uint8_t* dst_argb,
-                        const uint8_t* shuffler,
-                        int width);
+void ARGBShuffleRow_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width);
 void ARGBShuffleRow_Any_SSSE3(const uint8_t* src_ptr,
                               uint8_t* dst_ptr,
                               const uint8_t* param,
@@ -1874,15 +2733,16 @@ void ARGBShuffleRow_Any_MSA(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             const uint8_t* param,
                             int width);
-void ARGBShuffleRow_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_ptr,
-                            const uint8_t* param,
-                            int width);
+void ARGBShuffleRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             const uint8_t* param,
+                             int width);
 
 void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
                           uint8_t* dst_argb,
                           int width);
 void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
 void RAWToRGB24Row_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
 void RGB565ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void ARGB1555ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
@@ -1901,42 +2761,57 @@ void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
                          uint8_t* dst_argb,
                          int width);
 void RGB24ToARGBRow_MSA(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
-void RGB24ToARGBRow_MMI(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RGB24ToARGBRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
+void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width);
 void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
 void RAWToARGBRow_MSA(const uint8_t* src_raw, uint8_t* dst_argb, int width);
-void RAWToARGBRow_MMI(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToARGBRow_LSX(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width);
 void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
 void RAWToRGB24Row_MSA(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
-void RAWToRGB24Row_MMI(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
+void RAWToRGB24Row_LSX(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
 void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
                           uint8_t* dst_argb,
                           int width);
 void RGB565ToARGBRow_MSA(const uint8_t* src_rgb565,
                          uint8_t* dst_argb,
                          int width);
-void RGB565ToARGBRow_MMI(const uint8_t* src_rgb565,
+void RGB565ToARGBRow_LSX(const uint8_t* src_rgb565,
                          uint8_t* dst_argb,
                          int width);
+void RGB565ToARGBRow_LASX(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width);
 void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
                             uint8_t* dst_argb,
                             int width);
 void ARGB1555ToARGBRow_MSA(const uint8_t* src_argb1555,
                            uint8_t* dst_argb,
                            int width);
-void ARGB1555ToARGBRow_MMI(const uint8_t* src_argb1555,
+void ARGB1555ToARGBRow_LSX(const uint8_t* src_argb1555,
                            uint8_t* dst_argb,
                            int width);
+void ARGB1555ToARGBRow_LASX(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width);
 void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
                             uint8_t* dst_argb,
                             int width);
 void ARGB4444ToARGBRow_MSA(const uint8_t* src_argb4444,
                            uint8_t* dst_argb,
                            int width);
-void ARGB4444ToARGBRow_MMI(const uint8_t* src_argb4444,
+void ARGB4444ToARGBRow_LSX(const uint8_t* src_argb4444,
                            uint8_t* dst_argb,
                            int width);
+void ARGB4444ToARGBRow_LASX(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width);
 void RGB24ToARGBRow_C(const uint8_t* src_rgb24, uint8_t* dst_argb, int width);
 void RAWToARGBRow_C(const uint8_t* src_raw, uint8_t* dst_argb, int width);
+void RAWToRGBARow_C(const uint8_t* src_raw, uint8_t* dst_rgba, int width);
 void RAWToRGB24Row_C(const uint8_t* src_raw, uint8_t* dst_rgb24, int width);
 void RGB565ToARGBRow_C(const uint8_t* src_rgb565, uint8_t* dst_argb, int width);
 void ARGB1555ToARGBRow_C(const uint8_t* src_argb1555,
@@ -1956,6 +2831,9 @@ void RGB24ToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
 void RAWToARGBRow_Any_SSSE3(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
+void RAWToRGBARow_Any_SSSE3(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
 void RAWToRGB24Row_Any_SSSE3(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
@@ -1985,45 +2863,59 @@ void RGB24ToARGBRow_Any_NEON(const uint8_t* src_ptr,
 void RGB24ToARGBRow_Any_MSA(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
-void RGB24ToARGBRow_Any_MMI(const uint8_t* src_ptr,
+void RGB24ToARGBRow_Any_LSX(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
+void RGB24ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
 void RAWToARGBRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToRGBARow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RAWToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RAWToARGBRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToARGBRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToARGBRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RAWToRGB24Row_Any_NEON(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
 void RAWToRGB24Row_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void RAWToRGB24Row_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void RAWToRGB24Row_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void RGB565ToARGBRow_Any_NEON(const uint8_t* src_ptr,
                               uint8_t* dst_ptr,
                               int width);
 void RGB565ToARGBRow_Any_MSA(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
-void RGB565ToARGBRow_Any_MMI(const uint8_t* src_ptr,
+void RGB565ToARGBRow_Any_LSX(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
+void RGB565ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
 void ARGB1555ToARGBRow_Any_NEON(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int width);
 void ARGB1555ToARGBRow_Any_MSA(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
-void ARGB1555ToARGBRow_Any_MMI(const uint8_t* src_ptr,
+void ARGB4444ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+void ARGB1555ToARGBRow_Any_LSX(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
-void ARGB4444ToARGBRow_Any_NEON(const uint8_t* src_ptr,
+void ARGB1555ToARGBRow_Any_LASX(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int width);
 
 void ARGB4444ToARGBRow_Any_MSA(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
-void ARGB4444ToARGBRow_Any_MMI(const uint8_t* src_ptr,
+void ARGB4444ToARGBRow_Any_LSX(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int width);
+void ARGB4444ToARGBRow_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
 
 void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width);
 void ARGBToRAWRow_SSSE3(const uint8_t* src, uint8_t* dst, int width);
@@ -2091,20 +2983,20 @@ void ARGBToRGB565DitherRow_MSA(const uint8_t* src_argb,
                                uint8_t* dst_rgb,
                                const uint32_t dither4,
                                int width);
+void ARGBToRGB565DitherRow_LASX(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width);
 
-void ARGBToRGB24Row_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
-void ARGBToRAWRow_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
-void ARGBToRGB565Row_MMI(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
-void ARGBToARGB1555Row_MMI(const uint8_t* src_argb,
-                           uint8_t* dst_rgb,
-                           int width);
-void ARGBToARGB4444Row_MMI(const uint8_t* src_argb,
-                           uint8_t* dst_rgb,
-                           int width);
-void ARGBToRGB565DitherRow_MMI(const uint8_t* src_argb,
-                               uint8_t* dst_rgb,
-                               const uint32_t dither4,
-                               int width);
+void ARGBToRGB24Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRAWRow_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToRGB565Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
+void ARGBToARGB1555Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
+void ARGBToARGB4444Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width);
 
 void ARGBToRGBARow_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
 void ARGBToRGB24Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
@@ -2115,11 +3007,76 @@ void ARGBToARGB4444Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width);
 void ABGRToAR30Row_C(const uint8_t* src_abgr, uint8_t* dst_ar30, int width);
 void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width);
 
+void ARGBToAR64Row_C(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_C(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_C(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_C(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void AR64ShuffleRow_C(const uint8_t* src_ar64,
+                      uint8_t* dst_ar64,
+                      const uint8_t* shuffler,
+                      int width);
+void ARGBToAR64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ar64,
+                         int width);
+void ARGBToAB64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ab64,
+                         int width);
+void AR64ToARGBRow_SSSE3(const uint16_t* src_ar64,
+                         uint8_t* dst_argb,
+                         int width);
+void AB64ToARGBRow_SSSE3(const uint16_t* src_ab64,
+                         uint8_t* dst_argb,
+                         int width);
+void ARGBToAR64Row_AVX2(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_AVX2(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_AVX2(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_AVX2(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void ARGBToAR64Row_NEON(const uint8_t* src_argb, uint16_t* dst_ar64, int width);
+void ARGBToAB64Row_NEON(const uint8_t* src_argb, uint16_t* dst_ab64, int width);
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64, uint8_t* dst_argb, int width);
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64, uint8_t* dst_argb, int width);
+void ARGBToAR64Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int width);
+void ARGBToAB64Row_Any_SSSE3(const uint8_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int width);
+void AR64ToARGBRow_Any_SSSE3(const uint16_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void AB64ToARGBRow_Any_SSSE3(const uint16_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int width);
+void ARGBToAR64Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void ARGBToAB64Row_Any_AVX2(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void AR64ToARGBRow_Any_AVX2(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void AB64ToARGBRow_Any_AVX2(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void ARGBToAR64Row_Any_NEON(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void ARGBToAB64Row_Any_NEON(const uint8_t* src_ptr,
+                            uint16_t* dst_ptr,
+                            int width);
+void AR64ToARGBRow_Any_NEON(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+void AB64ToARGBRow_Any_NEON(const uint16_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int width);
+
 void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width);
-void J400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width);
 void J400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
@@ -2131,7 +3088,7 @@ void J400ToARGBRow_Any_NEON(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int width);
 void J400ToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void J400ToARGBRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void J400ToARGBRow_Any_LSX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 
 void I444ToARGBRow_C(const uint8_t* src_y,
                      const uint8_t* src_u,
@@ -2163,6 +3120,51 @@ void I210ToARGBRow_C(const uint16_t* src_y,
                      uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
+void I212ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I212ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I210AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I410AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
+void I444AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width);
 void I422AlphaToARGBRow_C(const uint8_t* src_y,
                           const uint8_t* src_u,
                           const uint8_t* src_v,
@@ -2207,6 +3209,27 @@ void UYVYToARGBRow_C(const uint8_t* src_uyvy,
                      uint8_t* rgb_buf,
                      const struct YuvConstants* yuvconstants,
                      int width);
+void P210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void P410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+
 void I422ToRGBARow_C(const uint8_t* src_y,
                      const uint8_t* src_u,
                      const uint8_t* src_v,
@@ -2243,6 +3266,12 @@ void I422ToARGBRow_AVX2(const uint8_t* y_buf,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
+void I422ToARGBRow_AVX512BW(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422ToRGBARow_AVX2(const uint8_t* y_buf,
                         const uint8_t* u_buf,
                         const uint8_t* v_buf,
@@ -2261,18 +3290,6 @@ void I444ToARGBRow_AVX2(const uint8_t* y_buf,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
-void I444ToARGBRow_SSSE3(const uint8_t* y_buf,
-                         const uint8_t* u_buf,
-                         const uint8_t* v_buf,
-                         uint8_t* dst_argb,
-                         const struct YuvConstants* yuvconstants,
-                         int width);
-void I444ToARGBRow_AVX2(const uint8_t* y_buf,
-                        const uint8_t* u_buf,
-                        const uint8_t* v_buf,
-                        uint8_t* dst_argb,
-                        const struct YuvConstants* yuvconstants,
-                        int width);
 void I422ToARGBRow_SSSE3(const uint8_t* y_buf,
                          const uint8_t* u_buf,
                          const uint8_t* v_buf,
@@ -2298,6 +3315,44 @@ void I210ToARGBRow_SSSE3(const uint16_t* y_buf,
                          uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width);
+void I212ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I212ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* u_buf,
+                         const uint16_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void I210AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                              const uint16_t* u_buf,
+                              const uint16_t* v_buf,
+                              const uint16_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I410AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                              const uint16_t* u_buf,
+                              const uint16_t* v_buf,
+                              const uint16_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
 void I422ToAR30Row_AVX2(const uint8_t* y_buf,
                         const uint8_t* u_buf,
                         const uint8_t* v_buf,
@@ -2316,6 +3371,58 @@ void I210ToAR30Row_AVX2(const uint16_t* y_buf,
                         uint8_t* dst_ar30,
                         const struct YuvConstants* yuvconstants,
                         int width);
+void I212ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I212ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I410ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I410ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* u_buf,
+                        const uint16_t* v_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I210AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void I444AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
 void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
                               const uint8_t* u_buf,
                               const uint8_t* v_buf,
@@ -2365,6 +3472,10 @@ void NV21ToRGB24Row_AVX2(const uint8_t* src_y,
                          uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width);
+void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_yuv24,
+                          int width);
 void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
                          const uint8_t* src_vu,
                          uint8_t* dst_yuv24,
@@ -2400,6 +3511,48 @@ void UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width);
+
+void P210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                         const uint16_t* uv_buf,
+                         uint8_t* dst_ar30,
+                         const struct YuvConstants* yuvconstants,
+                         int width);
+void P210ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToARGBRow_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P210ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToAR30Row_AVX2(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+
 void I422ToRGBARow_SSSE3(const uint8_t* y_buf,
                          const uint8_t* u_buf,
                          const uint8_t* v_buf,
@@ -2460,6 +3613,12 @@ void I422ToARGBRow_Any_AVX2(const uint8_t* y_buf,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I422ToARGBRow_Any_AVX512BW(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
 void I422ToRGBARow_Any_AVX2(const uint8_t* y_buf,
                             const uint8_t* u_buf,
                             const uint8_t* v_buf,
@@ -2502,6 +3661,44 @@ void I210ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
                              uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
+void I212ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I212ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I410ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* u_buf,
+                             const uint16_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void I210AlphaToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                                  const uint16_t* u_buf,
+                                  const uint16_t* v_buf,
+                                  const uint16_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
+void I410AlphaToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                                  const uint16_t* u_buf,
+                                  const uint16_t* v_buf,
+                                  const uint16_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
 void I422ToAR30Row_Any_AVX2(const uint8_t* y_buf,
                             const uint8_t* u_buf,
                             const uint8_t* v_buf,
@@ -2520,6 +3717,58 @@ void I210ToAR30Row_Any_AVX2(const uint16_t* y_buf,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I212ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I212ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I410ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I410ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* u_buf,
+                            const uint16_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void I210AlphaToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                                 const uint16_t* u_buf,
+                                 const uint16_t* v_buf,
+                                 const uint16_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I410AlphaToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                                 const uint16_t* u_buf,
+                                 const uint16_t* v_buf,
+                                 const uint16_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
+void I444AlphaToARGBRow_Any_SSSE3(const uint8_t* y_buf,
+                                  const uint8_t* u_buf,
+                                  const uint8_t* v_buf,
+                                  const uint8_t* a_buf,
+                                  uint8_t* dst_ptr,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width);
+void I444AlphaToARGBRow_Any_AVX2(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
 void I422AlphaToARGBRow_Any_SSSE3(const uint8_t* y_buf,
                                   const uint8_t* u_buf,
                                   const uint8_t* v_buf,
@@ -2574,6 +3823,10 @@ void NV21ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
                              uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
+void NV21ToYUV24Row_Any_SSSE3(const uint8_t* src_y,
+                              const uint8_t* src_vu,
+                              uint8_t* dst_yuv24,
+                              int width);
 void NV21ToYUV24Row_Any_AVX2(const uint8_t* src_y,
                              const uint8_t* src_vu,
                              uint8_t* dst_yuv24,
@@ -2604,6 +3857,46 @@ void UYVYToARGBRow_Any_AVX2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void P210ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P410ToARGBRow_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P210ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P410ToAR30Row_Any_SSSE3(const uint16_t* y_buf,
+                             const uint16_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void P210ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToARGBRow_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P210ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToAR30Row_Any_AVX2(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422ToRGBARow_Any_SSSE3(const uint8_t* y_buf,
                              const uint8_t* u_buf,
                              const uint8_t* v_buf,
@@ -2659,30 +3952,57 @@ void I422ToRGB24Row_Any_AVX2(const uint8_t* y_buf,
                              const struct YuvConstants* yuvconstants,
                              int width);
 
-void I400ToARGBRow_C(const uint8_t* src_y, uint8_t* rgb_buf, int width);
-void I400ToARGBRow_SSE2(const uint8_t* y_buf, uint8_t* dst_argb, int width);
-void I400ToARGBRow_AVX2(const uint8_t* y_buf, uint8_t* dst_argb, int width);
-void I400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width);
-void I400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width);
-void I400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* dst_argb, int width);
+void I400ToARGBRow_C(const uint8_t* src_y,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width);
+void I400ToARGBRow_SSE2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_AVX2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void I400ToARGBRow_MSA(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
+void I400ToARGBRow_LSX(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width);
 void I400ToARGBRow_Any_SSE2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
                             int width);
 void I400ToARGBRow_Any_AVX2(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
                             int width);
 void I400ToARGBRow_Any_NEON(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
+                            const struct YuvConstants* param,
                             int width);
-void I400ToARGBRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void I400ToARGBRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void I400ToARGBRow_Any_MSA(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void I400ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
 
 // ARGB preattenuated alpha blend.
-void ARGBBlendRow_SSSE3(const uint8_t* src_argb0,
+void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
                         const uint8_t* src_argb1,
                         uint8_t* dst_argb,
                         int width);
-void ARGBBlendRow_NEON(const uint8_t* src_argb0,
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width);
@@ -2690,11 +4010,11 @@ void ARGBBlendRow_MSA(const uint8_t* src_argb0,
                       const uint8_t* src_argb1,
                       uint8_t* dst_argb,
                       int width);
-void ARGBBlendRow_MMI(const uint8_t* src_argb0,
+void ARGBBlendRow_LSX(const uint8_t* src_argb0,
                       const uint8_t* src_argb1,
                       uint8_t* dst_argb,
                       int width);
-void ARGBBlendRow_C(const uint8_t* src_argb0,
+void ARGBBlendRow_C(const uint8_t* src_argb,
                     const uint8_t* src_argb1,
                     uint8_t* dst_argb,
                     int width);
@@ -2720,16 +4040,6 @@ void BlendPlaneRow_Any_AVX2(const uint8_t* y_buf,
                             const uint8_t* v_buf,
                             uint8_t* dst_ptr,
                             int width);
-void BlendPlaneRow_MMI(const uint8_t* src0,
-                       const uint8_t* src1,
-                       const uint8_t* alpha,
-                       uint8_t* dst,
-                       int width);
-void BlendPlaneRow_Any_MMI(const uint8_t* y_buf,
-                           const uint8_t* u_buf,
-                           const uint8_t* v_buf,
-                           uint8_t* dst_ptr,
-                           int width);
 void BlendPlaneRow_C(const uint8_t* src0,
                      const uint8_t* src1,
                      const uint8_t* alpha,
@@ -2738,11 +4048,11 @@ void BlendPlaneRow_C(const uint8_t* src0,
 
 // ARGB multiply images. Same API as Blend, but these require
 // pointer and width alignment for SSE2.
-void ARGBMultiplyRow_C(const uint8_t* src_argb0,
+void ARGBMultiplyRow_C(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width);
-void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
+void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2750,7 +4060,7 @@ void ARGBMultiplyRow_Any_SSE2(const uint8_t* y_buf,
                               const uint8_t* uv_buf,
                               uint8_t* dst_ptr,
                               int width);
-void ARGBMultiplyRow_AVX2(const uint8_t* src_argb0,
+void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2758,7 +4068,7 @@ void ARGBMultiplyRow_Any_AVX2(const uint8_t* y_buf,
                               const uint8_t* uv_buf,
                               uint8_t* dst_ptr,
                               int width);
-void ARGBMultiplyRow_NEON(const uint8_t* src_argb0,
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2774,21 +4084,21 @@ void ARGBMultiplyRow_Any_MSA(const uint8_t* y_buf,
                              const uint8_t* uv_buf,
                              uint8_t* dst_ptr,
                              int width);
-void ARGBMultiplyRow_MMI(const uint8_t* src_argb0,
-                         const uint8_t* src_argb1,
-                         uint8_t* dst_argb,
-                         int width);
-void ARGBMultiplyRow_Any_MMI(const uint8_t* y_buf,
-                             const uint8_t* uv_buf,
-                             uint8_t* dst_ptr,
-                             int width);
+void ARGBMultiplyRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBMultiplyRow_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
 
 // ARGB add images.
-void ARGBAddRow_C(const uint8_t* src_argb0,
+void ARGBAddRow_C(const uint8_t* src_argb,
                   const uint8_t* src_argb1,
                   uint8_t* dst_argb,
                   int width);
-void ARGBAddRow_SSE2(const uint8_t* src_argb0,
+void ARGBAddRow_SSE2(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width);
@@ -2796,7 +4106,7 @@ void ARGBAddRow_Any_SSE2(const uint8_t* y_buf,
                          const uint8_t* uv_buf,
                          uint8_t* dst_ptr,
                          int width);
-void ARGBAddRow_AVX2(const uint8_t* src_argb0,
+void ARGBAddRow_AVX2(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width);
@@ -2804,7 +4114,7 @@ void ARGBAddRow_Any_AVX2(const uint8_t* y_buf,
                          const uint8_t* uv_buf,
                          uint8_t* dst_ptr,
                          int width);
-void ARGBAddRow_NEON(const uint8_t* src_argb0,
+void ARGBAddRow_NEON(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width);
@@ -2820,22 +4130,22 @@ void ARGBAddRow_Any_MSA(const uint8_t* y_buf,
                         const uint8_t* uv_buf,
                         uint8_t* dst_ptr,
                         int width);
-void ARGBAddRow_MMI(const uint8_t* src_argb0,
-                    const uint8_t* src_argb1,
-                    uint8_t* dst_argb,
-                    int width);
-void ARGBAddRow_Any_MMI(const uint8_t* y_buf,
-                        const uint8_t* uv_buf,
-                        uint8_t* dst_ptr,
-                        int width);
+void ARGBAddRow_LASX(const uint8_t* src_argb0,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width);
+void ARGBAddRow_Any_LASX(const uint8_t* y_buf,
+                         const uint8_t* uv_buf,
+                         uint8_t* dst_ptr,
+                         int width);
 
 // ARGB subtract images. Same API as Blend, but these require
 // pointer and width alignment for SSE2.
-void ARGBSubtractRow_C(const uint8_t* src_argb0,
+void ARGBSubtractRow_C(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width);
-void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
+void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2843,7 +4153,7 @@ void ARGBSubtractRow_Any_SSE2(const uint8_t* y_buf,
                               const uint8_t* uv_buf,
                               uint8_t* dst_ptr,
                               int width);
-void ARGBSubtractRow_AVX2(const uint8_t* src_argb0,
+void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2851,7 +4161,7 @@ void ARGBSubtractRow_Any_AVX2(const uint8_t* y_buf,
                               const uint8_t* uv_buf,
                               uint8_t* dst_ptr,
                               int width);
-void ARGBSubtractRow_NEON(const uint8_t* src_argb0,
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width);
@@ -2867,14 +4177,14 @@ void ARGBSubtractRow_Any_MSA(const uint8_t* y_buf,
                              const uint8_t* uv_buf,
                              uint8_t* dst_ptr,
                              int width);
-void ARGBSubtractRow_MMI(const uint8_t* src_argb0,
-                         const uint8_t* src_argb1,
-                         uint8_t* dst_argb,
-                         int width);
-void ARGBSubtractRow_Any_MMI(const uint8_t* y_buf,
-                             const uint8_t* uv_buf,
-                             uint8_t* dst_ptr,
-                             int width);
+void ARGBSubtractRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width);
+void ARGBSubtractRow_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              int width);
 
 void ARGBToRGB24Row_Any_SSSE3(const uint8_t* src_ptr,
                               uint8_t* dst_ptr,
@@ -2963,24 +4273,24 @@ void ARGBToRGB565DitherRow_Any_MSA(const uint8_t* src_ptr,
                                    uint8_t* dst_ptr,
                                    const uint32_t param,
                                    int width);
+void ARGBToRGB565DitherRow_Any_LASX(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    const uint32_t param,
+                                    int width);
 
-void ARGBToRGB24Row_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_ptr,
-                            int width);
-void ARGBToRAWRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void ARGBToRGB565Row_Any_MMI(const uint8_t* src_ptr,
+void ARGBToRGB24Row_Any_LASX(const uint8_t* src_ptr,
                              uint8_t* dst_ptr,
                              int width);
-void ARGBToARGB1555Row_Any_MMI(const uint8_t* src_ptr,
-                               uint8_t* dst_ptr,
-                               int width);
-void ARGBToARGB4444Row_Any_MMI(const uint8_t* src_ptr,
-                               uint8_t* dst_ptr,
-                               int width);
-void ARGBToRGB565DitherRow_Any_MMI(const uint8_t* src_ptr,
-                                   uint8_t* dst_ptr,
-                                   const uint32_t param,
-                                   int width);
+void ARGBToRAWRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void ARGBToRGB565Row_Any_LASX(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int width);
+void ARGBToARGB1555Row_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
+void ARGBToARGB4444Row_Any_LASX(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int width);
 
 void I444ToARGBRow_Any_NEON(const uint8_t* y_buf,
                             const uint8_t* u_buf,
@@ -2994,6 +4304,13 @@ void I422ToARGBRow_Any_NEON(const uint8_t* y_buf,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I444AlphaToARGBRow_Any_NEON(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
 void I422AlphaToARGBRow_Any_NEON(const uint8_t* y_buf,
                                  const uint8_t* u_buf,
                                  const uint8_t* v_buf,
@@ -3051,9 +4368,9 @@ void NV21ToRGB24Row_Any_NEON(const uint8_t* y_buf,
                              uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
-void NV21ToYUV24Row_Any_NEON(const uint8_t* src_y,
-                             const uint8_t* src_vu,
-                             uint8_t* dst_yuv24,
+void NV21ToYUV24Row_Any_NEON(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
                              int width);
 void NV12ToRGB565Row_Any_NEON(const uint8_t* y_buf,
                               const uint8_t* uv_buf,
@@ -3068,24 +4385,82 @@ void UYVYToARGBRow_Any_NEON(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void P210ToARGBRow_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToARGBRow_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P210ToAR30Row_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P410ToAR30Row_NEON(const uint16_t* y_buf,
+                        const uint16_t* uv_buf,
+                        uint8_t* dst_ar30,
+                        const struct YuvConstants* yuvconstants,
+                        int width);
+void P210ToARGBRow_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToARGBRow_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_argb,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P210ToAR30Row_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ar30,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void P410ToAR30Row_Any_NEON(const uint16_t* y_buf,
+                            const uint16_t* uv_buf,
+                            uint8_t* dst_ar30,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I444ToARGBRow_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
                            uint8_t* dst_ptr,
                            const struct YuvConstants* yuvconstants,
                            int width);
+void I444ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* u_buf,
+                           const uint8_t* v_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
 void I422ToARGBRow_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
                            uint8_t* dst_ptr,
                            const struct YuvConstants* yuvconstants,
                            int width);
+void I422ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422ToRGBARow_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
                            uint8_t* dst_ptr,
                            const struct YuvConstants* yuvconstants,
                            int width);
+void I422ToRGBARow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
 void I422AlphaToARGBRow_Any_MSA(const uint8_t* y_buf,
                                 const uint8_t* u_buf,
                                 const uint8_t* v_buf,
@@ -3093,30 +4468,61 @@ void I422AlphaToARGBRow_Any_MSA(const uint8_t* y_buf,
                                 uint8_t* dst_ptr,
                                 const struct YuvConstants* yuvconstants,
                                 int width);
+void I422AlphaToARGBRow_Any_LASX(const uint8_t* y_buf,
+                                 const uint8_t* u_buf,
+                                 const uint8_t* v_buf,
+                                 const uint8_t* a_buf,
+                                 uint8_t* dst_ptr,
+                                 const struct YuvConstants* yuvconstants,
+                                 int width);
 void I422ToRGB24Row_Any_MSA(const uint8_t* y_buf,
                             const uint8_t* u_buf,
                             const uint8_t* v_buf,
                             uint8_t* dst_ptr,
                             const struct YuvConstants* yuvconstants,
                             int width);
+void I422ToRGB24Row_Any_LASX(const uint8_t* y_buf,
+                             const uint8_t* u_buf,
+                             const uint8_t* v_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
 void I422ToRGB565Row_Any_MSA(const uint8_t* y_buf,
                              const uint8_t* u_buf,
                              const uint8_t* v_buf,
                              uint8_t* dst_ptr,
                              const struct YuvConstants* yuvconstants,
                              int width);
+void I422ToRGB565Row_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
 void I422ToARGB4444Row_Any_MSA(const uint8_t* y_buf,
                                const uint8_t* u_buf,
                                const uint8_t* v_buf,
                                uint8_t* dst_ptr,
                                const struct YuvConstants* yuvconstants,
                                int width);
+void I422ToARGB4444Row_Any_LASX(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
 void I422ToARGB1555Row_Any_MSA(const uint8_t* y_buf,
                                const uint8_t* u_buf,
                                const uint8_t* v_buf,
                                uint8_t* dst_ptr,
                                const struct YuvConstants* yuvconstants,
                                int width);
+void I422ToARGB1555Row_Any_LASX(const uint8_t* y_buf,
+                                const uint8_t* u_buf,
+                                const uint8_t* v_buf,
+                                uint8_t* dst_ptr,
+                                const struct YuvConstants* yuvconstants,
+                                int width);
 void NV12ToARGBRow_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* uv_buf,
                            uint8_t* dst_ptr,
@@ -3141,6 +4547,45 @@ void UYVYToARGBRow_Any_MSA(const uint8_t* src_ptr,
                            const struct YuvConstants* yuvconstants,
                            int width);
 
+void NV12ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void NV12ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void NV12ToRGB565Row_Any_LSX(const uint8_t* y_buf,
+                             const uint8_t* uv_buf,
+                             uint8_t* dst_ptr,
+                             const struct YuvConstants* yuvconstants,
+                             int width);
+void NV12ToRGB565Row_Any_LASX(const uint8_t* y_buf,
+                              const uint8_t* uv_buf,
+                              uint8_t* dst_ptr,
+                              const struct YuvConstants* yuvconstants,
+                              int width);
+void NV21ToARGBRow_Any_LSX(const uint8_t* y_buf,
+                           const uint8_t* uv_buf,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void NV21ToARGBRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* uv_buf,
+                            uint8_t* dst_ptr,
+                            const struct YuvConstants* yuvconstants,
+                            int width);
+void YUY2ToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+void UYVYToARGBRow_Any_LSX(const uint8_t* src_ptr,
+                           uint8_t* dst_ptr,
+                           const struct YuvConstants* yuvconstants,
+                           int width);
+
 void YUY2ToYRow_AVX2(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
 void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2,
                       int stride_yuy2,
@@ -3172,25 +4617,25 @@ void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
                          uint8_t* dst_v,
                          int width);
 void YUY2ToYRow_MSA(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
-void YUY2ToYRow_MMI(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
+void YUY2ToYRow_LASX(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
 void YUY2ToUVRow_MSA(const uint8_t* src_yuy2,
                      int src_stride_yuy2,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void YUY2ToUVRow_MMI(const uint8_t* src_yuy2,
-                     int src_stride_yuy2,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width);
+void YUY2ToUVRow_LASX(const uint8_t* src_yuy2,
+                      int src_stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
 void YUY2ToUV422Row_MSA(const uint8_t* src_yuy2,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
-void YUY2ToUV422Row_MMI(const uint8_t* src_yuy2,
-                        uint8_t* dst_u,
-                        uint8_t* dst_v,
-                        int width);
+void YUY2ToUV422Row_LASX(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
 void YUY2ToYRow_C(const uint8_t* src_yuy2, uint8_t* dst_y, int width);
 void YUY2ToUVRow_C(const uint8_t* src_yuy2,
                    int src_stride_yuy2,
@@ -3203,7 +4648,7 @@ void YUY2ToUV422Row_C(const uint8_t* src_yuy2,
                       int width);
 void YUY2ToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void YUY2ToUVRow_Any_AVX2(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3213,7 +4658,7 @@ void YUY2ToUV422Row_Any_AVX2(const uint8_t* src_ptr,
                              int width);
 void YUY2ToYRow_Any_SSE2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void YUY2ToUVRow_Any_SSE2(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3223,7 +4668,7 @@ void YUY2ToUV422Row_Any_SSE2(const uint8_t* src_ptr,
                              int width);
 void YUY2ToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void YUY2ToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3232,25 +4677,25 @@ void YUY2ToUV422Row_Any_NEON(const uint8_t* src_ptr,
                              uint8_t* dst_v,
                              int width);
 void YUY2ToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void YUY2ToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void YUY2ToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void YUY2ToUVRow_Any_MSA(const uint8_t* src_ptr,
                          int src_stride_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void YUY2ToUVRow_Any_MMI(const uint8_t* src_ptr,
-                         int src_stride_ptr,
-                         uint8_t* dst_u,
-                         uint8_t* dst_v,
-                         int width);
+void YUY2ToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
 void YUY2ToUV422Row_Any_MSA(const uint8_t* src_ptr,
                             uint8_t* dst_u,
                             uint8_t* dst_v,
                             int width);
-void YUY2ToUV422Row_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_u,
-                            uint8_t* dst_v,
-                            int width);
+void YUY2ToUV422Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
 void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
 void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
                       int stride_uyvy,
@@ -3292,25 +4737,25 @@ void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
                          uint8_t* dst_v,
                          int width);
 void UYVYToYRow_MSA(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
-void UYVYToYRow_MMI(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
+void UYVYToYRow_LASX(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
 void UYVYToUVRow_MSA(const uint8_t* src_uyvy,
                      int src_stride_uyvy,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width);
-void UYVYToUVRow_MMI(const uint8_t* src_uyvy,
-                     int src_stride_uyvy,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width);
+void UYVYToUVRow_LASX(const uint8_t* src_uyvy,
+                      int src_stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width);
 void UYVYToUV422Row_MSA(const uint8_t* src_uyvy,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width);
-void UYVYToUV422Row_MMI(const uint8_t* src_uyvy,
-                        uint8_t* dst_u,
-                        uint8_t* dst_v,
-                        int width);
+void UYVYToUV422Row_LASX(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width);
 
 void UYVYToYRow_C(const uint8_t* src_uyvy, uint8_t* dst_y, int width);
 void UYVYToUVRow_C(const uint8_t* src_uyvy,
@@ -3324,7 +4769,7 @@ void UYVYToUV422Row_C(const uint8_t* src_uyvy,
                       int width);
 void UYVYToYRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void UYVYToUVRow_Any_AVX2(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3334,7 +4779,7 @@ void UYVYToUV422Row_Any_AVX2(const uint8_t* src_ptr,
                              int width);
 void UYVYToYRow_Any_SSE2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void UYVYToUVRow_Any_SSE2(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3344,7 +4789,7 @@ void UYVYToUV422Row_Any_SSE2(const uint8_t* src_ptr,
                              int width);
 void UYVYToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void UYVYToUVRow_Any_NEON(const uint8_t* src_ptr,
-                          int src_stride_ptr,
+                          int src_stride,
                           uint8_t* dst_u,
                           uint8_t* dst_v,
                           int width);
@@ -3353,53 +4798,57 @@ void UYVYToUV422Row_Any_NEON(const uint8_t* src_ptr,
                              uint8_t* dst_v,
                              int width);
 void UYVYToYRow_Any_MSA(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
-void UYVYToYRow_Any_MMI(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void UYVYToYRow_Any_LASX(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void UYVYToUVRow_Any_MSA(const uint8_t* src_ptr,
                          int src_stride_ptr,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width);
-void UYVYToUVRow_Any_MMI(const uint8_t* src_ptr,
-                         int src_stride_ptr,
-                         uint8_t* dst_u,
-                         uint8_t* dst_v,
-                         int width);
+void UYVYToUVRow_Any_LASX(const uint8_t* src_ptr,
+                          int src_stride_ptr,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width);
 void UYVYToUV422Row_Any_MSA(const uint8_t* src_ptr,
                             uint8_t* dst_u,
                             uint8_t* dst_v,
                             int width);
-void UYVYToUV422Row_Any_MMI(const uint8_t* src_ptr,
-                            uint8_t* dst_u,
-                            uint8_t* dst_v,
-                            int width);
-void UVToVURow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width);
-void UVToVURow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width);
-void UVToVURow_Any_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void UYVYToUV422Row_Any_LASX(const uint8_t* src_ptr,
+                             uint8_t* dst_u,
+                             uint8_t* dst_v,
+                             int width);
+void SwapUVRow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void SwapUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_SSSE3(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void SwapUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_vu, int width);
+void SwapUVRow_Any_AVX2(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
 void AYUVToYRow_C(const uint8_t* src_ayuv, uint8_t* dst_y, int width);
 void AYUVToUVRow_C(const uint8_t* src_ayuv,
-                   int stride_ayuv,
+                   int src_stride_ayuv,
                    uint8_t* dst_uv,
                    int width);
 void AYUVToVURow_C(const uint8_t* src_ayuv,
-                   int stride_ayuv,
+                   int src_stride_ayuv,
                    uint8_t* dst_vu,
                    int width);
 void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width);
 void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
-                      int stride_ayuv,
+                      int src_stride_ayuv,
                       uint8_t* dst_uv,
                       int width);
 void AYUVToVURow_NEON(const uint8_t* src_ayuv,
-                      int stride_ayuv,
+                      int src_stride_ayuv,
                       uint8_t* dst_vu,
                       int width);
-void AYUVToYRow_Any_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width);
-void AYUVToUVRow_Any_NEON(const uint8_t* src_ayuv,
-                          int stride_ayuv,
-                          uint8_t* dst_uv,
+void AYUVToYRow_Any_NEON(const uint8_t* src_ptr, uint8_t* dst_ptr, int width);
+void AYUVToUVRow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
+                          uint8_t* dst_vu,
                           int width);
-void AYUVToVURow_Any_NEON(const uint8_t* src_ayuv,
-                          int stride_ayuv,
+void AYUVToVURow_Any_NEON(const uint8_t* src_ptr,
+                          int src_stride,
                           uint8_t* dst_vu,
                           int width);
 
@@ -3478,41 +4927,41 @@ void I422ToYUY2Row_MSA(const uint8_t* src_y,
                        const uint8_t* src_v,
                        uint8_t* dst_yuy2,
                        int width);
-void I422ToYUY2Row_MMI(const uint8_t* src_y,
-                       const uint8_t* src_u,
-                       const uint8_t* src_v,
-                       uint8_t* dst_yuy2,
-                       int width);
+void I422ToYUY2Row_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width);
 void I422ToUYVYRow_MSA(const uint8_t* src_y,
                        const uint8_t* src_u,
                        const uint8_t* src_v,
                        uint8_t* dst_uyvy,
                        int width);
-void I422ToUYVYRow_MMI(const uint8_t* src_y,
-                       const uint8_t* src_u,
-                       const uint8_t* src_v,
-                       uint8_t* dst_uyvy,
-                       int width);
+void I422ToUYVYRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width);
 void I422ToYUY2Row_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
                            uint8_t* dst_ptr,
                            int width);
-void I422ToYUY2Row_Any_MMI(const uint8_t* y_buf,
-                           const uint8_t* u_buf,
-                           const uint8_t* v_buf,
-                           uint8_t* dst_ptr,
-                           int width);
+void I422ToYUY2Row_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
 void I422ToUYVYRow_Any_MSA(const uint8_t* y_buf,
                            const uint8_t* u_buf,
                            const uint8_t* v_buf,
                            uint8_t* dst_ptr,
                            int width);
-void I422ToUYVYRow_Any_MMI(const uint8_t* y_buf,
-                           const uint8_t* u_buf,
-                           const uint8_t* v_buf,
-                           uint8_t* dst_ptr,
-                           int width);
+void I422ToUYVYRow_Any_LASX(const uint8_t* y_buf,
+                            const uint8_t* u_buf,
+                            const uint8_t* v_buf,
+                            uint8_t* dst_ptr,
+                            int width);
 
 // Effects related row functions.
 void ARGBAttenuateRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width);
@@ -3528,9 +4977,9 @@ void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
 void ARGBAttenuateRow_MSA(const uint8_t* src_argb,
                           uint8_t* dst_argb,
                           int width);
-void ARGBAttenuateRow_MMI(const uint8_t* src_argb,
-                          uint8_t* dst_argb,
-                          int width);
+void ARGBAttenuateRow_LASX(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width);
 void ARGBAttenuateRow_Any_SSSE3(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int width);
@@ -3543,9 +4992,9 @@ void ARGBAttenuateRow_Any_NEON(const uint8_t* src_ptr,
 void ARGBAttenuateRow_Any_MSA(const uint8_t* src_ptr,
                               uint8_t* dst_ptr,
                               int width);
-void ARGBAttenuateRow_Any_MMI(const uint8_t* src_ptr,
-                              uint8_t* dst_ptr,
-                              int width);
+void ARGBAttenuateRow_Any_LASX(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int width);
 
 // Inverse table for unattenuate, shared by C and SSE2.
 extern const uint32_t fixed_invtbl8[256];
@@ -3569,13 +5018,13 @@ void ARGBGrayRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width);
 void ARGBGrayRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_argb, int width);
 void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width);
 void ARGBGrayRow_MSA(const uint8_t* src_argb, uint8_t* dst_argb, int width);
-void ARGBGrayRow_MMI(const uint8_t* src_argb, uint8_t* dst_argb, int width);
+void ARGBGrayRow_LASX(const uint8_t* src_argb, uint8_t* dst_argb, int width);
 
 void ARGBSepiaRow_C(uint8_t* dst_argb, int width);
 void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width);
 void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width);
 void ARGBSepiaRow_MSA(uint8_t* dst_argb, int width);
-void ARGBSepiaRow_MMI(uint8_t* dst_argb, int width);
+void ARGBSepiaRow_LASX(uint8_t* dst_argb, int width);
 
 void ARGBColorMatrixRow_C(const uint8_t* src_argb,
                           uint8_t* dst_argb,
@@ -3593,7 +5042,7 @@ void ARGBColorMatrixRow_MSA(const uint8_t* src_argb,
                             uint8_t* dst_argb,
                             const int8_t* matrix_argb,
                             int width);
-void ARGBColorMatrixRow_MMI(const uint8_t* src_argb,
+void ARGBColorMatrixRow_LSX(const uint8_t* src_argb,
                             uint8_t* dst_argb,
                             const int8_t* matrix_argb,
                             int width);
@@ -3632,6 +5081,11 @@ void ARGBQuantizeRow_MSA(uint8_t* dst_argb,
                          int interval_size,
                          int interval_offset,
                          int width);
+void ARGBQuantizeRow_LSX(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width);
 
 void ARGBShadeRow_C(const uint8_t* src_argb,
                     uint8_t* dst_argb,
@@ -3649,10 +5103,10 @@ void ARGBShadeRow_MSA(const uint8_t* src_argb,
                       uint8_t* dst_argb,
                       int width,
                       uint32_t value);
-void ARGBShadeRow_MMI(const uint8_t* src_argb,
-                      uint8_t* dst_argb,
-                      int width,
-                      uint32_t value);
+void ARGBShadeRow_LASX(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value);
 
 // Used for blur.
 void CumulativeSumToAverageRow_SSE2(const int32_t* topleft,
@@ -3666,11 +5120,6 @@ void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
                                   const int32_t* previous_cumsum,
                                   int width);
 
-void ComputeCumulativeSumRow_MMI(const uint8_t* row,
-                                 int32_t* cumsum,
-                                 const int32_t* previous_cumsum,
-                                 int width);
-
 void CumulativeSumToAverageRow_C(const int32_t* tl,
                                  const int32_t* bl,
                                  int w,
@@ -3721,7 +5170,7 @@ void InterpolateRow_MSA(uint8_t* dst_ptr,
                         ptrdiff_t src_stride,
                         int width,
                         int source_y_fraction);
-void InterpolateRow_MMI(uint8_t* dst_ptr,
+void InterpolateRow_LSX(uint8_t* dst_ptr,
                         const uint8_t* src_ptr,
                         ptrdiff_t src_stride,
                         int width,
@@ -3746,7 +5195,7 @@ void InterpolateRow_Any_MSA(uint8_t* dst_ptr,
                             ptrdiff_t src_stride_ptr,
                             int width,
                             int source_y_fraction);
-void InterpolateRow_Any_MMI(uint8_t* dst_ptr,
+void InterpolateRow_Any_LSX(uint8_t* dst_ptr,
                             const uint8_t* src_ptr,
                             ptrdiff_t src_stride_ptr,
                             int width,
@@ -3757,6 +5206,47 @@ void InterpolateRow_16_C(uint16_t* dst_ptr,
                          ptrdiff_t src_stride,
                          int width,
                          int source_y_fraction);
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int width,
+                            int source_y_fraction);
+void InterpolateRow_16_Any_NEON(uint16_t* dst_ptr,
+                                const uint16_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int width,
+                                int source_y_fraction);
+
+void InterpolateRow_16To8_C(uint8_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int scale,
+                            int width,
+                            int source_y_fraction);
+void InterpolateRow_16To8_NEON(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction);
+void InterpolateRow_16To8_Any_NEON(uint8_t* dst_ptr,
+                                   const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int scale,
+                                   int width,
+                                   int source_y_fraction);
+void InterpolateRow_16To8_AVX2(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction);
+void InterpolateRow_16To8_Any_AVX2(uint8_t* dst_ptr,
+                                   const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int scale,
+                                   int width,
+                                   int source_y_fraction);
 
 // Sobel images.
 void SobelXRow_C(const uint8_t* src_y0,
@@ -3779,11 +5269,6 @@ void SobelXRow_MSA(const uint8_t* src_y0,
                    const uint8_t* src_y2,
                    uint8_t* dst_sobelx,
                    int width);
-void SobelXRow_MMI(const uint8_t* src_y0,
-                   const uint8_t* src_y1,
-                   const uint8_t* src_y2,
-                   uint8_t* dst_sobelx,
-                   int width);
 void SobelYRow_C(const uint8_t* src_y0,
                  const uint8_t* src_y1,
                  uint8_t* dst_sobely,
@@ -3800,10 +5285,6 @@ void SobelYRow_MSA(const uint8_t* src_y0,
                    const uint8_t* src_y1,
                    uint8_t* dst_sobely,
                    int width);
-void SobelYRow_MMI(const uint8_t* src_y0,
-                   const uint8_t* src_y1,
-                   uint8_t* dst_sobely,
-                   int width);
 void SobelRow_C(const uint8_t* src_sobelx,
                 const uint8_t* src_sobely,
                 uint8_t* dst_argb,
@@ -3820,7 +5301,7 @@ void SobelRow_MSA(const uint8_t* src_sobelx,
                   const uint8_t* src_sobely,
                   uint8_t* dst_argb,
                   int width);
-void SobelRow_MMI(const uint8_t* src_sobelx,
+void SobelRow_LSX(const uint8_t* src_sobelx,
                   const uint8_t* src_sobely,
                   uint8_t* dst_argb,
                   int width);
@@ -3840,7 +5321,7 @@ void SobelToPlaneRow_MSA(const uint8_t* src_sobelx,
                          const uint8_t* src_sobely,
                          uint8_t* dst_y,
                          int width);
-void SobelToPlaneRow_MMI(const uint8_t* src_sobelx,
+void SobelToPlaneRow_LSX(const uint8_t* src_sobelx,
                          const uint8_t* src_sobely,
                          uint8_t* dst_y,
                          int width);
@@ -3860,7 +5341,7 @@ void SobelXYRow_MSA(const uint8_t* src_sobelx,
                     const uint8_t* src_sobely,
                     uint8_t* dst_argb,
                     int width);
-void SobelXYRow_MMI(const uint8_t* src_sobelx,
+void SobelXYRow_LSX(const uint8_t* src_sobelx,
                     const uint8_t* src_sobely,
                     uint8_t* dst_argb,
                     int width);
@@ -3876,7 +5357,7 @@ void SobelRow_Any_MSA(const uint8_t* y_buf,
                       const uint8_t* uv_buf,
                       uint8_t* dst_ptr,
                       int width);
-void SobelRow_Any_MMI(const uint8_t* y_buf,
+void SobelRow_Any_LSX(const uint8_t* y_buf,
                       const uint8_t* uv_buf,
                       uint8_t* dst_ptr,
                       int width);
@@ -3892,7 +5373,7 @@ void SobelToPlaneRow_Any_MSA(const uint8_t* y_buf,
                              const uint8_t* uv_buf,
                              uint8_t* dst_ptr,
                              int width);
-void SobelToPlaneRow_Any_MMI(const uint8_t* y_buf,
+void SobelToPlaneRow_Any_LSX(const uint8_t* y_buf,
                              const uint8_t* uv_buf,
                              uint8_t* dst_ptr,
                              int width);
@@ -3908,7 +5389,7 @@ void SobelXYRow_Any_MSA(const uint8_t* y_buf,
                         const uint8_t* uv_buf,
                         uint8_t* dst_ptr,
                         int width);
-void SobelXYRow_Any_MMI(const uint8_t* y_buf,
+void SobelXYRow_Any_LSX(const uint8_t* y_buf,
                         const uint8_t* uv_buf,
                         uint8_t* dst_ptr,
                         int width);
@@ -3984,6 +5465,14 @@ void HalfFloatRow_Any_MSA(const uint16_t* src_ptr,
                           uint16_t* dst_ptr,
                           float param,
                           int width);
+void HalfFloatRow_LSX(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width);
+void HalfFloatRow_Any_LSX(const uint16_t* src_ptr,
+                          uint16_t* dst_ptr,
+                          float param,
+                          int width);
 void ByteToFloatRow_C(const uint8_t* src, float* dst, float scale, int width);
 void ByteToFloatRow_NEON(const uint8_t* src,
                          float* dst,
@@ -4018,16 +5507,24 @@ float ScaleSumSamples_NEON(const float* src,
 void ScaleSamples_C(const float* src, float* dst, float scale, int width);
 void ScaleSamples_NEON(const float* src, float* dst, float scale, int width);
 
-void FloatDivToByteRow_C(const float* src_weights,
-                         const float* src_values,
-                         uint8_t* dst_out,
-                         uint8_t* dst_mask,
-                         int width);
-void FloatDivToByteRow_NEON(const float* src_weights,
-                            const float* src_values,
-                            uint8_t* dst_out,
-                            uint8_t* dst_mask,
-                            int width);
+void GaussRow_F32_NEON(const float* src, float* dst, int width);
+void GaussRow_F32_C(const float* src, float* dst, int width);
+
+void GaussCol_F32_NEON(const float* src0,
+                       const float* src1,
+                       const float* src2,
+                       const float* src3,
+                       const float* src4,
+                       float* dst,
+                       int width);
+
+void GaussCol_F32_C(const float* src0,
+                    const float* src1,
+                    const float* src2,
+                    const float* src3,
+                    const float* src4,
+                    float* dst,
+                    int width);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/files/include/libyuv/scale.h b/files/include/libyuv/scale.h
index 23ba1634..443f89c2 100644
--- a/files/include/libyuv/scale.h
+++ b/files/include/libyuv/scale.h
@@ -49,6 +49,18 @@ void ScalePlane_16(const uint16_t* src,
                    int dst_height,
                    enum FilterMode filtering);
 
+// Sample is expected to be in the low 12 bits.
+LIBYUV_API
+void ScalePlane_12(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
+                   enum FilterMode filtering);
+
 // Scales a YUV 4:2:0 image from the src width and height to the
 // dst width and height.
 // If filtering is kFilterNone, a simple nearest-neighbor algorithm is
@@ -97,6 +109,25 @@ int I420Scale_16(const uint16_t* src_y,
                  int dst_height,
                  enum FilterMode filtering);
 
+LIBYUV_API
+int I420Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
 // Scales a YUV 4:4:4 image from the src width and height to the
 // dst width and height.
 // If filtering is kFilterNone, a simple nearest-neighbor algorithm is
@@ -145,6 +176,116 @@ int I444Scale_16(const uint16_t* src_y,
                  int dst_height,
                  enum FilterMode filtering);
 
+LIBYUV_API
+int I444Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+// Scales a YUV 4:2:2 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+LIBYUV_API
+int I422Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering);
+
+LIBYUV_API
+int I422Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+LIBYUV_API
+int I422Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering);
+
+// Scales an NV12 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// kFilterBox is not supported for the UV channel and will be treated as
+// bilinear.
+// Returns 0 if successful.
+
+LIBYUV_API
+int NV12Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_uv,
+              int dst_stride_uv,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering);
+
 #ifdef __cplusplus
 // Legacy API.  Deprecated.
 LIBYUV_API
diff --git a/files/include/libyuv/scale_rgb.h b/files/include/libyuv/scale_rgb.h
new file mode 100644
index 00000000..d17c39fd
--- /dev/null
+++ b/files/include/libyuv/scale_rgb.h
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_RGB_H_
+#define INCLUDE_LIBYUV_SCALE_RGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// RGB can be RAW, RGB24 or YUV24
+// RGB scales 24 bit images by converting a row at a time to ARGB
+// and using ARGB row functions to scale, then convert to RGB.
+// TODO(fbarchard): Allow input/output formats to be specified.
+LIBYUV_API
+int RGBScale(const uint8_t* src_rgb,
+             int src_stride_rgb,
+             int src_width,
+             int src_height,
+             uint8_t* dst_rgb,
+             int dst_stride_rgb,
+             int dst_width,
+             int dst_height,
+             enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_UV_H_
diff --git a/files/include/libyuv/scale_row.h b/files/include/libyuv/scale_row.h
index 6e207a9c..6cb5e128 100644
--- a/files/include/libyuv/scale_row.h
+++ b/files/include/libyuv/scale_row.h
@@ -72,6 +72,43 @@ extern "C" {
 #define HAS_SCALEROWDOWN4_SSSE3
 #endif
 
+// The following are available for gcc/clang x86 platforms:
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+#define HAS_SCALEUVROWDOWN2BOX_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_SSE2
+#define HAS_SCALEROWUP2_LINEAR_SSSE3
+#define HAS_SCALEROWUP2_BILINEAR_SSE2
+#define HAS_SCALEROWUP2_BILINEAR_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_12_SSSE3
+#define HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+#define HAS_SCALEROWUP2_LINEAR_16_SSE2
+#define HAS_SCALEROWUP2_BILINEAR_16_SSE2
+#define HAS_SCALEUVROWUP2_LINEAR_SSSE3
+#define HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+#define HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+#define HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+#endif
+
+// The following are available for gcc/clang x86 platforms, but
+// require clang 3.4 or gcc 4.7.
+// TODO(fbarchard): Port to Visual C
+#if !defined(LIBYUV_DISABLE_X86) &&               \
+    (defined(__x86_64__) || defined(__i386__)) && \
+    (defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#define HAS_SCALEUVROWDOWN2BOX_AVX2
+#define HAS_SCALEROWUP2_LINEAR_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_AVX2
+#define HAS_SCALEROWUP2_LINEAR_12_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_12_AVX2
+#define HAS_SCALEROWUP2_LINEAR_16_AVX2
+#define HAS_SCALEROWUP2_BILINEAR_16_AVX2
+#define HAS_SCALEUVROWUP2_LINEAR_AVX2
+#define HAS_SCALEUVROWUP2_BILINEAR_AVX2
+#define HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+#define HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+#endif
+
 // The following are available on all x86 platforms, but
 // require VS2012, clang 3.4 or gcc 4.7.
 // The code supports NaCL but requires a new compiler and validator.
@@ -96,6 +133,18 @@ extern "C" {
 #define HAS_SCALEROWDOWN34_NEON
 #define HAS_SCALEROWDOWN38_NEON
 #define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEUVROWDOWN2BOX_NEON
+#define HAS_SCALEUVROWDOWNEVEN_NEON
+#define HAS_SCALEROWUP2_LINEAR_NEON
+#define HAS_SCALEROWUP2_BILINEAR_NEON
+#define HAS_SCALEROWUP2_LINEAR_12_NEON
+#define HAS_SCALEROWUP2_BILINEAR_12_NEON
+#define HAS_SCALEROWUP2_LINEAR_16_NEON
+#define HAS_SCALEROWUP2_BILINEAR_16_NEON
+#define HAS_SCALEUVROWUP2_LINEAR_NEON
+#define HAS_SCALEUVROWUP2_BILINEAR_NEON
+#define HAS_SCALEUVROWUP2_LINEAR_16_NEON
+#define HAS_SCALEUVROWUP2_BILINEAR_16_NEON
 #endif
 
 #if !defined(LIBYUV_DISABLE_MSA) && defined(__mips_msa)
@@ -111,21 +160,17 @@ extern "C" {
 #define HAS_SCALEROWDOWN4_MSA
 #endif
 
-#if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
-#define HAS_FIXEDDIV1_MIPS
-#define HAS_FIXEDDIV_MIPS
-#define HAS_SCALEADDROW_16_MMI
-#define HAS_SCALEADDROW_MMI
-#define HAS_SCALEARGBCOLS_MMI
-#define HAS_SCALEARGBCOLSUP2_MMI
-#define HAS_SCALEARGBROWDOWN2_MMI
-#define HAS_SCALEARGBROWDOWNEVEN_MMI
-#define HAS_SCALECOLS_16_MMI
-#define HAS_SCALECOLS_MMI
-#define HAS_SCALEROWDOWN2_16_MMI
-#define HAS_SCALEROWDOWN2_MMI
-#define HAS_SCALEROWDOWN4_16_MMI
-#define HAS_SCALEROWDOWN4_MMI
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#define HAS_SCALEARGBROWDOWN2_LSX
+#define HAS_SCALEARGBROWDOWNEVEN_LSX
+#define HAS_SCALEROWDOWN2_LSX
+#define HAS_SCALEROWDOWN4_LSX
+#define HAS_SCALEROWDOWN38_LSX
+#define HAS_SCALEFILTERCOLS_LSX
+#define HAS_SCALEADDROW_LSX
+#define HAS_SCALEARGBCOLS_LSX
+#define HAS_SCALEARGBFILTERCOLS_LSX
+#define HAS_SCALEROWDOWN34_LSX
 #endif
 
 // Scale ARGB vertically with bilinear interpolation.
@@ -155,6 +200,20 @@ void ScalePlaneVertical_16(int src_height,
                            int wpp,
                            enum FilterMode filtering);
 
+void ScalePlaneVertical_16To8(int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint16_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int y,
+                              int dy,
+                              int wpp,
+                              int scale,
+                              enum FilterMode filtering);
+
 // Simplify the filtering based on scale factors.
 enum FilterMode ScaleFilterReduce(int src_width,
                                   int src_height,
@@ -260,6 +319,40 @@ void ScaleRowDown34_1_Box_16_C(const uint16_t* src_ptr,
                                ptrdiff_t src_stride,
                                uint16_t* d,
                                int dst_width);
+
+void ScaleRowUp2_Linear_C(const uint8_t* src_ptr,
+                          uint8_t* dst_ptr,
+                          int dst_width);
+void ScaleRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            ptrdiff_t dst_stride,
+                            int dst_width);
+void ScaleRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_Any_C(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowUp2_Bilinear_Any_C(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width);
+void ScaleRowUp2_Linear_16_Any_C(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_C(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+
 void ScaleCols_C(uint8_t* dst_ptr,
                  const uint8_t* src_ptr,
                  int dst_width,
@@ -375,6 +468,87 @@ void ScaleARGBFilterCols64_C(uint8_t* dst_argb,
                              int dst_width,
                              int x32,
                              int dx);
+void ScaleUVRowDown2_C(const uint8_t* src_uv,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst_uv,
+                       int dst_width);
+void ScaleUVRowDown2Linear_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDown2Box_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_uv,
+                          int dst_width);
+void ScaleUVRowDownEven_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          int src_stepx,
+                          uint8_t* dst_uv,
+                          int dst_width);
+void ScaleUVRowDownEvenBox_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width);
+
+void ScaleUVRowUp2_Linear_C(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleUVRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              ptrdiff_t dst_stride,
+                              int dst_width);
+void ScaleUVRowUp2_Linear_Any_C(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_C(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleUVRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint16_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_C(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_C(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+
+void ScaleUVCols_C(uint8_t* dst_uv,
+                   const uint8_t* src_uv,
+                   int dst_width,
+                   int x,
+                   int dx);
+void ScaleUVCols64_C(uint8_t* dst_uv,
+                     const uint8_t* src_uv,
+                     int dst_width,
+                     int x32,
+                     int dx);
+void ScaleUVColsUp2_C(uint8_t* dst_uv,
+                      const uint8_t* src_uv,
+                      int dst_width,
+                      int,
+                      int);
+void ScaleUVFilterCols_C(uint8_t* dst_uv,
+                         const uint8_t* src_uv,
+                         int dst_width,
+                         int x,
+                         int dx);
+void ScaleUVFilterCols64_C(uint8_t* dst_uv,
+                           const uint8_t* src_uv,
+                           int dst_width,
+                           int x32,
+                           int dx);
 
 // Specialized scalers for x86.
 void ScaleRowDown2_SSSE3(const uint8_t* src_ptr,
@@ -442,6 +616,120 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8_t* src_ptr,
                                 ptrdiff_t src_stride,
                                 uint8_t* dst_ptr,
                                 int dst_width);
+
+void ScaleRowUp2_Linear_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_SSE2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_SSSE3(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_12_SSSE3(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_16_SSE2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_SSE2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width);
+void ScaleRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_12_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_Any_SSE2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_SSE2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_SSSE3(const uint16_t* src_ptr,
+                                     uint16_t* dst_ptr,
+                                     int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_SSSE3(const uint16_t* src_ptr,
+                                       ptrdiff_t src_stride,
+                                       uint16_t* dst_ptr,
+                                       ptrdiff_t dst_stride,
+                                       int dst_width);
+void ScaleRowUp2_Linear_16_Any_SSE2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_SSE2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowUp2_Bilinear_Any_SSSE3(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleRowUp2_Linear_Any_AVX2(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_AVX2(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_AVX2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_AVX2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_16_Any_AVX2(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_AVX2(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+
 void ScaleRowDown2_Any_SSSE3(const uint8_t* src_ptr,
                              ptrdiff_t src_stride,
                              uint8_t* dst_ptr,
@@ -592,16 +880,6 @@ void ScaleARGBCols_Any_MSA(uint8_t* dst_ptr,
                            int dst_width,
                            int x,
                            int dx);
-void ScaleARGBCols_MMI(uint8_t* dst_argb,
-                       const uint8_t* src_argb,
-                       int dst_width,
-                       int x,
-                       int dx);
-void ScaleARGBCols_Any_MMI(uint8_t* dst_ptr,
-                           const uint8_t* src_ptr,
-                           int dst_width,
-                           int x,
-                           int dx);
 
 // ARGB Row functions
 void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
@@ -640,15 +918,15 @@ void ScaleARGBRowDown2Box_MSA(const uint8_t* src_argb,
                               ptrdiff_t src_stride,
                               uint8_t* dst_argb,
                               int dst_width);
-void ScaleARGBRowDown2_MMI(const uint8_t* src_argb,
+void ScaleARGBRowDown2_LSX(const uint8_t* src_argb,
                            ptrdiff_t src_stride,
                            uint8_t* dst_argb,
                            int dst_width);
-void ScaleARGBRowDown2Linear_MMI(const uint8_t* src_argb,
+void ScaleARGBRowDown2Linear_LSX(const uint8_t* src_argb,
                                  ptrdiff_t src_stride,
                                  uint8_t* dst_argb,
                                  int dst_width);
-void ScaleARGBRowDown2Box_MMI(const uint8_t* src_argb,
+void ScaleARGBRowDown2Box_LSX(const uint8_t* src_argb,
                               ptrdiff_t src_stride,
                               uint8_t* dst_argb,
                               int dst_width);
@@ -688,15 +966,15 @@ void ScaleARGBRowDown2Box_Any_MSA(const uint8_t* src_ptr,
                                   ptrdiff_t src_stride,
                                   uint8_t* dst_ptr,
                                   int dst_width);
-void ScaleARGBRowDown2_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBRowDown2_Any_LSX(const uint8_t* src_ptr,
                                ptrdiff_t src_stride,
                                uint8_t* dst_ptr,
                                int dst_width);
-void ScaleARGBRowDown2Linear_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBRowDown2Linear_Any_LSX(const uint8_t* src_ptr,
                                      ptrdiff_t src_stride,
                                      uint8_t* dst_ptr,
                                      int dst_width);
-void ScaleARGBRowDown2Box_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBRowDown2Box_Any_LSX(const uint8_t* src_ptr,
                                   ptrdiff_t src_stride,
                                   uint8_t* dst_ptr,
                                   int dst_width);
@@ -730,12 +1008,12 @@ void ScaleARGBRowDownEvenBox_MSA(const uint8_t* src_argb,
                                  int src_stepx,
                                  uint8_t* dst_argb,
                                  int dst_width);
-void ScaleARGBRowDownEven_MMI(const uint8_t* src_argb,
+void ScaleARGBRowDownEven_LSX(const uint8_t* src_argb,
                               ptrdiff_t src_stride,
                               int32_t src_stepx,
                               uint8_t* dst_argb,
                               int dst_width);
-void ScaleARGBRowDownEvenBox_MMI(const uint8_t* src_argb,
+void ScaleARGBRowDownEvenBox_LSX(const uint8_t* src_argb,
                                  ptrdiff_t src_stride,
                                  int src_stepx,
                                  uint8_t* dst_argb,
@@ -770,16 +1048,255 @@ void ScaleARGBRowDownEvenBox_Any_MSA(const uint8_t* src_ptr,
                                      int src_stepx,
                                      uint8_t* dst_ptr,
                                      int dst_width);
-void ScaleARGBRowDownEven_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBRowDownEven_Any_LSX(const uint8_t* src_ptr,
                                   ptrdiff_t src_stride,
                                   int32_t src_stepx,
                                   uint8_t* dst_ptr,
                                   int dst_width);
-void ScaleARGBRowDownEvenBox_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBRowDownEvenBox_Any_LSX(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     int src_stepx,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+
+// UV Row functions
+void ScaleUVRowDown2_SSSE3(const uint8_t* src_ptr,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_uv,
+                           int dst_width);
+void ScaleUVRowDown2Linear_SSSE3(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_uv,
+                                 int dst_width);
+void ScaleUVRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_uv,
+                              int dst_width);
+void ScaleUVRowDown2Box_AVX2(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDown2_NEON(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width);
+void ScaleUVRowDown2Linear_NEON(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_uv,
+                                int dst_width);
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width);
+void ScaleUVRowDown2_MSA(const uint8_t* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8_t* dst_uv,
+                         int dst_width);
+void ScaleUVRowDown2Linear_MSA(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_uv,
+                               int dst_width);
+void ScaleUVRowDown2Box_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_uv,
+                            int dst_width);
+void ScaleUVRowDown2_Any_SSSE3(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowDown2Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     int dst_width);
+void ScaleUVRowDown2Box_Any_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowDown2Box_Any_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDown2_Any_NEON(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleUVRowDown2Linear_Any_NEON(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowDown2Box_Any_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDown2_Any_MSA(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleUVRowDown2Linear_Any_MSA(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowDown2Box_Any_MSA(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowDownEven_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              int src_stepx,
+                              uint8_t* dst_uv,
+                              int dst_width);
+void ScaleUVRowDownEvenBox_SSSE3(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_uv,
+                                 int dst_width);
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width);
+void ScaleUVRowDownEvenBox_NEON(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int src_stepx,
+                                uint8_t* dst_uv,
+                                int dst_width);
+void ScaleUVRowDownEven_MSA(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int32_t src_stepx,
+                            uint8_t* dst_uv,
+                            int dst_width);
+void ScaleUVRowDownEvenBox_MSA(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8_t* dst_uv,
+                               int dst_width);
+void ScaleUVRowDownEven_Any_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowDownEvenBox_Any_SSSE3(const uint8_t* src_ptr,
                                      ptrdiff_t src_stride,
                                      int src_stepx,
                                      uint8_t* dst_ptr,
                                      int dst_width);
+void ScaleUVRowDownEven_Any_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleUVRowDownEvenBox_Any_NEON(const uint8_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    int src_stepx,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowDownEven_Any_MSA(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                int32_t src_stepx,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowDownEvenBox_Any_MSA(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+
+void ScaleUVRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width);
+void ScaleUVRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleUVRowUp2_Linear_Any_SSSE3(const uint8_t* src_ptr,
+                                    uint8_t* dst_ptr,
+                                    int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_SSSE3(const uint8_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleUVRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_Any_AVX2(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_AVX2(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width);
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width);
+void ScaleUVRowUp2_Linear_Any_NEON(const uint8_t* src_ptr,
+                                   uint8_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_Any_NEON(const uint8_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_16_SSE41(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width);
+void ScaleUVRowUp2_Bilinear_16_SSE41(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_SSE41(const uint16_t* src_ptr,
+                                       uint16_t* dst_ptr,
+                                       int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_SSE41(const uint16_t* src_ptr,
+                                         ptrdiff_t src_stride,
+                                         uint16_t* dst_ptr,
+                                         ptrdiff_t dst_stride,
+                                         int dst_width);
+void ScaleUVRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_AVX2(const uint16_t* src_ptr,
+                                      uint16_t* dst_ptr,
+                                      int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_AVX2(const uint16_t* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint16_t* dst_ptr,
+                                        ptrdiff_t dst_stride,
+                                        int dst_width);
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width);
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width);
+void ScaleUVRowUp2_Linear_16_Any_NEON(const uint16_t* src_ptr,
+                                      uint16_t* dst_ptr,
+                                      int dst_width);
+void ScaleUVRowUp2_Bilinear_16_Any_NEON(const uint16_t* src_ptr,
+                                        ptrdiff_t src_stride,
+                                        uint16_t* dst_ptr,
+                                        ptrdiff_t dst_stride,
+                                        int dst_width);
 
 // ScaleRowDown2Box also used by planar functions
 // NEON downscalers with interpolation.
@@ -891,6 +1408,55 @@ void ScaleRowDown38_2_Box_Any_NEON(const uint8_t* src_ptr,
                                    uint8_t* dst_ptr,
                                    int dst_width);
 
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width);
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width);
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width);
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width);
+void ScaleRowUp2_Linear_Any_NEON(const uint8_t* src_ptr,
+                                 uint8_t* dst_ptr,
+                                 int dst_width);
+void ScaleRowUp2_Bilinear_Any_NEON(const uint8_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint8_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width);
+void ScaleRowUp2_Linear_12_Any_NEON(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_12_Any_NEON(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+void ScaleRowUp2_Linear_16_Any_NEON(const uint16_t* src_ptr,
+                                    uint16_t* dst_ptr,
+                                    int dst_width);
+void ScaleRowUp2_Bilinear_16_Any_NEON(const uint16_t* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint16_t* dst_ptr,
+                                      ptrdiff_t dst_stride,
+                                      int dst_width);
+
 void ScaleAddRow_NEON(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
 void ScaleAddRow_Any_NEON(const uint8_t* src_ptr,
                           uint16_t* dst_ptr,
@@ -1012,93 +1578,129 @@ void ScaleRowDown34_1_Box_Any_MSA(const uint8_t* src_ptr,
                                   uint8_t* dst_ptr,
                                   int dst_width);
 
-void ScaleRowDown2_MMI(const uint8_t* src_ptr,
+void ScaleRowDown2_LSX(const uint8_t* src_ptr,
                        ptrdiff_t src_stride,
                        uint8_t* dst,
                        int dst_width);
-void ScaleRowDown2_16_MMI(const uint16_t* src_ptr,
-                          ptrdiff_t src_stride,
-                          uint16_t* dst,
-                          int dst_width);
-void ScaleRowDown2Linear_MMI(const uint8_t* src_ptr,
+void ScaleRowDown2Linear_LSX(const uint8_t* src_ptr,
                              ptrdiff_t src_stride,
                              uint8_t* dst,
                              int dst_width);
-void ScaleRowDown2Linear_16_MMI(const uint16_t* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint16_t* dst,
-                                int dst_width);
-void ScaleRowDown2Box_MMI(const uint8_t* src_ptr,
+void ScaleRowDown2Box_LSX(const uint8_t* src_ptr,
                           ptrdiff_t src_stride,
                           uint8_t* dst,
                           int dst_width);
-void ScaleRowDown2Box_16_MMI(const uint16_t* src_ptr,
-                             ptrdiff_t src_stride,
-                             uint16_t* dst,
-                             int dst_width);
-void ScaleRowDown2Box_Odd_MMI(const uint8_t* src_ptr,
-                              ptrdiff_t src_stride,
-                              uint8_t* dst,
-                              int dst_width);
-void ScaleRowDown4_MMI(const uint8_t* src_ptr,
+void ScaleRowDown4_LSX(const uint8_t* src_ptr,
                        ptrdiff_t src_stride,
                        uint8_t* dst,
                        int dst_width);
-void ScaleRowDown4_16_MMI(const uint16_t* src_ptr,
-                          ptrdiff_t src_stride,
-                          uint16_t* dst,
-                          int dst_width);
-void ScaleRowDown4Box_MMI(const uint8_t* src_ptr,
+void ScaleRowDown4Box_LSX(const uint8_t* src_ptr,
                           ptrdiff_t src_stride,
                           uint8_t* dst,
                           int dst_width);
-void ScaleRowDown4Box_16_MMI(const uint16_t* src_ptr,
-                             ptrdiff_t src_stride,
-                             uint16_t* dst,
-                             int dst_width);
-void ScaleAddRow_MMI(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
-void ScaleAddRow_16_MMI(const uint16_t* src_ptr,
-                        uint32_t* dst_ptr,
-                        int src_width);
-void ScaleColsUp2_MMI(uint8_t* dst_ptr,
-                      const uint8_t* src_ptr,
-                      int dst_width,
-                      int x,
-                      int dx);
-void ScaleColsUp2_16_MMI(uint16_t* dst_ptr,
-                         const uint16_t* src_ptr,
+void ScaleRowDown38_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown38_2_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleRowDown38_3_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width);
+void ScaleAddRow_LSX(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width);
+void ScaleFilterCols_LSX(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
                          int dst_width,
                          int x,
                          int dx);
-void ScaleARGBColsUp2_MMI(uint8_t* dst_argb,
-                          const uint8_t* src_argb,
-                          int dst_width,
-                          int x,
-                          int dx);
-
-void ScaleRowDown2_Any_MMI(const uint8_t* src_ptr,
+void ScaleARGBFilterCols_LSX(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_LSX(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx);
+void ScaleRowDown34_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width);
+void ScaleRowDown34_0_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+void ScaleRowDown34_1_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width);
+void ScaleRowDown2_Any_LSX(const uint8_t* src_ptr,
                            ptrdiff_t src_stride,
                            uint8_t* dst_ptr,
                            int dst_width);
-void ScaleRowDown2Linear_Any_MMI(const uint8_t* src_ptr,
+void ScaleRowDown2Linear_Any_LSX(const uint8_t* src_ptr,
                                  ptrdiff_t src_stride,
                                  uint8_t* dst_ptr,
                                  int dst_width);
-void ScaleRowDown2Box_Any_MMI(const uint8_t* src_ptr,
+void ScaleRowDown2Box_Any_LSX(const uint8_t* src_ptr,
                               ptrdiff_t src_stride,
                               uint8_t* dst_ptr,
                               int dst_width);
-void ScaleRowDown4_Any_MMI(const uint8_t* src_ptr,
+void ScaleRowDown4_Any_LSX(const uint8_t* src_ptr,
                            ptrdiff_t src_stride,
                            uint8_t* dst_ptr,
                            int dst_width);
-void ScaleRowDown4Box_Any_MMI(const uint8_t* src_ptr,
+void ScaleRowDown4Box_Any_LSX(const uint8_t* src_ptr,
                               ptrdiff_t src_stride,
                               uint8_t* dst_ptr,
                               int dst_width);
-void ScaleAddRow_Any_MMI(const uint8_t* src_ptr,
+void ScaleRowDown38_Any_LSX(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown38_2_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown38_3_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleAddRow_Any_LSX(const uint8_t* src_ptr,
                          uint16_t* dst_ptr,
                          int src_width);
+void ScaleFilterCols_Any_LSX(uint8_t* dst_ptr,
+                             const uint8_t* src_ptr,
+                             int dst_width,
+                             int x,
+                             int dx);
+void ScaleARGBCols_Any_LSX(uint8_t* dst_ptr,
+                           const uint8_t* src_ptr,
+                           int dst_width,
+                           int x,
+                           int dx);
+void ScaleARGBFilterCols_Any_LSX(uint8_t* dst_ptr,
+                                 const uint8_t* src_ptr,
+                                 int dst_width,
+                                 int x,
+                                 int dx);
+void ScaleRowDown34_Any_LSX(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            int dst_width);
+void ScaleRowDown34_0_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+void ScaleRowDown34_1_Box_Any_LSX(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  int dst_width);
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/include/libyuv/scale_uv.h b/files/include/libyuv/scale_uv.h
new file mode 100644
index 00000000..8e74e319
--- /dev/null
+++ b/files/include/libyuv/scale_uv.h
@@ -0,0 +1,51 @@
+/*
+ *  Copyright 2020 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_UV_H_
+#define INCLUDE_LIBYUV_SCALE_UV_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int UVScale(const uint8_t* src_uv,
+            int src_stride_uv,
+            int src_width,
+            int src_height,
+            uint8_t* dst_uv,
+            int dst_stride_uv,
+            int dst_width,
+            int dst_height,
+            enum FilterMode filtering);
+
+// Scale a 16 bit UV image.
+// This function is currently incomplete, it can't handle all cases.
+LIBYUV_API
+int UVScale_16(const uint16_t* src_uv,
+               int src_stride_uv,
+               int src_width,
+               int src_height,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int dst_width,
+               int dst_height,
+               enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_UV_H_
diff --git a/files/include/libyuv/version.h b/files/include/libyuv/version.h
index 741ef34d..a85be048 100644
--- a/files/include/libyuv/version.h
+++ b/files/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1732
+#define LIBYUV_VERSION 1837
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_
diff --git a/files/include/libyuv/video_common.h b/files/include/libyuv/video_common.h
index ffcbdbf1..32b8a521 100644
--- a/files/include/libyuv/video_common.h
+++ b/files/include/libyuv/video_common.h
@@ -50,7 +50,7 @@ extern "C" {
 // Secondary formats are converted in 2 steps.
 // Auxilliary formats call primary converters.
 enum FourCC {
-  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  // 10 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
   FOURCC_I420 = FOURCC('I', '4', '2', '0'),
   FOURCC_I422 = FOURCC('I', '4', '2', '2'),
   FOURCC_I444 = FOURCC('I', '4', '4', '4'),
@@ -59,17 +59,20 @@ enum FourCC {
   FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
   FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
   FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
-  FOURCC_H010 = FOURCC('H', '0', '1', '0'),  // unofficial fourcc. 10 bit lsb
+  FOURCC_I010 = FOURCC('I', '0', '1', '0'),  // bt.601 10 bit 420
+  FOURCC_I210 = FOURCC('I', '2', '1', '0'),  // bt.601 10 bit 422
 
-  // 1 Secondary YUV format: row biplanar.
+  // 1 Secondary YUV format: row biplanar.  deprecated.
   FOURCC_M420 = FOURCC('M', '4', '2', '0'),
 
-  // 11 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc
+  // 13 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp, 1 10 bpc 2 64 bpp
   FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
   FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
   FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
   FOURCC_AR30 = FOURCC('A', 'R', '3', '0'),  // 10 bit per channel. 2101010.
   FOURCC_AB30 = FOURCC('A', 'B', '3', '0'),  // ABGR version of 10 bit
+  FOURCC_AR64 = FOURCC('A', 'R', '6', '4'),  // 16 bit per channel.
+  FOURCC_AB64 = FOURCC('A', 'B', '6', '4'),  // ABGR version of 16 bit
   FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
   FOURCC_RAW = FOURCC('r', 'a', 'w', ' '),
   FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
@@ -80,15 +83,36 @@ enum FourCC {
   // 1 Primary Compressed YUV format.
   FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
 
-  // 8 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  // 14 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
   FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
   FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
   FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
   FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
-  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
-  FOURCC_J400 = FOURCC('J', '4', '0', '0'),  // unofficial fourcc
-  FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // unofficial fourcc
-  FOURCC_H422 = FOURCC('H', '4', '2', '2'),  // unofficial fourcc
+  FOURCC_J420 =
+      FOURCC('J', '4', '2', '0'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J422 =
+      FOURCC('J', '4', '2', '2'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J444 =
+      FOURCC('J', '4', '4', '4'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_J400 =
+      FOURCC('J', '4', '0', '0'),  // jpeg (bt.601 full), unofficial fourcc
+  FOURCC_F420 = FOURCC('F', '4', '2', '0'),  // bt.709 full, unofficial fourcc
+  FOURCC_F422 = FOURCC('F', '4', '2', '2'),  // bt.709 full, unofficial fourcc
+  FOURCC_F444 = FOURCC('F', '4', '4', '4'),  // bt.709 full, unofficial fourcc
+  FOURCC_H420 = FOURCC('H', '4', '2', '0'),  // bt.709, unofficial fourcc
+  FOURCC_H422 = FOURCC('H', '4', '2', '2'),  // bt.709, unofficial fourcc
+  FOURCC_H444 = FOURCC('H', '4', '4', '4'),  // bt.709, unofficial fourcc
+  FOURCC_U420 = FOURCC('U', '4', '2', '0'),  // bt.2020, unofficial fourcc
+  FOURCC_U422 = FOURCC('U', '4', '2', '2'),  // bt.2020, unofficial fourcc
+  FOURCC_U444 = FOURCC('U', '4', '4', '4'),  // bt.2020, unofficial fourcc
+  FOURCC_F010 = FOURCC('F', '0', '1', '0'),  // bt.709 full range 10 bit 420
+  FOURCC_H010 = FOURCC('H', '0', '1', '0'),  // bt.709 10 bit 420
+  FOURCC_U010 = FOURCC('U', '0', '1', '0'),  // bt.2020 10 bit 420
+  FOURCC_F210 = FOURCC('F', '2', '1', '0'),  // bt.709 full range 10 bit 422
+  FOURCC_H210 = FOURCC('H', '2', '1', '0'),  // bt.709 10 bit 422
+  FOURCC_U210 = FOURCC('U', '2', '1', '0'),  // bt.2020 10 bit 422
+  FOURCC_P010 = FOURCC('P', '0', '1', '0'),
+  FOURCC_P210 = FOURCC('P', '2', '1', '0'),
 
   // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
   FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
@@ -133,7 +157,7 @@ enum FourCCBpp {
   FOURCC_BPP_NV12 = 12,
   FOURCC_BPP_YUY2 = 16,
   FOURCC_BPP_UYVY = 16,
-  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_M420 = 12,  // deprecated
   FOURCC_BPP_Q420 = 12,
   FOURCC_BPP_ARGB = 32,
   FOURCC_BPP_BGRA = 32,
@@ -141,6 +165,8 @@ enum FourCCBpp {
   FOURCC_BPP_RGBA = 32,
   FOURCC_BPP_AR30 = 32,
   FOURCC_BPP_AB30 = 32,
+  FOURCC_BPP_AR64 = 64,
+  FOURCC_BPP_AB64 = 64,
   FOURCC_BPP_24BG = 24,
   FOURCC_BPP_RAW = 24,
   FOURCC_BPP_RGBP = 16,
@@ -158,7 +184,12 @@ enum FourCCBpp {
   FOURCC_BPP_J400 = 8,
   FOURCC_BPP_H420 = 12,
   FOURCC_BPP_H422 = 16,
-  FOURCC_BPP_H010 = 24,
+  FOURCC_BPP_I010 = 15,
+  FOURCC_BPP_I210 = 20,
+  FOURCC_BPP_H010 = 15,
+  FOURCC_BPP_H210 = 20,
+  FOURCC_BPP_P010 = 15,
+  FOURCC_BPP_P210 = 20,
   FOURCC_BPP_MJPG = 0,  // 0 means unknown.
   FOURCC_BPP_H264 = 0,
   FOURCC_BPP_IYUV = 12,
diff --git a/files/infra/config/PRESUBMIT.py b/files/infra/config/PRESUBMIT.py
index 89eaa519..01ec0eed 100644
--- a/files/infra/config/PRESUBMIT.py
+++ b/files/infra/config/PRESUBMIT.py
@@ -1,10 +1,6 @@
-# Copyright 2018 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
+# Copyright 2018 The PDFium Authors. All rights reserved.
+# Use of this source code is governed by a BSD-style license that can be
+# found in the LICENSE file.
 
 
 def CheckChangeOnUpload(input_api, output_api):
diff --git a/files/infra/config/README.md b/files/infra/config/README.md
index c036d610..e5e3b5f8 100644
--- a/files/infra/config/README.md
+++ b/files/infra/config/README.md
@@ -1 +1,2 @@
-This directory contains configuration files for infra services.
+This folder contains libyuv project-wide configurations
+for chrome-infra services.
diff --git a/files/infra/config/codereview.settings b/files/infra/config/codereview.settings
new file mode 100644
index 00000000..6d742273
--- /dev/null
+++ b/files/infra/config/codereview.settings
@@ -0,0 +1,6 @@
+# This file is used by gcl and git-cl to get repository specific information.
+CODE_REVIEW_SERVER: codereview.chromium.org
+PROJECT: libyuv
+GERRIT_HOST: True
+VIEW_VC: https://chromium.googlesource.com/libyuv/libyuv/+/
+
diff --git a/files/infra/config/commit-queue.cfg b/files/infra/config/commit-queue.cfg
new file mode 100644
index 00000000..4a8d77f4
--- /dev/null
+++ b/files/infra/config/commit-queue.cfg
@@ -0,0 +1,143 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see Config message:
+#   https://luci-config.appspot.com/schemas/projects:commit-queue.cfg
+
+cq_status_host: "chromium-cq-status.appspot.com"
+submit_options {
+  max_burst: 4
+  burst_delay {
+    seconds: 480
+  }
+}
+config_groups {
+  name: "config"
+  gerrit {
+    url: "https://chromium-review.googlesource.com"
+    projects {
+      name: "libyuv/libyuv"
+      ref_regexp: "refs/heads/infra/config"
+    }
+  }
+  verifiers {
+    gerrit_cq_ability {
+      committer_list: "project-libyuv-committers"
+      dry_run_access_list: "project-libyuv-tryjob-access"
+    }
+    tryjob {
+      builders {
+        name: "libyuv/try/presubmit"
+      }
+      retry_config {
+        single_quota: 1
+        global_quota: 2
+        failure_weight: 1
+        transient_failure_weight: 1
+        timeout_weight: 2
+      }
+    }
+  }
+}
+config_groups {
+  name: "master"
+  gerrit {
+    url: "https://chromium-review.googlesource.com"
+    projects {
+      name: "libyuv/libyuv"
+      ref_regexp: "refs/heads/main"
+      ref_regexp: "refs/heads/master"
+    }
+  }
+  verifiers {
+    gerrit_cq_ability {
+      committer_list: "project-libyuv-committers"
+      dry_run_access_list: "project-libyuv-tryjob-access"
+    }
+    tryjob {
+      builders {
+        name: "libyuv/try/android"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_arm64"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_rel"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/android_x64"
+      }
+      builders {
+        name: "libyuv/try/android_x86"
+      }
+      builders {
+        name: "libyuv/try/ios_arm64"
+      }
+      builders {
+        name: "libyuv/try/ios_arm64_rel"
+      }
+      builders {
+        name: "libyuv/try/linux"
+      }
+      builders {
+        name: "libyuv/try/linux_asan"
+      }
+      builders {
+        name: "libyuv/try/linux_gcc"
+        experiment_percentage: 100
+      }
+      builders {
+        name: "libyuv/try/linux_msan"
+      }
+      builders {
+        name: "libyuv/try/linux_rel"
+      }
+      builders {
+        name: "libyuv/try/linux_tsan2"
+      }
+      builders {
+        name: "libyuv/try/linux_ubsan"
+      }
+      builders {
+        name: "libyuv/try/linux_ubsan_vptr"
+      }
+      builders {
+        name: "libyuv/try/mac"
+      }
+      builders {
+        name: "libyuv/try/mac_asan"
+      }
+      builders {
+        name: "libyuv/try/mac_rel"
+      }
+      builders {
+        name: "libyuv/try/win"
+      }
+      builders {
+        name: "libyuv/try/win_clang"
+      }
+      builders {
+        name: "libyuv/try/win_clang_rel"
+      }
+      builders {
+        name: "libyuv/try/win_rel"
+      }
+      builders {
+        name: "libyuv/try/win_x64_clang_rel"
+      }
+      builders {
+        name: "libyuv/try/win_x64_rel"
+      }
+      retry_config {
+        single_quota: 1
+        global_quota: 2
+        failure_weight: 1
+        transient_failure_weight: 1
+        timeout_weight: 2
+      }
+    }
+  }
+}
diff --git a/files/infra/config/cq.cfg b/files/infra/config/cq.cfg
deleted file mode 100644
index 7bcc0595..00000000
--- a/files/infra/config/cq.cfg
+++ /dev/null
@@ -1,51 +0,0 @@
-# Commit Queue configuration file. The documentation of the format can be found
-# at http://luci-config.appspot.com/schemas/projects/refs:cq.cfg.
-
-version: 1
-cq_status_url: "https://chromium-cq-status.appspot.com"
-git_repo_url: "https://chromium.googlesource.com/libyuv/libyuv.git"
-
-gerrit {}
-
-verifiers {
-  gerrit_cq_ability {
-    committer_list: "project-libyuv-committers"
-    dry_run_access_list: "project-libyuv-tryjob-access"
-  }
-
-  try_job {
-    buckets {
-      name: "luci.libyuv.try"
-      builders { name: "win" }
-      builders { name: "win_rel" }
-      builders { name: "win_x64_rel" }
-      builders { name: "win_clang" }
-      builders { name: "win_clang_rel" }
-      builders { name: "win_x64_clang_rel" }
-      builders { name: "mac" }
-      builders { name: "mac_rel" }
-      builders { name: "mac_asan" }
-      builders { name: "ios" }
-      builders { name: "ios_rel" }
-      builders { name: "ios_arm64" }
-      builders { name: "ios_arm64_rel" }
-      builders { name: "linux" }
-      builders { name: "linux_rel" }
-      builders {
-        name: "linux_gcc"
-        experiment_percentage: 100
-      }
-      builders { name: "linux_memcheck" }
-      builders { name: "linux_tsan2" }
-      builders { name: "linux_asan" }
-      builders { name: "linux_msan" }
-      builders { name: "linux_ubsan" }
-      builders { name: "linux_ubsan_vptr" }
-      builders { name: "android" }
-      builders { name: "android_rel" }
-      builders { name: "android_arm64" }
-      builders { name: "android_x86" }
-      builders { name: "android_x64" }
-    }
-  }
-}
diff --git a/files/infra/config/cr-buildbucket.cfg b/files/infra/config/cr-buildbucket.cfg
new file mode 100644
index 00000000..061cf33b
--- /dev/null
+++ b/files/infra/config/cr-buildbucket.cfg
@@ -0,0 +1,1743 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see BuildbucketCfg message:
+#   https://luci-config.appspot.com/schemas/projects:buildbucket.cfg
+
+buckets {
+  name: "ci"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  swarming {
+    builders {
+      name: "Android ARM64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM32 Debug (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM32 Release (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android Tester ARM64 Debug (Nexus 5X)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android32 x86 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Android64 x64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux Asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux MSan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux Tsan v2"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux UBSan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux UBSan vptr"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux32 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux32 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Linux64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac Asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Mac64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Debug (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win32 Release (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Debug (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "Win64 Release (Clang)"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "iOS ARM64 Debug"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "iOS ARM64 Release"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.ci"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "client.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
+buckets {
+  name: "cron"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  swarming {
+    builders {
+      name: "DEPS Autoroller"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Linux"
+      dimensions: "pool:luci.webrtc.cron"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "recipe": "libyuv/roll_deps"'
+        '}'
+      execution_timeout_secs: 7200
+      build_numbers: YES
+      service_account: "libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
+buckets {
+  name: "try"
+  acls {
+    role: WRITER
+    group: "project-libyuv-admins"
+  }
+  acls {
+    group: "all"
+  }
+  acls {
+    role: SCHEDULER
+    group: "project-libyuv-tryjob-access"
+  }
+  acls {
+    role: SCHEDULER
+    group: "service-account-cq"
+  }
+  swarming {
+    builders {
+      name: "android"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_arm64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "device_type:bullhead"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_x64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "android_x86"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "ios_arm64"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "ios_arm64_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_gcc"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_msan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_tsan2"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_ubsan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "linux_ubsan_vptr"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac_asan"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "mac_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Mac-10.15"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "presubmit"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Ubuntu-18.04"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": true,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "run_presubmit",'
+        '  "repo_name": "libyuv",'
+        '  "runhooks": true'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_clang"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_clang_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_x64_clang_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+    builders {
+      name: "win_x64_rel"
+      swarming_host: "chromium-swarm.appspot.com"
+      swarming_tags: "vpython:native-python-wrapper"
+      dimensions: "cores:8"
+      dimensions: "cpu:x86-64"
+      dimensions: "os:Windows-10"
+      dimensions: "pool:luci.flex.try"
+      exe {
+        cipd_package: "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build"
+        cipd_version: "refs/heads/main"
+        cmd: "luciexe"
+      }
+      properties:
+        '{'
+        '  "$build/goma": {'
+        '    "enable_ats": false,'
+        '    "server_host": "goma.chromium.org",'
+        '    "use_luci_auth": true'
+        '  },'
+        '  "builder_group": "tryserver.libyuv",'
+        '  "recipe": "libyuv/libyuv"'
+        '}'
+      execution_timeout_secs: 10800
+      build_numbers: YES
+      service_account: "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+      experiments {
+        key: "luci.recipes.use_python3"
+        value: 100
+      }
+    }
+  }
+}
diff --git a/files/infra/config/luci-logdog.cfg b/files/infra/config/luci-logdog.cfg
new file mode 100644
index 00000000..adc75bef
--- /dev/null
+++ b/files/infra/config/luci-logdog.cfg
@@ -0,0 +1,9 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectConfig message:
+#   https://luci-config.appspot.com/schemas/projects:luci-logdog.cfg
+
+reader_auth_groups: "all"
+writer_auth_groups: "luci-logdog-chromium-writers"
+archive_gs_bucket: "chromium-luci-logdog"
diff --git a/files/infra/config/luci-milo.cfg b/files/infra/config/luci-milo.cfg
new file mode 100644
index 00000000..baf786f2
--- /dev/null
+++ b/files/infra/config/luci-milo.cfg
@@ -0,0 +1,246 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see Project message:
+#   https://luci-config.appspot.com/schemas/projects:luci-milo.cfg
+
+consoles {
+  id: "main"
+  name: "libyuv Main Console"
+  repo_url: "https://chromium.googlesource.com/libyuv/libyuv"
+  refs: "regexp:refs/heads/main"
+  manifest_name: "REVISION"
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android ARM64 Debug"
+    category: "Android|Builder"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Debug"
+    category: "Android|Builder"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Release"
+    category: "Android|Builder"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android32 x86 Debug"
+    category: "Android|Builder|x86"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android64 x64 Debug"
+    category: "Android|Builder|x64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM32 Debug (Nexus 5X)"
+    category: "Android|Tester|ARM 32"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM32 Release (Nexus 5X)"
+    category: "Android|Tester|ARM 32"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Android Tester ARM64 Debug (Nexus 5X)"
+    category: "Android|Tester|ARM 64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux Asan"
+    category: "Linux"
+    short_name: "asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux MSan"
+    category: "Linux"
+    short_name: "msan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux Tsan v2"
+    category: "Linux"
+    short_name: "tsan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux UBSan"
+    category: "Linux|UBSan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux UBSan vptr"
+    category: "Linux|UBSan"
+    short_name: "vptr"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux32 Debug"
+    category: "Linux|32"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux32 Release"
+    category: "Linux|32"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux64 Debug"
+    category: "Linux|64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Linux64 Release"
+    category: "Linux|64"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac Asan"
+    category: "Mac"
+    short_name: "asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac64 Debug"
+    category: "Mac"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Mac64 Release"
+    category: "Mac"
+    short_name: "rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Debug"
+    category: "Win|32|Debug"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Debug (Clang)"
+    category: "Win|32|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Release"
+    category: "Win|32|Release"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win32 Release (Clang)"
+    category: "Win|32|Release"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Debug"
+    category: "Win|64|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Debug (Clang)"
+    category: "Win|64|Debug"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Release"
+    category: "Win|64|Release"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/Win64 Release (Clang)"
+    category: "Win|64|Release"
+    short_name: "clg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/iOS ARM64 Debug"
+    category: "iOS|ARM64"
+    short_name: "dbg"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.ci/iOS ARM64 Release"
+    category: "iOS|ARM64"
+    short_name: "rel"
+  }
+  include_experimental_builds: true
+}
+consoles {
+  id: "cron"
+  name: "Cron"
+  builders {
+    name: "buildbucket/luci.libyuv.cron/DEPS Autoroller"
+  }
+  builder_view_only: true
+}
+consoles {
+  id: "try"
+  name: "libyuv Try Builders"
+  builders {
+    name: "buildbucket/luci.libyuv.try/android"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_arm64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_x64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/android_x86"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/ios_arm64"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/ios_arm64_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_gcc"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_msan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_tsan2"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_ubsan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/linux_ubsan_vptr"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac_asan"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/mac_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_clang"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_clang_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_x64_clang_rel"
+  }
+  builders {
+    name: "buildbucket/luci.libyuv.try/win_x64_rel"
+  }
+  builder_view_only: true
+}
+logo_url: "https://storage.googleapis.com/chrome-infra-public/logo/libyuv-logo.png"
diff --git a/files/infra/config/luci-scheduler.cfg b/files/infra/config/luci-scheduler.cfg
new file mode 100644
index 00000000..0ec5dd0e
--- /dev/null
+++ b/files/infra/config/luci-scheduler.cfg
@@ -0,0 +1,385 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectConfig message:
+#   https://luci-config.appspot.com/schemas/projects:luci-scheduler.cfg
+
+job {
+  id: "Android ARM64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android ARM64 Debug"
+  }
+}
+job {
+  id: "Android Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Debug"
+  }
+}
+job {
+  id: "Android Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Release"
+  }
+}
+job {
+  id: "Android Tester ARM32 Debug (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM32 Debug (Nexus 5X)"
+  }
+}
+job {
+  id: "Android Tester ARM32 Release (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM32 Release (Nexus 5X)"
+  }
+}
+job {
+  id: "Android Tester ARM64 Debug (Nexus 5X)"
+  realm: "ci"
+  acls {
+    role: TRIGGERER
+    granted_to: "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android Tester ARM64 Debug (Nexus 5X)"
+  }
+}
+job {
+  id: "Android32 x86 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android32 x86 Debug"
+  }
+}
+job {
+  id: "Android64 x64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Android64 x64 Debug"
+  }
+}
+job {
+  id: "DEPS Autoroller"
+  realm: "cron"
+  schedule: "0 14 * * *"
+  acl_sets: "cron"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "cron"
+    builder: "DEPS Autoroller"
+  }
+}
+job {
+  id: "Linux Asan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux Asan"
+  }
+}
+job {
+  id: "Linux MSan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux MSan"
+  }
+}
+job {
+  id: "Linux Tsan v2"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux Tsan v2"
+  }
+}
+job {
+  id: "Linux UBSan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux UBSan"
+  }
+}
+job {
+  id: "Linux UBSan vptr"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux UBSan vptr"
+  }
+}
+job {
+  id: "Linux32 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux32 Debug"
+  }
+}
+job {
+  id: "Linux32 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux32 Release"
+  }
+}
+job {
+  id: "Linux64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux64 Debug"
+  }
+}
+job {
+  id: "Linux64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Linux64 Release"
+  }
+}
+job {
+  id: "Mac Asan"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac Asan"
+  }
+}
+job {
+  id: "Mac64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac64 Debug"
+  }
+}
+job {
+  id: "Mac64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Mac64 Release"
+  }
+}
+job {
+  id: "Win32 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Debug"
+  }
+}
+job {
+  id: "Win32 Debug (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Debug (Clang)"
+  }
+}
+job {
+  id: "Win32 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Release"
+  }
+}
+job {
+  id: "Win32 Release (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win32 Release (Clang)"
+  }
+}
+job {
+  id: "Win64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Debug"
+  }
+}
+job {
+  id: "Win64 Debug (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Debug (Clang)"
+  }
+}
+job {
+  id: "Win64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Release"
+  }
+}
+job {
+  id: "Win64 Release (Clang)"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "Win64 Release (Clang)"
+  }
+}
+job {
+  id: "iOS ARM64 Debug"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "iOS ARM64 Debug"
+  }
+}
+job {
+  id: "iOS ARM64 Release"
+  realm: "ci"
+  acl_sets: "ci"
+  buildbucket {
+    server: "cr-buildbucket.appspot.com"
+    bucket: "ci"
+    builder: "iOS ARM64 Release"
+  }
+}
+trigger {
+  id: "master-gitiles-trigger"
+  realm: "ci"
+  acl_sets: "ci"
+  triggers: "Android ARM64 Debug"
+  triggers: "Android Debug"
+  triggers: "Android Release"
+  triggers: "Android32 x86 Debug"
+  triggers: "Android64 x64 Debug"
+  triggers: "Linux Asan"
+  triggers: "Linux MSan"
+  triggers: "Linux Tsan v2"
+  triggers: "Linux UBSan"
+  triggers: "Linux UBSan vptr"
+  triggers: "Linux32 Debug"
+  triggers: "Linux32 Release"
+  triggers: "Linux64 Debug"
+  triggers: "Linux64 Release"
+  triggers: "Mac Asan"
+  triggers: "Mac64 Debug"
+  triggers: "Mac64 Release"
+  triggers: "Win32 Debug"
+  triggers: "Win32 Debug (Clang)"
+  triggers: "Win32 Release"
+  triggers: "Win32 Release (Clang)"
+  triggers: "Win64 Debug"
+  triggers: "Win64 Debug (Clang)"
+  triggers: "Win64 Release"
+  triggers: "Win64 Release (Clang)"
+  triggers: "iOS ARM64 Debug"
+  triggers: "iOS ARM64 Release"
+  gitiles {
+    repo: "https://chromium.googlesource.com/libyuv/libyuv"
+    refs: "regexp:refs/heads/main"
+  }
+}
+acl_sets {
+  name: "ci"
+  acls {
+    role: OWNER
+    granted_to: "group:project-libyuv-admins"
+  }
+  acls {
+    granted_to: "group:all"
+  }
+}
+acl_sets {
+  name: "cron"
+  acls {
+    role: OWNER
+    granted_to: "group:project-libyuv-admins"
+  }
+  acls {
+    granted_to: "group:all"
+  }
+}
diff --git a/files/infra/config/main.star b/files/infra/config/main.star
new file mode 100755
index 00000000..b922ca02
--- /dev/null
+++ b/files/infra/config/main.star
@@ -0,0 +1,361 @@
+#!/usr/bin/env lucicfg
+# https://chromium.googlesource.com/infra/luci/luci-go/+/master/lucicfg/doc/
+
+"""LUCI project configuration for libyuv CQ and CI."""
+
+lucicfg.check_version("1.30.9")
+
+LIBYUV_GIT = "https://chromium.googlesource.com/libyuv/libyuv"
+LIBYUV_GERRIT = "https://chromium-review.googlesource.com/libyuv/libyuv"
+
+GOMA_BACKEND_RBE_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+}
+
+GOMA_BACKEND_RBE_ATS_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+    "enable_ats": True,
+}
+
+# Disable ATS on Windows CQ/try.
+GOMA_BACKEND_RBE_NO_ATS_PROD = {
+    "server_host": "goma.chromium.org",
+    "use_luci_auth": True,
+    "enable_ats": False,
+}
+
+# Use LUCI Scheduler BBv2 names and add Scheduler realms configs.
+lucicfg.enable_experiment("crbug.com/1182002")
+
+luci.builder.defaults.experiments.set(
+    {
+        "luci.recipes.use_python3": 100,
+    },
+)
+
+lucicfg.config(
+    lint_checks = ["default"],
+    config_dir = ".",
+    tracked_files = [
+        "commit-queue.cfg",
+        "cr-buildbucket.cfg",
+        "luci-logdog.cfg",
+        "luci-milo.cfg",
+        "luci-scheduler.cfg",
+        "project.cfg",
+        "realms.cfg",
+    ],
+)
+
+# Generates project.cfg
+
+luci.project(
+    name = "libyuv",
+    buildbucket = "cr-buildbucket.appspot.com",
+    logdog = "luci-logdog.appspot.com",
+    milo = "luci-milo.appspot.com",
+    notify = "luci-notify.appspot.com",
+    scheduler = "luci-scheduler.appspot.com",
+    swarming = "chromium-swarm.appspot.com",
+    acls = [
+        acl.entry(acl.PROJECT_CONFIGS_READER, groups = ["all"]),
+        acl.entry(acl.LOGDOG_READER, groups = ["all"]),
+        acl.entry(acl.LOGDOG_WRITER, groups = ["luci-logdog-chromium-writers"]),
+        acl.entry(acl.SCHEDULER_READER, groups = ["all"]),
+        acl.entry(acl.SCHEDULER_OWNER, groups = ["project-libyuv-admins"]),
+        acl.entry(acl.BUILDBUCKET_READER, groups = ["all"]),
+        acl.entry(acl.BUILDBUCKET_OWNER, groups = ["project-libyuv-admins"]),
+    ],
+    bindings = [
+        luci.binding(
+            roles = "role/configs.validator",
+            users = "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com",
+        ),
+    ],
+)
+
+# Generates luci-logdog.cfg
+
+luci.logdog(
+    gs_bucket = "chromium-luci-logdog",
+)
+
+# Generates luci-scheduler.cfg
+
+luci.gitiles_poller(
+    name = "master-gitiles-trigger",
+    bucket = "ci",
+    repo = LIBYUV_GIT,
+)
+
+# Generates luci-milo.cfg
+
+luci.milo(
+    logo = "https://storage.googleapis.com/chrome-infra-public/logo/libyuv-logo.png",
+)
+
+def libyuv_ci_view(name, category, short_name):
+    return luci.console_view_entry(
+        console_view = "main",
+        builder = name,
+        category = category,
+        short_name = short_name,
+    )
+
+def libyuv_try_view(name):
+    return luci.list_view_entry(
+        list_view = "try",
+        builder = name,
+    )
+
+luci.console_view(
+    name = "main",
+    title = "libyuv Main Console",
+    include_experimental_builds = True,
+    repo = LIBYUV_GIT,
+)
+
+luci.list_view(
+    name = "cron",
+    title = "Cron",
+    entries = ["DEPS Autoroller"],
+)
+
+luci.list_view(
+    name = "try",
+    title = "libyuv Try Builders",
+)
+
+# Generates commit-queue.cfg
+
+def libyuv_try_job_verifier(name, cq_group, experiment_percentage):
+    return luci.cq_tryjob_verifier(
+        builder = name,
+        cq_group = cq_group,
+        experiment_percentage = experiment_percentage,
+    )
+
+luci.cq(
+    status_host = "chromium-cq-status.appspot.com",
+    submit_max_burst = 4,
+    submit_burst_delay = 8 * time.minute,
+)
+
+luci.cq_group(
+    name = "master",
+    watch = [
+        cq.refset(
+            repo = LIBYUV_GERRIT,
+            refs = ["refs/heads/main", "refs/heads/master"],
+        ),
+    ],
+    acls = [
+        acl.entry(acl.CQ_COMMITTER, groups = ["project-libyuv-committers"]),
+        acl.entry(acl.CQ_DRY_RUNNER, groups = ["project-libyuv-tryjob-access"]),
+    ],
+    retry_config = cq.RETRY_ALL_FAILURES,
+    cancel_stale_tryjobs = True,
+)
+
+luci.cq_group(
+    name = "config",
+    watch = [
+        cq.refset(
+            repo = LIBYUV_GERRIT,
+            refs = ["refs/heads/infra/config"],
+        ),
+    ],
+    acls = [
+        acl.entry(acl.CQ_COMMITTER, groups = ["project-libyuv-committers"]),
+        acl.entry(acl.CQ_DRY_RUNNER, groups = ["project-libyuv-tryjob-access"]),
+    ],
+    retry_config = cq.RETRY_ALL_FAILURES,
+    cancel_stale_tryjobs = True,
+)
+
+# Generates cr-buildbucket.cfg
+
+luci.bucket(
+    name = "ci",
+)
+luci.bucket(
+    name = "try",
+    acls = [
+        acl.entry(acl.BUILDBUCKET_TRIGGERER, groups = [
+            "project-libyuv-tryjob-access",
+            "service-account-cq",
+        ]),
+    ],
+)
+luci.bucket(
+    name = "cron",
+)
+
+def get_os_dimensions(os):
+    if os == "android":
+        return {"device_type": "bullhead"}
+    if os == "ios" or os == "mac":
+        return {"os": "Mac-10.15", "cpu": "x86-64"}
+    elif os == "win":
+        return {"os": "Windows-10", "cores": "8", "cpu": "x86-64"}
+    elif os == "linux":
+        return {"os": "Ubuntu-18.04", "cores": "8", "cpu": "x86-64"}
+    return {}
+
+def get_os_properties(os, try_builder = False):
+    if os == "android":
+        return {"$build/goma": GOMA_BACKEND_RBE_PROD}
+    elif os in ("ios", "mac"):
+        return {"$build/goma": GOMA_BACKEND_RBE_PROD}
+    elif os == "win" and try_builder:
+        return {"$build/goma": GOMA_BACKEND_RBE_NO_ATS_PROD}
+    elif os == "win":
+        return {"$build/goma": GOMA_BACKEND_RBE_ATS_PROD}
+    elif os == "linux":
+        return {"$build/goma": GOMA_BACKEND_RBE_ATS_PROD}
+    return {}
+
+def libyuv_ci_builder(name, dimensions, properties, triggered_by):
+    return luci.builder(
+        name = name,
+        dimensions = dimensions,
+        properties = properties,
+        bucket = "ci",
+        service_account = "libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com",
+        triggered_by = triggered_by,
+        swarming_tags = ["vpython:native-python-wrapper"],
+        execution_timeout = 180 * time.minute,
+        build_numbers = True,
+        executable = luci.recipe(
+            name = "libyuv/libyuv",
+            cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+            use_python3 = True,
+        ),
+    )
+
+def libyuv_try_builder(name, dimensions, properties, recipe_name = "libyuv/libyuv"):
+    return luci.builder(
+        name = name,
+        dimensions = dimensions,
+        properties = properties,
+        bucket = "try",
+        service_account = "libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com",
+        swarming_tags = ["vpython:native-python-wrapper"],
+        execution_timeout = 180 * time.minute,
+        build_numbers = True,
+        executable = luci.recipe(
+            name = recipe_name,
+            cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+            use_python3 = True,
+        ),
+    )
+
+def ci_builder(name, os, category, short_name = None):
+    dimensions = get_os_dimensions(os)
+    properties = get_os_properties(os)
+
+    dimensions["pool"] = "luci.flex.ci"
+    properties["builder_group"] = "client.libyuv"
+
+    triggered_by = ["master-gitiles-trigger" if os != "android" else "Android Debug"]
+    libyuv_ci_view(name, category, short_name)
+    return libyuv_ci_builder(name, dimensions, properties, triggered_by)
+
+def try_builder(name, os, experiment_percentage = None):
+    dimensions = get_os_dimensions(os)
+    properties = get_os_properties(os, try_builder = True)
+
+    dimensions["pool"] = "luci.flex.try"
+    properties["builder_group"] = "tryserver.libyuv"
+
+    if name == "presubmit":
+        recipe_name = "run_presubmit"
+        properties["repo_name"] = "libyuv"
+        properties["runhooks"] = True
+        libyuv_try_job_verifier(name, "config", experiment_percentage)
+        return libyuv_try_builder(name, dimensions, properties, recipe_name)
+
+    libyuv_try_job_verifier(name, "master", experiment_percentage)
+    libyuv_try_view(name)
+    return libyuv_try_builder(name, dimensions, properties)
+
+luci.builder(
+    name = "DEPS Autoroller",
+    bucket = "cron",
+    service_account = "libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com",
+    dimensions = {
+        "pool": "luci.webrtc.cron",
+        "os": "Linux",
+        "cpu": "x86-64",
+    },
+    swarming_tags = ["vpython:native-python-wrapper"],
+    execution_timeout = 120 * time.minute,
+    build_numbers = True,
+    schedule = "0 14 * * *",  # Every 2 hours.
+    executable = luci.recipe(
+        name = "libyuv/roll_deps",
+        cipd_package = "infra/recipe_bundles/chromium.googlesource.com/chromium/tools/build",
+        use_python3 = True,
+    ),
+)
+
+ci_builder("Android ARM64 Debug", "linux", "Android|Builder", "dbg")
+ci_builder("Android Debug", "linux", "Android|Builder", "dbg")
+ci_builder("Android Release", "linux", "Android|Builder", "rel")
+ci_builder("Android32 x86 Debug", "linux", "Android|Builder|x86", "dbg")
+ci_builder("Android64 x64 Debug", "linux", "Android|Builder|x64", "dbg")
+ci_builder("Android Tester ARM32 Debug (Nexus 5X)", "android", "Android|Tester|ARM 32", "dbg")
+ci_builder("Android Tester ARM32 Release (Nexus 5X)", "android", "Android|Tester|ARM 32", "rel")
+ci_builder("Android Tester ARM64 Debug (Nexus 5X)", "android", "Android|Tester|ARM 64", "dbg")
+ci_builder("Linux Asan", "linux", "Linux", "asan")
+ci_builder("Linux MSan", "linux", "Linux", "msan")
+ci_builder("Linux Tsan v2", "linux", "Linux", "tsan")
+ci_builder("Linux UBSan", "linux", "Linux|UBSan")
+ci_builder("Linux UBSan vptr", "linux", "Linux|UBSan", "vptr")
+ci_builder("Linux32 Debug", "linux", "Linux|32", "dbg")
+ci_builder("Linux32 Release", "linux", "Linux|32", "rel")
+ci_builder("Linux64 Debug", "linux", "Linux|64", "dbg")
+ci_builder("Linux64 Release", "linux", "Linux|64", "rel")
+ci_builder("Mac Asan", "mac", "Mac", "asan")
+ci_builder("Mac64 Debug", "mac", "Mac", "dbg")
+ci_builder("Mac64 Release", "mac", "Mac", "rel")
+ci_builder("Win32 Debug", "win", "Win|32|Debug")
+ci_builder("Win32 Debug (Clang)", "win", "Win|32|Debug", "clg")
+ci_builder("Win32 Release", "win", "Win|32|Release")
+ci_builder("Win32 Release (Clang)", "win", "Win|32|Release", "clg")
+ci_builder("Win64 Debug", "win", "Win|64|Debug", "clg")
+ci_builder("Win64 Debug (Clang)", "win", "Win|64|Debug", "clg")
+ci_builder("Win64 Release", "win", "Win|64|Release")
+ci_builder("Win64 Release (Clang)", "win", "Win|64|Release", "clg")
+ci_builder("iOS ARM64 Debug", "ios", "iOS|ARM64", "dbg")
+ci_builder("iOS ARM64 Release", "ios", "iOS|ARM64", "rel")
+
+# TODO(crbug.com/1242847): make this not experimental.
+try_builder("android", "android", experiment_percentage = 100)
+try_builder("android_arm64", "android", experiment_percentage = 100)
+try_builder("android_rel", "android", experiment_percentage = 100)
+
+try_builder("android_x64", "linux")
+try_builder("android_x86", "linux")
+try_builder("ios_arm64", "ios")
+try_builder("ios_arm64_rel", "ios")
+try_builder("linux", "linux")
+try_builder("linux_asan", "linux")
+try_builder("linux_gcc", "linux", experiment_percentage = 100)
+try_builder("linux_msan", "linux")
+try_builder("linux_rel", "linux")
+try_builder("linux_tsan2", "linux")
+try_builder("linux_ubsan", "linux")
+try_builder("linux_ubsan_vptr", "linux")
+try_builder("mac", "mac")
+try_builder("mac_asan", "mac")
+try_builder("mac_rel", "mac")
+try_builder("win", "win")
+try_builder("win_clang", "win")
+try_builder("win_clang_rel", "win")
+try_builder("win_rel", "win")
+try_builder("win_x64_clang_rel", "win")
+try_builder("win_x64_rel", "win")
+try_builder("presubmit", "linux")
diff --git a/files/infra/config/project.cfg b/files/infra/config/project.cfg
new file mode 100644
index 00000000..700226ad
--- /dev/null
+++ b/files/infra/config/project.cfg
@@ -0,0 +1,15 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see ProjectCfg message:
+#   https://luci-config.appspot.com/schemas/projects:project.cfg
+
+name: "libyuv"
+access: "group:all"
+lucicfg {
+  version: "1.30.9"
+  package_dir: "."
+  config_dir: "."
+  entry_point: "main.star"
+  experiments: "crbug.com/1182002"
+}
diff --git a/files/infra/config/realms.cfg b/files/infra/config/realms.cfg
new file mode 100644
index 00000000..ae04529e
--- /dev/null
+++ b/files/infra/config/realms.cfg
@@ -0,0 +1,79 @@
+# Auto-generated by lucicfg.
+# Do not modify manually.
+#
+# For the schema of this file, see RealmsCfg message:
+#   https://luci-config.appspot.com/schemas/projects:realms.cfg
+
+realms {
+  name: "@root"
+  bindings {
+    role: "role/buildbucket.owner"
+    principals: "group:project-libyuv-admins"
+  }
+  bindings {
+    role: "role/buildbucket.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/configs.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/configs.validator"
+    principals: "user:libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/logdog.reader"
+    principals: "group:all"
+  }
+  bindings {
+    role: "role/logdog.writer"
+    principals: "group:luci-logdog-chromium-writers"
+  }
+  bindings {
+    role: "role/scheduler.owner"
+    principals: "group:project-libyuv-admins"
+  }
+  bindings {
+    role: "role/scheduler.reader"
+    principals: "group:all"
+  }
+}
+realms {
+  name: "ci"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/scheduler.triggerer"
+    principals: "user:libyuv-ci-builder@chops-service-accounts.iam.gserviceaccount.com"
+    conditions {
+      restrict {
+        attribute: "scheduler.job.name"
+        values: "Android Tester ARM32 Debug (Nexus 5X)"
+        values: "Android Tester ARM32 Release (Nexus 5X)"
+        values: "Android Tester ARM64 Debug (Nexus 5X)"
+      }
+    }
+  }
+}
+realms {
+  name: "cron"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-ci-autoroll-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+}
+realms {
+  name: "try"
+  bindings {
+    role: "role/buildbucket.builderServiceAccount"
+    principals: "user:libyuv-try-builder@chops-service-accounts.iam.gserviceaccount.com"
+  }
+  bindings {
+    role: "role/buildbucket.triggerer"
+    principals: "group:project-libyuv-tryjob-access"
+    principals: "group:service-account-cq"
+  }
+}
diff --git a/files/libyuv.gypi b/files/libyuv.gypi
index 18b2feca..48936aa7 100644
--- a/files/libyuv.gypi
+++ b/files/libyuv.gypi
@@ -27,7 +27,9 @@
       'include/libyuv/row.h',
       'include/libyuv/scale.h',
       'include/libyuv/scale_argb.h',
+      'include/libyuv/scale_rgb.h',
       'include/libyuv/scale_row.h',
+      'include/libyuv/scale_uv.h',
       'include/libyuv/version.h',
       'include/libyuv/video_common.h',
 
@@ -35,6 +37,7 @@
       'source/compare.cc',
       'source/compare_common.cc',
       'source/compare_gcc.cc',
+      'source/compare_msa.cc',
       'source/compare_neon.cc',
       'source/compare_neon64.cc',
       'source/compare_win.cc',
@@ -54,7 +57,6 @@
       'source/rotate_argb.cc',
       'source/rotate_common.cc',
       'source/rotate_gcc.cc',
-      'source/rotate_dspr2.cc',
       'source/rotate_msa.cc',
       'source/rotate_neon.cc',
       'source/rotate_neon64.cc',
@@ -62,7 +64,6 @@
       'source/row_any.cc',
       'source/row_common.cc',
       'source/row_gcc.cc',
-      'source/row_dspr2.cc',
       'source/row_msa.cc',
       'source/row_neon.cc',
       'source/row_neon64.cc',
@@ -72,10 +73,11 @@
       'source/scale_argb.cc',
       'source/scale_common.cc',
       'source/scale_gcc.cc',
-      'source/scale_dspr2.cc',
       'source/scale_msa.cc',
       'source/scale_neon.cc',
       'source/scale_neon64.cc',
+      'source/scale_rgb.cc',
+      'source/scale_uv.cc',
       'source/scale_win.cc',
       'source/video_common.cc',
     ],
diff --git a/files/libyuv_nacl.gyp b/files/libyuv_nacl.gyp
deleted file mode 100644
index b8fe57ee..00000000
--- a/files/libyuv_nacl.gyp
+++ /dev/null
@@ -1,37 +0,0 @@
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-{
-  'includes': [
-    'libyuv.gypi',
-    '../../native_client/build/untrusted.gypi',
-  ],
-  'targets': [
-    {
-      'target_name': 'libyuv_nacl',
-      'type': 'none',
-      'variables': {
-        'nlib_target': 'libyuv_nacl.a',
-        'build_glibc': 0,
-        'build_newlib': 0,
-        'build_pnacl_newlib': 1,
-      },
-      'include_dirs': [
-        'include',
-      ],
-      'direct_dependent_settings': {
-        'include_dirs': [
-          'include',
-        ],
-      },
-      'sources': [
-        '<@(libyuv_sources)',
-      ],
-    },  # target libyuv_nacl
-  ]
-}
diff --git a/files/libyuv_test.gyp b/files/libyuv_test.gyp
deleted file mode 100644
index 87e7a5bb..00000000
--- a/files/libyuv_test.gyp
+++ /dev/null
@@ -1,203 +0,0 @@
-# Copyright 2011 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-{
-  'variables': {
-    # Can be enabled if your jpeg has GYP support.
-    'libyuv_disable_jpeg%': 1,
-    'mips_msa%': 0,  # Default to msa off.
-  },
-  'targets': [
-    {
-      'target_name': 'libyuv_unittest',
-      'type': '<(gtest_target_type)',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-        'testing/gtest.gyp:gtest',
-        'third_party/gflags/gflags.gyp:gflags',
-      ],
-      'direct_dependent_settings': {
-        'defines': [
-          'GTEST_RELATIVE_PATH',
-        ],
-      },
-      'export_dependent_settings': [
-        '<(DEPTH)/testing/gtest.gyp:gtest',
-      ],
-      'sources': [
-        # headers
-        'unit_test/unit_test.h',
-
-        # sources
-        'unit_test/basictypes_test.cc',
-        'unit_test/compare_test.cc',
-        'unit_test/color_test.cc',
-        'unit_test/convert_test.cc',
-        'unit_test/cpu_test.cc',
-        'unit_test/math_test.cc',
-        'unit_test/planar_test.cc',
-        'unit_test/rotate_argb_test.cc',
-        'unit_test/rotate_test.cc',
-        'unit_test/scale_argb_test.cc',
-        'unit_test/scale_test.cc',
-        'unit_test/unit_test.cc',
-        'unit_test/video_common_test.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-        [ 'OS == "ios"', {
-          'xcode_settings': {
-            'DEBUGGING_SYMBOLS': 'YES',
-            'DEBUG_INFORMATION_FORMAT' : 'dwarf-with-dsym',
-            # Work around compile issue with isosim.mm, see
-            # https://code.google.com/p/libyuv/issues/detail?id=548 for details.
-            'WARNING_CFLAGS': [
-              '-Wno-sometimes-uninitialized',
-            ],
-          },
-          'cflags': [
-            '-Wno-sometimes-uninitialized',
-          ],
-        }],
-        [ 'OS != "ios" and libyuv_disable_jpeg != 1', {
-          'defines': [
-            'HAVE_JPEG',
-          ],
-        }],
-        ['OS=="android"', {
-          'dependencies': [
-            '<(DEPTH)/testing/android/native_test.gyp:native_test_native_code',
-          ],
-        }],
-        # TODO(YangZhang): These lines can be removed when high accuracy
-        # YUV to RGB to Neon is ported.
-        [ '(target_arch == "armv7" or target_arch == "armv7s" \
-          or (target_arch == "arm" and arm_version >= 7) \
-          or target_arch == "arm64") \
-          and (arm_neon == 1 or arm_neon_optional == 1)', {
-          'defines': [
-            'LIBYUV_NEON'
-          ],
-        }],
-        [ '(target_arch == "mipsel" or target_arch == "mips64el") \
-          and (mips_msa == 1)', {
-          'defines': [
-            'LIBYUV_MSA'
-          ],
-        }],
-      ], # conditions
-      'defines': [
-        # Enable the following 3 macros to turn off assembly for specified CPU.
-        # 'LIBYUV_DISABLE_X86',
-        # 'LIBYUV_DISABLE_NEON',
-        # 'LIBYUV_DISABLE_DSPR2',
-        # Enable the following macro to build libyuv as a shared library (dll).
-        # 'LIBYUV_USING_SHARED_LIBRARY',
-      ],
-    },
-    {
-      'target_name': 'compare',
-      'type': 'executable',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'sources': [
-        # sources
-        'util/compare.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-      ], # conditions
-    },
-    {
-      'target_name': 'yuvconvert',
-      'type': 'executable',
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'sources': [
-        # sources
-        'util/yuvconvert.cc',
-      ],
-      'conditions': [
-        ['OS=="linux"', {
-          'cflags': [
-            '-fexceptions',
-          ],
-        }],
-      ], # conditions
-    },
-    # TODO(fbarchard): Enable SSE2 and OpenMP for better performance.
-    {
-      'target_name': 'psnr',
-      'type': 'executable',
-      'sources': [
-        # sources
-        'util/psnr_main.cc',
-        'util/psnr.cc',
-        'util/ssim.cc',
-      ],
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-      'conditions': [
-        [ 'OS != "ios" and libyuv_disable_jpeg != 1', {
-          'defines': [
-            'HAVE_JPEG',
-          ],
-        }],
-      ], # conditions
-    },
-
-    {
-      'target_name': 'cpuid',
-      'type': 'executable',
-      'sources': [
-        # sources
-        'util/cpuid.c',
-      ],
-      'dependencies': [
-        'libyuv.gyp:libyuv',
-      ],
-    },
-  ], # targets
-  'conditions': [
-    ['OS=="android"', {
-      'targets': [
-        {
-          'target_name': 'yuv_unittest_apk',
-          'type': 'none',
-          'variables': {
-            'test_suite_name': 'yuv_unittest',
-            'input_shlib_path': '<(SHARED_LIB_DIR)/(SHARED_LIB_PREFIX)libyuv_unittest<(SHARED_LIB_SUFFIX)',
-          },
-          'includes': [
-            'build/apk_test.gypi',
-          ],
-          'dependencies': [
-            'libyuv_unittest',
-          ],
-        },
-      ],
-    }],
-  ],
-}
-
-# Local Variables:
-# tab-width:2
-# indent-tabs-mode:nil
-# End:
-# vim: set expandtab tabstop=2 shiftwidth=2:
diff --git a/files/linux.mk b/files/linux.mk
index e9a26a79..b541b47c 100644
--- a/files/linux.mk
+++ b/files/linux.mk
@@ -13,15 +13,14 @@ LOCAL_OBJ_FILES := \
 	source/compare.o           \
 	source/compare_common.o    \
 	source/compare_gcc.o       \
-	source/compare_mmi.o       \
 	source/compare_msa.o       \
-	source/compare_neon64.o    \
 	source/compare_neon.o      \
+	source/compare_neon64.o    \
 	source/compare_win.o       \
-	source/convert_argb.o      \
 	source/convert.o           \
-	source/convert_from_argb.o \
+	source/convert_argb.o      \
 	source/convert_from.o      \
+	source/convert_from_argb.o \
 	source/convert_jpeg.o      \
 	source/convert_to_argb.o   \
 	source/convert_to_i420.o   \
@@ -29,33 +28,32 @@ LOCAL_OBJ_FILES := \
 	source/mjpeg_decoder.o     \
 	source/mjpeg_validate.o    \
 	source/planar_functions.o  \
+	source/rotate.o            \
 	source/rotate_any.o        \
 	source/rotate_argb.o       \
-	source/rotate.o            \
 	source/rotate_common.o     \
 	source/rotate_gcc.o        \
-	source/rotate_mmi.o        \
 	source/rotate_msa.o        \
-	source/rotate_neon64.o     \
 	source/rotate_neon.o       \
+	source/rotate_neon64.o     \
 	source/rotate_win.o        \
 	source/row_any.o           \
 	source/row_common.o        \
 	source/row_gcc.o           \
-	source/row_mmi.o           \
 	source/row_msa.o           \
-	source/row_neon64.o        \
 	source/row_neon.o          \
+	source/row_neon64.o        \
 	source/row_win.o           \
+	source/scale.o             \
 	source/scale_any.o         \
 	source/scale_argb.o        \
-	source/scale.o             \
 	source/scale_common.o      \
 	source/scale_gcc.o         \
-	source/scale_mmi.o         \
 	source/scale_msa.o         \
-	source/scale_neon64.o      \
 	source/scale_neon.o        \
+	source/scale_neon64.o      \
+	source/scale_rgb.o         \
+	source/scale_uv.o          \
 	source/scale_win.o         \
 	source/video_common.o
 
@@ -65,7 +63,7 @@ LOCAL_OBJ_FILES := \
 .c.o:
 	$(CC) -c $(CFLAGS) $*.c -o $*.o
 
-all: libyuv.a yuvconvert cpuid psnr
+all: libyuv.a i444tonv12_eg yuvconvert yuvconstants cpuid psnr
 
 libyuv.a: $(LOCAL_OBJ_FILES)
 	$(AR) $(ARFLAGS) $@ $(LOCAL_OBJ_FILES)
@@ -74,10 +72,18 @@ libyuv.a: $(LOCAL_OBJ_FILES)
 yuvconvert: util/yuvconvert.cc libyuv.a
 	$(CXX) $(CXXFLAGS) -Iutil/ -o $@ util/yuvconvert.cc libyuv.a
 
+# A C test utility that generates yuvconstants for yuv to rgb.
+yuvconstants: util/yuvconstants.c libyuv.a
+	$(CXX) $(CXXFLAGS) -Iutil/ -lm -o $@ util/yuvconstants.c libyuv.a
+
 # A standalone test utility
 psnr: util/psnr.cc
 	$(CXX) $(CXXFLAGS) -Iutil/ -o $@ util/psnr.cc util/psnr_main.cc util/ssim.cc
 
+# A simple conversion example.
+i444tonv12_eg: util/i444tonv12_eg.cc libyuv.a
+	$(CXX) $(CXXFLAGS) -o $@ util/i444tonv12_eg.cc libyuv.a
+
 # A C test utility that uses libyuv conversion from C.
 # gcc 4.4 and older require -fno-exceptions to avoid link error on __gxx_personality_v0
 # CC=gcc-4.4 CXXFLAGS=-fno-exceptions CXX=g++-4.4 make -f linux.mk
@@ -85,4 +91,4 @@ cpuid: util/cpuid.c libyuv.a
 	$(CC) $(CFLAGS) -o $@ util/cpuid.c libyuv.a
 
 clean:
-	/bin/rm -f source/*.o *.ii *.s libyuv.a yuvconvert cpuid psnr
+	/bin/rm -f source/*.o *.ii *.s libyuv.a i444tonv12_eg yuvconvert yuvconstants cpuid psnr
diff --git a/files/setup_links.py b/files/setup_links.py
deleted file mode 100755
index b2b459e6..00000000
--- a/files/setup_links.py
+++ /dev/null
@@ -1,497 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Setup links to a Chromium checkout for WebRTC.
-
-WebRTC standalone shares a lot of dependencies and build tools with Chromium.
-To do this, many of the paths of a Chromium checkout is emulated by creating
-symlinks to files and directories. This script handles the setup of symlinks to
-achieve this.
-
-It also handles cleanup of the legacy Subversion-based approach that was used
-before Chrome switched over their master repo from Subversion to Git.
-"""
-
-
-import ctypes
-import errno
-import logging
-import optparse
-import os
-import shelve
-import shutil
-import subprocess
-import sys
-import textwrap
-
-
-DIRECTORIES = [
-  'build',
-  'buildtools',
-  'mojo',  # TODO(kjellander): Remove, see webrtc:5629.
-  'native_client',
-  'net',
-  'testing',
-  'third_party/binutils',
-  'third_party/drmemory',
-  'third_party/instrumented_libraries',
-  'third_party/libjpeg',
-  'third_party/libjpeg_turbo',
-  'third_party/llvm-build',
-  'third_party/lss',
-  'third_party/yasm',
-  'third_party/WebKit',  # TODO(kjellander): Remove, see webrtc:5629.
-  'tools/clang',
-  'tools/gn',
-  'tools/gyp',
-  'tools/memory',
-  'tools/python',
-  'tools/swarming_client',
-  'tools/valgrind',
-  'tools/vim',
-  'tools/win',
-]
-
-from sync_chromium import get_target_os_list
-target_os = get_target_os_list()
-if 'android' in target_os:
-  DIRECTORIES += [
-    'base',
-    'third_party/android_platform',
-    'third_party/android_tools',
-    'third_party/appurify-python',
-    'third_party/ashmem',
-    'third_party/catapult',
-    'third_party/icu',
-    'third_party/ijar',
-    'third_party/jsr-305',
-    'third_party/junit',
-    'third_party/libxml',
-    'third_party/mockito',
-    'third_party/modp_b64',
-    'third_party/protobuf',
-    'third_party/requests',
-    'third_party/robolectric',
-    'tools/android',
-    'tools/grit',
-  ]
-if 'ios' in target_os:
-  DIRECTORIES.append('third_party/class-dump')
-
-FILES = {
-  'tools/isolate_driver.py': None,
-  'third_party/BUILD.gn': None,
-}
-
-ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-CHROMIUM_CHECKOUT = os.path.join('chromium', 'src')
-LINKS_DB = 'links'
-
-# Version management to make future upgrades/downgrades easier to support.
-SCHEMA_VERSION = 1
-
-
-def query_yes_no(question, default=False):
-  """Ask a yes/no question via raw_input() and return their answer.
-
-  Modified from http://stackoverflow.com/a/3041990.
-  """
-  prompt = " [%s/%%s]: "
-  prompt = prompt % ('Y' if default is True  else 'y')
-  prompt = prompt % ('N' if default is False else 'n')
-
-  if default is None:
-    default = 'INVALID'
-
-  while True:
-    sys.stdout.write(question + prompt)
-    choice = raw_input().lower()
-    if choice == '' and default != 'INVALID':
-      return default
-
-    if 'yes'.startswith(choice):
-      return True
-    elif 'no'.startswith(choice):
-      return False
-
-    print "Please respond with 'yes' or 'no' (or 'y' or 'n')."
-
-
-# Actions
-class Action(object):
-  def __init__(self, dangerous):
-    self.dangerous = dangerous
-
-  def announce(self, planning):
-    """Log a description of this action.
-
-    Args:
-      planning - True iff we're in the planning stage, False if we're in the
-                 doit stage.
-    """
-    pass
-
-  def doit(self, links_db):
-    """Execute the action, recording what we did to links_db, if necessary."""
-    pass
-
-
-class Remove(Action):
-  def __init__(self, path, dangerous):
-    super(Remove, self).__init__(dangerous)
-    self._priority = 0
-    self._path = path
-
-  def announce(self, planning):
-    log = logging.warn
-    filesystem_type = 'file'
-    if not self.dangerous:
-      log = logging.info
-      filesystem_type = 'link'
-    if planning:
-      log('Planning to remove %s: %s', filesystem_type, self._path)
-    else:
-      log('Removing %s: %s', filesystem_type, self._path)
-
-  def doit(self, _):
-    os.remove(self._path)
-
-
-class Rmtree(Action):
-  def __init__(self, path):
-    super(Rmtree, self).__init__(dangerous=True)
-    self._priority = 0
-    self._path = path
-
-  def announce(self, planning):
-    if planning:
-      logging.warn('Planning to remove directory: %s', self._path)
-    else:
-      logging.warn('Removing directory: %s', self._path)
-
-  def doit(self, _):
-    if sys.platform.startswith('win'):
-      # shutil.rmtree() doesn't work on Windows if any of the directories are
-      # read-only, which svn repositories are.
-      subprocess.check_call(['rd', '/q', '/s', self._path], shell=True)
-    else:
-      shutil.rmtree(self._path)
-
-
-class Makedirs(Action):
-  def __init__(self, path):
-    super(Makedirs, self).__init__(dangerous=False)
-    self._priority = 1
-    self._path = path
-
-  def doit(self, _):
-    try:
-      os.makedirs(self._path)
-    except OSError as e:
-      if e.errno != errno.EEXIST:
-        raise
-
-
-class Symlink(Action):
-  def __init__(self, source_path, link_path):
-    super(Symlink, self).__init__(dangerous=False)
-    self._priority = 2
-    self._source_path = source_path
-    self._link_path = link_path
-
-  def announce(self, planning):
-    if planning:
-      logging.info(
-          'Planning to create link from %s to %s', self._link_path,
-          self._source_path)
-    else:
-      logging.debug(
-          'Linking from %s to %s', self._link_path, self._source_path)
-
-  def doit(self, links_db):
-    # Files not in the root directory need relative path calculation.
-    # On Windows, use absolute paths instead since NTFS doesn't seem to support
-    # relative paths for symlinks.
-    if sys.platform.startswith('win'):
-      source_path = os.path.abspath(self._source_path)
-    else:
-      if os.path.dirname(self._link_path) != self._link_path:
-        source_path = os.path.relpath(self._source_path,
-                                      os.path.dirname(self._link_path))
-
-    os.symlink(source_path, os.path.abspath(self._link_path))
-    links_db[self._source_path] = self._link_path
-
-
-class LinkError(IOError):
-  """Failed to create a link."""
-  pass
-
-
-# Handles symlink creation on the different platforms.
-if sys.platform.startswith('win'):
-  def symlink(source_path, link_path):
-    flag = 1 if os.path.isdir(source_path) else 0
-    if not ctypes.windll.kernel32.CreateSymbolicLinkW(
-        unicode(link_path), unicode(source_path), flag):
-      raise OSError('Failed to create symlink to %s. Notice that only NTFS '
-                    'version 5.0 and up has all the needed APIs for '
-                    'creating symlinks.' % source_path)
-  os.symlink = symlink
-
-
-class WebRTCLinkSetup(object):
-  def __init__(self, links_db, force=False, dry_run=False, prompt=False):
-    self._force = force
-    self._dry_run = dry_run
-    self._prompt = prompt
-    self._links_db = links_db
-
-  def CreateLinks(self, on_bot):
-    logging.debug('CreateLinks')
-    # First, make a plan of action
-    actions = []
-
-    for source_path, link_path in FILES.iteritems():
-      actions += self._ActionForPath(
-          source_path, link_path, check_fn=os.path.isfile, check_msg='files')
-    for source_dir in DIRECTORIES:
-      actions += self._ActionForPath(
-          source_dir, None, check_fn=os.path.isdir,
-          check_msg='directories')
-
-    if not on_bot and self._force:
-      # When making the manual switch from legacy SVN checkouts to the new
-      # Git-based Chromium DEPS, the .gclient_entries file that contains cached
-      # URLs for all DEPS entries must be removed to avoid future sync problems.
-      entries_file = os.path.join(os.path.dirname(ROOT_DIR), '.gclient_entries')
-      if os.path.exists(entries_file):
-        actions.append(Remove(entries_file, dangerous=True))
-
-    actions.sort()
-
-    if self._dry_run:
-      for action in actions:
-        action.announce(planning=True)
-      logging.info('Not doing anything because dry-run was specified.')
-      sys.exit(0)
-
-    if any(a.dangerous for a in actions):
-      logging.warn('Dangerous actions:')
-      for action in (a for a in actions if a.dangerous):
-        action.announce(planning=True)
-      print
-
-      if not self._force:
-        logging.error(textwrap.dedent("""\
-        @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
-                              A C T I O N     R E Q I R E D
-        @@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
-
-        Because chromium/src is transitioning to Git (from SVN), we needed to
-        change the way that the WebRTC standalone checkout works. Instead of
-        individually syncing subdirectories of Chromium in SVN, we're now
-        syncing Chromium (and all of its DEPS, as defined by its own DEPS file),
-        into the `chromium/src` directory.
-
-        As such, all Chromium directories which are currently pulled by DEPS are
-        now replaced with a symlink into the full Chromium checkout.
-
-        To avoid disrupting developers, we've chosen to not delete your
-        directories forcibly, in case you have some work in progress in one of
-        them :).
-
-        ACTION REQUIRED:
-        Before running `gclient sync|runhooks` again, you must run:
-        %s%s --force
-
-        Which will replace all directories which now must be symlinks, after
-        prompting with a summary of the work-to-be-done.
-        """), 'python ' if sys.platform.startswith('win') else '', sys.argv[0])
-        sys.exit(1)
-      elif self._prompt:
-        if not query_yes_no('Would you like to perform the above plan?'):
-          sys.exit(1)
-
-    for action in actions:
-      action.announce(planning=False)
-      action.doit(self._links_db)
-
-    if not on_bot and self._force:
-      logging.info('Completed!\n\nNow run `gclient sync|runhooks` again to '
-                   'let the remaining hooks (that probably were interrupted) '
-                   'execute.')
-
-  def CleanupLinks(self):
-    logging.debug('CleanupLinks')
-    for source, link_path  in self._links_db.iteritems():
-      if source == 'SCHEMA_VERSION':
-        continue
-      if os.path.islink(link_path) or sys.platform.startswith('win'):
-        # os.path.islink() always returns false on Windows
-        # See http://bugs.python.org/issue13143.
-        logging.debug('Removing link to %s at %s', source, link_path)
-        if not self._dry_run:
-          if os.path.exists(link_path):
-            if sys.platform.startswith('win') and os.path.isdir(link_path):
-              subprocess.check_call(['rmdir', '/q', '/s', link_path],
-                                    shell=True)
-            else:
-              os.remove(link_path)
-          del self._links_db[source]
-
-  @staticmethod
-  def _ActionForPath(source_path, link_path=None, check_fn=None,
-                     check_msg=None):
-    """Create zero or more Actions to link to a file or directory.
-
-    This will be a symlink on POSIX platforms. On Windows this requires
-    that NTFS is version 5.0 or higher (Vista or newer).
-
-    Args:
-      source_path: Path relative to the Chromium checkout root.
-        For readability, the path may contain slashes, which will
-        automatically be converted to the right path delimiter on Windows.
-      link_path: The location for the link to create. If omitted it will be the
-        same path as source_path.
-      check_fn: A function returning true if the type of filesystem object is
-        correct for the attempted call. Otherwise an error message with
-        check_msg will be printed.
-      check_msg: String used to inform the user of an invalid attempt to create
-        a file.
-    Returns:
-      A list of Action objects.
-    """
-    def fix_separators(path):
-      if sys.platform.startswith('win'):
-        return path.replace(os.altsep, os.sep)
-      else:
-        return path
-
-    assert check_fn
-    assert check_msg
-    link_path = link_path or source_path
-    link_path = fix_separators(link_path)
-
-    source_path = fix_separators(source_path)
-    source_path = os.path.join(CHROMIUM_CHECKOUT, source_path)
-    if os.path.exists(source_path) and not check_fn:
-      raise LinkError('_LinkChromiumPath can only be used to link to %s: '
-                      'Tried to link to: %s' % (check_msg, source_path))
-
-    if not os.path.exists(source_path):
-      logging.debug('Silently ignoring missing source: %s. This is to avoid '
-                    'errors on platform-specific dependencies.', source_path)
-      return []
-
-    actions = []
-
-    if os.path.exists(link_path) or os.path.islink(link_path):
-      if os.path.islink(link_path):
-        actions.append(Remove(link_path, dangerous=False))
-      elif os.path.isfile(link_path):
-        actions.append(Remove(link_path, dangerous=True))
-      elif os.path.isdir(link_path):
-        actions.append(Rmtree(link_path))
-      else:
-        raise LinkError('Don\'t know how to plan: %s' % link_path)
-
-    # Create parent directories to the target link if needed.
-    target_parent_dirs = os.path.dirname(link_path)
-    if (target_parent_dirs and
-        target_parent_dirs != link_path and
-        not os.path.exists(target_parent_dirs)):
-      actions.append(Makedirs(target_parent_dirs))
-
-    actions.append(Symlink(source_path, link_path))
-
-    return actions
-
-def _initialize_database(filename):
-  links_database = shelve.open(filename)
-
-  # Wipe the database if this version of the script ends up looking at a
-  # newer (future) version of the links db, just to be sure.
-  version = links_database.get('SCHEMA_VERSION')
-  if version and version != SCHEMA_VERSION:
-    logging.info('Found database with schema version %s while this script only '
-                 'supports %s. Wiping previous database contents.', version,
-                 SCHEMA_VERSION)
-    links_database.clear()
-  links_database['SCHEMA_VERSION'] = SCHEMA_VERSION
-  return links_database
-
-
-def main():
-  on_bot = os.environ.get('CHROME_HEADLESS') == '1'
-
-  parser = optparse.OptionParser()
-  parser.add_option('-d', '--dry-run', action='store_true', default=False,
-                    help='Print what would be done, but don\'t perform any '
-                         'operations. This will automatically set logging to '
-                         'verbose.')
-  parser.add_option('-c', '--clean-only', action='store_true', default=False,
-                    help='Only clean previously created links, don\'t create '
-                         'new ones. This will automatically set logging to '
-                         'verbose.')
-  parser.add_option('-f', '--force', action='store_true', default=on_bot,
-                    help='Force link creation. CAUTION: This deletes existing '
-                         'folders and files in the locations where links are '
-                         'about to be created.')
-  parser.add_option('-n', '--no-prompt', action='store_false', dest='prompt',
-                    default=(not on_bot),
-                    help='Prompt if we\'re planning to do a dangerous action')
-  parser.add_option('-v', '--verbose', action='store_const',
-                    const=logging.DEBUG, default=logging.INFO,
-                    help='Print verbose output for debugging.')
-  options, _ = parser.parse_args()
-
-  if options.dry_run or options.force or options.clean_only:
-    options.verbose = logging.DEBUG
-  logging.basicConfig(format='%(message)s', level=options.verbose)
-
-  # Work from the root directory of the checkout.
-  script_dir = os.path.dirname(os.path.abspath(__file__))
-  os.chdir(script_dir)
-
-  if sys.platform.startswith('win'):
-    def is_admin():
-      try:
-        return os.getuid() == 0
-      except AttributeError:
-        return ctypes.windll.shell32.IsUserAnAdmin() != 0
-    if not is_admin():
-      logging.error('On Windows, you now need to have administrator '
-                    'privileges for the shell running %s (or '
-                    '`gclient sync|runhooks`).\nPlease start another command '
-                    'prompt as Administrator and try again.', sys.argv[0])
-      return 1
-
-  if not os.path.exists(CHROMIUM_CHECKOUT):
-    logging.error('Cannot find a Chromium checkout at %s. Did you run "gclient '
-                  'sync" before running this script?', CHROMIUM_CHECKOUT)
-    return 2
-
-  links_database = _initialize_database(LINKS_DB)
-  try:
-    symlink_creator = WebRTCLinkSetup(links_database, options.force,
-                                      options.dry_run, options.prompt)
-    symlink_creator.CleanupLinks()
-    if not options.clean_only:
-      symlink_creator.CreateLinks(on_bot)
-  except LinkError as e:
-    print >> sys.stderr, e.message
-    return 3
-  finally:
-    links_database.close()
-  return 0
-
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/files/source/compare.cc b/files/source/compare.cc
index 5aa3a4db..d4713b60 100644
--- a/files/source/compare.cc
+++ b/files/source/compare.cc
@@ -69,13 +69,13 @@ static uint32_t ARGBDetectRow_C(const uint8_t* argb, int width) {
     if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
       return FOURCC_BGRA;
     }
-    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+    if (argb[3] != 255) {  // Fourth byte is not Alpha of 255, so not BGRA.
       return FOURCC_ARGB;
     }
     if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
       return FOURCC_BGRA;
     }
-    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+    if (argb[7] != 255) {  // Second pixel fourth byte is not Alpha of 255.
       return FOURCC_ARGB;
     }
     argb += 8;
@@ -154,11 +154,6 @@ uint64_t ComputeHammingDistance(const uint8_t* src_a,
     HammingDistance = HammingDistance_MSA;
   }
 #endif
-#if defined(HAS_HAMMINGDISTANCE_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    HammingDistance = HammingDistance_MMI;
-  }
-#endif
 
 #ifdef _OPENMP
 #pragma omp parallel for reduction(+ : diff)
@@ -216,11 +211,6 @@ uint64_t ComputeSumSquareError(const uint8_t* src_a,
     SumSquareError = SumSquareError_MSA;
   }
 #endif
-#if defined(HAS_SUMSQUAREERROR_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SumSquareError = SumSquareError_MMI;
-  }
-#endif
 #ifdef _OPENMP
 #pragma omp parallel for reduction(+ : sse)
 #endif
diff --git a/files/source/compare_common.cc b/files/source/compare_common.cc
index d4b170ad..d1cab8d2 100644
--- a/files/source/compare_common.cc
+++ b/files/source/compare_common.cc
@@ -17,36 +17,6 @@ namespace libyuv {
 extern "C" {
 #endif
 
-#if ORIGINAL_OPT
-uint32_t HammingDistance_C1(const uint8_t* src_a,
-                            const uint8_t* src_b,
-                            int count) {
-  uint32_t diff = 0u;
-
-  int i;
-  for (i = 0; i < count; ++i) {
-    int x = src_a[i] ^ src_b[i];
-    if (x & 1)
-      ++diff;
-    if (x & 2)
-      ++diff;
-    if (x & 4)
-      ++diff;
-    if (x & 8)
-      ++diff;
-    if (x & 16)
-      ++diff;
-    if (x & 32)
-      ++diff;
-    if (x & 64)
-      ++diff;
-    if (x & 128)
-      ++diff;
-  }
-  return diff;
-}
-#endif
-
 // Hakmem method for hamming distance.
 uint32_t HammingDistance_C(const uint8_t* src_a,
                            const uint8_t* src_b,
diff --git a/files/source/compare_gcc.cc b/files/source/compare_gcc.cc
index 676527c1..b834b42a 100644
--- a/files/source/compare_gcc.cc
+++ b/files/source/compare_gcc.cc
@@ -19,8 +19,7 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 #if defined(__x86_64__)
 uint32_t HammingDistance_SSE42(const uint8_t* src_a,
@@ -29,38 +28,38 @@ uint32_t HammingDistance_SSE42(const uint8_t* src_a,
   uint64_t diff = 0u;
 
   asm volatile(
-      "xor        %3,%3                          \n"
-      "xor        %%r8,%%r8                      \n"
-      "xor        %%r9,%%r9                      \n"
-      "xor        %%r10,%%r10                    \n"
+      "xor         %3,%3                         \n"
+      "xor         %%r8,%%r8                     \n"
+      "xor         %%r9,%%r9                     \n"
+      "xor         %%r10,%%r10                   \n"
 
       // Process 32 bytes per loop.
       LABELALIGN
       "1:                                        \n"
-      "mov        (%0),%%rcx                     \n"
-      "mov        0x8(%0),%%rdx                  \n"
-      "xor        (%1),%%rcx                     \n"
-      "xor        0x8(%1),%%rdx                  \n"
-      "popcnt     %%rcx,%%rcx                    \n"
-      "popcnt     %%rdx,%%rdx                    \n"
-      "mov        0x10(%0),%%rsi                 \n"
-      "mov        0x18(%0),%%rdi                 \n"
-      "xor        0x10(%1),%%rsi                 \n"
-      "xor        0x18(%1),%%rdi                 \n"
-      "popcnt     %%rsi,%%rsi                    \n"
-      "popcnt     %%rdi,%%rdi                    \n"
-      "add        $0x20,%0                       \n"
-      "add        $0x20,%1                       \n"
-      "add        %%rcx,%3                       \n"
-      "add        %%rdx,%%r8                     \n"
-      "add        %%rsi,%%r9                     \n"
-      "add        %%rdi,%%r10                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "mov         (%0),%%rcx                    \n"
+      "mov         0x8(%0),%%rdx                 \n"
+      "xor         (%1),%%rcx                    \n"
+      "xor         0x8(%1),%%rdx                 \n"
+      "popcnt      %%rcx,%%rcx                   \n"
+      "popcnt      %%rdx,%%rdx                   \n"
+      "mov         0x10(%0),%%rsi                \n"
+      "mov         0x18(%0),%%rdi                \n"
+      "xor         0x10(%1),%%rsi                \n"
+      "xor         0x18(%1),%%rdi                \n"
+      "popcnt      %%rsi,%%rsi                   \n"
+      "popcnt      %%rdi,%%rdi                   \n"
+      "add         $0x20,%0                      \n"
+      "add         $0x20,%1                      \n"
+      "add         %%rcx,%3                      \n"
+      "add         %%rdx,%%r8                    \n"
+      "add         %%rsi,%%r9                    \n"
+      "add         %%rdi,%%r10                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
 
-      "add        %%r8, %3                       \n"
-      "add        %%r9, %3                       \n"
-      "add        %%r10, %3                      \n"
+      "add         %%r8, %3                      \n"
+      "add         %%r9, %3                      \n"
+      "add         %%r10, %3                     \n"
       : "+r"(src_a),  // %0
         "+r"(src_b),  // %1
         "+r"(count),  // %2
@@ -80,26 +79,26 @@ uint32_t HammingDistance_SSE42(const uint8_t* src_a,
       // Process 16 bytes per loop.
       LABELALIGN
       "1:                                        \n"
-      "mov        (%0),%%ecx                     \n"
-      "mov        0x4(%0),%%edx                  \n"
-      "xor        (%1),%%ecx                     \n"
-      "xor        0x4(%1),%%edx                  \n"
-      "popcnt     %%ecx,%%ecx                    \n"
-      "add        %%ecx,%3                       \n"
-      "popcnt     %%edx,%%edx                    \n"
-      "add        %%edx,%3                       \n"
-      "mov        0x8(%0),%%ecx                  \n"
-      "mov        0xc(%0),%%edx                  \n"
-      "xor        0x8(%1),%%ecx                  \n"
-      "xor        0xc(%1),%%edx                  \n"
-      "popcnt     %%ecx,%%ecx                    \n"
-      "add        %%ecx,%3                       \n"
-      "popcnt     %%edx,%%edx                    \n"
-      "add        %%edx,%3                       \n"
-      "add        $0x10,%0                       \n"
-      "add        $0x10,%1                       \n"
-      "sub        $0x10,%2                       \n"
-      "jg         1b                             \n"
+      "mov         (%0),%%ecx                    \n"
+      "mov         0x4(%0),%%edx                 \n"
+      "xor         (%1),%%ecx                    \n"
+      "xor         0x4(%1),%%edx                 \n"
+      "popcnt      %%ecx,%%ecx                   \n"
+      "add         %%ecx,%3                      \n"
+      "popcnt      %%edx,%%edx                   \n"
+      "add         %%edx,%3                      \n"
+      "mov         0x8(%0),%%ecx                 \n"
+      "mov         0xc(%0),%%edx                 \n"
+      "xor         0x8(%1),%%ecx                 \n"
+      "xor         0xc(%1),%%edx                 \n"
+      "popcnt      %%ecx,%%ecx                   \n"
+      "add         %%ecx,%3                      \n"
+      "popcnt      %%edx,%%edx                   \n"
+      "add         %%edx,%3                      \n"
+      "add         $0x10,%0                      \n"
+      "add         $0x10,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_a),  // %0
         "+r"(src_b),  // %1
         "+r"(count),  // %2
@@ -121,46 +120,46 @@ uint32_t HammingDistance_SSSE3(const uint8_t* src_a,
   uint32_t diff = 0u;
 
   asm volatile(
-      "movdqa     %4,%%xmm2                      \n"
-      "movdqa     %5,%%xmm3                      \n"
-      "pxor       %%xmm0,%%xmm0                  \n"
-      "pxor       %%xmm1,%%xmm1                  \n"
-      "sub        %0,%1                          \n"
+      "movdqa      %4,%%xmm2                     \n"
+      "movdqa      %5,%%xmm3                     \n"
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqa     (%0),%%xmm4                    \n"
-      "movdqa     0x10(%0), %%xmm5               \n"
-      "pxor       (%0,%1), %%xmm4                \n"
-      "movdqa     %%xmm4,%%xmm6                  \n"
-      "pand       %%xmm2,%%xmm6                  \n"
-      "psrlw      $0x4,%%xmm4                    \n"
-      "movdqa     %%xmm3,%%xmm7                  \n"
-      "pshufb     %%xmm6,%%xmm7                  \n"
-      "pand       %%xmm2,%%xmm4                  \n"
-      "movdqa     %%xmm3,%%xmm6                  \n"
-      "pshufb     %%xmm4,%%xmm6                  \n"
-      "paddb      %%xmm7,%%xmm6                  \n"
-      "pxor       0x10(%0,%1),%%xmm5             \n"
-      "add        $0x20,%0                       \n"
-      "movdqa     %%xmm5,%%xmm4                  \n"
-      "pand       %%xmm2,%%xmm5                  \n"
-      "psrlw      $0x4,%%xmm4                    \n"
-      "movdqa     %%xmm3,%%xmm7                  \n"
-      "pshufb     %%xmm5,%%xmm7                  \n"
-      "pand       %%xmm2,%%xmm4                  \n"
-      "movdqa     %%xmm3,%%xmm5                  \n"
-      "pshufb     %%xmm4,%%xmm5                  \n"
-      "paddb      %%xmm7,%%xmm5                  \n"
-      "paddb      %%xmm5,%%xmm6                  \n"
-      "psadbw     %%xmm1,%%xmm6                  \n"
-      "paddd      %%xmm6,%%xmm0                  \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "movdqa      (%0),%%xmm4                   \n"
+      "movdqa      0x10(%0), %%xmm5              \n"
+      "pxor        (%0,%1), %%xmm4               \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "pand        %%xmm2,%%xmm6                 \n"
+      "psrlw       $0x4,%%xmm4                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "pshufb      %%xmm6,%%xmm7                 \n"
+      "pand        %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "paddb       %%xmm7,%%xmm6                 \n"
+      "pxor        0x10(%0,%1),%%xmm5            \n"
+      "add         $0x20,%0                      \n"
+      "movdqa      %%xmm5,%%xmm4                 \n"
+      "pand        %%xmm2,%%xmm5                 \n"
+      "psrlw       $0x4,%%xmm4                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "pshufb      %%xmm5,%%xmm7                 \n"
+      "pand        %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufb      %%xmm4,%%xmm5                 \n"
+      "paddb       %%xmm7,%%xmm5                 \n"
+      "paddb       %%xmm5,%%xmm6                 \n"
+      "psadbw      %%xmm1,%%xmm6                 \n"
+      "paddd       %%xmm6,%%xmm0                 \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
 
-      "pshufd     $0xaa,%%xmm0,%%xmm1            \n"
-      "paddd      %%xmm1,%%xmm0                  \n"
-      "movd       %%xmm0, %3                     \n"
+      "pshufd      $0xaa,%%xmm0,%%xmm1           \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0, %3                    \n"
       : "+r"(src_a),       // %0
         "+r"(src_b),       // %1
         "+r"(count),       // %2
@@ -182,40 +181,40 @@ uint32_t HammingDistance_AVX2(const uint8_t* src_a,
   asm volatile(
       "vbroadcastf128 %4,%%ymm2                  \n"
       "vbroadcastf128 %5,%%ymm3                  \n"
-      "vpxor      %%ymm0,%%ymm0,%%ymm0           \n"
-      "vpxor      %%ymm1,%%ymm1,%%ymm1           \n"
-      "sub        %0,%1                          \n"
+      "vpxor       %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpxor       %%ymm1,%%ymm1,%%ymm1          \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqa    (%0),%%ymm4                    \n"
-      "vmovdqa    0x20(%0), %%ymm5               \n"
-      "vpxor      (%0,%1), %%ymm4, %%ymm4        \n"
-      "vpand      %%ymm2,%%ymm4,%%ymm6           \n"
-      "vpsrlw     $0x4,%%ymm4,%%ymm4             \n"
-      "vpshufb    %%ymm6,%%ymm3,%%ymm6           \n"
-      "vpand      %%ymm2,%%ymm4,%%ymm4           \n"
-      "vpshufb    %%ymm4,%%ymm3,%%ymm4           \n"
-      "vpaddb     %%ymm4,%%ymm6,%%ymm6           \n"
-      "vpxor      0x20(%0,%1),%%ymm5,%%ymm4      \n"
-      "add        $0x40,%0                       \n"
-      "vpand      %%ymm2,%%ymm4,%%ymm5           \n"
-      "vpsrlw     $0x4,%%ymm4,%%ymm4             \n"
-      "vpshufb    %%ymm5,%%ymm3,%%ymm5           \n"
-      "vpand      %%ymm2,%%ymm4,%%ymm4           \n"
-      "vpshufb    %%ymm4,%%ymm3,%%ymm4           \n"
-      "vpaddb     %%ymm5,%%ymm4,%%ymm4           \n"
-      "vpaddb     %%ymm6,%%ymm4,%%ymm4           \n"
-      "vpsadbw    %%ymm1,%%ymm4,%%ymm4           \n"
-      "vpaddd     %%ymm0,%%ymm4,%%ymm0           \n"
-      "sub        $0x40,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqa     (%0),%%ymm4                   \n"
+      "vmovdqa     0x20(%0), %%ymm5              \n"
+      "vpxor       (%0,%1), %%ymm4, %%ymm4       \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm6          \n"
+      "vpsrlw      $0x4,%%ymm4,%%ymm4            \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm6          \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm4          \n"
+      "vpshufb     %%ymm4,%%ymm3,%%ymm4          \n"
+      "vpaddb      %%ymm4,%%ymm6,%%ymm6          \n"
+      "vpxor       0x20(%0,%1),%%ymm5,%%ymm4     \n"
+      "add         $0x40,%0                      \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm5          \n"
+      "vpsrlw      $0x4,%%ymm4,%%ymm4            \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm5          \n"
+      "vpand       %%ymm2,%%ymm4,%%ymm4          \n"
+      "vpshufb     %%ymm4,%%ymm3,%%ymm4          \n"
+      "vpaddb      %%ymm5,%%ymm4,%%ymm4          \n"
+      "vpaddb      %%ymm6,%%ymm4,%%ymm4          \n"
+      "vpsadbw     %%ymm1,%%ymm4,%%ymm4          \n"
+      "vpaddd      %%ymm0,%%ymm4,%%ymm0          \n"
+      "sub         $0x40,%2                      \n"
+      "jg          1b                            \n"
 
-      "vpermq     $0xb1,%%ymm0,%%ymm1            \n"
-      "vpaddd     %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xaa,%%ymm0,%%ymm1            \n"
-      "vpaddd     %%ymm1,%%ymm0,%%ymm0           \n"
-      "vmovd      %%xmm0, %3                     \n"
+      "vpermq      $0xb1,%%ymm0,%%ymm1           \n"
+      "vpaddd      %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xaa,%%ymm0,%%ymm1           \n"
+      "vpaddd      %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovd       %%xmm0, %3                    \n"
       "vzeroupper                                \n"
       : "+r"(src_a),       // %0
         "+r"(src_b),       // %1
@@ -234,34 +233,34 @@ uint32_t SumSquareError_SSE2(const uint8_t* src_a,
                              int count) {
   uint32_t sse;
   asm volatile(
-      "pxor      %%xmm0,%%xmm0                   \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqu    (%1),%%xmm2                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "movdqa    %%xmm1,%%xmm3                   \n"
-      "psubusb   %%xmm2,%%xmm1                   \n"
-      "psubusb   %%xmm3,%%xmm2                   \n"
-      "por       %%xmm2,%%xmm1                   \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "punpckhbw %%xmm5,%%xmm2                   \n"
-      "pmaddwd   %%xmm1,%%xmm1                   \n"
-      "pmaddwd   %%xmm2,%%xmm2                   \n"
-      "paddd     %%xmm1,%%xmm0                   \n"
-      "paddd     %%xmm2,%%xmm0                   \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "psubusb     %%xmm2,%%xmm1                 \n"
+      "psubusb     %%xmm3,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpckhbw   %%xmm5,%%xmm2                 \n"
+      "pmaddwd     %%xmm1,%%xmm1                 \n"
+      "pmaddwd     %%xmm2,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
 
-      "pshufd    $0xee,%%xmm0,%%xmm1             \n"
-      "paddd     %%xmm1,%%xmm0                   \n"
-      "pshufd    $0x1,%%xmm0,%%xmm1              \n"
-      "paddd     %%xmm1,%%xmm0                   \n"
-      "movd      %%xmm0,%3                       \n"
+      "pshufd      $0xee,%%xmm0,%%xmm1           \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "pshufd      $0x1,%%xmm0,%%xmm1            \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0,%3                     \n"
 
       : "+r"(src_a),  // %0
         "+r"(src_b),  // %1
@@ -301,44 +300,44 @@ static const uvec32 kHashMul3 = {
 uint32_t HashDjb2_SSE41(const uint8_t* src, int count, uint32_t seed) {
   uint32_t hash;
   asm volatile(
-      "movd      %2,%%xmm0                       \n"
-      "pxor      %%xmm7,%%xmm7                   \n"
-      "movdqa    %4,%%xmm6                       \n"
+      "movd        %2,%%xmm0                     \n"
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "movdqa      %4,%%xmm6                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "pmulld    %%xmm6,%%xmm0                   \n"
-      "movdqa    %5,%%xmm5                       \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "punpcklbw %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm2,%%xmm3                   \n"
-      "punpcklwd %%xmm7,%%xmm3                   \n"
-      "pmulld    %%xmm5,%%xmm3                   \n"
-      "movdqa    %6,%%xmm5                       \n"
-      "movdqa    %%xmm2,%%xmm4                   \n"
-      "punpckhwd %%xmm7,%%xmm4                   \n"
-      "pmulld    %%xmm5,%%xmm4                   \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "punpckhbw %%xmm7,%%xmm1                   \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "punpcklwd %%xmm7,%%xmm2                   \n"
-      "pmulld    %%xmm5,%%xmm2                   \n"
-      "movdqa    %8,%%xmm5                       \n"
-      "punpckhwd %%xmm7,%%xmm1                   \n"
-      "pmulld    %%xmm5,%%xmm1                   \n"
-      "paddd     %%xmm4,%%xmm3                   \n"
-      "paddd     %%xmm2,%%xmm1                   \n"
-      "paddd     %%xmm3,%%xmm1                   \n"
-      "pshufd    $0xe,%%xmm1,%%xmm2              \n"
-      "paddd     %%xmm2,%%xmm1                   \n"
-      "pshufd    $0x1,%%xmm1,%%xmm2              \n"
-      "paddd     %%xmm2,%%xmm1                   \n"
-      "paddd     %%xmm1,%%xmm0                   \n"
-      "sub       $0x10,%1                        \n"
-      "jg        1b                              \n"
-      "movd      %%xmm0,%3                       \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "pmulld      %%xmm6,%%xmm0                 \n"
+      "movdqa      %5,%%xmm5                     \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm7,%%xmm3                 \n"
+      "pmulld      %%xmm5,%%xmm3                 \n"
+      "movdqa      %6,%%xmm5                     \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "punpckhwd   %%xmm7,%%xmm4                 \n"
+      "pmulld      %%xmm5,%%xmm4                 \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "punpckhbw   %%xmm7,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"
+      "pmulld      %%xmm5,%%xmm2                 \n"
+      "movdqa      %8,%%xmm5                     \n"
+      "punpckhwd   %%xmm7,%%xmm1                 \n"
+      "pmulld      %%xmm5,%%xmm1                 \n"
+      "paddd       %%xmm4,%%xmm3                 \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm1                 \n"
+      "pshufd      $0xe,%%xmm1,%%xmm2            \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "pshufd      $0x1,%%xmm1,%%xmm2            \n"
+      "paddd       %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"
+      "sub         $0x10,%1                      \n"
+      "jg          1b                            \n"
+      "movd        %%xmm0,%3                     \n"
       : "+r"(src),        // %0
         "+r"(count),      // %1
         "+rm"(seed),      // %2
diff --git a/files/source/compare_neon.cc b/files/source/compare_neon.cc
index 2a2181e0..afdd6012 100644
--- a/files/source/compare_neon.cc
+++ b/files/source/compare_neon.cc
@@ -29,24 +29,24 @@ uint32_t HammingDistance_NEON(const uint8_t* src_a,
   uint32_t diff;
 
   asm volatile(
-      "vmov.u16   q4, #0                         \n"  // accumulator
+      "vmov.u16    q4, #0                        \n"  // accumulator
 
       "1:                                        \n"
-      "vld1.8     {q0, q1}, [%0]!                \n"
-      "vld1.8     {q2, q3}, [%1]!                \n"
-      "veor.32    q0, q0, q2                     \n"
-      "veor.32    q1, q1, q3                     \n"
-      "vcnt.i8    q0, q0                         \n"
-      "vcnt.i8    q1, q1                         \n"
-      "subs       %2, %2, #32                    \n"
-      "vadd.u8    q0, q0, q1                     \n"  // 16 byte counts
-      "vpadal.u8  q4, q0                         \n"  // 8 shorts
-      "bgt        1b                             \n"
+      "vld1.8      {q0, q1}, [%0]!               \n"
+      "vld1.8      {q2, q3}, [%1]!               \n"
+      "veor.32     q0, q0, q2                    \n"
+      "veor.32     q1, q1, q3                    \n"
+      "vcnt.i8     q0, q0                        \n"
+      "vcnt.i8     q1, q1                        \n"
+      "subs        %2, %2, #32                   \n"
+      "vadd.u8     q0, q0, q1                    \n"  // 16 byte counts
+      "vpadal.u8   q4, q0                        \n"  // 8 shorts
+      "bgt         1b                            \n"
 
-      "vpaddl.u16 q0, q4                         \n"  // 4 ints
-      "vpadd.u32  d0, d0, d1                     \n"
-      "vpadd.u32  d0, d0, d0                     \n"
-      "vmov.32    %3, d0[0]                      \n"
+      "vpaddl.u16  q0, q4                        \n"  // 4 ints
+      "vpadd.u32   d0, d0, d1                    \n"
+      "vpadd.u32   d0, d0, d0                    \n"
+      "vmov.32     %3, d0[0]                     \n"
 
       : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
       :
@@ -59,29 +59,29 @@ uint32_t SumSquareError_NEON(const uint8_t* src_a,
                              int count) {
   uint32_t sse;
   asm volatile(
-      "vmov.u8    q8, #0                         \n"
-      "vmov.u8    q10, #0                        \n"
-      "vmov.u8    q9, #0                         \n"
-      "vmov.u8    q11, #0                        \n"
+      "vmov.u8     q8, #0                        \n"
+      "vmov.u8     q10, #0                       \n"
+      "vmov.u8     q9, #0                        \n"
+      "vmov.u8     q11, #0                       \n"
 
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"
-      "vld1.8     {q1}, [%1]!                    \n"
-      "subs       %2, %2, #16                    \n"
-      "vsubl.u8   q2, d0, d2                     \n"
-      "vsubl.u8   q3, d1, d3                     \n"
-      "vmlal.s16  q8, d4, d4                     \n"
-      "vmlal.s16  q9, d6, d6                     \n"
-      "vmlal.s16  q10, d5, d5                    \n"
-      "vmlal.s16  q11, d7, d7                    \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q1}, [%1]!                   \n"
+      "subs        %2, %2, #16                   \n"
+      "vsubl.u8    q2, d0, d2                    \n"
+      "vsubl.u8    q3, d1, d3                    \n"
+      "vmlal.s16   q8, d4, d4                    \n"
+      "vmlal.s16   q9, d6, d6                    \n"
+      "vmlal.s16   q10, d5, d5                   \n"
+      "vmlal.s16   q11, d7, d7                   \n"
+      "bgt         1b                            \n"
 
-      "vadd.u32   q8, q8, q9                     \n"
-      "vadd.u32   q10, q10, q11                  \n"
-      "vadd.u32   q11, q8, q10                   \n"
-      "vpaddl.u32 q1, q11                        \n"
-      "vadd.u64   d0, d2, d3                     \n"
-      "vmov.32    %3, d0[0]                      \n"
+      "vadd.u32    q8, q8, q9                    \n"
+      "vadd.u32    q10, q10, q11                 \n"
+      "vadd.u32    q11, q8, q10                  \n"
+      "vpaddl.u32  q1, q11                       \n"
+      "vadd.u64    d0, d2, d3                    \n"
+      "vmov.32     %3, d0[0]                     \n"
       : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
       :
       : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
diff --git a/files/source/compare_neon64.cc b/files/source/compare_neon64.cc
index 6e8f672a..70fb9b91 100644
--- a/files/source/compare_neon64.cc
+++ b/files/source/compare_neon64.cc
@@ -27,22 +27,24 @@ uint32_t HammingDistance_NEON(const uint8_t* src_a,
                               int count) {
   uint32_t diff;
   asm volatile(
-      "movi       v4.8h, #0                      \n"
+      "movi        v4.8h, #0                     \n"
 
       "1:                                        \n"
-      "ld1        {v0.16b, v1.16b}, [%0], #32    \n"
-      "ld1        {v2.16b, v3.16b}, [%1], #32    \n"
-      "eor        v0.16b, v0.16b, v2.16b         \n"
-      "eor        v1.16b, v1.16b, v3.16b         \n"
-      "cnt        v0.16b, v0.16b                 \n"
-      "cnt        v1.16b, v1.16b                 \n"
-      "subs       %w2, %w2, #32                  \n"
-      "add        v0.16b, v0.16b, v1.16b         \n"
-      "uadalp     v4.8h, v0.16b                  \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b, v1.16b}, [%0], #32   \n"
+      "ld1         {v2.16b, v3.16b}, [%1], #32   \n"
+      "eor         v0.16b, v0.16b, v2.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "eor         v1.16b, v1.16b, v3.16b        \n"
+      "cnt         v0.16b, v0.16b                \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "cnt         v1.16b, v1.16b                \n"
+      "subs        %w2, %w2, #32                 \n"
+      "add         v0.16b, v0.16b, v1.16b        \n"
+      "uadalp      v4.8h, v0.16b                 \n"
+      "b.gt        1b                            \n"
 
-      "uaddlv     s4, v4.8h                      \n"
-      "fmov       %w3, s4                        \n"
+      "uaddlv      s4, v4.8h                     \n"
+      "fmov        %w3, s4                       \n"
       : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(diff)
       :
       : "cc", "v0", "v1", "v2", "v3", "v4");
@@ -54,28 +56,30 @@ uint32_t SumSquareError_NEON(const uint8_t* src_a,
                              int count) {
   uint32_t sse;
   asm volatile(
-      "eor        v16.16b, v16.16b, v16.16b      \n"
-      "eor        v18.16b, v18.16b, v18.16b      \n"
-      "eor        v17.16b, v17.16b, v17.16b      \n"
-      "eor        v19.16b, v19.16b, v19.16b      \n"
+      "eor         v16.16b, v16.16b, v16.16b     \n"
+      "eor         v18.16b, v18.16b, v18.16b     \n"
+      "eor         v17.16b, v17.16b, v17.16b     \n"
+      "eor         v19.16b, v19.16b, v19.16b     \n"
 
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"
-      "ld1        {v1.16b}, [%1], #16            \n"
-      "subs       %w2, %w2, #16                  \n"
-      "usubl      v2.8h, v0.8b, v1.8b            \n"
-      "usubl2     v3.8h, v0.16b, v1.16b          \n"
-      "smlal      v16.4s, v2.4h, v2.4h           \n"
-      "smlal      v17.4s, v3.4h, v3.4h           \n"
-      "smlal2     v18.4s, v2.8h, v2.8h           \n"
-      "smlal2     v19.4s, v3.8h, v3.8h           \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"
+      "ld1         {v1.16b}, [%1], #16           \n"
+      "subs        %w2, %w2, #16                 \n"
+      "usubl       v2.8h, v0.8b, v1.8b           \n"
+      "usubl2      v3.8h, v0.16b, v1.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "smlal       v16.4s, v2.4h, v2.4h          \n"
+      "smlal       v17.4s, v3.4h, v3.4h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "smlal2      v18.4s, v2.8h, v2.8h          \n"
+      "smlal2      v19.4s, v3.8h, v3.8h          \n"
+      "b.gt        1b                            \n"
 
-      "add        v16.4s, v16.4s, v17.4s         \n"
-      "add        v18.4s, v18.4s, v19.4s         \n"
-      "add        v19.4s, v16.4s, v18.4s         \n"
-      "addv       s0, v19.4s                     \n"
-      "fmov       %w3, s0                        \n"
+      "add         v16.4s, v16.4s, v17.4s        \n"
+      "add         v18.4s, v18.4s, v19.4s        \n"
+      "add         v19.4s, v16.4s, v18.4s        \n"
+      "addv        s0, v19.4s                    \n"
+      "fmov        %w3, s0                       \n"
       : "+r"(src_a), "+r"(src_b), "+r"(count), "=r"(sse)
       :
       : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
diff --git a/files/source/compare_win.cc b/files/source/compare_win.cc
index d57d3d9d..9bb27f1d 100644
--- a/files/source/compare_win.cc
+++ b/files/source/compare_win.cc
@@ -22,8 +22,9 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
 
 uint32_t HammingDistance_SSE42(const uint8_t* src_a,
                                const uint8_t* src_b,
@@ -77,8 +78,7 @@ __declspec(naked) uint32_t
   }
 }
 
-// Visual C 2012 required for AVX2.
-#if _MSC_VER >= 1700
+#ifdef HAS_SUMSQUAREERROR_AVX2
 // C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
 #pragma warning(disable : 4752)
 __declspec(naked) uint32_t
@@ -118,7 +118,7 @@ __declspec(naked) uint32_t
     ret
   }
 }
-#endif  // _MSC_VER >= 1700
+#endif  // HAS_SUMSQUAREERROR_AVX2
 
 uvec32 kHash16x33 = {0x92d9e201, 0, 0, 0};  // 33 ^ 16
 uvec32 kHashMul0 = {
@@ -196,7 +196,7 @@ __declspec(naked) uint32_t
 }
 
 // Visual C 2012 required for AVX2.
-#if _MSC_VER >= 1700
+#ifdef HAS_HASHDJB2_AVX2
 __declspec(naked) uint32_t
     HashDjb2_AVX2(const uint8_t* src, int count, uint32_t seed) {
   __asm {
@@ -231,7 +231,7 @@ __declspec(naked) uint32_t
     ret
   }
 }
-#endif  // _MSC_VER >= 1700
+#endif  // HAS_HASHDJB2_AVX2
 
 #endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
 
diff --git a/files/source/convert.cc b/files/source/convert.cc
index 614fa482..7178580f 100644
--- a/files/source/convert.cc
+++ b/files/source/convert.cc
@@ -15,7 +15,9 @@
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
 #include "libyuv/row.h"
-#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/scale.h"      // For ScalePlane()
+#include "libyuv/scale_row.h"  // For FixedDiv
+#include "libyuv/scale_uv.h"   // For UVScale()
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -48,7 +50,7 @@ static int I4xxToI420(const uint8_t* src_y,
   const int dst_y_height = Abs(src_y_height);
   const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
   const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
-  if (src_uv_width == 0 || src_uv_height == 0) {
+  if (src_uv_width <= 0 || src_uv_height == 0) {
     return -1;
   }
   if (dst_y) {
@@ -82,7 +84,8 @@ int I420Copy(const uint8_t* src_y,
              int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -124,7 +127,8 @@ int I010Copy(const uint16_t* src_y,
              int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -148,6 +152,53 @@ int I010Copy(const uint16_t* src_y,
   return 0;
 }
 
+static int Planar16bitTo8bit(const uint16_t* src_y,
+                             int src_stride_y,
+                             const uint16_t* src_u,
+                             int src_stride_u,
+                             const uint16_t* src_v,
+                             int src_stride_v,
+                             uint8_t* dst_y,
+                             int dst_stride_y,
+                             uint8_t* dst_u,
+                             int dst_stride_u,
+                             uint8_t* dst_v,
+                             int dst_stride_v,
+                             int width,
+                             int height,
+                             int subsample_x,
+                             int subsample_y,
+                             int depth) {
+  int uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  int uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  int scale = 1 << (24 - depth);
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    uv_height = -uv_height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (uv_height - 1) * src_stride_u;
+    src_v = src_v + (uv_height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, scale, width,
+                    height);
+  // Convert UV planes.
+  Convert16To8Plane(src_u, src_stride_u, dst_u, dst_stride_u, scale, uv_width,
+                    uv_height);
+  Convert16To8Plane(src_v, src_stride_v, dst_v, dst_stride_v, scale, uv_width,
+                    uv_height);
+  return 0;
+}
+
 // Convert 10 bit YUV to 8 bit.
 LIBYUV_API
 int I010ToI420(const uint16_t* src_y,
@@ -164,34 +215,344 @@ int I010ToI420(const uint16_t* src_y,
                int dst_stride_v,
                int width,
                int height) {
-  int halfwidth = (width + 1) >> 1;
-  int halfheight = (height + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           1, 10);
+}
+
+LIBYUV_API
+int I210ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  const int depth = 10;
+  const int scale = 1 << (24 - depth);
+
+  if (width <= 0 || height == 0) {
     return -1;
   }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    halfheight = (height + 1) >> 1;
     src_y = src_y + (height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
     src_stride_y = -src_stride_y;
     src_stride_u = -src_stride_u;
     src_stride_v = -src_stride_v;
   }
 
-  // Convert Y plane.
-  Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, 16384, width,
-                    height);
-  // Convert UV planes.
-  Convert16To8Plane(src_u, src_stride_u, dst_u, dst_stride_u, 16384, halfwidth,
-                    halfheight);
-  Convert16To8Plane(src_v, src_stride_v, dst_v, dst_stride_v, 16384, halfwidth,
-                    halfheight);
+  {
+    const int uv_width = SUBSAMPLE(width, 1, 1);
+    const int uv_height = SUBSAMPLE(height, 1, 1);
+    const int dy = FixedDiv(height, uv_height);
+
+    Convert16To8Plane(src_y, src_stride_y, dst_y, dst_stride_y, scale, width,
+                      height);
+    ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_u,
+                             dst_stride_u, src_u, dst_u, 0, 32768, dy,
+                             /*bpp=*/1, scale, kFilterBilinear);
+    ScalePlaneVertical_16To8(height, uv_width, uv_height, src_stride_v,
+                             dst_stride_v, src_v, dst_v, 0, 32768, dy,
+                             /*bpp=*/1, scale, kFilterBilinear);
+  }
   return 0;
 }
 
+LIBYUV_API
+int I210ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           0, 10);
+}
+
+LIBYUV_API
+int I410ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 0,
+                           0, 10);
+}
+
+LIBYUV_API
+int I012ToI420(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           1, 12);
+}
+
+LIBYUV_API
+int I212ToI422(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 1,
+                           0, 12);
+}
+
+LIBYUV_API
+int I412ToI444(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Planar16bitTo8bit(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_y, dst_stride_y, dst_u,
+                           dst_stride_u, dst_v, dst_stride_v, width, height, 0,
+                           0, 12);
+}
+
+// Any Ix10 To I010 format with mirroring.
+static int Ix10ToI010(const uint16_t* src_y,
+                      int src_stride_y,
+                      const uint16_t* src_u,
+                      int src_stride_u,
+                      const uint16_t* src_v,
+                      int src_stride_v,
+                      uint16_t* dst_y,
+                      int dst_stride_y,
+                      uint16_t* dst_u,
+                      int dst_stride_u,
+                      uint16_t* dst_v,
+                      int dst_stride_v,
+                      int width,
+                      int height,
+                      int subsample_x,
+                      int subsample_y) {
+  const int dst_y_width = Abs(width);
+  const int dst_y_height = Abs(height);
+  const int src_uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  const int src_uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  dst_y_width, dst_y_height, kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, src_uv_width, src_uv_height, dst_u,
+                dst_stride_u, dst_uv_width, dst_uv_height, kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, src_uv_width, src_uv_height, dst_v,
+                dst_stride_v, dst_uv_width, dst_uv_height, kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int I410ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Ix10ToI010(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, 0, 0);
+}
+
+LIBYUV_API
+int I210ToI010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  return Ix10ToI010(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                    dst_v, dst_stride_v, width, height, 1, 0);
+}
+
+// Any I[420]1[02] to P[420]1[02] format with mirroring.
+static int IxxxToPxxx(const uint16_t* src_y,
+                      int src_stride_y,
+                      const uint16_t* src_u,
+                      int src_stride_u,
+                      const uint16_t* src_v,
+                      int src_stride_v,
+                      uint16_t* dst_y,
+                      int dst_stride_y,
+                      uint16_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height,
+                      int subsample_x,
+                      int subsample_y,
+                      int depth) {
+  const int uv_width = SUBSAMPLE(width, subsample_x, subsample_x);
+  const int uv_height = SUBSAMPLE(height, subsample_y, subsample_y);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  ConvertToMSBPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height,
+                       depth);
+  MergeUVPlane_16(src_u, src_stride_u, src_v, src_stride_v, dst_uv,
+                  dst_stride_uv, uv_width, uv_height, depth);
+  return 0;
+}
+
+LIBYUV_API
+int I010ToP010(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 1, 10);
+}
+
+LIBYUV_API
+int I210ToP210(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 0, 10);
+}
+
+LIBYUV_API
+int I012ToP012(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 1, 12);
+}
+
+LIBYUV_API
+int I212ToP212(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  return IxxxToPxxx(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                    src_stride_v, dst_y, dst_stride_y, dst_uv, dst_stride_uv,
+                    width, height, 1, 0, 12);
+}
+
 // 422 chroma is 1/2 width, 1x height
 // 420 chroma is 1/2 width, 1/2 height
 LIBYUV_API
@@ -215,6 +576,48 @@ int I422ToI420(const uint8_t* src_y,
                     dst_v, dst_stride_v, width, height, src_uv_width, height);
 }
 
+LIBYUV_API
+int I422ToI210(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert8To16Plane(src_y, src_stride_y, dst_y, dst_stride_y, 1024, width,
+                    height);
+  // Convert UV planes.
+  Convert8To16Plane(src_u, src_stride_u, dst_u, dst_stride_u, 1024, halfwidth,
+                    height);
+  Convert8To16Plane(src_v, src_stride_v, dst_v, dst_stride_v, 1024, halfwidth,
+                    height);
+  return 0;
+}
+
 // TODO(fbarchard): Implement row conversion.
 LIBYUV_API
 int I422ToNV21(const uint8_t* src_y,
@@ -256,6 +659,60 @@ int I422ToNV21(const uint8_t* src_y,
   return 0;
 }
 
+LIBYUV_API
+int MM21ToNV12(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (!src_uv || !dst_uv || width <= 0) {
+    return -1;
+  }
+
+  int sign = height < 0 ? -1 : 1;
+
+  if (dst_y) {
+    DetilePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height, 32);
+  }
+  DetilePlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, (width + 1) & ~1,
+              (height + sign) / 2, 16);
+
+  return 0;
+}
+
+LIBYUV_API
+int MM21ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int sign = height < 0 ? -1 : 1;
+
+  if (!src_uv || !dst_u || !dst_v || width <= 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    DetilePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height, 32);
+  }
+  DetileSplitUVPlane(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                     dst_stride_v, (width + 1) & ~1, (height + sign) / 2, 16);
+
+  return 0;
+}
+
 #ifdef I422TONV21_ROW_VERSION
 // Unittest fails for this version.
 // 422 chroma is 1/2 width, 1x height
@@ -328,11 +785,11 @@ int I422ToNV21(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_MERGEUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeUVRow = MergeUVRow_Any_MMI;
-    if (IS_ALIGNED(halfwidth, 8)) {
-      MergeUVRow = MergeUVRow_MMI;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow = MergeUVRow_LSX;
     }
   }
 #endif
@@ -368,11 +825,11 @@ int I422ToNV21(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -426,9 +883,8 @@ int I444ToI420(const uint8_t* src_y,
                     dst_v, dst_stride_v, width, height, width, height);
 }
 
-// TODO(fbarchard): Implement row conversion.
 LIBYUV_API
-int I444ToNV21(const uint8_t* src_y,
+int I444ToNV12(const uint8_t* src_y,
                int src_stride_y,
                const uint8_t* src_u,
                int src_stride_u,
@@ -436,16 +892,16 @@ int I444ToNV21(const uint8_t* src_y,
                int src_stride_v,
                uint8_t* dst_y,
                int dst_stride_y,
-               uint8_t* dst_vu,
-               int dst_stride_vu,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
                int width,
                int height) {
-  int halfwidth = (width + 1) >> 1;
-  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    halfheight = (height + 1) >> 1;
     src_y = src_y + (height - 1) * src_stride_y;
     src_u = src_u + (height - 1) * src_stride_u;
     src_v = src_v + (height - 1) * src_stride_v;
@@ -453,19 +909,32 @@ int I444ToNV21(const uint8_t* src_y,
     src_stride_u = -src_stride_u;
     src_stride_v = -src_stride_v;
   }
-  // Allocate u and v buffers
-  align_buffer_64(plane_u, halfwidth * halfheight * 2);
-  uint8_t* plane_v = plane_u + halfwidth * halfheight;
-
-  I444ToI420(src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
-             dst_y, dst_stride_y, plane_u, halfwidth, plane_v, halfwidth, width,
-             height);
-  MergeUVPlane(plane_v, halfwidth, plane_u, halfwidth, dst_vu, dst_stride_vu,
-               halfwidth, halfheight);
-  free_aligned_buffer_64(plane_u);
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  HalfMergeUVPlane(src_u, src_stride_u, src_v, src_stride_v, dst_uv,
+                   dst_stride_uv, width, height);
   return 0;
 }
 
+LIBYUV_API
+int I444ToNV21(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  return I444ToNV12(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                    src_stride_u, dst_y, dst_stride_y, dst_vu, dst_stride_vu,
+                    width, height);
+}
+
 // I400 is greyscale typically used in MJPG
 LIBYUV_API
 int I400ToI420(const uint8_t* src_y,
@@ -527,70 +996,21 @@ int I400ToNV21(const uint8_t* src_y,
   return 0;
 }
 
-static void CopyPlane2(const uint8_t* src,
-                       int src_stride_0,
-                       int src_stride_1,
-                       uint8_t* dst,
-                       int dst_stride,
-                       int width,
-                       int height) {
-  int y;
-  void (*CopyRow)(const uint8_t* src, uint8_t* dst, int width) = CopyRow_C;
-#if defined(HAS_COPYROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
-  }
-#endif
-#if defined(HAS_COPYROW_AVX)
-  if (TestCpuFlag(kCpuHasAVX)) {
-    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
-  }
-#endif
-#if defined(HAS_COPYROW_ERMS)
-  if (TestCpuFlag(kCpuHasERMS)) {
-    CopyRow = CopyRow_ERMS;
-  }
-#endif
-#if defined(HAS_COPYROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
-  }
-#endif
-
-  // Copy plane
-  for (y = 0; y < height - 1; y += 2) {
-    CopyRow(src, dst, width);
-    CopyRow(src + src_stride_0, dst + dst_stride, width);
-    src += src_stride_0 + src_stride_1;
-    dst += dst_stride * 2;
-  }
-  if (height & 1) {
-    CopyRow(src, dst, width);
-  }
-}
-
-// Support converting from FOURCC_M420
-// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
-// easy conversion to I420.
-// M420 format description:
-// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
-// Chroma is half width / half height. (420)
-// src_stride_m420 is row planar. Normally this will be the width in pixels.
-//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
-//   this as well as the two Y planes.
-static int X420ToI420(const uint8_t* src_y,
-                      int src_stride_y0,
-                      int src_stride_y1,
-                      const uint8_t* src_uv,
-                      int src_stride_uv,
-                      uint8_t* dst_y,
-                      int dst_stride_y,
-                      uint8_t* dst_u,
-                      int dst_stride_u,
-                      uint8_t* dst_v,
-                      int dst_stride_v,
-                      int width,
-                      int height) {
+// Convert NV12 to I420.
+// TODO(fbarchard): Consider inverting destination. Faster on ARM with prfm.
+LIBYUV_API
+int NV12ToI420(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
   if (!src_uv || !dst_u || !dst_v || width <= 0 || height == 0) {
@@ -600,21 +1020,16 @@ static int X420ToI420(const uint8_t* src_y,
   if (height < 0) {
     height = -height;
     halfheight = (height + 1) >> 1;
-    if (dst_y) {
-      dst_y = dst_y + (height - 1) * dst_stride_y;
-    }
-    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
-    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
-    dst_stride_y = -dst_stride_y;
-    dst_stride_u = -dst_stride_u;
-    dst_stride_v = -dst_stride_v;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
   }
   // Coalesce rows.
-  if (src_stride_y0 == width && src_stride_y1 == width &&
-      dst_stride_y == width) {
+  if (src_stride_y == width && dst_stride_y == width) {
     width *= height;
     height = 1;
-    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
+    src_stride_y = dst_stride_y = 0;
   }
   // Coalesce rows.
   if (src_stride_uv == halfwidth * 2 && dst_stride_u == halfwidth &&
@@ -625,12 +1040,7 @@ static int X420ToI420(const uint8_t* src_y,
   }
 
   if (dst_y) {
-    if (src_stride_y0 == src_stride_y1) {
-      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
-    } else {
-      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
-                 width, height);
-    }
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
   }
 
   // Split UV plane - NV12 / NV21
@@ -640,12 +1050,12 @@ static int X420ToI420(const uint8_t* src_y,
   return 0;
 }
 
-// Convert NV12 to I420.
+// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
 LIBYUV_API
-int NV12ToI420(const uint8_t* src_y,
+int NV21ToI420(const uint8_t* src_y,
                int src_stride_y,
-               const uint8_t* src_uv,
-               int src_stride_uv,
+               const uint8_t* src_vu,
+               int src_stride_vu,
                uint8_t* dst_y,
                int dst_stride_y,
                uint8_t* dst_u,
@@ -654,46 +1064,107 @@ int NV12ToI420(const uint8_t* src_y,
                int dst_stride_v,
                int width,
                int height) {
-  return X420ToI420(src_y, src_stride_y, src_stride_y, src_uv, src_stride_uv,
-                    dst_y, dst_stride_y, dst_u, dst_stride_u, dst_v,
-                    dst_stride_v, width, height);
+  return NV12ToI420(src_y, src_stride_y, src_vu, src_stride_vu, dst_y,
+                    dst_stride_y, dst_v, dst_stride_v, dst_u, dst_stride_u,
+                    width, height);
 }
 
-// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
 LIBYUV_API
-int NV21ToI420(const uint8_t* src_y,
+int NV12ToNV24(const uint8_t* src_y,
                int src_stride_y,
-               const uint8_t* src_vu,
-               int src_stride_vu,
+               const uint8_t* src_uv,
+               int src_stride_uv,
                uint8_t* dst_y,
                int dst_stride_y,
-               uint8_t* dst_u,
-               int dst_stride_u,
-               uint8_t* dst_v,
-               int dst_stride_v,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
                int width,
                int height) {
-  return X420ToI420(src_y, src_stride_y, src_stride_y, src_vu, src_stride_vu,
-                    dst_y, dst_stride_y, dst_v, dst_stride_v, dst_u,
-                    dst_stride_u, width, height);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1),
+          SUBSAMPLE(height, 1, 1), dst_uv, dst_stride_uv, Abs(width),
+          Abs(height), kFilterBilinear);
+  return 0;
 }
 
-// Convert M420 to I420.
 LIBYUV_API
-int M420ToI420(const uint8_t* src_m420,
-               int src_stride_m420,
+int NV16ToNV24(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
                uint8_t* dst_y,
                int dst_stride_y,
-               uint8_t* dst_u,
-               int dst_stride_u,
-               uint8_t* dst_v,
-               int dst_stride_v,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
                int width,
                int height) {
-  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
-                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3, dst_y,
-                    dst_stride_y, dst_u, dst_stride_u, dst_v, dst_stride_v,
-                    width, height);
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1), height, dst_uv,
+          dst_stride_uv, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int P010ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_16(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale_16(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1),
+             SUBSAMPLE(height, 1, 1), dst_uv, dst_stride_uv, Abs(width),
+             Abs(height), kFilterBilinear);
+  return 0;
+}
+
+LIBYUV_API
+int P210ToP410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_uv,
+               int src_stride_uv,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  if (width <= 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_16(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  UVScale_16(src_uv, src_stride_uv, SUBSAMPLE(width, 1, 1), height, dst_uv,
+             dst_stride_uv, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
 }
 
 // Convert YUY2 to I420.
@@ -750,7 +1221,7 @@ int YUY2ToI420(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_YUY2TOYROW_MSA)
+#if defined(HAS_YUY2TOYROW_MSA) && defined(HAS_YUY2TOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     YUY2ToYRow = YUY2ToYRow_Any_MSA;
     YUY2ToUVRow = YUY2ToUVRow_Any_MSA;
@@ -760,15 +1231,13 @@ int YUY2ToI420(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_YUY2TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    YUY2ToYRow = YUY2ToYRow_Any_MMI;
-    YUY2ToUVRow = YUY2ToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      YUY2ToYRow = YUY2ToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        YUY2ToUVRow = YUY2ToUVRow_MMI;
-      }
+#if defined(HAS_YUY2TOYROW_LASX) && defined(HAS_YUY2TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    YUY2ToYRow = YUY2ToYRow_Any_LASX;
+    YUY2ToUVRow = YUY2ToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_LASX;
+      YUY2ToUVRow = YUY2ToUVRow_LASX;
     }
   }
 #endif
@@ -853,13 +1322,13 @@ int UYVYToI420(const uint8_t* src_uyvy,
     }
   }
 #endif
-#if defined(HAS_UYVYTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    UYVYToYRow = UYVYToYRow_Any_MMI;
-    UYVYToUVRow = UYVYToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      UYVYToYRow = UYVYToYRow_MMI;
-      UYVYToUVRow = UYVYToUVRow_MMI;
+#if defined(HAS_UYVYTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    UYVYToYRow = UYVYToYRow_Any_LASX;
+    UYVYToUVRow = UYVYToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_LASX;
+      UYVYToUVRow = UYVYToUVRow_LASX;
     }
   }
 #endif
@@ -1045,30 +1514,10 @@ int ARGBToI420(const uint8_t* src_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
-    ARGBToYRow = ARGBToYRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
-      ARGBToYRow = ARGBToYRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
-    ARGBToYRow = ARGBToYRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
-      ARGBToYRow = ARGBToYRow_AVX2;
-    }
-  }
-#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -1081,35 +1530,57 @@ int ARGBToI420(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToYRow = ARGBToYRow_Any_MSA;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_MSA;
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_MSA;
+      ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -1170,7 +1641,7 @@ int BGRAToI420(const uint8_t* src_bgra,
 #if defined(HAS_BGRATOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     BGRAToYRow = BGRAToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       BGRAToYRow = BGRAToYRow_NEON;
     }
   }
@@ -1183,35 +1654,23 @@ int BGRAToI420(const uint8_t* src_bgra,
     }
   }
 #endif
-#if defined(HAS_BGRATOYROW_MSA)
+#if defined(HAS_BGRATOYROW_MSA) && defined(HAS_BGRATOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     BGRAToYRow = BGRAToYRow_Any_MSA;
-    if (IS_ALIGNED(width, 16)) {
-      BGRAToYRow = BGRAToYRow_MSA;
-    }
-  }
-#endif
-#if defined(HAS_BGRATOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
     BGRAToUVRow = BGRAToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
+      BGRAToYRow = BGRAToYRow_MSA;
       BGRAToUVRow = BGRAToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_BGRATOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    BGRAToYRow = BGRAToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      BGRAToYRow = BGRAToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_BGRATOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    BGRAToUVRow = BGRAToUVRow_Any_MMI;
+#if defined(HAS_BGRATOYROW_LSX) && defined(HAS_BGRATOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    BGRAToYRow = BGRAToYRow_Any_LSX;
+    BGRAToUVRow = BGRAToUVRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
-      BGRAToUVRow = BGRAToUVRow_MMI;
+      BGRAToYRow = BGRAToYRow_LSX;
+      BGRAToUVRow = BGRAToUVRow_LSX;
     }
   }
 #endif
@@ -1259,20 +1718,42 @@ int ABGRToI420(const uint8_t* src_abgr,
     src_abgr = src_abgr + (height - 1) * src_stride_abgr;
     src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+#if defined(HAS_ABGRTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
     ABGRToYRow = ABGRToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_SSSE3;
       ABGRToYRow = ABGRToYRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYRow = ABGRToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ABGRTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ABGRToYRow = ABGRToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ABGRToYRow = ABGRToYRow_NEON;
     }
   }
@@ -1285,35 +1766,23 @@ int ABGRToI420(const uint8_t* src_abgr,
     }
   }
 #endif
-#if defined(HAS_ABGRTOYROW_MSA)
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ABGRToYRow = ABGRToYRow_Any_MSA;
-    if (IS_ALIGNED(width, 16)) {
-      ABGRToYRow = ABGRToYRow_MSA;
-    }
-  }
-#endif
-#if defined(HAS_ABGRTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
     ABGRToUVRow = ABGRToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_MSA;
       ABGRToUVRow = ABGRToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ABGRTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ABGRToYRow = ABGRToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ABGRToYRow = ABGRToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ABGRTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ABGRToUVRow = ABGRToUVRow_Any_MMI;
+#if defined(HAS_ABGRTOYROW_LSX) && defined(HAS_ABGRTOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    ABGRToUVRow = ABGRToUVRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_MMI;
+      ABGRToYRow = ABGRToYRow_LSX;
+      ABGRToUVRow = ABGRToUVRow_LSX;
     }
   }
 #endif
@@ -1361,20 +1830,26 @@ int RGBAToI420(const uint8_t* src_rgba,
     src_rgba = src_rgba + (height - 1) * src_stride_rgba;
     src_stride_rgba = -src_stride_rgba;
   }
-#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+#if defined(HAS_RGBATOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
     RGBAToYRow = RGBAToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      RGBAToUVRow = RGBAToUVRow_SSSE3;
       RGBAToYRow = RGBAToYRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_RGBATOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGBAToYRow = RGBAToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGBAToYRow = RGBAToYRow_NEON;
     }
   }
@@ -1387,35 +1862,23 @@ int RGBAToI420(const uint8_t* src_rgba,
     }
   }
 #endif
-#if defined(HAS_RGBATOYROW_MSA)
+#if defined(HAS_RGBATOYROW_MSA) && defined(HAS_RGBATOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGBAToYRow = RGBAToYRow_Any_MSA;
-    if (IS_ALIGNED(width, 16)) {
-      RGBAToYRow = RGBAToYRow_MSA;
-    }
-  }
-#endif
-#if defined(HAS_RGBATOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
     RGBAToUVRow = RGBAToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
+      RGBAToYRow = RGBAToYRow_MSA;
       RGBAToUVRow = RGBAToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_RGBATOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGBAToYRow = RGBAToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RGBAToYRow = RGBAToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_RGBATOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGBAToUVRow = RGBAToUVRow_Any_MMI;
+#if defined(HAS_RGBATOYROW_LSX) && defined(HAS_RGBATOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGBAToYRow = RGBAToYRow_Any_LSX;
+    RGBAToUVRow = RGBAToUVRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
-      RGBAToUVRow = RGBAToUVRow_MMI;
+      RGBAToYRow = RGBAToYRow_LSX;
+      RGBAToUVRow = RGBAToUVRow_LSX;
     }
   }
 #endif
@@ -1436,6 +1899,12 @@ int RGBAToI420(const uint8_t* src_rgba,
   return 0;
 }
 
+// Enabled if 1 pass is available
+#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
+     defined(HAS_RGB24TOYROW_LSX))
+#define HAS_RGB24TOYROW
+#endif
+
 // Convert RGB24 to I420.
 LIBYUV_API
 int RGB24ToI420(const uint8_t* src_rgb24,
@@ -1449,8 +1918,7 @@ int RGB24ToI420(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
   void (*RGB24ToUVRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
                        uint8_t* dst_u, uint8_t* dst_v, int width) =
       RGB24ToUVRow_C;
@@ -1475,19 +1943,20 @@ int RGB24ToI420(const uint8_t* src_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
 
+#if defined(HAS_RGB24TOYROW)
+
 // Neon version does direct RGB24 to YUV.
-#if defined(HAS_RGB24TOYROW_NEON)
+#if defined(HAS_RGB24TOYROW_NEON) && defined(HAS_RGB24TOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
     RGB24ToYRow = RGB24ToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGB24ToYRow = RGB24ToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVRow = RGB24ToUVRow_NEON;
-      }
+      RGB24ToUVRow = RGB24ToUVRow_NEON;
     }
   }
-#elif defined(HAS_RGB24TOYROW_MSA)
+#endif
+#if defined(HAS_RGB24TOYROW_MSA) && defined(HAS_RGB24TOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGB24ToUVRow = RGB24ToUVRow_Any_MSA;
     RGB24ToYRow = RGB24ToYRow_Any_MSA;
@@ -1496,19 +1965,31 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       RGB24ToUVRow = RGB24ToUVRow_MSA;
     }
   }
-#elif defined(HAS_RGB24TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB24ToUVRow = RGB24ToUVRow_Any_MMI;
-    RGB24ToYRow = RGB24ToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RGB24ToYRow = RGB24ToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVRow = RGB24ToUVRow_MMI;
-      }
+#endif
+#if defined(HAS_RGB24TOYROW_LSX) && defined(HAS_RGB24TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_LSX;
+    RGB24ToYRow = RGB24ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYRow = RGB24ToYRow_LSX;
+      RGB24ToUVRow = RGB24ToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYROW_LASX) && defined(HAS_RGB24TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_LASX;
+    RGB24ToYRow = RGB24ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYRow = RGB24ToYRow_LASX;
+      RGB24ToUVRow = RGB24ToUVRow_LASX;
     }
   }
+#endif
+
 // Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
+#else  // HAS_RGB24TOYROW
+
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
@@ -1517,39 +1998,49 @@ int RGB24ToI420(const uint8_t* src_rgb24,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
 #endif
+#endif  // HAS_RGB24TOYROW
 
   {
-#if !(defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-      defined(HAS_RGB24TOYROW_MMI))
+#if !defined(HAS_RGB24TOYROW)
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
       RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
@@ -1566,8 +2057,7 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-     defined(HAS_RGB24TOYROW_MMI))
+#if defined(HAS_RGB24TOYROW)
       RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
       RGB24ToYRow(src_rgb24, dst_y, width);
 #else
@@ -1576,15 +2066,19 @@ int RGB24ToI420(const uint8_t* src_rgb24,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_RGB24TOYROW_NEON) || defined(HAS_RGB24TOYROW_MSA) || \
-      defined(HAS_RGB24TOYROW_MMI))
+#if !defined(HAS_RGB24TOYROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RGB24TOYROW
+
+// Enabled if 1 pass is available
+#if defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA)
+#define HAS_RGB24TOYJROW
+#endif
 
-// TODO(fbarchard): Use Matrix version to implement I420 and J420.
 // Convert RGB24 to J420.
 LIBYUV_API
 int RGB24ToJ420(const uint8_t* src_rgb24,
@@ -1598,8 +2092,7 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
+#if defined(HAS_RGB24TOYJROW)
   void (*RGB24ToUVJRow)(const uint8_t* src_rgb24, int src_stride_rgb24,
                         uint8_t* dst_u, uint8_t* dst_v, int width) =
       RGB24ToUVJRow_C;
@@ -1624,19 +2117,20 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
     src_stride_rgb24 = -src_stride_rgb24;
   }
 
+#if defined(HAS_RGB24TOYJROW)
+
 // Neon version does direct RGB24 to YUV.
-#if defined(HAS_RGB24TOYJROW_NEON)
+#if defined(HAS_RGB24TOYJROW_NEON) && defined(HAS_RGB24TOUVJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToUVJRow = RGB24ToUVJRow_Any_NEON;
     RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGB24ToYJRow = RGB24ToYJRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVJRow = RGB24ToUVJRow_NEON;
-      }
+      RGB24ToUVJRow = RGB24ToUVJRow_NEON;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MSA)
+#endif
+#if defined(HAS_RGB24TOYJROW_MSA) && defined(HAS_RGB24TOUVJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGB24ToUVJRow = RGB24ToUVJRow_Any_MSA;
     RGB24ToYJRow = RGB24ToYJRow_Any_MSA;
@@ -1645,19 +2139,11 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       RGB24ToUVJRow = RGB24ToUVJRow_MSA;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB24ToUVJRow = RGB24ToUVJRow_Any_MMI;
-    RGB24ToYJRow = RGB24ToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RGB24ToYJRow = RGB24ToYJRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        RGB24ToUVJRow = RGB24ToUVJRow_MMI;
-      }
-    }
-  }
+#endif
+
 // Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
+#else  // HAS_RGB24TOYJROW
+
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
@@ -1666,39 +2152,49 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#if defined(HAS_ARGBTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
       ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_AVX2) && defined(HAS_ARGBTOUVJROW_AVX2)
+#if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVJRow = ARGBToUVJRow_AVX2;
       ARGBToYJRow = ARGBToYJRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
 #endif
+#endif  // HAS_RGB24TOYJROW
 
   {
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
+#if !defined(HAS_RGB24TOYJROW)
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
+#if defined(HAS_RGB24TOYJROW)
       RGB24ToUVJRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
       RGB24ToYJRow(src_rgb24, dst_y, width);
       RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
@@ -1715,8 +2211,7 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
+#if defined(HAS_RGB24TOYJROW)
       RGB24ToUVJRow(src_rgb24, 0, dst_u, dst_v, width);
       RGB24ToYJRow(src_rgb24, dst_y, width);
 #else
@@ -1725,13 +2220,19 @@ int RGB24ToJ420(const uint8_t* src_rgb24,
       ARGBToYJRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
+#if !defined(HAS_RGB24TOYJROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RGB24TOYJROW
+
+// Enabled if 1 pass is available
+#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
+     defined(HAS_RAWTOYROW_LSX))
+#define HAS_RAWTOYROW
+#endif
 
 // Convert RAW to I420.
 LIBYUV_API
@@ -1746,8 +2247,7 @@ int RAWToI420(const uint8_t* src_raw,
               int width,
               int height) {
   int y;
-#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-     defined(HAS_RAWTOYROW_MMI))
+#if defined(HAS_RAWTOYROW)
   void (*RAWToUVRow)(const uint8_t* src_raw, int src_stride_raw, uint8_t* dst_u,
                      uint8_t* dst_v, int width) = RAWToUVRow_C;
   void (*RAWToYRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
@@ -1771,19 +2271,20 @@ int RAWToI420(const uint8_t* src_raw,
     src_stride_raw = -src_stride_raw;
   }
 
+#if defined(HAS_RAWTOYROW)
+
 // Neon version does direct RAW to YUV.
-#if defined(HAS_RAWTOYROW_NEON)
+#if defined(HAS_RAWTOYROW_NEON) && defined(HAS_RAWTOUVROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RAWToUVRow = RAWToUVRow_Any_NEON;
     RAWToYRow = RAWToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RAWToYRow = RAWToYRow_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        RAWToUVRow = RAWToUVRow_NEON;
-      }
+      RAWToUVRow = RAWToUVRow_NEON;
     }
   }
-#elif defined(HAS_RAWTOYROW_MSA)
+#endif
+#if defined(HAS_RAWTOYROW_MSA) && defined(HAS_RAWTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RAWToUVRow = RAWToUVRow_Any_MSA;
     RAWToYRow = RAWToYRow_Any_MSA;
@@ -1792,19 +2293,31 @@ int RAWToI420(const uint8_t* src_raw,
       RAWToUVRow = RAWToUVRow_MSA;
     }
   }
-#elif defined(HAS_RAWTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RAWToUVRow = RAWToUVRow_Any_MMI;
-    RAWToYRow = RAWToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RAWToYRow = RAWToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        RAWToUVRow = RAWToUVRow_MMI;
-      }
+#endif
+#if defined(HAS_RAWTOYROW_LSX) && defined(HAS_RAWTOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToUVRow = RAWToUVRow_Any_LSX;
+    RAWToYRow = RAWToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYRow = RAWToYRow_LSX;
+      RAWToUVRow = RAWToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYROW_LASX) && defined(HAS_RAWTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToUVRow = RAWToUVRow_Any_LASX;
+    RAWToYRow = RAWToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYRow = RAWToYRow_LASX;
+      RAWToUVRow = RAWToUVRow_LASX;
     }
   }
+#endif
+
 // Other platforms do intermediate conversion from RAW to ARGB.
-#else
+#else  // HAS_RAWTOYROW
+
 #if defined(HAS_RAWTOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
@@ -1813,39 +2326,49 @@ int RAWToI420(const uint8_t* src_raw,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RAWTOYROW
 
   {
-#if !(defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-      defined(HAS_RAWTOYROW_MMI))
+#if !defined(HAS_RAWTOYROW)
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-     defined(HAS_RAWTOYROW_MMI))
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
       RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
@@ -1862,8 +2385,7 @@ int RAWToI420(const uint8_t* src_raw,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-     defined(HAS_RAWTOYROW_MMI))
+#if defined(HAS_RAWTOYROW)
       RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
       RAWToYRow(src_raw, dst_y, width);
 #else
@@ -1872,13 +2394,167 @@ int RAWToI420(const uint8_t* src_raw,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_RAWTOYROW_NEON) || defined(HAS_RAWTOYROW_MSA) || \
-      defined(HAS_RAWTOYROW_MMI))
+#if !defined(HAS_RAWTOYROW)
     free_aligned_buffer_64(row);
 #endif
   }
   return 0;
 }
+#undef HAS_RAWTOYROW
+
+// Enabled if 1 pass is available
+#if defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA)
+#define HAS_RAWTOYJROW
+#endif
+
+// Convert RAW to J420.
+LIBYUV_API
+int RAWToJ420(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int width,
+              int height) {
+  int y;
+#if defined(HAS_RAWTOYJROW)
+  void (*RAWToUVJRow)(const uint8_t* src_raw, int src_stride_raw,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RAWToUVJRow_C;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
+      RAWToYJRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
+      ARGBToYJRow_C;
+#endif
+  if (!src_raw || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYJROW)
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVJRow = RAWToUVJRow_Any_NEON;
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
+      RAWToUVJRow = RAWToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_MSA) && defined(HAS_RAWTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToUVJRow = RAWToUVJRow_Any_MSA;
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
+      RAWToUVJRow = RAWToUVJRow_MSA;
+    }
+  }
+#endif
+
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else  // HAS_RAWTOYJROW
+
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RAWTOYJROW
+
+  {
+#if !defined(HAS_RAWTOYJROW)
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+      RAWToYJRow(src_raw, dst_y, width);
+      RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+      ARGBToUVJRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+      ARGBToYJRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, 0, dst_u, dst_v, width);
+      RAWToYJRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVJRow(row, 0, dst_u, dst_v, width);
+      ARGBToYJRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYJROW)
+    free_aligned_buffer_64(row);
+#endif
+  }
+  return 0;
+}
+#undef HAS_RAWTOYJROW
 
 // Convert RGB565 to I420.
 LIBYUV_API
@@ -1894,7 +2570,7 @@ int RGB565ToI420(const uint8_t* src_rgb565,
                  int height) {
   int y;
 #if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
-     defined(HAS_RGB565TOYROW_MMI))
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
   void (*RGB565ToUVRow)(const uint8_t* src_rgb565, int src_stride_rgb565,
                         uint8_t* dst_u, uint8_t* dst_v, int width) =
       RGB565ToUVRow_C;
@@ -1931,7 +2607,10 @@ int RGB565ToI420(const uint8_t* src_rgb565,
       }
     }
   }
-#elif defined(HAS_RGB565TOYROW_MSA)
+// MSA version does direct RGB565 to YUV.
+#elif (defined(HAS_RGB565TOYROW_MSA) || defined(HAS_RGB565TOYROW_LSX) || \
+       defined(HAS_RGB565TOYROW_LASX))
+#if defined(HAS_RGB565TOYROW_MSA) && defined(HAS_RGB565TOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGB565ToUVRow = RGB565ToUVRow_Any_MSA;
     RGB565ToYRow = RGB565ToYRow_Any_MSA;
@@ -1940,17 +2619,27 @@ int RGB565ToI420(const uint8_t* src_rgb565,
       RGB565ToUVRow = RGB565ToUVRow_MSA;
     }
   }
-#elif defined(HAS_RGB565TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB565ToUVRow = RGB565ToUVRow_Any_MMI;
-    RGB565ToYRow = RGB565ToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RGB565ToYRow = RGB565ToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        RGB565ToUVRow = RGB565ToUVRow_MMI;
-      }
+#endif
+#if defined(HAS_RGB565TOYROW_LSX) && defined(HAS_RGB565TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_LSX;
+    RGB565ToYRow = RGB565ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToYRow = RGB565ToYRow_LSX;
+      RGB565ToUVRow = RGB565ToUVRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOYROW_LASX) && defined(HAS_RGB565TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_LASX;
+    RGB565ToYRow = RGB565ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB565ToYRow = RGB565ToYRow_LASX;
+      RGB565ToUVRow = RGB565ToUVRow_LASX;
     }
   }
+#endif
 // Other platforms do intermediate conversion from RGB565 to ARGB.
 #else
 #if defined(HAS_RGB565TOARGBROW_SSE2)
@@ -1969,37 +2658,49 @@ int RGB565ToI420(const uint8_t* src_rgb565,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #endif
   {
 #if !(defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
-      defined(HAS_RGB565TOYROW_MMI))
+      defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
 #endif
     for (y = 0; y < height - 1; y += 2) {
 #if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
-     defined(HAS_RGB565TOYROW_MMI))
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
       RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
       RGB565ToYRow(src_rgb565, dst_y, width);
       RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
@@ -2017,7 +2718,7 @@ int RGB565ToI420(const uint8_t* src_rgb565,
     }
     if (height & 1) {
 #if (defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
-     defined(HAS_RGB565TOYROW_MMI))
+     defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
       RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
       RGB565ToYRow(src_rgb565, dst_y, width);
 #else
@@ -2027,7 +2728,7 @@ int RGB565ToI420(const uint8_t* src_rgb565,
 #endif
     }
 #if !(defined(HAS_RGB565TOYROW_NEON) || defined(HAS_RGB565TOYROW_MSA) || \
-      defined(HAS_RGB565TOYROW_MMI))
+      defined(HAS_RGB565TOYROW_LSX) || defined(HAS_RGB565TOYROW_LASX))
     free_aligned_buffer_64(row);
 #endif
   }
@@ -2048,7 +2749,7 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
                    int height) {
   int y;
 #if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
-     defined(HAS_ARGB1555TOYROW_MMI))
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
   void (*ARGB1555ToUVRow)(const uint8_t* src_argb1555, int src_stride_argb1555,
                           uint8_t* dst_u, uint8_t* dst_v, int width) =
       ARGB1555ToUVRow_C;
@@ -2086,7 +2787,10 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
       }
     }
   }
-#elif defined(HAS_ARGB1555TOYROW_MSA)
+// MSA version does direct ARGB1555 to YUV.
+#elif (defined(HAS_ARGB1555TOYROW_MSA) || defined(HAS_ARGB1555TOYROW_LSX) || \
+       defined(HAS_ARGB1555TOYROW_LASX))
+#if defined(HAS_ARGB1555TOYROW_MSA) && defined(HAS_ARGB1555TOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGB1555ToUVRow = ARGB1555ToUVRow_Any_MSA;
     ARGB1555ToYRow = ARGB1555ToYRow_Any_MSA;
@@ -2095,17 +2799,27 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
       ARGB1555ToUVRow = ARGB1555ToUVRow_MSA;
     }
   }
-#elif defined(HAS_ARGB1555TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_MMI;
-    ARGB1555ToYRow = ARGB1555ToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGB1555ToYRow = ARGB1555ToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        ARGB1555ToUVRow = ARGB1555ToUVRow_MMI;
-      }
+#endif
+#if defined(HAS_ARGB1555TOYROW_LSX) && defined(HAS_ARGB1555TOUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_LSX;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_LSX;
+      ARGB1555ToUVRow = ARGB1555ToUVRow_LSX;
     }
   }
+#endif
+#if defined(HAS_ARGB1555TOYROW_LASX) && defined(HAS_ARGB1555TOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_LASX;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_LASX;
+      ARGB1555ToUVRow = ARGB1555ToUVRow_LASX;
+    }
+  }
+#endif
 // Other platforms do intermediate conversion from ARGB1555 to ARGB.
 #else
 #if defined(HAS_ARGB1555TOARGBROW_SSE2)
@@ -2124,30 +2838,42 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #endif
   {
 #if !(defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
-      defined(HAS_ARGB1555TOYROW_MMI))
+      defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
@@ -2155,7 +2881,7 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
 
     for (y = 0; y < height - 1; y += 2) {
 #if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
-     defined(HAS_ARGB1555TOYROW_MMI))
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
       ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
       ARGB1555ToYRow(src_argb1555, dst_y, width);
       ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
@@ -2175,7 +2901,7 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
     }
     if (height & 1) {
 #if (defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
-     defined(HAS_ARGB1555TOYROW_MMI))
+     defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
       ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
       ARGB1555ToYRow(src_argb1555, dst_y, width);
 #else
@@ -2185,7 +2911,7 @@ int ARGB1555ToI420(const uint8_t* src_argb1555,
 #endif
     }
 #if !(defined(HAS_ARGB1555TOYROW_NEON) || defined(HAS_ARGB1555TOYROW_MSA) || \
-      defined(HAS_ARGB1555TOYROW_MMI))
+      defined(HAS_ARGB1555TOYROW_LSX) || defined(HAS_ARGB1555TOYROW_LASX))
     free_aligned_buffer_64(row);
 #endif
   }
@@ -2205,7 +2931,7 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
                    int width,
                    int height) {
   int y;
-#if (defined(HAS_ARGB4444TOYROW_NEON) || defined(HAS_ARGB4444TOYROW_MMI))
+#if defined(HAS_ARGB4444TOYROW_NEON)
   void (*ARGB4444ToUVRow)(const uint8_t* src_argb4444, int src_stride_argb4444,
                           uint8_t* dst_u, uint8_t* dst_v, int width) =
       ARGB4444ToUVRow_C;
@@ -2243,17 +2969,6 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
       }
     }
   }
-#elif defined(HAS_ARGB4444TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGB4444ToUVRow = ARGB4444ToUVRow_Any_MMI;
-    ARGB4444ToYRow = ARGB4444ToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGB4444ToYRow = ARGB4444ToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        ARGB4444ToUVRow = ARGB4444ToUVRow_MMI;
-      }
-    }
-  }
 // Other platforms do intermediate conversion from ARGB4444 to ARGB.
 #else
 #if defined(HAS_ARGB4444TOARGBROW_SSE2)
@@ -2280,27 +2995,55 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGB4444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToUVRow = ARGBToUVRow_Any_MSA;
     ARGBToYRow = ARGBToYRow_Any_MSA;
@@ -2312,29 +3055,27 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUVRow = ARGBToUVRow_MMI;
-      }
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
 #endif
 
   {
-#if !(defined(HAS_ARGB4444TOYROW_NEON) || defined(HAS_ARGB4444TOYROW_MMI))
+#if !(defined(HAS_ARGB4444TOYROW_NEON))
     // Allocate 2 rows of ARGB.
     const int kRowSize = (width * 4 + 31) & ~31;
     align_buffer_64(row, kRowSize * 2);
 #endif
 
     for (y = 0; y < height - 1; y += 2) {
-#if (defined(HAS_ARGB4444TOYROW_NEON) || defined(HAS_ARGB4444TOYROW_MMI))
+#if defined(HAS_ARGB4444TOYROW_NEON)
       ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width);
       ARGB4444ToYRow(src_argb4444, dst_y, width);
       ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y,
@@ -2353,7 +3094,7 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
       dst_v += dst_stride_v;
     }
     if (height & 1) {
-#if (defined(HAS_ARGB4444TOYROW_NEON) || defined(HAS_ARGB4444TOYROW_MMI))
+#if defined(HAS_ARGB4444TOYROW_NEON)
       ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width);
       ARGB4444ToYRow(src_argb4444, dst_y, width);
 #else
@@ -2362,7 +3103,7 @@ int ARGB4444ToI420(const uint8_t* src_argb4444,
       ARGBToYRow(row, dst_y, width);
 #endif
     }
-#if !(defined(HAS_ARGB4444TOYROW_NEON) || defined(HAS_ARGB4444TOYROW_MMI))
+#if !(defined(HAS_ARGB4444TOYROW_NEON))
     free_aligned_buffer_64(row);
 #endif
   }
@@ -2378,125 +3119,129 @@ int RGB24ToJ400(const uint8_t* src_rgb24,
                 int width,
                 int height) {
   int y;
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
   void (*RGB24ToYJRow)(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) =
       RGB24ToYJRow_C;
-#else
-  void (*RGB24ToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
-      RGB24ToARGBRow_C;
-  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_yj, int width) =
-      ARGBToYJRow_C;
-#endif
   if (!src_rgb24 || !dst_yj || width <= 0 || height == 0) {
     return -1;
   }
-  // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
     src_stride_rgb24 = -src_stride_rgb24;
   }
-
-// Neon version does direct RGB24 to YUV.
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_yj = 0;
+  }
+#if defined(HAS_RGB24TOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToYJRow = RGB24ToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB24ToYJRow = RGB24ToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToYJRow = RGB24ToYJRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_RGB24TOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     RGB24ToYJRow = RGB24ToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       RGB24ToYJRow = RGB24ToYJRow_NEON;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MSA)
+#endif
+#if defined(HAS_RGB24TOYJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     RGB24ToYJRow = RGB24ToYJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       RGB24ToYJRow = RGB24ToYJRow_MSA;
     }
   }
-#elif defined(HAS_RGB24TOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB24ToYJRow = RGB24ToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      RGB24ToYJRow = RGB24ToYJRow_MMI;
-    }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB24ToYJRow(src_rgb24, dst_yj, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_yj += dst_stride_yj;
   }
-// Other platforms do intermediate conversion from RGB24 to ARGB.
-#else
-#if defined(HAS_RGB24TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
-    }
+  return 0;
+}
+
+// Convert RAW to J400.
+LIBYUV_API
+int RAWToJ400(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_yj,
+              int dst_stride_yj,
+              int width,
+              int height) {
+  int y;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_yj, int width) =
+      RAWToYJRow_C;
+  if (!src_raw || !dst_yj || width <= 0 || height == 0) {
+    return -1;
   }
-#endif
-#if defined(HAS_ARGBTOYJROW_SSSE3)
+
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_yj = 0;
+  }
+
+#if defined(HAS_RAWTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    RAWToYJRow = RAWToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYJRow = ARGBToYJRow_SSSE3;
+      RAWToYJRow = RAWToYJRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_AVX2)
+#if defined(HAS_RAWTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    RAWToYJRow = RAWToYJRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToYJRow = ARGBToYJRow_AVX2;
+      RAWToYJRow = RAWToYJRow_AVX2;
     }
   }
 #endif
-#endif
-
-  {
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
-    // Allocate 2 rows of ARGB.
-    const int kRowSize = (width * 4 + 31) & ~31;
-    align_buffer_64(row, kRowSize * 2);
-#endif
-
-    for (y = 0; y < height - 1; y += 2) {
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
-      RGB24ToYJRow(src_rgb24, dst_yj, width);
-      RGB24ToYJRow(src_rgb24 + src_stride_rgb24, dst_yj + dst_stride_yj, width);
-#else
-      RGB24ToARGBRow(src_rgb24, row, width);
-      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
-      ARGBToYJRow(row, dst_yj, width);
-      ARGBToYJRow(row + kRowSize, dst_yj + dst_stride_yj, width);
-#endif
-      src_rgb24 += src_stride_rgb24 * 2;
-      dst_yj += dst_stride_yj * 2;
+#if defined(HAS_RAWTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
     }
-    if (height & 1) {
-#if (defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-     defined(HAS_RGB24TOYJROW_MMI))
-      RGB24ToYJRow(src_rgb24, dst_yj, width);
-#else
-      RGB24ToARGBRow(src_rgb24, row, width);
-      ARGBToYJRow(row, dst_yj, width);
+  }
 #endif
+#if defined(HAS_RAWTOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
     }
-#if !(defined(HAS_RGB24TOYJROW_NEON) || defined(HAS_RGB24TOYJROW_MSA) || \
-      defined(HAS_RGB24TOYJROW_MMI))
-    free_aligned_buffer_64(row);
-#endif
   }
-  return 0;
-}
+#endif
 
-static void SplitPixels(const uint8_t* src_u,
-                        int src_pixel_stride_uv,
-                        uint8_t* dst_u,
-                        int width) {
-  int i;
-  for (i = 0; i < width; ++i) {
-    *dst_u = *src_u;
-    ++dst_u;
-    src_u += src_pixel_stride_uv;
+  for (y = 0; y < height; ++y) {
+    RAWToYJRow(src_raw, dst_yj, width);
+    src_raw += src_stride_raw;
+    dst_yj += dst_stride_yj;
   }
+  return 0;
 }
 
 // Convert Android420 to I420.
@@ -2516,58 +3261,10 @@ int Android420ToI420(const uint8_t* src_y,
                      int dst_stride_v,
                      int width,
                      int height) {
-  int y;
-  const ptrdiff_t vu_off = src_v - src_u;
-  int halfwidth = (width + 1) >> 1;
-  int halfheight = (height + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    halfheight = (height + 1) >> 1;
-    src_y = src_y + (height - 1) * src_stride_y;
-    src_u = src_u + (halfheight - 1) * src_stride_u;
-    src_v = src_v + (halfheight - 1) * src_stride_v;
-    src_stride_y = -src_stride_y;
-    src_stride_u = -src_stride_u;
-    src_stride_v = -src_stride_v;
-  }
-
-  if (dst_y) {
-    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-  }
-
-  // Copy UV planes as is - I420
-  if (src_pixel_stride_uv == 1) {
-    CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
-    CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
-    return 0;
-    // Split UV planes - NV21
-  }
-  if (src_pixel_stride_uv == 2 && vu_off == -1 &&
-      src_stride_u == src_stride_v) {
-    SplitUVPlane(src_v, src_stride_v, dst_v, dst_stride_v, dst_u, dst_stride_u,
-                 halfwidth, halfheight);
-    return 0;
-    // Split UV planes - NV12
-  }
-  if (src_pixel_stride_uv == 2 && vu_off == 1 && src_stride_u == src_stride_v) {
-    SplitUVPlane(src_u, src_stride_u, dst_u, dst_stride_u, dst_v, dst_stride_v,
-                 halfwidth, halfheight);
-    return 0;
-  }
-
-  for (y = 0; y < halfheight; ++y) {
-    SplitPixels(src_u, src_pixel_stride_uv, dst_u, halfwidth);
-    SplitPixels(src_v, src_pixel_stride_uv, dst_v, halfwidth);
-    src_u += src_stride_u;
-    src_v += src_stride_v;
-    dst_u += dst_stride_u;
-    dst_v += dst_stride_v;
-  }
-  return 0;
+  return Android420ToI420Rotate(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                                src_stride_v, src_pixel_stride_uv, dst_y,
+                                dst_stride_y, dst_u, dst_stride_u, dst_v,
+                                dst_stride_v, width, height, kRotate0);
 }
 
 #ifdef __cplusplus
diff --git a/files/source/convert_argb.cc b/files/source/convert_argb.cc
index 54050333..71ef8c10 100644
--- a/files/source/convert_argb.cc
+++ b/files/source/convert_argb.cc
@@ -7,7 +7,6 @@
  *  in the file PATENTS. All contributing project authors may
  *  be found in the AUTHORS file in the root of the source tree.
  */
-
 #include "libyuv/convert_argb.h"
 
 #include "libyuv/cpu_id.h"
@@ -17,6 +16,7 @@
 #include "libyuv/planar_functions.h"  // For CopyPlane and ARGBShuffle.
 #include "libyuv/rotate_argb.h"
 #include "libyuv/row.h"
+#include "libyuv/scale_row.h"  // For ScaleRowUp2_Linear and ScaleRowUp2_Bilinear
 #include "libyuv/video_common.h"
 
 #ifdef __cplusplus
@@ -47,18 +47,19 @@ int ARGBCopy(const uint8_t* src_argb,
   return 0;
 }
 
-// Convert I420 to ARGB with matrix
-static int I420ToARGBMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert I420 to ARGB with matrix.
+LIBYUV_API
+int I420ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
                         const uint8_t* v_buf, uint8_t* rgb_buf,
@@ -89,6 +90,15 @@ static int I420ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -105,6 +115,14 @@ static int I420ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
@@ -226,18 +244,55 @@ int H420ToABGR(const uint8_t* src_y,
                           width, height);
 }
 
-// Convert I422 to ARGB with matrix
-static int I422ToARGBMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert U420 to ARGB.
+LIBYUV_API
+int U420ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U420 to ABGR.
+LIBYUV_API
+int U420ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I420ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I422 to ARGB with matrix.
+LIBYUV_API
+int I422ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
                         const uint8_t* v_buf, uint8_t* rgb_buf,
@@ -275,6 +330,15 @@ static int I422ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -291,6 +355,14 @@ static int I422ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I422ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
@@ -410,20 +482,286 @@ int H422ToABGR(const uint8_t* src_y,
                           width, height);
 }
 
-// Convert 10 bit YUV to ARGB with matrix
+// Convert U422 to ARGB.
+LIBYUV_API
+int U422ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U422 to ABGR.
+LIBYUV_API
+int U422ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I422ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I444 to ARGB with matrix.
+LIBYUV_API
+int I444ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && src_stride_u == width && src_stride_v == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I444 to ABGR.
+LIBYUV_API
+int I444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert J444 to ARGB.
+LIBYUV_API
+int J444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvJPEGConstants, width, height);
+}
+
+// Convert J444 to ABGR.
+LIBYUV_API
+int J444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuJPEGConstants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert H444 to ARGB.
+LIBYUV_API
+int H444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert H444 to ABGR.
+LIBYUV_API
+int H444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuH709Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert U444 to ARGB.
+LIBYUV_API
+int U444ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U444 to ABGR.
+LIBYUV_API
+int U444ToABGR(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
 // TODO(fbarchard): Consider passing scale multiplier to I210ToARGB to
 // multiply 10 bit yuv into high bits to allow any number of bits.
-static int I010ToAR30Matrix(const uint16_t* src_y,
-                            int src_stride_y,
-                            const uint16_t* src_u,
-                            int src_stride_u,
-                            const uint16_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_ar30,
-                            int dst_stride_ar30,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+LIBYUV_API
+int I010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*I210ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
                         const uint16_t* v_buf, uint8_t* rgb_buf,
@@ -500,6 +838,23 @@ int H010ToAR30(const uint16_t* src_y,
                           &kYuvH709Constants, width, height);
 }
 
+// Convert U010 to AR30.
+LIBYUV_API
+int U010ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuv2020Constants, width, height);
+}
+
 // Convert I010 to AB30.
 LIBYUV_API
 int I010ToAB30(const uint16_t* src_y,
@@ -534,18 +889,302 @@ int H010ToAB30(const uint16_t* src_y,
                           &kYvuH709Constants, width, height);
 }
 
-// Convert 10 bit YUV to ARGB with matrix
-static int I010ToARGBMatrix(const uint16_t* src_y,
-                            int src_stride_y,
-                            const uint16_t* src_u,
-                            int src_stride_u,
-                            const uint16_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert U010 to AB30.
+LIBYUV_API
+int U010ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I010ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert 12 bit YUV to ARGB with matrix.
+// TODO(fbarchard): Consider passing scale multiplier to I212ToARGB to
+// multiply 12 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I012ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I212ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I212ToAR30Row_C;
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I212TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I212ToAR30Row = I212ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I212ToAR30Row = I212ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I212TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I212ToAR30Row = I212ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I212ToAR30Row = I212ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I212ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
+// TODO(fbarchard): Consider passing scale multiplier to I210ToARGB to
+// multiply 10 bit yuv into high bits to allow any number of bits.
+LIBYUV_API
+int I210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToAR30Row_C;
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToAR30Row = I210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210ToAR30Row = I210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToAR30Row = I210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToAR30Row = I210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I210ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I210 to AR30.
+LIBYUV_API
+int I210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert H210 to AR30.
+LIBYUV_API
+int H210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvH709Constants, width, height);
+}
+
+// Convert U210 to AR30.
+LIBYUV_API
+int U210ToAR30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert I210 to AB30.
+LIBYUV_API
+int I210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuI601Constants, width, height);
+}
+
+// Convert H210 to AB30.
+LIBYUV_API
+int H210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuH709Constants, width, height);
+}
+
+// Convert U210 to AB30.
+LIBYUV_API
+int U210ToAB30(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I210ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYuv2020Constants, width, height);
+}
+
+LIBYUV_API
+int I410ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I410ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert 10 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*I210ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
                         const uint16_t* v_buf, uint8_t* rgb_buf,
@@ -660,23 +1299,60 @@ int H010ToABGR(const uint16_t* src_y,
                           width, height);
 }
 
-// Convert I444 to ARGB with matrix
-static int I444ToARGBMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert U010 to ARGB.
+LIBYUV_API
+int U010ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U010 to ABGR.
+LIBYUV_API
+int U010ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I010ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert 12 bit YUV to ARGB with matrix.
+LIBYUV_API
+int I012ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
-  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
-                        const uint8_t* v_buf, uint8_t* rgb_buf,
+  void (*I212ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
                         const struct YuvConstants* yuvconstants, int width) =
-      I444ToARGBRow_C;
+      I212ToARGBRow_C;
   if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
     return -1;
   }
@@ -686,48 +1362,79 @@ static int I444ToARGBMatrix(const uint8_t* src_y,
     dst_argb = dst_argb + (height - 1) * dst_stride_argb;
     dst_stride_argb = -dst_stride_argb;
   }
-  // Coalesce rows.
-  if (src_stride_y == width && src_stride_u == width && src_stride_v == width &&
-      dst_stride_argb == width * 4) {
-    width *= height;
-    height = 1;
-    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
-  }
-#if defined(HAS_I444TOARGBROW_SSSE3)
+#if defined(HAS_I212TOARGBROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    I212ToARGBRow = I212ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I444ToARGBRow = I444ToARGBRow_SSSE3;
+      I212ToARGBRow = I212ToARGBRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_I444TOARGBROW_AVX2)
+#if defined(HAS_I212TOARGBROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    I212ToARGBRow = I212ToARGBRow_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
-      I444ToARGBRow = I444ToARGBRow_AVX2;
+      I212ToARGBRow = I212ToARGBRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_I444TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+  for (y = 0; y < height; ++y) {
+    I212ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert 10 bit 422 YUV to ARGB with matrix.
+LIBYUV_API
+int I210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I210ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I210ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210ToARGBRow = I210ToARGBRow_Any_SSSE3;
     if (IS_ALIGNED(width, 8)) {
-      I444ToARGBRow = I444ToARGBRow_NEON;
+      I210ToARGBRow = I210ToARGBRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_I444TOARGBROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I444ToARGBRow = I444ToARGBRow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I444ToARGBRow = I444ToARGBRow_MSA;
+#if defined(HAS_I210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210ToARGBRow = I210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210ToARGBRow = I210ToARGBRow_AVX2;
     }
   }
 #endif
-
   for (y = 0; y < height; ++y) {
-    I444ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    I210ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
     src_u += src_stride_u;
@@ -736,74 +1443,378 @@ static int I444ToARGBMatrix(const uint8_t* src_y,
   return 0;
 }
 
-// Convert I444 to ARGB.
+// Convert I210 to ARGB.
 LIBYUV_API
-int I444ToARGB(const uint8_t* src_y,
+int I210ToARGB(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
                           src_stride_v, dst_argb, dst_stride_argb,
                           &kYuvI601Constants, width, height);
 }
 
-// Convert I444 to ABGR.
+// Convert I210 to ABGR.
 LIBYUV_API
-int I444ToABGR(const uint8_t* src_y,
+int I210ToABGR(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_abgr,
                int dst_stride_abgr,
                int width,
                int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y, src_v,
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
                           src_stride_v,  // Swap U and V
                           src_u, src_stride_u, dst_abgr, dst_stride_abgr,
                           &kYvuI601Constants,  // Use Yvu matrix
                           width, height);
 }
 
-// Convert J444 to ARGB.
+// Convert H210 to ARGB.
 LIBYUV_API
-int J444ToARGB(const uint8_t* src_y,
+int H210ToARGB(const uint16_t* src_y,
                int src_stride_y,
-               const uint8_t* src_u,
+               const uint16_t* src_u,
                int src_stride_u,
-               const uint8_t* src_v,
+               const uint16_t* src_v,
                int src_stride_v,
                uint8_t* dst_argb,
                int dst_stride_argb,
                int width,
                int height) {
-  return I444ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
                           src_stride_v, dst_argb, dst_stride_argb,
-                          &kYuvJPEGConstants, width, height);
+                          &kYuvH709Constants, width, height);
 }
 
-// Convert I420 with Alpha to preattenuated ARGB.
-static int I420AlphaToARGBMatrix(const uint8_t* src_y,
-                                 int src_stride_y,
-                                 const uint8_t* src_u,
-                                 int src_stride_u,
-                                 const uint8_t* src_v,
-                                 int src_stride_v,
-                                 const uint8_t* src_a,
-                                 int src_stride_a,
-                                 uint8_t* dst_argb,
-                                 int dst_stride_argb,
-                                 const struct YuvConstants* yuvconstants,
-                                 int width,
-                                 int height,
-                                 int attenuate) {
+// Convert H210 to ABGR.
+LIBYUV_API
+int H210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvuH709Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert U210 to ARGB.
+LIBYUV_API
+int U210ToARGB(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_argb, dst_stride_argb,
+                          &kYuv2020Constants, width, height);
+}
+
+// Convert U210 to ABGR.
+LIBYUV_API
+int U210ToABGR(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_abgr,
+               int dst_stride_abgr,
+               int width,
+               int height) {
+  return I210ToARGBMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_abgr, dst_stride_abgr,
+                          &kYvu2020Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+LIBYUV_API
+int I410ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    I410ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P010ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToARGBRow = P210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToARGBRow = P210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToARGBRow = P210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToARGBRow = P210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P210ToARGBMatrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P210TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToARGBRow = P210ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToARGBRow = P210ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToARGBRow = P210ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToARGBRow = P210ToARGBRow_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P010ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToAR30Row_C;
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToAR30Row = P210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToAR30Row = P210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToAR30Row = P210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToAR30Row = P210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToAR30Row(src_y, src_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+int P210ToAR30Matrix(const uint16_t* src_y,
+                     int src_stride_y,
+                     const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*P210ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P210ToAR30Row_C;
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P210TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P210ToAR30Row = P210ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P210ToAR30Row = P210ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P210TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P210ToAR30Row = P210ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P210ToAR30Row = P210ToAR30Row_AVX2;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    P210ToAR30Row(src_y, src_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert I420 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I420AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
   int y;
   void (*I422AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
                              const uint8_t* v_buf, const uint8_t* a_buf,
@@ -812,7 +1823,8 @@ static int I420AlphaToARGBMatrix(const uint8_t* src_y,
                              int width) = I422AlphaToARGBRow_C;
   void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
                            int width) = ARGBAttenuateRow_C;
-  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -853,6 +1865,14 @@ static int I420AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I422ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBATTENUATEROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
@@ -885,14 +1905,6 @@ static int I420AlphaToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_ARGBATTENUATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBAttenuateRow = ARGBAttenuateRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBAttenuateRow = ARGBAttenuateRow_MMI;
-    }
-  }
-#endif
 
   for (y = 0; y < height; ++y) {
     I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
@@ -911,6 +1923,242 @@ static int I420AlphaToARGBMatrix(const uint8_t* src_y,
   return 0;
 }
 
+// Convert I422 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I422AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I422AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I422AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I422ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422AlphaToARGBRow = I422AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I422AlphaToARGBRow = I422AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I444 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I444AlphaToARGBMatrix(const uint8_t* src_y,
+                          int src_stride_y,
+                          const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          const uint8_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
 // Convert I420 with Alpha to ARGB.
 LIBYUV_API
 int I420AlphaToARGB(const uint8_t* src_y,
@@ -954,16 +2202,400 @@ int I420AlphaToABGR(const uint8_t* src_y,
       width, height, attenuate);
 }
 
-// Convert I400 to ARGB.
+// Convert I422 with Alpha to ARGB.
 LIBYUV_API
-int I400ToARGB(const uint8_t* src_y,
-               int src_stride_y,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
-               int width,
-               int height) {
+int I422AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I422AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, src_a, src_stride_a, dst_argb,
+                               dst_stride_argb, &kYuvI601Constants, width,
+                               height, attenuate);
+}
+
+// Convert I422 with Alpha to ABGR.
+LIBYUV_API
+int I422AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I422AlphaToARGBMatrix(
+      src_y, src_stride_y, src_v, src_stride_v,  // Swap U and V
+      src_u, src_stride_u, src_a, src_stride_a, dst_abgr, dst_stride_abgr,
+      &kYvuI601Constants,  // Use Yvu matrix
+      width, height, attenuate);
+}
+
+// Convert I444 with Alpha to ARGB.
+LIBYUV_API
+int I444AlphaToARGB(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I444AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                               src_stride_v, src_a, src_stride_a, dst_argb,
+                               dst_stride_argb, &kYuvI601Constants, width,
+                               height, attenuate);
+}
+
+// Convert I444 with Alpha to ABGR.
+LIBYUV_API
+int I444AlphaToABGR(const uint8_t* src_y,
+                    int src_stride_y,
+                    const uint8_t* src_u,
+                    int src_stride_u,
+                    const uint8_t* src_v,
+                    int src_stride_v,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_abgr,
+                    int dst_stride_abgr,
+                    int width,
+                    int height,
+                    int attenuate) {
+  return I444AlphaToARGBMatrix(
+      src_y, src_stride_y, src_v, src_stride_v,  // Swap U and V
+      src_u, src_stride_u, src_a, src_stride_a, dst_abgr, dst_stride_abgr,
+      &kYvuI601Constants,  // Use Yvu matrix
+      width, height, attenuate);
+}
+
+// Convert I010 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I010AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
   int y;
-  void (*I400ToARGBRow)(const uint8_t* y_buf, uint8_t* rgb_buf, int width) =
+  void (*I210AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I210AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I210 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I210AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I210AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I210AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I210ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I210AlphaToARGBRow = I210AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I210AlphaToARGBRow = I210AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I210AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I410 with Alpha to preattenuated ARGB with matrix.
+LIBYUV_API
+int I410AlphaToARGBMatrix(const uint16_t* src_y,
+                          int src_stride_y,
+                          const uint16_t* src_u,
+                          int src_stride_u,
+                          const uint16_t* src_v,
+                          int src_stride_v,
+                          const uint16_t* src_a,
+                          int src_stride_a,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          const struct YuvConstants* yuvconstants,
+                          int width,
+                          int height,
+                          int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I410AlphaToARGBRow(src_y, src_u, src_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB with matrix.
+LIBYUV_API
+int I400ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I400ToARGBRow)(const uint8_t* y_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
       I400ToARGBRow_C;
   if (!src_y || !dst_argb || width <= 0 || height == 0) {
     return -1;
@@ -1012,23 +2644,35 @@ int I400ToARGB(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_I400TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I400ToARGBRow = I400ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I400ToARGBRow = I400ToARGBRow_MMI;
+#if defined(HAS_I400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I400ToARGBRow = I400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_LSX;
     }
   }
 #endif
 
   for (y = 0; y < height; ++y) {
-    I400ToARGBRow(src_y, dst_argb, width);
+    I400ToARGBRow(src_y, dst_argb, yuvconstants, width);
     dst_argb += dst_stride_argb;
     src_y += src_stride_y;
   }
   return 0;
 }
 
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8_t* src_y,
+               int src_stride_y,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  return I400ToARGBMatrix(src_y, src_stride_y, dst_argb, dst_stride_argb,
+                          &kYuvI601Constants, width, height);
+}
+
 // Convert J400 to ARGB.
 LIBYUV_API
 int J400ToARGB(const uint8_t* src_y,
@@ -1087,11 +2731,11 @@ int J400ToARGB(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_J400TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    J400ToARGBRow = J400ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      J400ToARGBRow = J400ToARGBRow_MMI;
+#if defined(HAS_J400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    J400ToARGBRow = J400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_LSX;
     }
   }
 #endif
@@ -1115,6 +2759,10 @@ static const uvec8 kShuffleMaskABGRToARGB = {
 static const uvec8 kShuffleMaskRGBAToARGB = {
     1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u};
 
+// Shuffle table for converting AR64 to AB64.
+static const uvec8 kShuffleMaskAR64ToAB64 = {
+    4u, 5u, 2u, 3u, 0u, 1u, 6u, 7u, 12u, 13u, 10u, 11u, 8u, 9u, 14u, 15u};
+
 // Convert BGRA to ARGB.
 LIBYUV_API
 int BGRAToARGB(const uint8_t* src_bgra,
@@ -1124,7 +2772,7 @@ int BGRAToARGB(const uint8_t* src_bgra,
                int width,
                int height) {
   return ARGBShuffle(src_bgra, src_stride_bgra, dst_argb, dst_stride_argb,
-                     (const uint8_t*)(&kShuffleMaskBGRAToARGB), width, height);
+                     (const uint8_t*)&kShuffleMaskBGRAToARGB, width, height);
 }
 
 // Convert ARGB to BGRA (same as BGRAToARGB).
@@ -1136,7 +2784,7 @@ int ARGBToBGRA(const uint8_t* src_bgra,
                int width,
                int height) {
   return ARGBShuffle(src_bgra, src_stride_bgra, dst_argb, dst_stride_argb,
-                     (const uint8_t*)(&kShuffleMaskBGRAToARGB), width, height);
+                     (const uint8_t*)&kShuffleMaskBGRAToARGB, width, height);
 }
 
 // Convert ABGR to ARGB.
@@ -1148,7 +2796,7 @@ int ABGRToARGB(const uint8_t* src_abgr,
                int width,
                int height) {
   return ARGBShuffle(src_abgr, src_stride_abgr, dst_argb, dst_stride_argb,
-                     (const uint8_t*)(&kShuffleMaskABGRToARGB), width, height);
+                     (const uint8_t*)&kShuffleMaskABGRToARGB, width, height);
 }
 
 // Convert ARGB to ABGR to (same as ABGRToARGB).
@@ -1160,7 +2808,7 @@ int ARGBToABGR(const uint8_t* src_abgr,
                int width,
                int height) {
   return ARGBShuffle(src_abgr, src_stride_abgr, dst_argb, dst_stride_argb,
-                     (const uint8_t*)(&kShuffleMaskABGRToARGB), width, height);
+                     (const uint8_t*)&kShuffleMaskABGRToARGB, width, height);
 }
 
 // Convert RGBA to ARGB.
@@ -1172,7 +2820,19 @@ int RGBAToARGB(const uint8_t* src_rgba,
                int width,
                int height) {
   return ARGBShuffle(src_rgba, src_stride_rgba, dst_argb, dst_stride_argb,
-                     (const uint8_t*)(&kShuffleMaskRGBAToARGB), width, height);
+                     (const uint8_t*)&kShuffleMaskRGBAToARGB, width, height);
+}
+
+// Convert AR64 To AB64.
+LIBYUV_API
+int AR64ToAB64(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height) {
+  return AR64Shuffle(src_ar64, src_stride_ar64, dst_ab64, dst_stride_ab64,
+                     (const uint8_t*)&kShuffleMaskAR64ToAB64, width, height);
 }
 
 // Convert RGB24 to ARGB.
@@ -1225,11 +2885,19 @@ int RGB24ToARGB(const uint8_t* src_rgb24,
     }
   }
 #endif
-#if defined(HAS_RGB24TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB24ToARGBRow = RGB24ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      RGB24ToARGBRow = RGB24ToARGBRow_MMI;
+#if defined(HAS_RGB24TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LASX;
     }
   }
 #endif
@@ -1292,11 +2960,19 @@ int RAWToARGB(const uint8_t* src_raw,
     }
   }
 #endif
-#if defined(HAS_RAWTOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RAWToARGBRow = RAWToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      RAWToARGBRow = RAWToARGBRow_MMI;
+#if defined(HAS_RAWTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToARGBRow = RAWToARGBRow_LASX;
     }
   }
 #endif
@@ -1309,6 +2985,57 @@ int RAWToARGB(const uint8_t* src_raw,
   return 0;
 }
 
+// Convert RAW to RGBA.
+LIBYUV_API
+int RAWToRGBA(const uint8_t* src_raw,
+              int src_stride_raw,
+              uint8_t* dst_rgba,
+              int dst_stride_rgba,
+              int width,
+              int height) {
+  int y;
+  void (*RAWToRGBARow)(const uint8_t* src_rgb, uint8_t* dst_rgba, int width) =
+      RAWToRGBARow_C;
+  if (!src_raw || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 && dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_rgba = 0;
+  }
+#if defined(HAS_RAWTORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToRGBARow = RAWToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGBARow = RAWToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToRGBARow = RAWToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToRGBARow = RAWToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToRGBARow(src_raw, dst_rgba, width);
+    src_raw += src_stride_raw;
+    dst_rgba += dst_stride_rgba;
+  }
+  return 0;
+}
+
 // Convert RGB565 to ARGB.
 LIBYUV_API
 int RGB565ToARGB(const uint8_t* src_rgb565,
@@ -1367,11 +3094,19 @@ int RGB565ToARGB(const uint8_t* src_rgb565,
     }
   }
 #endif
-#if defined(HAS_RGB565TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RGB565ToARGBRow = RGB565ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      RGB565ToARGBRow = RGB565ToARGBRow_MMI;
+#if defined(HAS_RGB565TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LASX;
     }
   }
 #endif
@@ -1442,11 +3177,19 @@ int ARGB1555ToARGB(const uint8_t* src_argb1555,
     }
   }
 #endif
-#if defined(HAS_ARGB1555TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGB1555ToARGBRow = ARGB1555ToARGBRow_MMI;
+#if defined(HAS_ARGB1555TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LASX;
     }
   }
 #endif
@@ -1517,11 +3260,19 @@ int ARGB4444ToARGB(const uint8_t* src_argb4444,
     }
   }
 #endif
-#if defined(HAS_ARGB4444TOARGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGB4444ToARGBRow = ARGB4444ToARGBRow_MMI;
+#if defined(HAS_ARGB4444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LASX;
     }
   }
 #endif
@@ -1630,16 +3381,135 @@ int AR30ToAB30(const uint8_t* src_ar30,
   return 0;
 }
 
-// Convert NV12 to ARGB with matrix
-static int NV12ToARGBMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_uv,
-                            int src_stride_uv,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert AR64 to ARGB.
+LIBYUV_API
+int AR64ToARGB(const uint16_t* src_ar64,
+               int src_stride_ar64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  int y;
+  void (*AR64ToARGBRow)(const uint16_t* src_ar64, uint8_t* dst_argb,
+                        int width) = AR64ToARGBRow_C;
+  if (!src_ar64 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar64 = src_ar64 + (height - 1) * src_stride_ar64;
+    src_stride_ar64 = -src_stride_ar64;
+  }
+  // Coalesce rows.
+  if (src_stride_ar64 == width * 4 && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar64 = dst_stride_argb = 0;
+  }
+#if defined(HAS_AR64TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      AR64ToARGBRow = AR64ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_AR64TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ToARGBRow = AR64ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_AR64TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AR64ToARGBRow = AR64ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ToARGBRow = AR64ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AR64ToARGBRow(src_ar64, dst_argb, width);
+    src_ar64 += src_stride_ar64;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert AB64 to ARGB.
+LIBYUV_API
+int AB64ToARGB(const uint16_t* src_ab64,
+               int src_stride_ab64,
+               uint8_t* dst_argb,
+               int dst_stride_argb,
+               int width,
+               int height) {
+  int y;
+  void (*AB64ToARGBRow)(const uint16_t* src_ar64, uint8_t* dst_argb,
+                        int width) = AB64ToARGBRow_C;
+  if (!src_ab64 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ab64 = src_ab64 + (height - 1) * src_stride_ab64;
+    src_stride_ab64 = -src_stride_ab64;
+  }
+  // Coalesce rows.
+  if (src_stride_ab64 == width * 4 && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ab64 = dst_stride_argb = 0;
+  }
+#if defined(HAS_AB64TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      AB64ToARGBRow = AB64ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_AB64TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      AB64ToARGBRow = AB64ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_AB64TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AB64ToARGBRow = AB64ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      AB64ToARGBRow = AB64ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AB64ToARGBRow(src_ab64, dst_argb, width);
+    src_ab64 += src_stride_ab64;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB with matrix.
+LIBYUV_API
+int NV12ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_uv,
+                     int src_stride_uv,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*NV12ToARGBRow)(
       const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
@@ -1685,6 +3555,22 @@ static int NV12ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV12TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV12ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
@@ -1697,16 +3583,17 @@ static int NV12ToARGBMatrix(const uint8_t* src_y,
   return 0;
 }
 
-// Convert NV21 to ARGB with matrix
-static int NV21ToARGBMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_vu,
-                            int src_stride_vu,
-                            uint8_t* dst_argb,
-                            int dst_stride_argb,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
+// Convert NV21 to ARGB with matrix.
+LIBYUV_API
+int NV21ToARGBMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_vu,
+                     int src_stride_vu,
+                     uint8_t* dst_argb,
+                     int dst_stride_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
   int y;
   void (*NV21ToARGBRow)(
       const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
@@ -1752,6 +3639,22 @@ static int NV21ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV21TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV21ToARGBRow(src_y, src_vu, dst_argb, yuvconstants, width);
@@ -1823,16 +3726,17 @@ int NV21ToABGR(const uint8_t* src_y,
 }
 
 // TODO(fbarchard): Consider SSSE3 2 step conversion.
-// Convert NV12 to RGB24 with matrix
-static int NV12ToRGB24Matrix(const uint8_t* src_y,
-                             int src_stride_y,
-                             const uint8_t* src_uv,
-                             int src_stride_uv,
-                             uint8_t* dst_rgb24,
-                             int dst_stride_rgb24,
-                             const struct YuvConstants* yuvconstants,
-                             int width,
-                             int height) {
+// Convert NV12 to RGB24 with matrix.
+LIBYUV_API
+int NV12ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_uv,
+                      int src_stride_uv,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
   int y;
   void (*NV12ToRGB24Row)(
       const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
@@ -1882,16 +3786,17 @@ static int NV12ToRGB24Matrix(const uint8_t* src_y,
   return 0;
 }
 
-// Convert NV21 to RGB24 with matrix
-static int NV21ToRGB24Matrix(const uint8_t* src_y,
-                             int src_stride_y,
-                             const uint8_t* src_vu,
-                             int src_stride_vu,
-                             uint8_t* dst_rgb24,
-                             int dst_stride_rgb24,
-                             const struct YuvConstants* yuvconstants,
-                             int width,
-                             int height) {
+// Convert NV21 to RGB24 with matrix.
+LIBYUV_API
+int NV21ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_vu,
+                      int src_stride_vu,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
   int y;
   void (*NV21ToRGB24Row)(
       const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
@@ -2028,6 +3933,14 @@ int NV21ToYUV24(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV21TOYUV24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToYUV24Row = NV21ToYUV24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToYUV24Row = NV21ToYUV24Row_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_NV21TOYUV24ROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     NV21ToYUV24Row = NV21ToYUV24Row_Any_AVX2;
@@ -2047,75 +3960,6 @@ int NV21ToYUV24(const uint8_t* src_y,
   return 0;
 }
 
-// Convert M420 to ARGB.
-LIBYUV_API
-int M420ToARGB(const uint8_t* src_m420,
-               int src_stride_m420,
-               uint8_t* dst_argb,
-               int dst_stride_argb,
-               int width,
-               int height) {
-  int y;
-  void (*NV12ToARGBRow)(
-      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
-      const struct YuvConstants* yuvconstants, int width) = NV12ToARGBRow_C;
-  if (!src_m420 || !dst_argb || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
-    dst_stride_argb = -dst_stride_argb;
-  }
-#if defined(HAS_NV12TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_NV12TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      NV12ToARGBRow = NV12ToARGBRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_NV12TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_NV12TOARGBROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    NV12ToARGBRow = NV12ToARGBRow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToARGBRow = NV12ToARGBRow_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height - 1; y += 2) {
-    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb,
-                  &kYuvI601Constants, width);
-    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
-                  dst_argb + dst_stride_argb, &kYuvI601Constants, width);
-    dst_argb += dst_stride_argb * 2;
-    src_m420 += src_stride_m420 * 3;
-  }
-  if (height & 1) {
-    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb,
-                  &kYuvI601Constants, width);
-  }
-  return 0;
-}
-
 // Convert YUY2 to ARGB.
 LIBYUV_API
 int YUY2ToARGB(const uint8_t* src_yuy2,
@@ -2175,6 +4019,14 @@ int YUY2ToARGB(const uint8_t* src_yuy2,
     }
   }
 #endif
+#if defined(HAS_YUY2TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     YUY2ToARGBRow(src_yuy2, dst_argb, &kYuvI601Constants, width);
     src_yuy2 += src_stride_yuy2;
@@ -2242,6 +4094,14 @@ int UYVYToARGB(const uint8_t* src_uyvy,
     }
   }
 #endif
+#if defined(HAS_UYVYTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     UYVYToARGBRow(src_uyvy, dst_argb, &kYuvI601Constants, width);
     src_uyvy += src_stride_uyvy;
@@ -2264,7 +4124,7 @@ static void WeavePixels(const uint8_t* src_u,
   }
 }
 
-// Convert Android420 to ARGB.
+// Convert Android420 to ARGB with matrix.
 LIBYUV_API
 int Android420ToARGBMatrix(const uint8_t* src_y,
                            int src_stride_y,
@@ -2365,6 +4225,3144 @@ int Android420ToABGR(const uint8_t* src_y,
                                 height);
 }
 
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I422ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGBARow = I422ToRGBARow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGBARow = I422ToRGBARow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGBARow = I422ToRGBARow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height) {
+  return I422ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_rgba, dst_stride_rgba,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height) {
+  return I422ToRGBAMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert NV12 to RGB565 with matrix.
+LIBYUV_API
+int NV12ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_uv,
+                       int src_stride_uv,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height) {
+  int y;
+  void (*NV12ToRGB565Row)(
+      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LSX;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, yuvconstants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return NV12ToRGB565Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                            dst_rgb565, dst_stride_rgb565, &kYuvI601Constants,
+                            width, height);
+}
+
+// Convert I422 to RGBA with matrix.
+LIBYUV_API
+int I420ToRGBAMatrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_rgba,
+                     int dst_stride_rgba,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGBARow = I422ToRGBARow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGBARow = I422ToRGBARow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGBARow = I422ToRGBARow_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_rgba,
+               int dst_stride_rgba,
+               int width,
+               int height) {
+  return I420ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_rgba, dst_stride_rgba,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_bgra,
+               int dst_stride_bgra,
+               int width,
+               int height) {
+  return I420ToRGBAMatrix(src_y, src_stride_y, src_v,
+                          src_stride_v,  // Swap U and V
+                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
+                          &kYvuI601Constants,  // Use Yvu matrix
+                          width, height);
+}
+
+// Convert I420 to RGB24 with matrix.
+LIBYUV_API
+int I420ToRGB24Matrix(const uint8_t* src_y,
+                      int src_stride_y,
+                      const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_rgb24,
+                      int dst_stride_rgb24,
+                      const struct YuvConstants* yuvconstants,
+                      int width,
+                      int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                         const uint8_t* v_buf, uint8_t* rgb_buf,
+                         const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB24Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB24Row = I422ToRGB24Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvI601Constants, width, height);
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuI601Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert J420 to RGB24.
+LIBYUV_API
+int J420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvJPEGConstants, width, height);
+}
+
+// Convert J420 to RAW.
+LIBYUV_API
+int J420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuJPEGConstants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert H420 to RGB24.
+LIBYUV_API
+int H420ToRGB24(const uint8_t* src_y,
+                int src_stride_y,
+                const uint8_t* src_u,
+                int src_stride_u,
+                const uint8_t* src_v,
+                int src_stride_v,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                           src_stride_v, dst_rgb24, dst_stride_rgb24,
+                           &kYuvH709Constants, width, height);
+}
+
+// Convert H420 to RAW.
+LIBYUV_API
+int H420ToRAW(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              uint8_t* dst_raw,
+              int dst_stride_raw,
+              int width,
+              int height) {
+  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
+                           src_stride_v,  // Swap U and V
+                           src_u, src_stride_u, dst_raw, dst_stride_raw,
+                           &kYvuH709Constants,  // Use Yvu matrix
+                           width, height);
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb1555,
+                   int dst_stride_argb1555,
+                   int width,
+                   int height) {
+  int y;
+  void (*I422ToARGB1555Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                            const uint8_t* v_buf, uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) = I422ToARGB1555Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb1555 || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, &kYuvI601Constants,
+                      width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8_t* src_y,
+                   int src_stride_y,
+                   const uint8_t* src_u,
+                   int src_stride_u,
+                   const uint8_t* src_v,
+                   int src_stride_v,
+                   uint8_t* dst_argb4444,
+                   int dst_stride_argb4444,
+                   int width,
+                   int height) {
+  int y;
+  void (*I422ToARGB4444Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                            const uint8_t* v_buf, uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) = I422ToARGB4444Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb4444 || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, &kYuvI601Constants,
+                      width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565 with specified color matrix.
+LIBYUV_API
+int I420ToRGB565Matrix(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const struct YuvConstants* yuvconstants,
+                       int width,
+                       int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                          const uint8_t* v_buf, uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB565Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB565Row = I422ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, yuvconstants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvI601Constants, width, height);
+}
+
+// Convert J420 to RGB565.
+LIBYUV_API
+int J420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvJPEGConstants, width, height);
+}
+
+// Convert H420 to RGB565.
+LIBYUV_API
+int H420ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                            src_stride_v, dst_rgb565, dst_stride_rgb565,
+                            &kYuvH709Constants, width, height);
+}
+
+// Convert I422 to RGB565.
+LIBYUV_API
+int I422ToRGB565(const uint8_t* src_y,
+                 int src_stride_y,
+                 const uint8_t* src_u,
+                 int src_stride_u,
+                 const uint8_t* src_v,
+                 int src_stride_v,
+                 uint8_t* dst_rgb565,
+                 int dst_stride_rgb565,
+                 int width,
+                 int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                          const uint8_t* v_buf, uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants, int width) =
+      I422ToRGB565Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToRGB565Row = I422ToRGB565Row_LASX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, &kYuvI601Constants, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8_t kDither565_4x4[16] = {
+    0, 4, 1, 5, 6, 2, 7, 3, 1, 5, 0, 4, 7, 3, 6, 2,
+};
+
+// Convert I420 to RGB565 with dithering.
+LIBYUV_API
+int I420ToRGB565Dither(const uint8_t* src_y,
+                       int src_stride_y,
+                       const uint8_t* src_u,
+                       int src_stride_u,
+                       const uint8_t* src_v,
+                       int src_stride_v,
+                       uint8_t* dst_rgb565,
+                       int dst_stride_rgb565,
+                       const uint8_t* dither4x4,
+                       int width,
+                       int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToARGBRow_C;
+  void (*ARGBToRGB565DitherRow)(const uint8_t* src_argb, uint8_t* dst_rgb,
+                                const uint32_t dither4, int width) =
+      ARGBToRGB565DitherRow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I422ToARGBRow = I422ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_LASX;
+    }
+  }
+#endif
+  {
+    // Allocate a row of argb.
+    align_buffer_64(row_argb, width * 4);
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row_argb, &kYuvI601Constants, width);
+      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
+                            *(const uint32_t*)(dither4x4 + ((y & 3) << 2)),
+                            width);
+      dst_rgb565 += dst_stride_rgb565;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row_argb);
+  }
+  return 0;
+}
+
+// Convert I420 to AR30 with matrix.
+LIBYUV_API
+int I420ToAR30Matrix(const uint8_t* src_y,
+                     int src_stride_y,
+                     const uint8_t* src_u,
+                     int src_stride_u,
+                     const uint8_t* src_v,
+                     int src_stride_v,
+                     uint8_t* dst_ar30,
+                     int dst_stride_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width,
+                     int height) {
+  int y;
+  void (*I422ToAR30Row)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I422ToAR30Row_C;
+
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+
+#if defined(HAS_I422TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToAR30Row = I422ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToAR30Row = I422ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToAR30Row = I422ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToAR30Row = I422ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to AR30.
+LIBYUV_API
+int I420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYuvI601Constants, width, height);
+}
+
+// Convert H420 to AR30.
+LIBYUV_API
+int H420ToAR30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ar30,
+               int dst_stride_ar30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                          src_stride_v, dst_ar30, dst_stride_ar30,
+                          &kYvuH709Constants, width, height);
+}
+
+// Convert I420 to AB30.
+LIBYUV_API
+int I420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuI601Constants, width, height);
+}
+
+// Convert H420 to AB30.
+LIBYUV_API
+int H420ToAB30(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_ab30,
+               int dst_stride_ab30,
+               int width,
+               int height) {
+  return I420ToAR30Matrix(src_y, src_stride_y, src_v, src_stride_v, src_u,
+                          src_stride_u, dst_ab30, dst_stride_ab30,
+                          &kYvuH709Constants, width, height);
+}
+
+static int I420ToARGBMatrixBilinear(const uint8_t* src_y,
+                                    int src_stride_y,
+                                    const uint8_t* src_u,
+                                    int src_stride_u,
+                                    const uint8_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToARGBRow = I444ToARGBRow_LASX;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSE2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 4);
+  uint8_t* temp_u_1 = row;
+  uint8_t* temp_u_2 = row + kRowSize;
+  uint8_t* temp_v_1 = row + kRowSize * 2;
+  uint8_t* temp_v_2 = row + kRowSize * 3;
+
+  Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+  Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+  I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_u, src_stride_u, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, src_stride_v, temp_v_1, kRowSize, width);
+    I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    I444ToARGBRow(src_y, temp_u_2, temp_v_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+    I444ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I422ToARGBMatrixLinear(const uint8_t* src_y,
+                                  int src_stride_y,
+                                  const uint8_t* src_u,
+                                  int src_stride_u,
+                                  const uint8_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                        const uint8_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I444ToARGBRow_C;
+  void (*ScaleRowUp)(const uint8_t* src_ptr, uint8_t* dst_ptr, int dst_width) =
+      ScaleRowUp2_Linear_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444ToARGBRow = I444ToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I444ToARGBRow = I444ToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+  uint8_t* temp_u = row;
+  uint8_t* temp_v = row + kRowSize;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_u, temp_u, width);
+    ScaleRowUp(src_v, temp_v, width);
+    I444ToARGBRow(src_y, temp_u, temp_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010ToAR30MatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_u,
+                                    int src_stride_u,
+                                    const uint16_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_ar30,
+                                    int dst_stride_ar30,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + kRowSize;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + kRowSize * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + kRowSize * 3;
+
+  Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+  Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+  I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+  dst_ar30 += dst_stride_ar30;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_u, src_stride_u, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, src_stride_v, temp_v_1, kRowSize, width);
+    I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    I410ToAR30Row(src_y, temp_u_2, temp_v_2, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+    I410ToAR30Row(src_y, temp_u_1, temp_v_1, dst_ar30, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+
+  return 0;
+}
+
+static int I210ToAR30MatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_u,
+                                  int src_stride_u,
+                                  const uint16_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_ar30,
+                                  int dst_stride_ar30,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I410ToAR30Row)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToAR30Row_C;
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_I410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToAR30Row = I410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToAR30Row = I410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToAR30Row = I410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToAR30Row = I410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + kRowSize;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_u, temp_u, width);
+    ScaleRowUp(src_v, temp_v, width);
+    I410ToAR30Row(src_y, temp_u, temp_v, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010ToARGBMatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_u,
+                                    int src_stride_u,
+                                    const uint16_t* src_v,
+                                    int src_stride_v,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + kRowSize;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + kRowSize * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + kRowSize * 3;
+
+  Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+  Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+  I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_u, src_stride_u, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, src_stride_v, temp_v_1, kRowSize, width);
+    I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    I410ToARGBRow(src_y, temp_u_2, temp_v_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+    I410ToARGBRow(src_y, temp_u_1, temp_v_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I210ToARGBMatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_u,
+                                  int src_stride_u,
+                                  const uint16_t* src_v,
+                                  int src_stride_v,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*I410ToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                        const uint16_t* v_buf, uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants, int width) =
+      I410ToARGBRow_C;
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410ToARGBRow = I410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410ToARGBRow = I410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410ToARGBRow = I410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410ToARGBRow = I410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + kRowSize;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_u, temp_u, width);
+    ScaleRowUp(src_v, temp_v, width);
+    I410ToARGBRow(src_y, temp_u, temp_v, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I420AlphaToARGBMatrixBilinear(
+    const uint8_t* src_y,
+    int src_stride_y,
+    const uint8_t* src_u,
+    int src_stride_u,
+    const uint8_t* src_v,
+    int src_stride_v,
+    const uint8_t* src_a,
+    int src_stride_a,
+    uint8_t* dst_argb,
+    int dst_stride_argb,
+    const struct YuvConstants* yuvconstants,
+    int width,
+    int height,
+    int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_Any_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSE2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 4);
+  uint8_t* temp_u_1 = row;
+  uint8_t* temp_u_2 = row + kRowSize;
+  uint8_t* temp_v_1 = row + kRowSize * 2;
+  uint8_t* temp_v_2 = row + kRowSize * 3;
+
+  Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+  Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+  I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                     width);
+  if (attenuate) {
+    ARGBAttenuateRow(dst_argb, dst_argb, width);
+  }
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+  src_a += src_stride_a;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_u, src_stride_u, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, src_stride_v, temp_v_1, kRowSize, width);
+    I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    I444AlphaToARGBRow(src_y, temp_u_2, temp_v_2, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    src_a += src_stride_a;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+    I444AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I422AlphaToARGBMatrixLinear(const uint8_t* src_y,
+                                       int src_stride_y,
+                                       const uint8_t* src_u,
+                                       int src_stride_u,
+                                       const uint8_t* src_v,
+                                       int src_stride_v,
+                                       const uint8_t* src_a,
+                                       int src_stride_a,
+                                       uint8_t* dst_argb,
+                                       int dst_stride_argb,
+                                       const struct YuvConstants* yuvconstants,
+                                       int width,
+                                       int height,
+                                       int attenuate) {
+  int y;
+  void (*I444AlphaToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
+                             const uint8_t* v_buf, const uint8_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I444AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*ScaleRowUp)(const uint8_t* src_ptr, uint8_t* dst_ptr, int dst_width) =
+      ScaleRowUp2_Linear_Any_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_I444ALPHATOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I444AlphaToARGBRow = I444AlphaToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      I444AlphaToARGBRow = I444AlphaToARGBRow_LASX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+  uint8_t* temp_u = row;
+  uint8_t* temp_v = row + kRowSize;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_u, temp_u, width);
+    ScaleRowUp(src_v, temp_v, width);
+    I444AlphaToARGBRow(src_y, temp_u, temp_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I010AlphaToARGBMatrixBilinear(
+    const uint16_t* src_y,
+    int src_stride_y,
+    const uint16_t* src_u,
+    int src_stride_u,
+    const uint16_t* src_v,
+    int src_stride_v,
+    const uint16_t* src_a,
+    int src_stride_a,
+    uint8_t* dst_argb,
+    int dst_stride_argb,
+    const struct YuvConstants* yuvconstants,
+    int width,
+    int height,
+    int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_AVX2;
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 4 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 4 * sizeof(uint16_t));
+  uint16_t* temp_u_1 = (uint16_t*)(row);
+  uint16_t* temp_u_2 = (uint16_t*)(row) + kRowSize;
+  uint16_t* temp_v_1 = (uint16_t*)(row) + kRowSize * 2;
+  uint16_t* temp_v_2 = (uint16_t*)(row) + kRowSize * 3;
+
+  Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+  Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+  I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                     width);
+  if (attenuate) {
+    ARGBAttenuateRow(dst_argb, dst_argb, width);
+  }
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+  src_a += src_stride_a;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_u, src_stride_u, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, src_stride_v, temp_v_1, kRowSize, width);
+    I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    I410AlphaToARGBRow(src_y, temp_u_2, temp_v_2, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_a += src_stride_a;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_u, 0, temp_u_1, kRowSize, width);
+    Scale2RowUp(src_v, 0, temp_v_1, kRowSize, width);
+    I410AlphaToARGBRow(src_y, temp_u_1, temp_v_1, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int I210AlphaToARGBMatrixLinear(const uint16_t* src_y,
+                                       int src_stride_y,
+                                       const uint16_t* src_u,
+                                       int src_stride_u,
+                                       const uint16_t* src_v,
+                                       int src_stride_v,
+                                       const uint16_t* src_a,
+                                       int src_stride_a,
+                                       uint8_t* dst_argb,
+                                       int dst_stride_argb,
+                                       const struct YuvConstants* yuvconstants,
+                                       int width,
+                                       int height,
+                                       int attenuate) {
+  int y;
+  void (*I410AlphaToARGBRow)(const uint16_t* y_buf, const uint16_t* u_buf,
+                             const uint16_t* v_buf, const uint16_t* a_buf,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) = I410AlphaToARGBRow_C;
+  void (*ARGBAttenuateRow)(const uint8_t* src_argb, uint8_t* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  if (!src_y || !src_u || !src_v || !src_a || !dst_argb || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I410ALPHATOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I410AlphaToARGBRow = I410AlphaToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I410AlphaToARGBRow = I410AlphaToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_MSA;
+    }
+  }
+#endif
+
+#if defined(HAS_SCALEROWUP2_LINEAR_12_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+#if defined(HAS_SCALEROWUP2_LINEAR_12_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2 * sizeof(uint16_t));
+  uint16_t* temp_u = (uint16_t*)(row);
+  uint16_t* temp_v = (uint16_t*)(row) + kRowSize;
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_u, temp_u, width);
+    ScaleRowUp(src_v, temp_v, width);
+    I410AlphaToARGBRow(src_y, temp_u, temp_v, src_a, dst_argb, yuvconstants,
+                       width);
+    if (attenuate) {
+      ARGBAttenuateRow(dst_argb, dst_argb, width);
+    }
+    dst_argb += dst_stride_argb;
+    src_a += src_stride_a;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P010ToARGBMatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_uv,
+                                    int src_stride_uv,
+                                    uint8_t* dst_argb,
+                                    int dst_stride_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*P410ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToARGBRow_C;
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_16_Any_C;
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToARGBRow = P410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToARGBRow = P410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToARGBRow = P410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToARGBRow = P410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (2 * width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2 * sizeof(uint16_t));
+  uint16_t* temp_uv_1 = (uint16_t*)(row);
+  uint16_t* temp_uv_2 = (uint16_t*)(row) + kRowSize;
+
+  Scale2RowUp(src_uv, 0, temp_uv_1, kRowSize, width);
+  P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+  dst_argb += dst_stride_argb;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_uv, src_stride_uv, temp_uv_1, kRowSize, width);
+    P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    P410ToARGBRow(src_y, temp_uv_2, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_uv, 0, temp_uv_1, kRowSize, width);
+    P410ToARGBRow(src_y, temp_uv_1, dst_argb, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P210ToARGBMatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_uv,
+                                  int src_stride_uv,
+                                  uint8_t* dst_argb,
+                                  int dst_stride_argb,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*P410ToARGBRow)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToARGBRow_C;
+  void (*ScaleRowUp)(const uint16_t* src_uv, uint16_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_16_Any_C;
+  if (!src_y || !src_uv || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_P410TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToARGBRow = P410ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToARGBRow = P410ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToARGBRow = P410ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToARGBRow = P410ToARGBRow_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  const int kRowSize = (2 * width + 31) & ~31;
+  align_buffer_64(row, kRowSize * sizeof(uint16_t));
+  uint16_t* temp_uv = (uint16_t*)(row);
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_uv, temp_uv, width);
+    P410ToARGBRow(src_y, temp_uv, dst_argb, yuvconstants, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P010ToAR30MatrixBilinear(const uint16_t* src_y,
+                                    int src_stride_y,
+                                    const uint16_t* src_uv,
+                                    int src_stride_uv,
+                                    uint8_t* dst_ar30,
+                                    int dst_stride_ar30,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width,
+                                    int height) {
+  int y;
+  void (*P410ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToAR30Row_C;
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_16_Any_C;
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToAR30Row = P410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToAR30Row = P410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToAR30Row = P410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToAR30Row = P410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  // alloc 2 lines temp
+  const int kRowSize = (2 * width + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2 * sizeof(uint16_t));
+  uint16_t* temp_uv_1 = (uint16_t*)(row);
+  uint16_t* temp_uv_2 = (uint16_t*)(row) + kRowSize;
+
+  Scale2RowUp(src_uv, 0, temp_uv_1, kRowSize, width);
+  P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+  dst_ar30 += dst_stride_ar30;
+  src_y += src_stride_y;
+
+  for (y = 0; y < height - 2; y += 2) {
+    Scale2RowUp(src_uv, src_stride_uv, temp_uv_1, kRowSize, width);
+    P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    P410ToAR30Row(src_y, temp_uv_2, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  if (!(height & 1)) {
+    Scale2RowUp(src_uv, 0, temp_uv_1, kRowSize, width);
+    P410ToAR30Row(src_y, temp_uv_1, dst_ar30, yuvconstants, width);
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+static int P210ToAR30MatrixLinear(const uint16_t* src_y,
+                                  int src_stride_y,
+                                  const uint16_t* src_uv,
+                                  int src_stride_uv,
+                                  uint8_t* dst_ar30,
+                                  int dst_stride_ar30,
+                                  const struct YuvConstants* yuvconstants,
+                                  int width,
+                                  int height) {
+  int y;
+  void (*P410ToAR30Row)(
+      const uint16_t* y_buf, const uint16_t* uv_buf, uint8_t* rgb_buf,
+      const struct YuvConstants* yuvconstants, int width) = P410ToAR30Row_C;
+  void (*ScaleRowUp)(const uint16_t* src_uv, uint16_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_16_Any_C;
+  if (!src_y || !src_uv || !dst_ar30 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+#if defined(HAS_P410TOAR30ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    P410ToAR30Row = P410ToAR30Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      P410ToAR30Row = P410ToAR30Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_P410TOAR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    P410ToAR30Row = P410ToAR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      P410ToAR30Row = P410ToAR30Row_AVX2;
+    }
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  const int kRowSize = (2 * width + 31) & ~31;
+  align_buffer_64(row, kRowSize * sizeof(uint16_t));
+  uint16_t* temp_uv = (uint16_t*)(row);
+
+  for (y = 0; y < height; ++y) {
+    ScaleRowUp(src_uv, temp_uv, width);
+    P410ToAR30Row(src_y, temp_uv, dst_ar30, yuvconstants, width);
+    dst_ar30 += dst_stride_ar30;
+    src_y += src_stride_y;
+    src_uv += src_stride_uv;
+  }
+
+  free_aligned_buffer_64(row);
+  return 0;
+}
+
+LIBYUV_API
+int I420ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I420ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I420ToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+    case kFilterLinear:
+      // Actually we can do this, but probably there's no usage.
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I422ToARGBMatrixFilter(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I422ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I422ToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_ar30, dst_stride_ar30,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010ToAR30MatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_ar30, dst_stride_ar30, yuvconstants, width, height);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_ar30, dst_stride_ar30,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210ToAR30MatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_ar30, dst_stride_ar30, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010ToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_u,
+                           int src_stride_u,
+                           const uint16_t* src_v,
+                           int src_stride_v,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210ToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                              src_stride_v, dst_argb, dst_stride_argb,
+                              yuvconstants, width, height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210ToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v,
+          dst_argb, dst_stride_argb, yuvconstants, width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I420AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I420AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I420AlphaToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I422AlphaToARGBMatrixFilter(const uint8_t* src_y,
+                                int src_stride_y,
+                                const uint8_t* src_u,
+                                int src_stride_u,
+                                const uint8_t* src_v,
+                                int src_stride_v,
+                                const uint8_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I422AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I422AlphaToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I010AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I010AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+      return I010AlphaToARGBMatrixBilinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int I210AlphaToARGBMatrixFilter(const uint16_t* src_y,
+                                int src_stride_y,
+                                const uint16_t* src_u,
+                                int src_stride_u,
+                                const uint16_t* src_v,
+                                int src_stride_v,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint8_t* dst_argb,
+                                int dst_stride_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width,
+                                int height,
+                                int attenuate,
+                                enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return I210AlphaToARGBMatrix(src_y, src_stride_y, src_u, src_stride_u,
+                                   src_v, src_stride_v, src_a, src_stride_a,
+                                   dst_argb, dst_stride_argb, yuvconstants,
+                                   width, height, attenuate);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return I210AlphaToARGBMatrixLinear(
+          src_y, src_stride_y, src_u, src_stride_u, src_v, src_stride_v, src_a,
+          src_stride_a, dst_argb, dst_stride_argb, yuvconstants, width, height,
+          attenuate);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P010ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P010ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_argb, dst_stride_argb, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return P010ToARGBMatrixBilinear(src_y, src_stride_y, src_uv,
+                                      src_stride_uv, dst_argb, dst_stride_argb,
+                                      yuvconstants, width, height);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P210ToARGBMatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_argb,
+                           int dst_stride_argb,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P210ToARGBMatrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_argb, dst_stride_argb, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return P210ToARGBMatrixLinear(src_y, src_stride_y, src_uv, src_stride_uv,
+                                    dst_argb, dst_stride_argb, yuvconstants,
+                                    width, height);
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P010ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P010ToAR30Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_ar30, dst_stride_ar30, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+      return P010ToAR30MatrixBilinear(src_y, src_stride_y, src_uv,
+                                      src_stride_uv, dst_ar30, dst_stride_ar30,
+                                      yuvconstants, width, height);
+    case kFilterLinear:
+      return -1;
+  }
+
+  return -1;
+}
+
+LIBYUV_API
+int P210ToAR30MatrixFilter(const uint16_t* src_y,
+                           int src_stride_y,
+                           const uint16_t* src_uv,
+                           int src_stride_uv,
+                           uint8_t* dst_ar30,
+                           int dst_stride_ar30,
+                           const struct YuvConstants* yuvconstants,
+                           int width,
+                           int height,
+                           enum FilterMode filter) {
+  switch (filter) {
+    case kFilterNone:
+      return P210ToAR30Matrix(src_y, src_stride_y, src_uv, src_stride_uv,
+                              dst_ar30, dst_stride_ar30, yuvconstants, width,
+                              height);
+    case kFilterBilinear:
+    case kFilterBox:
+    case kFilterLinear:
+      return P210ToAR30MatrixLinear(src_y, src_stride_y, src_uv, src_stride_uv,
+                                    dst_ar30, dst_stride_ar30, yuvconstants,
+                                    width, height);
+  }
+
+  return -1;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/convert_from.cc b/files/source/convert_from.cc
index 60140cb4..8bd07e4c 100644
--- a/files/source/convert_from.cc
+++ b/files/source/convert_from.cc
@@ -30,6 +30,8 @@ static __inline int Abs(int v) {
 }
 
 // I420 To any I4xx YUV format with mirroring.
+// TODO(fbarchard): Consider kFilterNone for Y, or CopyPlane
+
 static int I420ToI4xx(const uint8_t* src_y,
                       int src_stride_y,
                       const uint8_t* src_u,
@@ -83,7 +85,8 @@ int I420ToI010(const uint8_t* src_y,
                int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -109,6 +112,51 @@ int I420ToI010(const uint8_t* src_y,
   return 0;
 }
 
+// Convert 8 bit YUV to 12 bit.
+LIBYUV_API
+int I420ToI012(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  // Convert Y plane.
+  Convert8To16Plane(src_y, src_stride_y, dst_y, dst_stride_y, 4096, width,
+                    height);
+  // Convert UV planes.
+  Convert8To16Plane(src_u, src_stride_u, dst_u, dst_stride_u, 4096, halfwidth,
+                    halfheight);
+  Convert8To16Plane(src_v, src_stride_v, dst_v, dst_stride_v, 4096, halfwidth,
+                    halfheight);
+  return 0;
+}
+
 // 420 chroma is 1/2 width, 1/2 height
 // 422 chroma is 1/2 width, 1x height
 LIBYUV_API
@@ -159,6 +207,102 @@ int I420ToI444(const uint8_t* src_y,
                     dst_uv_height);
 }
 
+// 420 chroma to 444 chroma, 10/12 bit version
+LIBYUV_API
+int I010ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, SUBSAMPLE(width, 1, 1),
+                SUBSAMPLE(height, 1, 1), dst_u, dst_stride_u, Abs(width),
+                Abs(height), kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, SUBSAMPLE(width, 1, 1),
+                SUBSAMPLE(height, 1, 1), dst_v, dst_stride_v, Abs(width),
+                Abs(height), kFilterBilinear);
+  return 0;
+}
+
+// 422 chroma to 444 chroma, 10/12 bit version
+LIBYUV_API
+int I210ToI410(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane_12(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+                  Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane_12(src_u, src_stride_u, SUBSAMPLE(width, 1, 1), height, dst_u,
+                dst_stride_u, Abs(width), Abs(height), kFilterBilinear);
+  ScalePlane_12(src_v, src_stride_v, SUBSAMPLE(width, 1, 1), height, dst_v,
+                dst_stride_v, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I422ToI444(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height) {
+  if (width == 0 || height == 0) {
+    return -1;
+  }
+
+  if (dst_y) {
+    ScalePlane(src_y, src_stride_y, width, height, dst_y, dst_stride_y,
+               Abs(width), Abs(height), kFilterBilinear);
+  }
+  ScalePlane(src_u, src_stride_u, SUBSAMPLE(width, 1, 1), height, dst_u,
+             dst_stride_u, Abs(width), Abs(height), kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, SUBSAMPLE(width, 1, 1), height, dst_v,
+             dst_stride_v, Abs(width), Abs(height), kFilterBilinear);
+  return 0;
+}
+
 // Copy to I400. Source can be I420,422,444,400,NV12,NV21
 LIBYUV_API
 int I400Copy(const uint8_t* src_y,
@@ -302,11 +446,11 @@ int I420ToYUY2(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_I422TOYUY2ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I422ToYUY2Row = I422ToYUY2Row_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToYUY2Row = I422ToYUY2Row_MMI;
+#if defined(HAS_I422TOYUY2ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_LASX;
     }
   }
 #endif
@@ -389,11 +533,11 @@ int I422ToUYVY(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_I422TOUYVYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I422ToUYVYRow = I422ToUYVYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToUYVYRow = I422ToUYVYRow_MMI;
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
     }
   }
 #endif
@@ -464,11 +608,11 @@ int I420ToUYVY(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_I422TOUYVYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I422ToUYVYRow = I422ToUYVYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToUYVYRow = I422ToUYVYRow_MMI;
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
     }
   }
 #endif
@@ -488,7 +632,6 @@ int I420ToUYVY(const uint8_t* src_y,
   return 0;
 }
 
-// TODO(fbarchard): test negative height for invert.
 LIBYUV_API
 int I420ToNV12(const uint8_t* src_y,
                int src_stride_y,
@@ -502,12 +645,22 @@ int I420ToNV12(const uint8_t* src_y,
                int dst_stride_uv,
                int width,
                int height) {
-  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv || width <= 0 ||
-      height == 0) {
+  int halfwidth = (width + 1) / 2;
+  int halfheight = (height + 1) / 2;
+  if (!src_y || !src_u || !src_v || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
-  int halfwidth = (width + 1) / 2;
-  int halfheight = height > 0 ? (height + 1) / 2 : (height - 1) / 2;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
   if (dst_y) {
     CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
   }
@@ -534,807 +687,6 @@ int I420ToNV21(const uint8_t* src_y,
                     width, height);
 }
 
-// Convert I422 to RGBA with matrix
-static int I420ToRGBAMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_rgba,
-                            int dst_stride_rgba,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
-  int y;
-  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
-                        const uint8_t* v_buf, uint8_t* rgb_buf,
-                        const struct YuvConstants* yuvconstants, int width) =
-      I422ToRGBARow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
-    dst_stride_rgba = -dst_stride_rgba;
-  }
-#if defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGBARow = I422ToRGBARow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGBARow = I422ToRGBARow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToRGBARow = I422ToRGBARow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
-    dst_rgba += dst_stride_rgba;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGBA.
-LIBYUV_API
-int I420ToRGBA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_rgba,
-               int dst_stride_rgba,
-               int width,
-               int height) {
-  return I420ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                          src_stride_v, dst_rgba, dst_stride_rgba,
-                          &kYuvI601Constants, width, height);
-}
-
-// Convert I420 to BGRA.
-LIBYUV_API
-int I420ToBGRA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_bgra,
-               int dst_stride_bgra,
-               int width,
-               int height) {
-  return I420ToRGBAMatrix(src_y, src_stride_y, src_v,
-                          src_stride_v,  // Swap U and V
-                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
-                          &kYvuI601Constants,  // Use Yvu matrix
-                          width, height);
-}
-
-// Convert I420 to RGB24 with matrix
-static int I420ToRGB24Matrix(const uint8_t* src_y,
-                             int src_stride_y,
-                             const uint8_t* src_u,
-                             int src_stride_u,
-                             const uint8_t* src_v,
-                             int src_stride_v,
-                             uint8_t* dst_rgb24,
-                             int dst_stride_rgb24,
-                             const struct YuvConstants* yuvconstants,
-                             int width,
-                             int height) {
-  int y;
-  void (*I422ToRGB24Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                         const uint8_t* v_buf, uint8_t* rgb_buf,
-                         const struct YuvConstants* yuvconstants, int width) =
-      I422ToRGB24Row_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb24 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
-    dst_stride_rgb24 = -dst_stride_rgb24;
-  }
-#if defined(HAS_I422TORGB24ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB24ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      I422ToRGB24Row = I422ToRGB24Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB24ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB24Row = I422ToRGB24Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB24ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToRGB24Row = I422ToRGB24Row_Any_MSA;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB24Row = I422ToRGB24Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, yuvconstants, width);
-    dst_rgb24 += dst_stride_rgb24;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGB24.
-LIBYUV_API
-int I420ToRGB24(const uint8_t* src_y,
-                int src_stride_y,
-                const uint8_t* src_u,
-                int src_stride_u,
-                const uint8_t* src_v,
-                int src_stride_v,
-                uint8_t* dst_rgb24,
-                int dst_stride_rgb24,
-                int width,
-                int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                           src_stride_v, dst_rgb24, dst_stride_rgb24,
-                           &kYuvI601Constants, width, height);
-}
-
-// Convert I420 to RAW.
-LIBYUV_API
-int I420ToRAW(const uint8_t* src_y,
-              int src_stride_y,
-              const uint8_t* src_u,
-              int src_stride_u,
-              const uint8_t* src_v,
-              int src_stride_v,
-              uint8_t* dst_raw,
-              int dst_stride_raw,
-              int width,
-              int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
-                           src_stride_v,  // Swap U and V
-                           src_u, src_stride_u, dst_raw, dst_stride_raw,
-                           &kYvuI601Constants,  // Use Yvu matrix
-                           width, height);
-}
-
-// Convert H420 to RGB24.
-LIBYUV_API
-int H420ToRGB24(const uint8_t* src_y,
-                int src_stride_y,
-                const uint8_t* src_u,
-                int src_stride_u,
-                const uint8_t* src_v,
-                int src_stride_v,
-                uint8_t* dst_rgb24,
-                int dst_stride_rgb24,
-                int width,
-                int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                           src_stride_v, dst_rgb24, dst_stride_rgb24,
-                           &kYuvH709Constants, width, height);
-}
-
-// Convert H420 to RAW.
-LIBYUV_API
-int H420ToRAW(const uint8_t* src_y,
-              int src_stride_y,
-              const uint8_t* src_u,
-              int src_stride_u,
-              const uint8_t* src_v,
-              int src_stride_v,
-              uint8_t* dst_raw,
-              int dst_stride_raw,
-              int width,
-              int height) {
-  return I420ToRGB24Matrix(src_y, src_stride_y, src_v,
-                           src_stride_v,  // Swap U and V
-                           src_u, src_stride_u, dst_raw, dst_stride_raw,
-                           &kYvuH709Constants,  // Use Yvu matrix
-                           width, height);
-}
-
-// Convert I420 to ARGB1555.
-LIBYUV_API
-int I420ToARGB1555(const uint8_t* src_y,
-                   int src_stride_y,
-                   const uint8_t* src_u,
-                   int src_stride_u,
-                   const uint8_t* src_v,
-                   int src_stride_v,
-                   uint8_t* dst_argb1555,
-                   int dst_stride_argb1555,
-                   int width,
-                   int height) {
-  int y;
-  void (*I422ToARGB1555Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                            const uint8_t* v_buf, uint8_t* rgb_buf,
-                            const struct YuvConstants* yuvconstants,
-                            int width) = I422ToARGB1555Row_C;
-  if (!src_y || !src_u || !src_v || !dst_argb1555 || width <= 0 ||
-      height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
-    dst_stride_argb1555 = -dst_stride_argb1555;
-  }
-#if defined(HAS_I422TOARGB1555ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB1555ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB1555ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB1555ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToARGB1555Row = I422ToARGB1555Row_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB1555Row = I422ToARGB1555Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, &kYuvI601Constants,
-                      width);
-    dst_argb1555 += dst_stride_argb1555;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to ARGB4444.
-LIBYUV_API
-int I420ToARGB4444(const uint8_t* src_y,
-                   int src_stride_y,
-                   const uint8_t* src_u,
-                   int src_stride_u,
-                   const uint8_t* src_v,
-                   int src_stride_v,
-                   uint8_t* dst_argb4444,
-                   int dst_stride_argb4444,
-                   int width,
-                   int height) {
-  int y;
-  void (*I422ToARGB4444Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                            const uint8_t* v_buf, uint8_t* rgb_buf,
-                            const struct YuvConstants* yuvconstants,
-                            int width) = I422ToARGB4444Row_C;
-  if (!src_y || !src_u || !src_v || !dst_argb4444 || width <= 0 ||
-      height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
-    dst_stride_argb4444 = -dst_stride_argb4444;
-  }
-#if defined(HAS_I422TOARGB4444ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB4444ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB4444ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGB4444ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToARGB4444Row = I422ToARGB4444Row_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGB4444Row = I422ToARGB4444Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, &kYuvI601Constants,
-                      width);
-    dst_argb4444 += dst_stride_argb4444;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGB565 with specified color matrix.
-LIBYUV_API
-int I420ToRGB565Matrix(const uint8_t* src_y,
-                       int src_stride_y,
-                       const uint8_t* src_u,
-                       int src_stride_u,
-                       const uint8_t* src_v,
-                       int src_stride_v,
-                       uint8_t* dst_rgb565,
-                       int dst_stride_rgb565,
-                       const struct YuvConstants* yuvconstants,
-                       int width,
-                       int height) {
-  int y;
-  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                          const uint8_t* v_buf, uint8_t* rgb_buf,
-                          const struct YuvConstants* yuvconstants, int width) =
-      I422ToRGB565Row_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-#if defined(HAS_I422TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB565Row = I422ToRGB565Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, yuvconstants, width);
-    dst_rgb565 += dst_stride_rgb565;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to RGB565.
-LIBYUV_API
-int I420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height) {
-  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                            src_stride_v, dst_rgb565, dst_stride_rgb565,
-                            &kYuvI601Constants, width, height);
-}
-
-// Convert J420 to RGB565.
-LIBYUV_API
-int J420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height) {
-  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                            src_stride_v, dst_rgb565, dst_stride_rgb565,
-                            &kYuvJPEGConstants, width, height);
-}
-
-// Convert H420 to RGB565.
-LIBYUV_API
-int H420ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height) {
-  return I420ToRGB565Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                            src_stride_v, dst_rgb565, dst_stride_rgb565,
-                            &kYuvH709Constants, width, height);
-}
-
-// Convert I422 to RGB565.
-LIBYUV_API
-int I422ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_u,
-                 int src_stride_u,
-                 const uint8_t* src_v,
-                 int src_stride_v,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height) {
-  int y;
-  void (*I422ToRGB565Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                          const uint8_t* v_buf, uint8_t* rgb_buf,
-                          const struct YuvConstants* yuvconstants, int width) =
-      I422ToRGB565Row_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-#if defined(HAS_I422TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGB565Row = I422ToRGB565Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGB565ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToRGB565Row = I422ToRGB565Row_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGB565Row = I422ToRGB565Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, &kYuvI601Constants, width);
-    dst_rgb565 += dst_stride_rgb565;
-    src_y += src_stride_y;
-    src_u += src_stride_u;
-    src_v += src_stride_v;
-  }
-  return 0;
-}
-
-// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
-static const uint8_t kDither565_4x4[16] = {
-    0, 4, 1, 5, 6, 2, 7, 3, 1, 5, 0, 4, 7, 3, 6, 2,
-};
-
-// Convert I420 to RGB565 with dithering.
-LIBYUV_API
-int I420ToRGB565Dither(const uint8_t* src_y,
-                       int src_stride_y,
-                       const uint8_t* src_u,
-                       int src_stride_u,
-                       const uint8_t* src_v,
-                       int src_stride_v,
-                       uint8_t* dst_rgb565,
-                       int dst_stride_rgb565,
-                       const uint8_t* dither4x4,
-                       int width,
-                       int height) {
-  int y;
-  void (*I422ToARGBRow)(const uint8_t* y_buf, const uint8_t* u_buf,
-                        const uint8_t* v_buf, uint8_t* rgb_buf,
-                        const struct YuvConstants* yuvconstants, int width) =
-      I422ToARGBRow_C;
-  void (*ARGBToRGB565DitherRow)(const uint8_t* src_argb, uint8_t* dst_rgb,
-                                const uint32_t dither4, int width) =
-      ARGBToRGB565DitherRow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgb565 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-  if (!dither4x4) {
-    dither4x4 = kDither565_4x4;
-  }
-#if defined(HAS_I422TOARGBROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToARGBRow = I422ToARGBRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToARGBRow = I422ToARGBRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TOARGBROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToARGBRow = I422ToARGBRow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToARGBRow = I422ToARGBRow_MSA;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTORGB565DITHERROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MSA;
-    }
-  }
-#endif
-  {
-    // Allocate a row of argb.
-    align_buffer_64(row_argb, width * 4);
-    for (y = 0; y < height; ++y) {
-      I422ToARGBRow(src_y, src_u, src_v, row_argb, &kYuvI601Constants, width);
-      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
-                            *(const uint32_t*)(dither4x4 + ((y & 3) << 2)),
-                            width);
-      dst_rgb565 += dst_stride_rgb565;
-      src_y += src_stride_y;
-      if (y & 1) {
-        src_u += src_stride_u;
-        src_v += src_stride_v;
-      }
-    }
-    free_aligned_buffer_64(row_argb);
-  }
-  return 0;
-}
-
-// Convert I420 to AR30 with matrix
-static int I420ToAR30Matrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_ar30,
-                            int dst_stride_ar30,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
-  int y;
-  void (*I422ToAR30Row)(const uint8_t* y_buf, const uint8_t* u_buf,
-                        const uint8_t* v_buf, uint8_t* rgb_buf,
-                        const struct YuvConstants* yuvconstants, int width) =
-      I422ToAR30Row_C;
-
-  if (!src_y || !src_u || !src_v || !dst_ar30 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
-    dst_stride_ar30 = -dst_stride_ar30;
-  }
-
-#if defined(HAS_I422TOAR30ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToAR30Row = I422ToAR30Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToAR30Row = I422ToAR30Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TOAR30ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToAR30Row = I422ToAR30Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToAR30Row = I422ToAR30Row_AVX2;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToAR30Row(src_y, src_u, src_v, dst_ar30, yuvconstants, width);
-    dst_ar30 += dst_stride_ar30;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_u += src_stride_u;
-      src_v += src_stride_v;
-    }
-  }
-  return 0;
-}
-
-// Convert I420 to AR30.
-LIBYUV_API
-int I420ToAR30(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_ar30,
-               int dst_stride_ar30,
-               int width,
-               int height) {
-  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                          src_stride_v, dst_ar30, dst_stride_ar30,
-                          &kYuvI601Constants, width, height);
-}
-
-// Convert H420 to AR30.
-LIBYUV_API
-int H420ToAR30(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_ar30,
-               int dst_stride_ar30,
-               int width,
-               int height) {
-  return I420ToAR30Matrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                          src_stride_v, dst_ar30, dst_stride_ar30,
-                          &kYvuH709Constants, width, height);
-}
-
 // Convert I420 to specified format
 LIBYUV_API
 int ConvertFromI420(const uint8_t* y,
@@ -1421,7 +773,8 @@ int ConvertFromI420(const uint8_t* y,
                    height);
       break;
     case FOURCC_NV12: {
-      uint8_t* dst_uv = dst_sample + width * height;
+      int dst_y_stride = dst_sample_stride ? dst_sample_stride : width;
+      uint8_t* dst_uv = dst_sample + dst_y_stride * height;
       r = I420ToNV12(y, y_stride, u, u_stride, v, v_stride, dst_sample,
                      dst_sample_stride ? dst_sample_stride : width, dst_uv,
                      dst_sample_stride ? dst_sample_stride : width, width,
@@ -1429,14 +782,14 @@ int ConvertFromI420(const uint8_t* y,
       break;
     }
     case FOURCC_NV21: {
-      uint8_t* dst_vu = dst_sample + width * height;
+      int dst_y_stride = dst_sample_stride ? dst_sample_stride : width;
+      uint8_t* dst_vu = dst_sample + dst_y_stride * height;
       r = I420ToNV21(y, y_stride, u, u_stride, v, v_stride, dst_sample,
                      dst_sample_stride ? dst_sample_stride : width, dst_vu,
                      dst_sample_stride ? dst_sample_stride : width, width,
                      height);
       break;
     }
-    // TODO(fbarchard): Add M420.
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YV12: {
diff --git a/files/source/convert_from_argb.cc b/files/source/convert_from_argb.cc
index fbcd039d..e50c2af3 100644
--- a/files/source/convert_from_argb.cc
+++ b/files/source/convert_from_argb.cc
@@ -76,11 +76,11 @@ int ARGBToI444(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOUV444ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUV444Row = ARGBToUV444Row_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToUV444Row = ARGBToUV444Row_MMI;
+#if defined(HAS_ARGBTOUV444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUV444Row = ARGBToUV444Row_LASX;
     }
   }
 #endif
@@ -103,7 +103,7 @@ int ARGBToI444(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -116,11 +116,11 @@ int ARGBToI444(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
     }
   }
 #endif
@@ -170,30 +170,42 @@ int ARGBToI422(const uint8_t* src_argb,
     height = 1;
     src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -206,37 +218,26 @@ int ARGBToI422(const uint8_t* src_argb,
     }
   }
 #endif
-
-#if defined(HAS_ARGBTOYROW_MSA)
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVRow = ARGBToUVRow_MSA;
     }
   }
 #endif
 
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -279,30 +280,10 @@ int ARGBToNV12(const uint8_t* src_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
-    ARGBToYRow = ARGBToYRow_Any_SSSE3;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
-      ARGBToYRow = ARGBToYRow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
-    ARGBToYRow = ARGBToYRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
-      ARGBToYRow = ARGBToYRow_AVX2;
-    }
-  }
-#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -315,35 +296,57 @@ int ARGBToNV12(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToYRow = ARGBToYRow_Any_MSA;
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToYRow = ARGBToYRow_MSA;
+      ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_MSA;
+      ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+      ARGBToYRow = ARGBToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -379,11 +382,11 @@ int ARGBToNV12(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_MERGEUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeUVRow_ = MergeUVRow_Any_MMI;
-    if (IS_ALIGNED(halfwidth, 8)) {
-      MergeUVRow_ = MergeUVRow_MMI;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
     }
   }
 #endif
@@ -439,30 +442,42 @@ int ARGBToNV21(const uint8_t* src_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -475,39 +490,28 @@ int ARGBToNV21(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVRow = ARGBToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
-
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
@@ -540,11 +544,11 @@ int ARGBToNV21(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_MERGEUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeUVRow_ = MergeUVRow_Any_MMI;
-    if (IS_ALIGNED(halfwidth, 8)) {
-      MergeUVRow_ = MergeUVRow_MMI;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
     }
   }
 #endif
@@ -599,30 +603,42 @@ int ABGRToNV12(const uint8_t* src_abgr,
     src_abgr = src_abgr + (height - 1) * src_stride_abgr;
     src_stride_abgr = -src_stride_abgr;
   }
-#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+#if defined(HAS_ABGRTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
     ABGRToYRow = ABGRToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_SSSE3;
       ABGRToYRow = ABGRToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ABGRTOYROW_AVX2) && defined(HAS_ABGRTOUVROW_AVX2)
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
     ABGRToYRow = ABGRToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ABGRToUVRow = ABGRToUVRow_AVX2;
       ABGRToYRow = ABGRToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ABGRTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ABGRToYRow = ABGRToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ABGRToYRow = ABGRToYRow_NEON;
     }
   }
@@ -635,35 +651,167 @@ int ABGRToNV12(const uint8_t* src_abgr,
     }
   }
 #endif
-#if defined(HAS_ABGRTOYROW_MSA)
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ABGRToYRow = ABGRToYRow_Any_MSA;
+    ABGRToUVRow = ABGRToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ABGRToYRow = ABGRToYRow_MSA;
     }
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
   }
 #endif
-#if defined(HAS_ABGRTOUVROW_MSA)
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
-    ABGRToUVRow = ABGRToUVRow_Any_MSA;
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
+      ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+      src_abgr += src_stride_abgr * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ABGRToUVRow(src_abgr, 0, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ABGRToYRow(src_abgr, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ABGRToNV21(const uint8_t* src_abgr,
+               int src_stride_abgr,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ABGRToUVRow)(const uint8_t* src_abgr0, int src_stride_abgr,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8_t* src_abgr, uint8_t* dst_y, int width) =
+      ABGRToYRow_C;
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_vu, int width) = MergeUVRow_C;
+  if (!src_abgr || !dst_y || !dst_vu || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToYRow = ABGRToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ABGRToUVRow = ABGRToUVRow_MSA;
+      ABGRToYRow = ABGRToYRow_AVX2;
     }
   }
 #endif
-#if defined(HAS_ABGRTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ABGRToYRow = ABGRToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ABGRToYRow = ABGRToYRow_MMI;
+#if defined(HAS_ABGRTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ABGRToUVRow = ABGRToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_NEON;
     }
   }
 #endif
-#if defined(HAS_ABGRTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ABGRToUVRow = ABGRToUVRow_Any_MMI;
+#if defined(HAS_ABGRTOYROW_MSA) && defined(HAS_ABGRTOUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ABGRToYRow = ABGRToYRow_Any_MSA;
+    ABGRToUVRow = ABGRToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
-      ABGRToUVRow = ABGRToUVRow_MMI;
+      ABGRToYRow = ABGRToYRow_MSA;
+    }
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToUVRow = ABGRToUVRow_MSA;
     }
   }
 #endif
@@ -699,11 +847,11 @@ int ABGRToNV12(const uint8_t* src_abgr,
     }
   }
 #endif
-#if defined(HAS_MERGEUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeUVRow_ = MergeUVRow_Any_MMI;
-    if (IS_ALIGNED(halfwidth, 8)) {
-      MergeUVRow_ = MergeUVRow_MMI;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
     }
   }
 #endif
@@ -714,16 +862,16 @@ int ABGRToNV12(const uint8_t* src_abgr,
 
     for (y = 0; y < height - 1; y += 2) {
       ABGRToUVRow(src_abgr, src_stride_abgr, row_u, row_v, width);
-      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
       ABGRToYRow(src_abgr, dst_y, width);
       ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
       src_abgr += src_stride_abgr * 2;
       dst_y += dst_stride_y * 2;
-      dst_uv += dst_stride_uv;
+      dst_vu += dst_stride_vu;
     }
     if (height & 1) {
       ABGRToUVRow(src_abgr, 0, row_u, row_v, width);
-      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
       ABGRToYRow(src_abgr, dst_y, width);
     }
     free_aligned_buffer_64(row_u);
@@ -764,30 +912,42 @@ int ARGBToYUY2(const uint8_t* src_argb,
     height = 1;
     src_stride_argb = dst_stride_yuy2 = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -800,35 +960,25 @@ int ARGBToYUY2(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVRow = ARGBToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -864,11 +1014,11 @@ int ARGBToYUY2(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_I422TOYUY2ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I422ToYUY2Row = I422ToYUY2Row_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToYUY2Row = I422ToYUY2Row_MMI;
+#if defined(HAS_I422TOYUY2ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToYUY2Row = I422ToYUY2Row_LASX;
     }
   }
 #endif
@@ -925,30 +1075,42 @@ int ARGBToUYVY(const uint8_t* src_argb,
     height = 1;
     src_stride_argb = dst_stride_uyvy = 0;
   }
-#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+#if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_SSSE3;
       ARGBToYRow = ARGBToYRow_SSSE3;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+#if defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
     ARGBToYRow = ARGBToYRow_Any_AVX2;
     if (IS_ALIGNED(width, 32)) {
-      ARGBToUVRow = ARGBToUVRow_AVX2;
       ARGBToYRow = ARGBToYRow_AVX2;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -961,35 +1123,25 @@ int ARGBToUYVY(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MSA)
+#if defined(HAS_ARGBTOYROW_MSA) && defined(HAS_ARGBTOUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYRow = ARGBToYRow_Any_MSA;
+    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVRow = ARGBToUVRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVRow = ARGBToUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      ARGBToUVRow = ARGBToUVRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    ARGBToUVRow = ARGBToUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
+      ARGBToUVRow = ARGBToUVRow_LASX;
     }
   }
 #endif
@@ -1025,11 +1177,11 @@ int ARGBToUYVY(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_I422TOUYVYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    I422ToUYVYRow = I422ToUYVYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToUYVYRow = I422ToUYVYRow_MMI;
+#if defined(HAS_I422TOUYVYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      I422ToUYVYRow = I422ToUYVYRow_LASX;
     }
   }
 #endif
@@ -1097,7 +1249,7 @@ int ARGBToI400(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
   }
@@ -1110,11 +1262,11 @@ int ARGBToI400(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYRow = ARGBToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYRow = ARGBToYRow_MMI;
+#if defined(HAS_ARGBTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYRow = ARGBToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_LASX;
     }
   }
 #endif
@@ -1195,7 +1347,7 @@ int ARGBToRGB24(const uint8_t* src_argb,
 #if defined(HAS_ARGBTORGB24ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToRGB24Row = ARGBToRGB24Row_NEON;
     }
   }
@@ -1208,11 +1360,11 @@ int ARGBToRGB24(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTORGB24ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToRGB24Row = ARGBToRGB24Row_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRGB24Row = ARGBToRGB24Row_MMI;
+#if defined(HAS_ARGBTORGB24ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_LASX;
     }
   }
 #endif
@@ -1282,11 +1434,11 @@ int ARGBToRAW(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTORAWROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToRAWRow = ARGBToRAWRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRAWRow = ARGBToRAWRow_MMI;
+#if defined(HAS_ARGBTORAWROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToRAWRow = ARGBToRAWRow_LASX;
     }
   }
 #endif
@@ -1360,11 +1512,11 @@ int ARGBToRGB565Dither(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTORGB565DITHERROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_MMI;
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_LASX;
     }
   }
 #endif
@@ -1437,11 +1589,11 @@ int ARGBToRGB565(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTORGB565ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToRGB565Row = ARGBToRGB565Row_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToRGB565Row = ARGBToRGB565Row_MMI;
+#if defined(HAS_ARGBTORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_LASX;
     }
   }
 #endif
@@ -1511,11 +1663,11 @@ int ARGBToARGB1555(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOARGB1555ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToARGB1555Row = ARGBToARGB1555Row_MMI;
+#if defined(HAS_ARGBTOARGB1555ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_LASX;
     }
   }
 #endif
@@ -1585,11 +1737,11 @@ int ARGBToARGB4444(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOARGB4444ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      ARGBToARGB4444Row = ARGBToARGB4444Row_MMI;
+#if defined(HAS_ARGBTOARGB4444ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_LASX;
     }
   }
 #endif
@@ -1727,16 +1879,22 @@ int ARGBToJ420(const uint8_t* src_argb,
     src_argb = src_argb + (height - 1) * src_stride_argb;
     src_stride_argb = -src_stride_argb;
   }
-#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#if defined(HAS_ARGBTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
       ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
@@ -1748,7 +1906,7 @@ int ARGBToJ420(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
@@ -1761,35 +1919,35 @@ int ARGBToJ420(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_MSA)
+#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYJRow = ARGBToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVJROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVJRow = ARGBToUVJRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_MMI;
+#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_MMI;
+      ARGBToYJRow = ARGBToYJRow_LSX;
+      ARGBToUVJRow = ARGBToUVJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
+      ARGBToUVJRow = ARGBToUVJRow_LASX;
     }
   }
 #endif
@@ -1844,16 +2002,22 @@ int ARGBToJ422(const uint8_t* src_argb,
     height = 1;
     src_stride_argb = dst_stride_yj = dst_stride_u = dst_stride_v = 0;
   }
-#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+#if defined(HAS_ARGBTOYJROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
     ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
       ARGBToYJRow = ARGBToYJRow_SSSE3;
     }
   }
 #endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
     ARGBToYJRow = ARGBToYJRow_Any_AVX2;
@@ -1865,7 +2029,7 @@ int ARGBToJ422(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
@@ -1878,35 +2042,35 @@ int ARGBToJ422(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_MSA)
+#if defined(HAS_ARGBTOYJROW_MSA) && defined(HAS_ARGBTOUVJROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     ARGBToYJRow = ARGBToYJRow_Any_MSA;
+    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
     if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_MSA;
     }
-  }
-#endif
-#if defined(HAS_ARGBTOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYJRow = ARGBToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_MMI;
-    }
-  }
-#endif
-#if defined(HAS_ARGBTOUVJROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_MSA;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUVJRow = ARGBToUVJRow_MSA;
     }
   }
 #endif
-#if defined(HAS_ARGBTOUVJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToUVJRow = ARGBToUVJRow_Any_MMI;
+#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
-      ARGBToUVJRow = ARGBToUVJRow_MMI;
+      ARGBToYJRow = ARGBToYJRow_LSX;
+      ARGBToUVJRow = ARGBToUVJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
+      ARGBToUVJRow = ARGBToUVJRow_LASX;
     }
   }
 #endif
@@ -1922,6 +2086,124 @@ int ARGBToJ422(const uint8_t* src_argb,
   return 0;
 }
 
+// Convert ARGB to AR64.
+LIBYUV_API
+int ARGBToAR64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ar64,
+               int dst_stride_ar64,
+               int width,
+               int height) {
+  int y;
+  void (*ARGBToAR64Row)(const uint8_t* src_argb, uint16_t* dst_ar64,
+                        int width) = ARGBToAR64Row_C;
+  if (!src_argb || !dst_ar64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_ar64 = 0;
+  }
+#if defined(HAS_ARGBTOAR64ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToAR64Row = ARGBToAR64Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAR64Row = ARGBToAR64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToAR64Row = ARGBToAR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAR64Row = ARGBToAR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToAR64Row(src_argb, dst_ar64, width);
+    src_argb += src_stride_argb;
+    dst_ar64 += dst_stride_ar64;
+  }
+  return 0;
+}
+
+// Convert ARGB to AB64.
+LIBYUV_API
+int ARGBToAB64(const uint8_t* src_argb,
+               int src_stride_argb,
+               uint16_t* dst_ab64,
+               int dst_stride_ab64,
+               int width,
+               int height) {
+  int y;
+  void (*ARGBToAB64Row)(const uint8_t* src_argb, uint16_t* dst_ar64,
+                        int width) = ARGBToAB64Row_C;
+  if (!src_argb || !dst_ab64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 && dst_stride_ab64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_ab64 = 0;
+  }
+#if defined(HAS_ARGBTOAB64ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToAB64Row = ARGBToAB64Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAB64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAB64Row = ARGBToAB64Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOAB64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToAB64Row = ARGBToAB64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToAB64Row = ARGBToAB64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToAB64Row(src_argb, dst_ab64, width);
+    src_argb += src_stride_argb;
+    dst_ab64 += dst_stride_ab64;
+  }
+  return 0;
+}
+
 // Convert ARGB to J400.
 LIBYUV_API
 int ARGBToJ400(const uint8_t* src_argb,
@@ -1966,7 +2248,7 @@ int ARGBToJ400(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
@@ -1979,23 +2261,282 @@ int ARGBToJ400(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYJRow = ARGBToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_MMI;
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYJRow(src_argb, dst_yj, width);
+    src_argb += src_stride_argb;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+// Convert RGBA to J400.
+LIBYUV_API
+int RGBAToJ400(const uint8_t* src_rgba,
+               int src_stride_rgba,
+               uint8_t* dst_yj,
+               int dst_stride_yj,
+               int width,
+               int height) {
+  int y;
+  void (*RGBAToYJRow)(const uint8_t* src_rgba, uint8_t* dst_yj, int width) =
+      RGBAToYJRow_C;
+  if (!src_rgba || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_rgba == width * 4 && dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_rgba = dst_stride_yj = 0;
+  }
+#if defined(HAS_RGBATOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToYJRow = RGBAToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGBAToYJRow = RGBAToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      RGBAToYJRow = RGBAToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToYJRow = RGBAToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RGBAToYJRow = RGBAToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_MSA;
     }
   }
 #endif
 
   for (y = 0; y < height; ++y) {
-    ARGBToYJRow(src_argb, dst_yj, width);
-    src_argb += src_stride_argb;
+    RGBAToYJRow(src_rgba, dst_yj, width);
+    src_rgba += src_stride_rgba;
     dst_yj += dst_stride_yj;
   }
   return 0;
 }
 
+// Enabled if 1 pass is available
+#if defined(HAS_RAWTOYJROW_NEON) || defined(HAS_RAWTOYJROW_MSA)
+#define HAS_RAWTOYJROW
+#endif
+
+// RAW to JNV21 full range NV21
+LIBYUV_API
+int RAWToJNV21(const uint8_t* src_raw,
+               int src_stride_raw,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_vu,
+               int dst_stride_vu,
+               int width,
+               int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+#if defined(HAS_RAWTOYJROW)
+  void (*RAWToUVJRow)(const uint8_t* src_raw, int src_stride_raw,
+                      uint8_t* dst_u, uint8_t* dst_v, int width) =
+      RAWToUVJRow_C;
+  void (*RAWToYJRow)(const uint8_t* src_raw, uint8_t* dst_y, int width) =
+      RAWToYJRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8_t* src_rgb, uint8_t* dst_argb, int width) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVJRow)(const uint8_t* src_argb0, int src_stride_argb,
+                       uint8_t* dst_u, uint8_t* dst_v, int width) =
+      ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8_t* src_argb, uint8_t* dst_y, int width) =
+      ARGBToYJRow_C;
+#endif
+  void (*MergeUVRow_)(const uint8_t* src_u, const uint8_t* src_v,
+                      uint8_t* dst_vu, int width) = MergeUVRow_C;
+  if (!src_raw || !dst_y || !dst_vu || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+#if defined(HAS_RAWTOYJROW)
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYJROW_NEON) && defined(HAS_RAWTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVJRow = RAWToUVJRow_Any_NEON;
+    RAWToYJRow = RAWToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_NEON;
+      RAWToUVJRow = RAWToUVJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_MSA) && defined(HAS_RAWTOUVJROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    RAWToUVJRow = RAWToUVJRow_Any_MSA;
+    RAWToYJRow = RAWToYJRow_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_MSA;
+      RAWToUVJRow = RAWToUVJRow_MSA;
+    }
+  }
+#endif
+
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else  // HAS_RAWTOYJROW
+
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVJRow = ARGBToUVJRow_AVX2;
+    }
+  }
+#endif
+#endif  // HAS_RAWTOYJROW
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MergeUVRow_ = MergeUVRow_Any_MSA;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow_ = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_LSX;
+    }
+  }
+#endif
+  {
+    // Allocate a row of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8_t* row_v = row_u + ((halfwidth + 31) & ~31);
+#if !defined(HAS_RAWTOYJROW)
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, src_stride_raw, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      RAWToYJRow(src_raw, dst_y, width);
+      RAWToYJRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+      ARGBToUVJRow(row, kRowSize, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      ARGBToYJRow(row, dst_y, width);
+      ARGBToYJRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_vu += dst_stride_vu;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYJROW)
+      RAWToUVJRow(src_raw, 0, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      RAWToYJRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVJRow(row, 0, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_vu, halfwidth);
+      ARGBToYJRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYJROW)
+    free_aligned_buffer_64(row);
+#endif
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+#undef HAS_RAWTOYJROW
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/convert_jpeg.cc b/files/source/convert_jpeg.cc
index f440c7c2..d7556ee9 100644
--- a/files/source/convert_jpeg.cc
+++ b/files/source/convert_jpeg.cc
@@ -328,6 +328,140 @@ int MJPGToNV21(const uint8_t* src_mjpg,
   return ret ? 0 : 1;
 }
 
+static void JpegI420ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I420ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I422ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 with VU swapped.
+  I444ToNV21(data[0], strides[0], data[2], strides[2], data[1], strides[1],
+             dest->y, dest->y_stride, dest->vu, dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToNV12(void* opaque,
+                           const uint8_t* const* data,
+                           const int* strides,
+                           int rows) {
+  NV21Buffers* dest = (NV21Buffers*)(opaque);
+  // Use NV21 since there is no UV plane.
+  I400ToNV21(data[0], strides[0], dest->y, dest->y_stride, dest->vu,
+             dest->vu_stride, dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->vu += ((rows + 1) >> 1) * dest->vu_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPEG) to NV12.
+LIBYUV_API
+int MJPGToNV12(const uint8_t* sample,
+               size_t sample_size,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int src_width,
+               int src_height,
+               int dst_width,
+               int dst_height) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != src_width ||
+              mjpeg_decoder.GetHeight() != src_height)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    // Use NV21Buffers but with UV instead of VU.
+    NV21Buffers bufs = {dst_y,         dst_stride_y, dst_uv,
+                        dst_stride_uv, dst_width,    dst_height};
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() == MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToNV12, &bufs, dst_width,
+                                           dst_height);
+      // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToNV12, &bufs, dst_width,
+                                           dst_height);
+    } else {
+      // Unknown colorspace.
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
 struct ARGBBuffers {
   uint8_t* argb;
   int argb_stride;
diff --git a/files/source/convert_to_argb.cc b/files/source/convert_to_argb.cc
index bde1aa88..84df16c8 100644
--- a/files/source/convert_to_argb.cc
+++ b/files/source/convert_to_argb.cc
@@ -32,9 +32,6 @@ extern "C" {
 // TODO(fbarchard): Add the following:
 // H010ToARGB
 // I010ToARGB
-// J400ToARGB
-// J422ToARGB
-// J444ToARGB
 
 LIBYUV_API
 int ConvertToARGB(const uint8_t* sample,
@@ -161,6 +158,11 @@ int ConvertToARGB(const uint8_t* sample,
       r = I400ToARGB(src, src_width, dst_argb, dst_stride_argb, crop_width,
                      inv_crop_height);
       break;
+    case FOURCC_J400:
+      src = sample + src_width * crop_y + crop_x;
+      r = J400ToARGB(src, src_width, dst_argb, dst_stride_argb, crop_width,
+                     inv_crop_height);
+      break;
 
     // Biplanar formats
     case FOURCC_NV12:
@@ -178,12 +180,6 @@ int ConvertToARGB(const uint8_t* sample,
       r = NV21ToARGB(src, src_width, src_uv, aligned_src_width, dst_argb,
                      dst_stride_argb, crop_width, inv_crop_height);
       break;
-    case FOURCC_M420:
-      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
-      r = M420ToARGB(src, src_width, dst_argb, dst_stride_argb, crop_width,
-                     inv_crop_height);
-      break;
-
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YV12: {
@@ -208,6 +204,19 @@ int ConvertToARGB(const uint8_t* sample,
       break;
     }
 
+    case FOURCC_J420: {
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8_t* src_u = sample + src_width * abs_src_height +
+                             (halfwidth * crop_y + crop_x) / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
     case FOURCC_H420: {
       int halfwidth = (src_width + 1) / 2;
       int halfheight = (abs_src_height + 1) / 2;
@@ -221,7 +230,7 @@ int ConvertToARGB(const uint8_t* sample,
       break;
     }
 
-    case FOURCC_J420: {
+    case FOURCC_U420: {
       int halfwidth = (src_width + 1) / 2;
       int halfheight = (abs_src_height + 1) / 2;
       const uint8_t* src_y = sample + (src_width * crop_y + crop_x);
@@ -229,7 +238,7 @@ int ConvertToARGB(const uint8_t* sample,
                              (halfwidth * crop_y + crop_x) / 2;
       const uint8_t* src_v = sample + src_width * abs_src_height +
                              halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
-      r = J420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+      r = U420ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
                      dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
@@ -256,6 +265,18 @@ int ConvertToARGB(const uint8_t* sample,
       break;
     }
 
+    case FOURCC_J422: {
+      int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u =
+          sample + src_width * abs_src_height + halfwidth * crop_y + crop_x / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      r = J422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
     case FOURCC_H422: {
       int halfwidth = (src_width + 1) / 2;
       const uint8_t* src_y = sample + src_width * crop_y + crop_x;
@@ -268,6 +289,18 @@ int ConvertToARGB(const uint8_t* sample,
       break;
     }
 
+    case FOURCC_U422: {
+      int halfwidth = (src_width + 1) / 2;
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u =
+          sample + src_width * abs_src_height + halfwidth * crop_y + crop_x / 2;
+      const uint8_t* src_v = sample + src_width * abs_src_height +
+                             halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      r = H422ToARGB(src_y, src_width, src_u, halfwidth, src_v, halfwidth,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
     case FOURCC_I444:
     case FOURCC_YV24: {
       const uint8_t* src_y = sample + src_width * crop_y + crop_x;
@@ -284,6 +317,40 @@ int ConvertToARGB(const uint8_t* sample,
                      dst_argb, dst_stride_argb, crop_width, inv_crop_height);
       break;
     }
+
+    case FOURCC_J444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = J444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_H444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = H444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_U444: {
+      const uint8_t* src_y = sample + src_width * crop_y + crop_x;
+      const uint8_t* src_u;
+      const uint8_t* src_v;
+      src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+      src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      r = U444ToARGB(src_y, src_width, src_u, src_width, src_v, src_width,
+                     dst_argb, dst_stride_argb, crop_width, inv_crop_height);
+      break;
+    }
+
 #ifdef HAVE_JPEG
     case FOURCC_MJPG:
       r = MJPGToARGB(sample, sample_size, dst_argb, dst_stride_argb, src_width,
diff --git a/files/source/convert_to_i420.cc b/files/source/convert_to_i420.cc
index 584be0ac..5869ecd7 100644
--- a/files/source/convert_to_i420.cc
+++ b/files/source/convert_to_i420.cc
@@ -89,18 +89,26 @@ int ConvertToI420(const uint8_t* sample,
 
   switch (format) {
     // Single plane formats
-    case FOURCC_YUY2:
+    case FOURCC_YUY2: {  // TODO(fbarchard): Find better odd crop fix.
+      uint8_t* u = (crop_x & 1) ? dst_v : dst_u;
+      uint8_t* v = (crop_x & 1) ? dst_u : dst_v;
+      int stride_u = (crop_x & 1) ? dst_stride_v : dst_stride_u;
+      int stride_v = (crop_x & 1) ? dst_stride_u : dst_stride_v;
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = YUY2ToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, dst_u,
-                     dst_stride_u, dst_v, dst_stride_v, crop_width,
-                     inv_crop_height);
+      r = YUY2ToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, u,
+                     stride_u, v, stride_v, crop_width, inv_crop_height);
       break;
-    case FOURCC_UYVY:
+    }
+    case FOURCC_UYVY: {
+      uint8_t* u = (crop_x & 1) ? dst_v : dst_u;
+      uint8_t* v = (crop_x & 1) ? dst_u : dst_v;
+      int stride_u = (crop_x & 1) ? dst_stride_v : dst_stride_u;
+      int stride_v = (crop_x & 1) ? dst_stride_u : dst_stride_v;
       src = sample + (aligned_src_width * crop_y + crop_x) * 2;
-      r = UYVYToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, dst_u,
-                     dst_stride_u, dst_v, dst_stride_v, crop_width,
-                     inv_crop_height);
+      r = UYVYToI420(src, aligned_src_width * 2, dst_y, dst_stride_y, u,
+                     stride_u, v, stride_v, crop_width, inv_crop_height);
       break;
+    }
     case FOURCC_RGBP:
       src = sample + (src_width * crop_y + crop_x) * 2;
       r = RGB565ToI420(src, src_width * 2, dst_y, dst_stride_y, dst_u,
@@ -179,11 +187,6 @@ int ConvertToI420(const uint8_t* sample,
                            dst_stride_y, dst_v, dst_stride_v, dst_u,
                            dst_stride_u, crop_width, inv_crop_height, rotation);
       break;
-    case FOURCC_M420:
-      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
-      r = M420ToI420(src, src_width, dst_y, dst_stride_y, dst_u, dst_stride_u,
-                     dst_v, dst_stride_v, crop_width, inv_crop_height);
-      break;
     // Triplanar formats
     case FOURCC_I420:
     case FOURCC_YV12: {
diff --git a/files/source/cpu_id.cc b/files/source/cpu_id.cc
index 48e2b615..56fe60e4 100644
--- a/files/source/cpu_id.cc
+++ b/files/source/cpu_id.cc
@@ -20,7 +20,7 @@
 #endif
 
 // For ArmCpuCaps() but unittested on all platforms
-#include <stdio.h>
+#include <stdio.h>  // For fopen()
 #include <string.h>
 
 #ifdef __cplusplus
@@ -75,9 +75,9 @@ void CpuId(int info_eax, int info_ecx, int* cpu_info) {
   asm volatile(
 #if defined(__i386__) && defined(__PIC__)
       // Preserve ebx for fpic 32 bit.
-      "mov %%ebx, %%edi                          \n"
+      "mov         %%ebx, %%edi                  \n"
       "cpuid                                     \n"
-      "xchg %%edi, %%ebx                         \n"
+      "xchg        %%edi, %%ebx                  \n"
       : "=D"(info_ebx),
 #else
       "cpuid                                     \n"
@@ -133,7 +133,7 @@ int GetXCR0() {
 #pragma optimize("g", on)
 #endif
 
-// based on libvpx arm_cpudetect.c
+// Based on libvpx arm_cpudetect.c
 // For Arm, but public to allow testing on any CPU
 LIBYUV_API SAFEBUFFERS int ArmCpuCaps(const char* cpuinfo_name) {
   char cpuinfo_line[512];
@@ -163,45 +163,54 @@ LIBYUV_API SAFEBUFFERS int ArmCpuCaps(const char* cpuinfo_name) {
 }
 
 // TODO(fbarchard): Consider read_msa_ir().
-// TODO(fbarchard): Add unittest.
-LIBYUV_API SAFEBUFFERS int MipsCpuCaps(const char* cpuinfo_name,
-                                       const char ase[]) {
+LIBYUV_API SAFEBUFFERS int MipsCpuCaps(const char* cpuinfo_name) {
   char cpuinfo_line[512];
+  int flag = 0x0;
   FILE* f = fopen(cpuinfo_name, "r");
   if (!f) {
-    // ase enabled if /proc/cpuinfo is unavailable.
-    if (strcmp(ase, " msa") == 0) {
-      return kCpuHasMSA;
-    }
-    if (strcmp(ase, " mmi") == 0) {
-      return kCpuHasMMI;
-    }
+    // Assume nothing if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
     return 0;
   }
   while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
-    if (memcmp(cpuinfo_line, "ASEs implemented", 16) == 0) {
-      char* p = strstr(cpuinfo_line, ase);
-      if (p) {
-        fclose(f);
-        if (strcmp(ase, " msa") == 0) {
-          return kCpuHasMSA;
-        }
-        return 0;
+    if (memcmp(cpuinfo_line, "cpu model", 9) == 0) {
+      // Workaround early kernel without MSA in ASEs line.
+      if (strstr(cpuinfo_line, "Loongson-2K")) {
+        flag |= kCpuHasMSA;
       }
-    } else if (memcmp(cpuinfo_line, "cpu model", 9) == 0) {
-      char* p = strstr(cpuinfo_line, "Loongson-3");
-      if (p) {
-        fclose(f);
-        if (strcmp(ase, " mmi") == 0) {
-          return kCpuHasMMI;
-        }
-        return 0;
+    }
+    if (memcmp(cpuinfo_line, "ASEs implemented", 16) == 0) {
+      if (strstr(cpuinfo_line, "msa")) {
+        flag |= kCpuHasMSA;
       }
+      // ASEs is the last line, so we can break here.
+      break;
     }
   }
   fclose(f);
-  return 0;
+  return flag;
+}
+
+// TODO(fbarchard): Consider read_loongarch_ir().
+#define LOONGARCH_CFG2 0x2
+#define LOONGARCH_CFG2_LSX (1 << 6)
+#define LOONGARCH_CFG2_LASX (1 << 7)
+
+#if defined(__loongarch__)
+LIBYUV_API SAFEBUFFERS int LoongarchCpuCaps(void) {
+  int flag = 0x0;
+  uint32_t cfg2 = 0;
+
+  __asm__ volatile("cpucfg %0, %1 \n\t" : "+&r"(cfg2) : "r"(LOONGARCH_CFG2));
+
+  if (cfg2 & LOONGARCH_CFG2_LSX)
+    flag |= kCpuHasLSX;
+
+  if (cfg2 & LOONGARCH_CFG2_LASX)
+    flag |= kCpuHasLASX;
+  return flag;
 }
+#endif
 
 static SAFEBUFFERS int GetCpuFlags(void) {
   int cpu_info = 0;
@@ -235,6 +244,7 @@ static SAFEBUFFERS int GetCpuFlags(void) {
       cpu_info |= (cpu_info7[1] & 0x80000000) ? kCpuHasAVX512VL : 0;
       cpu_info |= (cpu_info7[2] & 0x00000002) ? kCpuHasAVX512VBMI : 0;
       cpu_info |= (cpu_info7[2] & 0x00000040) ? kCpuHasAVX512VBMI2 : 0;
+      cpu_info |= (cpu_info7[2] & 0x00000800) ? kCpuHasAVX512VNNI : 0;
       cpu_info |= (cpu_info7[2] & 0x00001000) ? kCpuHasAVX512VBITALG : 0;
       cpu_info |= (cpu_info7[2] & 0x00004000) ? kCpuHasAVX512VPOPCNTDQ : 0;
       cpu_info |= (cpu_info7[2] & 0x00000100) ? kCpuHasGFNI : 0;
@@ -242,13 +252,13 @@ static SAFEBUFFERS int GetCpuFlags(void) {
   }
 #endif
 #if defined(__mips__) && defined(__linux__)
-#if defined(__mips_msa)
-  cpu_info = MipsCpuCaps("/proc/cpuinfo", " msa");
-#elif defined(_MIPS_ARCH_LOONGSON3A)
-  cpu_info = MipsCpuCaps("/proc/cpuinfo", " mmi");
-#endif
+  cpu_info = MipsCpuCaps("/proc/cpuinfo");
   cpu_info |= kCpuHasMIPS;
 #endif
+#if defined(__loongarch__) && defined(__linux__)
+  cpu_info = LoongarchCpuCaps();
+  cpu_info |= kCpuHasLOONGARCH;
+#endif
 #if defined(__arm__) || defined(__aarch64__)
 // gcc -mfpu=neon defines __ARM_NEON__
 // __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
diff --git a/files/source/mjpeg_decoder.cc b/files/source/mjpeg_decoder.cc
index 5c5e5ead..4ccf00a3 100644
--- a/files/source/mjpeg_decoder.cc
+++ b/files/source/mjpeg_decoder.cc
@@ -417,7 +417,6 @@ void init_source(j_decompress_ptr cinfo) {
 boolean fill_input_buffer(j_decompress_ptr cinfo) {
   BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data);
   if (buf_vec->pos >= buf_vec->len) {
-    assert(0 && "No more data");
     // ERROR: No more data
     return FALSE;
   }
@@ -430,7 +429,7 @@ boolean fill_input_buffer(j_decompress_ptr cinfo) {
 void skip_input_data(j_decompress_ptr cinfo, long num_bytes) {  // NOLINT
   jpeg_source_mgr* src = cinfo->src;
   size_t bytes = static_cast<size_t>(num_bytes);
-  if(bytes > src->bytes_in_buffer) {
+  if (bytes > src->bytes_in_buffer) {
     src->next_input_byte = nullptr;
     src->bytes_in_buffer = 0;
   } else {
diff --git a/files/source/planar_functions.cc b/files/source/planar_functions.cc
index 9cab230f..169d4a8f 100644
--- a/files/source/planar_functions.cc
+++ b/files/source/planar_functions.cc
@@ -10,6 +10,7 @@
 
 #include "libyuv/planar_functions.h"
 
+#include <assert.h>
 #include <string.h>  // for memset()
 
 #include "libyuv/cpu_id.h"
@@ -34,6 +35,9 @@ void CopyPlane(const uint8_t* src_y,
                int height) {
   int y;
   void (*CopyRow)(const uint8_t* src, uint8_t* dst, int width) = CopyRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -80,8 +84,6 @@ void CopyPlane(const uint8_t* src_y,
   }
 }
 
-// TODO(fbarchard): Consider support for negative height.
-// TODO(fbarchard): Consider stride measured in bytes.
 LIBYUV_API
 void CopyPlane_16(const uint16_t* src_y,
                   int src_stride_y,
@@ -89,36 +91,8 @@ void CopyPlane_16(const uint16_t* src_y,
                   int dst_stride_y,
                   int width,
                   int height) {
-  int y;
-  void (*CopyRow)(const uint16_t* src, uint16_t* dst, int width) = CopyRow_16_C;
-  // Coalesce rows.
-  if (src_stride_y == width && dst_stride_y == width) {
-    width *= height;
-    height = 1;
-    src_stride_y = dst_stride_y = 0;
-  }
-#if defined(HAS_COPYROW_16_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_16_SSE2;
-  }
-#endif
-#if defined(HAS_COPYROW_16_ERMS)
-  if (TestCpuFlag(kCpuHasERMS)) {
-    CopyRow = CopyRow_16_ERMS;
-  }
-#endif
-#if defined(HAS_COPYROW_16_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
-    CopyRow = CopyRow_16_NEON;
-  }
-#endif
-
-  // Copy plane
-  for (y = 0; y < height; ++y) {
-    CopyRow(src_y, dst_y, width);
-    src_y += src_stride_y;
-    dst_y += dst_stride_y;
-  }
+  CopyPlane((const uint8_t*)src_y, src_stride_y * 2, (uint8_t*)dst_y,
+            dst_stride_y * 2, width * 2, height);
 }
 
 // Convert a plane of 16 bit data to 8 bit
@@ -134,6 +108,9 @@ void Convert16To8Plane(const uint16_t* src_y,
   void (*Convert16To8Row)(const uint16_t* src_y, uint8_t* dst_y, int scale,
                           int width) = Convert16To8Row_C;
 
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -146,6 +123,14 @@ void Convert16To8Plane(const uint16_t* src_y,
     height = 1;
     src_stride_y = dst_stride_y = 0;
   }
+#if defined(HAS_CONVERT16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Convert16To8Row = Convert16To8Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      Convert16To8Row = Convert16To8Row_NEON;
+    }
+  }
+#endif
 #if defined(HAS_CONVERT16TO8ROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     Convert16To8Row = Convert16To8Row_Any_SSSE3;
@@ -184,6 +169,9 @@ void Convert8To16Plane(const uint8_t* src_y,
   void (*Convert8To16Row)(const uint8_t* src_y, uint16_t* dst_y, int scale,
                           int width) = Convert8To16Row_C;
 
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -238,9 +226,12 @@ int I422Copy(const uint8_t* src_y,
              int width,
              int height) {
   int halfwidth = (width + 1) >> 1;
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -276,7 +267,8 @@ int I444Copy(const uint8_t* src_y,
              int dst_stride_v,
              int width,
              int height) {
-  if (!src_u || !src_v || !dst_u || !dst_v || width <= 0 || height == 0) {
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
     return -1;
   }
   // Negative height means invert the image.
@@ -298,6 +290,49 @@ int I444Copy(const uint8_t* src_y,
   return 0;
 }
 
+// Copy I210.
+LIBYUV_API
+int I210Copy(const uint16_t* src_y,
+             int src_stride_y,
+             const uint16_t* src_u,
+             int src_stride_u,
+             const uint16_t* src_v,
+             int src_stride_v,
+             uint16_t* dst_y,
+             int dst_stride_y,
+             uint16_t* dst_u,
+             int dst_stride_u,
+             uint16_t* dst_v,
+             int dst_stride_v,
+             int width,
+             int height) {
+  int halfwidth = (width + 1) >> 1;
+
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
 // Copy I400.
 LIBYUV_API
 int I400ToI400(const uint8_t* src_y,
@@ -349,6 +384,54 @@ int I420ToI400(const uint8_t* src_y,
   return 0;
 }
 
+// Copy NV12. Supports inverting.
+int NV12Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_uv,
+             int src_stride_uv,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_uv,
+             int dst_stride_uv,
+             int width,
+             int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+
+  if (!src_y || !dst_y || !src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, halfwidth * 2,
+            halfheight);
+  return 0;
+}
+
+// Copy NV21. Supports inverting.
+int NV21Copy(const uint8_t* src_y,
+             int src_stride_y,
+             const uint8_t* src_vu,
+             int src_stride_vu,
+             uint8_t* dst_y,
+             int dst_stride_y,
+             uint8_t* dst_vu,
+             int dst_stride_vu,
+             int width,
+             int height) {
+  return NV12Copy(src_y, src_stride_y, src_vu, src_stride_vu, dst_y,
+                  dst_stride_y, dst_vu, dst_stride_vu, width, height);
+}
+
 // Support function for NV12 etc UV channels.
 // Width and height are plane sizes (typically half pixel width).
 LIBYUV_API
@@ -363,6 +446,9 @@ void SplitUVPlane(const uint8_t* src_uv,
   int y;
   void (*SplitUVRow)(const uint8_t* src_uv, uint8_t* dst_u, uint8_t* dst_v,
                      int width) = SplitUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -410,11 +496,11 @@ void SplitUVPlane(const uint8_t* src_uv,
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SplitUVRow = SplitUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SplitUVRow = SplitUVRow_MMI;
+#if defined(HAS_SPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SplitUVRow = SplitUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_LSX;
     }
   }
 #endif
@@ -440,6 +526,9 @@ void MergeUVPlane(const uint8_t* src_u,
   int y;
   void (*MergeUVRow)(const uint8_t* src_u, const uint8_t* src_v,
                      uint8_t* dst_uv, int width) = MergeUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -485,11 +574,11 @@ void MergeUVPlane(const uint8_t* src_u,
     }
   }
 #endif
-#if defined(HAS_MERGEUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeUVRow = MergeUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      MergeUVRow = MergeUVRow_MMI;
+#if defined(HAS_MERGEUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    MergeUVRow = MergeUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow = MergeUVRow_LSX;
     }
   }
 #endif
@@ -503,6 +592,289 @@ void MergeUVPlane(const uint8_t* src_u,
   }
 }
 
+// Support function for P010 etc UV channels.
+// Width and height are plane sizes (typically half pixel width).
+LIBYUV_API
+void SplitUVPlane_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_u,
+                     int dst_stride_u,
+                     uint16_t* dst_v,
+                     int dst_stride_v,
+                     int width,
+                     int height,
+                     int depth) {
+  int y;
+  void (*SplitUVRow_16)(const uint16_t* src_uv, uint16_t* dst_u,
+                        uint16_t* dst_v, int depth, int width) =
+      SplitUVRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_u = dst_u + (height - 1) * dst_stride_u;
+    dst_v = dst_v + (height - 1) * dst_stride_v;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == width * 2 && dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow_16 = SplitUVRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow_16 = SplitUVRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow_16 = SplitUVRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SplitUVRow_16 = SplitUVRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Copy a row of UV.
+    SplitUVRow_16(src_uv, dst_u, dst_v, depth, width);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+}
+
+LIBYUV_API
+void MergeUVPlane_16(const uint16_t* src_u,
+                     int src_stride_u,
+                     const uint16_t* src_v,
+                     int src_stride_v,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height,
+                     int depth) {
+  int y;
+  void (*MergeUVRow_16)(const uint16_t* src_u, const uint16_t* src_v,
+                        uint16_t* dst_uv, int depth, int width) =
+      MergeUVRow_16_C;
+  assert(depth >= 8);
+  assert(depth <= 16);
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uv = dst_uv + (height - 1) * dst_stride_uv;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_u == width && src_stride_v == width &&
+      dst_stride_uv == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_16 = MergeUVRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeUVRow_16 = MergeUVRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_16 = MergeUVRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeUVRow_16 = MergeUVRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_16(src_u, src_v, dst_uv, depth, width);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+}
+
+// Convert plane from lsb to msb
+LIBYUV_API
+void ConvertToMSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth) {
+  int y;
+  int scale = 1 << (16 - depth);
+  void (*MultiplyRow_16)(const uint16_t* src_y, uint16_t* dst_y, int scale,
+                         int width) = MultiplyRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+
+#if defined(HAS_MULTIPLYROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MultiplyRow_16 = MultiplyRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MultiplyRow_16 = MultiplyRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MULTIPLYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MultiplyRow_16 = MultiplyRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MultiplyRow_16 = MultiplyRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MultiplyRow_16(src_y, dst_y, scale, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert plane from msb to lsb
+LIBYUV_API
+void ConvertToLSBPlane_16(const uint16_t* src_y,
+                          int src_stride_y,
+                          uint16_t* dst_y,
+                          int dst_stride_y,
+                          int width,
+                          int height,
+                          int depth) {
+  int y;
+  int scale = 1 << depth;
+  void (*DivideRow)(const uint16_t* src_y, uint16_t* dst_y, int scale,
+                    int width) = DivideRow_16_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+
+#if defined(HAS_DIVIDEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    DivideRow = DivideRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      DivideRow = DivideRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_DIVIDEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DivideRow = DivideRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DivideRow = DivideRow_16_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    DivideRow(src_y, dst_y, scale, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Swap U and V channels in interleaved UV plane.
+LIBYUV_API
+void SwapUVPlane(const uint8_t* src_uv,
+                 int src_stride_uv,
+                 uint8_t* dst_vu,
+                 int dst_stride_vu,
+                 int width,
+                 int height) {
+  int y;
+  void (*SwapUVRow)(const uint8_t* src_uv, uint8_t* dst_vu, int width) =
+      SwapUVRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == width * 2 && dst_stride_vu == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_uv = dst_stride_vu = 0;
+  }
+
+#if defined(HAS_SWAPUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SwapUVRow = SwapUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      SwapUVRow = SwapUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SWAPUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SwapUVRow = SwapUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SwapUVRow = SwapUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SWAPUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SwapUVRow = SwapUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SwapUVRow = SwapUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    SwapUVRow(src_uv, dst_vu, width);
+    src_uv += src_stride_uv;
+    dst_vu += dst_stride_vu;
+  }
+}
+
 // Convert NV21 to NV12.
 LIBYUV_API
 int NV21ToNV12(const uint8_t* src_y,
@@ -515,49 +887,150 @@ int NV21ToNV12(const uint8_t* src_y,
                int dst_stride_uv,
                int width,
                int height) {
-  int y;
-  void (*UVToVURow)(const uint8_t* src_uv, uint8_t* dst_vu, int width) =
-      UVToVURow_C;
-
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
+
   if (!src_vu || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
     halfheight = (height + 1) >> 1;
-    src_y = src_y + (height - 1) * src_stride_y;
     src_vu = src_vu + (halfheight - 1) * src_stride_vu;
-    src_stride_y = -src_stride_y;
     src_stride_vu = -src_stride_vu;
   }
-  // Coalesce rows.
-  if (src_stride_vu == halfwidth * 2 && dst_stride_uv == halfwidth * 2) {
-    halfwidth *= halfheight;
-    halfheight = 1;
-    src_stride_vu = dst_stride_uv = 0;
+
+  SwapUVPlane(src_vu, src_stride_vu, dst_uv, dst_stride_uv, halfwidth,
+              halfheight);
+  return 0;
+}
+
+// Detile a plane of data
+// tile width is 16 and assumed.
+// tile_height is 16 or 32 for MM21.
+// src_stride_y is bytes per row of source ignoring tiling. e.g. 640
+// TODO: More detile row functions.
+
+LIBYUV_API
+void DetilePlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height,
+                 int tile_height) {
+  const ptrdiff_t src_tile_stride = 16 * tile_height;
+  int y;
+  void (*DetileRow)(const uint8_t* src, ptrdiff_t src_tile_stride, uint8_t* dst,
+                    int width) = DetileRow_C;
+  assert(src_stride_y >= 0);
+  assert(tile_height > 0);
+  assert(src_stride_y > 0);
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
   }
 
-#if defined(HAS_UVToVUROW_NEON)
+#if defined(HAS_DETILEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    DetileRow = DetileRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow = DetileRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_DETILEROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    UVToVURow = UVToVURow_Any_NEON;
-    if (IS_ALIGNED(halfwidth, 16)) {
-      UVToVURow = UVToVURow_NEON;
+    DetileRow = DetileRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileRow = DetileRow_NEON;
     }
   }
 #endif
-  if (dst_y) {
-    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileRow(src_y, src_tile_stride, dst_y, width);
+    dst_y += dst_stride_y;
+    src_y += 16;
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_y = src_y - src_tile_stride + src_stride_y * tile_height;
+    }
   }
+}
 
-  for (y = 0; y < halfheight; ++y) {
-    UVToVURow(src_vu, dst_uv, halfwidth);
-    src_vu += src_stride_vu;
-    dst_uv += dst_stride_uv;
+LIBYUV_API
+void DetileSplitUVPlane(const uint8_t* src_uv,
+                        int src_stride_uv,
+                        uint8_t* dst_u,
+                        int dst_stride_u,
+                        uint8_t* dst_v,
+                        int dst_stride_v,
+                        int width,
+                        int height,
+                        int tile_height) {
+  const ptrdiff_t src_tile_stride = 16 * tile_height;
+  int y;
+  void (*DetileSplitUVRow)(const uint8_t* src, ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u, uint8_t* dst_v, int width) =
+      DetileSplitUVRow_C;
+  assert(src_stride_uv >= 0);
+  assert(tile_height > 0);
+  assert(src_stride_uv > 0);
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_u = dst_u + (height - 1) * dst_stride_u;
+    dst_stride_u = -dst_stride_u;
+    dst_v = dst_v + (height - 1) * dst_stride_v;
+    dst_stride_v = -dst_stride_v;
+  }
+
+#if defined(HAS_DETILESPLITUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    DetileSplitUVRow = DetileSplitUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      DetileSplitUVRow = DetileSplitUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_DETILESPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    DetileSplitUVRow = DetileSplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      DetileSplitUVRow = DetileSplitUVRow_NEON;
+    }
+  }
+#endif
+
+  // Detile plane
+  for (y = 0; y < height; ++y) {
+    DetileSplitUVRow(src_uv, src_tile_stride, dst_u, dst_v, width);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += 16;
+    // Advance to next row of tiles.
+    if ((y & (tile_height - 1)) == (tile_height - 1)) {
+      src_uv = src_uv - src_tile_stride + src_stride_uv * tile_height;
+    }
   }
-  return 0;
 }
 
 // Support function for NV12 etc RGB channels.
@@ -576,6 +1049,9 @@ void SplitRGBPlane(const uint8_t* src_rgb,
   int y;
   void (*SplitRGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
                       uint8_t* dst_b, int width) = SplitRGBRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -609,14 +1085,6 @@ void SplitRGBPlane(const uint8_t* src_rgb,
     }
   }
 #endif
-#if defined(HAS_SPLITRGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SplitRGBRow = SplitRGBRow_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      SplitRGBRow = SplitRGBRow_MMI;
-    }
-  }
-#endif
 
   for (y = 0; y < height; ++y) {
     // Copy a row of RGB.
@@ -643,6 +1111,9 @@ void MergeRGBPlane(const uint8_t* src_r,
   void (*MergeRGBRow)(const uint8_t* src_r, const uint8_t* src_g,
                       const uint8_t* src_b, uint8_t* dst_rgb, int width) =
       MergeRGBRow_C;
+  if (width <= 0 || height == 0) {
+    return;
+  }
   // Coalesce rows.
   // Negative height means invert the image.
   if (height < 0) {
@@ -673,86 +1144,673 @@ void MergeRGBPlane(const uint8_t* src_r,
     }
   }
 #endif
-#if defined(HAS_MERGERGBROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MergeRGBRow = MergeRGBRow_Any_MMI;
+
+  for (y = 0; y < height; ++y) {
+    // Merge a row of U and V into a row of RGB.
+    MergeRGBRow(src_r, src_g, src_b, dst_rgb, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_rgb += dst_stride_rgb;
+  }
+}
+
+LIBYUV_NOINLINE
+void SplitARGBPlaneAlpha(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_r,
+                         int dst_stride_r,
+                         uint8_t* dst_g,
+                         int dst_stride_g,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         int width,
+                         int height) {
+  int y;
+  void (*SplitARGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
+                       uint8_t* dst_b, uint8_t* dst_a, int width) =
+      SplitARGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_argb == width * 4 && dst_stride_r == width &&
+      dst_stride_g == width && dst_stride_b == width && dst_stride_a == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_r = dst_stride_g = dst_stride_b =
+        dst_stride_a = 0;
+  }
+
+#if defined(HAS_SPLITARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitARGBRow = SplitARGBRow_Any_SSE2;
     if (IS_ALIGNED(width, 8)) {
-      MergeRGBRow = MergeRGBRow_MMI;
+      SplitARGBRow = SplitARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SplitARGBRow = SplitARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      SplitARGBRow = SplitARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitARGBRow = SplitARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitARGBRow = SplitARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitARGBRow = SplitARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitARGBRow = SplitARGBRow_NEON;
     }
   }
 #endif
 
   for (y = 0; y < height; ++y) {
-    // Merge a row of U and V into a row of RGB.
-    MergeRGBRow(src_r, src_g, src_b, dst_rgb, width);
+    SplitARGBRow(src_argb, dst_r, dst_g, dst_b, dst_a, width);
+    dst_r += dst_stride_r;
+    dst_g += dst_stride_g;
+    dst_b += dst_stride_b;
+    dst_a += dst_stride_a;
+    src_argb += src_stride_argb;
+  }
+}
+
+LIBYUV_NOINLINE
+void SplitARGBPlaneOpaque(const uint8_t* src_argb,
+                          int src_stride_argb,
+                          uint8_t* dst_r,
+                          int dst_stride_r,
+                          uint8_t* dst_g,
+                          int dst_stride_g,
+                          uint8_t* dst_b,
+                          int dst_stride_b,
+                          int width,
+                          int height) {
+  int y;
+  void (*SplitXRGBRow)(const uint8_t* src_rgb, uint8_t* dst_r, uint8_t* dst_g,
+                       uint8_t* dst_b, int width) = SplitXRGBRow_C;
+  assert(height > 0);
+
+  if (src_stride_argb == width * 4 && dst_stride_r == width &&
+      dst_stride_g == width && dst_stride_b == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_r = dst_stride_g = dst_stride_b = 0;
+  }
+
+#if defined(HAS_SPLITXRGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitXRGBRow = SplitXRGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      SplitXRGBRow = SplitXRGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    SplitXRGBRow = SplitXRGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      SplitXRGBRow = SplitXRGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitXRGBRow = SplitXRGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitXRGBRow = SplitXRGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITXRGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitXRGBRow = SplitXRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitXRGBRow = SplitXRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    SplitXRGBRow(src_argb, dst_r, dst_g, dst_b, width);
+    dst_r += dst_stride_r;
+    dst_g += dst_stride_g;
+    dst_b += dst_stride_b;
+    src_argb += src_stride_argb;
+  }
+}
+
+LIBYUV_API
+void SplitARGBPlane(const uint8_t* src_argb,
+                    int src_stride_argb,
+                    uint8_t* dst_r,
+                    int dst_stride_r,
+                    uint8_t* dst_g,
+                    int dst_stride_g,
+                    uint8_t* dst_b,
+                    int dst_stride_b,
+                    uint8_t* dst_a,
+                    int dst_stride_a,
+                    int width,
+                    int height) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_r = dst_r + (height - 1) * dst_stride_r;
+    dst_g = dst_g + (height - 1) * dst_stride_g;
+    dst_b = dst_b + (height - 1) * dst_stride_b;
+    dst_a = dst_a + (height - 1) * dst_stride_a;
+    dst_stride_r = -dst_stride_r;
+    dst_stride_g = -dst_stride_g;
+    dst_stride_b = -dst_stride_b;
+    dst_stride_a = -dst_stride_a;
+  }
+
+  if (dst_a == NULL) {
+    SplitARGBPlaneOpaque(src_argb, src_stride_argb, dst_r, dst_stride_r, dst_g,
+                         dst_stride_g, dst_b, dst_stride_b, width, height);
+  } else {
+    SplitARGBPlaneAlpha(src_argb, src_stride_argb, dst_r, dst_stride_r, dst_g,
+                        dst_stride_g, dst_b, dst_stride_b, dst_a, dst_stride_a,
+                        width, height);
+  }
+}
+
+LIBYUV_NOINLINE
+void MergeARGBPlaneAlpha(const uint8_t* src_r,
+                         int src_stride_r,
+                         const uint8_t* src_g,
+                         int src_stride_g,
+                         const uint8_t* src_b,
+                         int src_stride_b,
+                         const uint8_t* src_a,
+                         int src_stride_a,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
+  int y;
+  void (*MergeARGBRow)(const uint8_t* src_r, const uint8_t* src_g,
+                       const uint8_t* src_b, const uint8_t* src_a,
+                       uint8_t* dst_argb, int width) = MergeARGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeARGBRow = MergeARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      MergeARGBRow = MergeARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeARGBRow = MergeARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGBRow = MergeARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeARGBRow = MergeARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGBRow = MergeARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeARGBRow(src_r, src_g, src_b, src_a, dst_argb, width);
     src_r += src_stride_r;
     src_g += src_stride_g;
     src_b += src_stride_b;
-    dst_rgb += dst_stride_rgb;
+    src_a += src_stride_a;
+    dst_argb += dst_stride_argb;
   }
 }
 
-// Mirror a plane of data.
-void MirrorPlane(const uint8_t* src_y,
-                 int src_stride_y,
-                 uint8_t* dst_y,
-                 int dst_stride_y,
-                 int width,
-                 int height) {
+LIBYUV_NOINLINE
+void MergeARGBPlaneOpaque(const uint8_t* src_r,
+                          int src_stride_r,
+                          const uint8_t* src_g,
+                          int src_stride_g,
+                          const uint8_t* src_b,
+                          int src_stride_b,
+                          uint8_t* dst_argb,
+                          int dst_stride_argb,
+                          int width,
+                          int height) {
   int y;
-  void (*MirrorRow)(const uint8_t* src, uint8_t* dst, int width) = MirrorRow_C;
+  void (*MergeXRGBRow)(const uint8_t* src_r, const uint8_t* src_g,
+                       const uint8_t* src_b, uint8_t* dst_argb, int width) =
+      MergeXRGBRow_C;
+
+  assert(height > 0);
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEXRGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeXRGBRow = MergeXRGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      MergeXRGBRow = MergeXRGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXRGBRow = MergeXRGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGBRow = MergeXRGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXRGBRow = MergeXRGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGBRow = MergeXRGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXRGBRow(src_r, src_g, src_b, dst_argb, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_API
+void MergeARGBPlane(const uint8_t* src_r,
+                    int src_stride_r,
+                    const uint8_t* src_g,
+                    int src_stride_g,
+                    const uint8_t* src_b,
+                    int src_stride_b,
+                    const uint8_t* src_a,
+                    int src_stride_a,
+                    uint8_t* dst_argb,
+                    int dst_stride_argb,
+                    int width,
+                    int height) {
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
-    src_y = src_y + (height - 1) * src_stride_y;
-    src_stride_y = -src_stride_y;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
   }
-#if defined(HAS_MIRRORROW_NEON)
+
+  if (src_a == NULL) {
+    MergeARGBPlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                         src_stride_b, dst_argb, dst_stride_argb, width,
+                         height);
+  } else {
+    MergeARGBPlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                        src_stride_b, src_a, src_stride_a, dst_argb,
+                        dst_stride_argb, width, height);
+  }
+}
+
+// TODO(yuan): Support 2 bit alpha channel.
+LIBYUV_API
+void MergeXR30Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    uint8_t* dst_ar30,
+                    int dst_stride_ar30,
+                    int width,
+                    int height,
+                    int depth) {
+  int y;
+  void (*MergeXR30Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, uint8_t* dst_ar30, int depth,
+                       int width) = MergeXR30Row_C;
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar30 = dst_ar30 + (height - 1) * dst_stride_ar30;
+    dst_stride_ar30 = -dst_stride_ar30;
+  }
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_ar30 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_ar30 = 0;
+  }
+#if defined(HAS_MERGEXR30ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXR30Row = MergeXR30Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXR30Row = MergeXR30Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXR30ROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    MirrorRow = MirrorRow_Any_NEON;
+    if (depth == 10) {
+      MergeXR30Row = MergeXR30Row_10_Any_NEON;
+      if (IS_ALIGNED(width, 8)) {
+        MergeXR30Row = MergeXR30Row_10_NEON;
+      }
+    } else {
+      MergeXR30Row = MergeXR30Row_Any_NEON;
+      if (IS_ALIGNED(width, 8)) {
+        MergeXR30Row = MergeXR30Row_NEON;
+      }
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXR30Row(src_r, src_g, src_b, dst_ar30, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_ar30 += dst_stride_ar30;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeAR64PlaneAlpha(const uint16_t* src_r,
+                                int src_stride_r,
+                                const uint16_t* src_g,
+                                int src_stride_g,
+                                const uint16_t* src_b,
+                                int src_stride_b,
+                                const uint16_t* src_a,
+                                int src_stride_a,
+                                uint16_t* dst_ar64,
+                                int dst_stride_ar64,
+                                int width,
+                                int height,
+                                int depth) {
+  int y;
+  void (*MergeAR64Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, const uint16_t* src_a,
+                       uint16_t* dst_argb, int depth, int width) =
+      MergeAR64Row_C;
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_ar64 = 0;
+  }
+#if defined(HAS_MERGEAR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeAR64Row = MergeAR64Row_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
-      MirrorRow = MirrorRow_NEON;
+      MergeAR64Row = MergeAR64Row_AVX2;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    MirrorRow = MirrorRow_Any_SSSE3;
+#if defined(HAS_MERGEAR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeAR64Row = MergeAR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeAR64Row = MergeAR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeAR64Row(src_r, src_g, src_b, src_a, dst_ar64, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    src_a += src_stride_a;
+    dst_ar64 += dst_stride_ar64;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeAR64PlaneOpaque(const uint16_t* src_r,
+                                 int src_stride_r,
+                                 const uint16_t* src_g,
+                                 int src_stride_g,
+                                 const uint16_t* src_b,
+                                 int src_stride_b,
+                                 uint16_t* dst_ar64,
+                                 int dst_stride_ar64,
+                                 int width,
+                                 int height,
+                                 int depth) {
+  int y;
+  void (*MergeXR64Row)(const uint16_t* src_r, const uint16_t* src_g,
+                       const uint16_t* src_b, uint16_t* dst_argb, int depth,
+                       int width) = MergeXR64Row_C;
+
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_ar64 = 0;
+  }
+#if defined(HAS_MERGEXR64ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXR64Row = MergeXR64Row_Any_AVX2;
     if (IS_ALIGNED(width, 16)) {
-      MirrorRow = MirrorRow_SSSE3;
+      MergeXR64Row = MergeXR64Row_AVX2;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_AVX2)
+#if defined(HAS_MERGEXR64ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXR64Row = MergeXR64Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeXR64Row = MergeXR64Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    MergeXR64Row(src_r, src_g, src_b, dst_ar64, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_ar64 += dst_stride_ar64;
+  }
+}
+
+LIBYUV_API
+void MergeAR64Plane(const uint16_t* src_r,
+                    int src_stride_r,
+                    const uint16_t* src_g,
+                    int src_stride_g,
+                    const uint16_t* src_b,
+                    int src_stride_b,
+                    const uint16_t* src_a,
+                    int src_stride_a,
+                    uint16_t* dst_ar64,
+                    int dst_stride_ar64,
+                    int width,
+                    int height,
+                    int depth) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_ar64 = dst_ar64 + (height - 1) * dst_stride_ar64;
+    dst_stride_ar64 = -dst_stride_ar64;
+  }
+
+  if (src_a == NULL) {
+    MergeAR64PlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                         src_stride_b, dst_ar64, dst_stride_ar64, width, height,
+                         depth);
+  } else {
+    MergeAR64PlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                        src_stride_b, src_a, src_stride_a, dst_ar64,
+                        dst_stride_ar64, width, height, depth);
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGB16To8PlaneAlpha(const uint16_t* src_r,
+                                     int src_stride_r,
+                                     const uint16_t* src_g,
+                                     int src_stride_g,
+                                     const uint16_t* src_b,
+                                     int src_stride_b,
+                                     const uint16_t* src_a,
+                                     int src_stride_a,
+                                     uint8_t* dst_argb,
+                                     int dst_stride_argb,
+                                     int width,
+                                     int height,
+                                     int depth) {
+  int y;
+  void (*MergeARGB16To8Row)(const uint16_t* src_r, const uint16_t* src_g,
+                            const uint16_t* src_b, const uint16_t* src_a,
+                            uint8_t* dst_argb, int depth, int width) =
+      MergeARGB16To8Row_C;
+
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      src_stride_a == width && dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = src_stride_a =
+        dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEARGB16TO8ROW_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    MirrorRow = MirrorRow_Any_AVX2;
-    if (IS_ALIGNED(width, 32)) {
-      MirrorRow = MirrorRow_AVX2;
+    MergeARGB16To8Row = MergeARGB16To8Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeARGB16To8Row = MergeARGB16To8Row_AVX2;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    MirrorRow = MirrorRow_Any_MSA;
-    if (IS_ALIGNED(width, 64)) {
-      MirrorRow = MirrorRow_MSA;
+#if defined(HAS_MERGEARGB16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeARGB16To8Row = MergeARGB16To8Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      MergeARGB16To8Row = MergeARGB16To8Row_NEON;
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MirrorRow = MirrorRow_Any_MMI;
+
+  for (y = 0; y < height; ++y) {
+    MergeARGB16To8Row(src_r, src_g, src_b, src_a, dst_argb, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    src_a += src_stride_a;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_NOINLINE
+static void MergeARGB16To8PlaneOpaque(const uint16_t* src_r,
+                                      int src_stride_r,
+                                      const uint16_t* src_g,
+                                      int src_stride_g,
+                                      const uint16_t* src_b,
+                                      int src_stride_b,
+                                      uint8_t* dst_argb,
+                                      int dst_stride_argb,
+                                      int width,
+                                      int height,
+                                      int depth) {
+  int y;
+  void (*MergeXRGB16To8Row)(const uint16_t* src_r, const uint16_t* src_g,
+                            const uint16_t* src_b, uint8_t* dst_argb, int depth,
+                            int width) = MergeXRGB16To8Row_C;
+
+  // Coalesce rows.
+  if (src_stride_r == width && src_stride_g == width && src_stride_b == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_r = src_stride_g = src_stride_b = dst_stride_argb = 0;
+  }
+#if defined(HAS_MERGEXRGB16TO8ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeXRGB16To8Row = MergeXRGB16To8Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MergeXRGB16To8Row = MergeXRGB16To8Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEXRGB16TO8ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeXRGB16To8Row = MergeXRGB16To8Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
-      MirrorRow = MirrorRow_MMI;
+      MergeXRGB16To8Row = MergeXRGB16To8Row_NEON;
     }
   }
 #endif
 
-  // Mirror plane
   for (y = 0; y < height; ++y) {
-    MirrorRow(src_y, dst_y, width);
-    src_y += src_stride_y;
-    dst_y += dst_stride_y;
+    MergeXRGB16To8Row(src_r, src_g, src_b, dst_argb, depth, width);
+    src_r += src_stride_r;
+    src_g += src_stride_g;
+    src_b += src_stride_b;
+    dst_argb += dst_stride_argb;
+  }
+}
+
+LIBYUV_API
+void MergeARGB16To8Plane(const uint16_t* src_r,
+                         int src_stride_r,
+                         const uint16_t* src_g,
+                         int src_stride_g,
+                         const uint16_t* src_b,
+                         int src_stride_b,
+                         const uint16_t* src_a,
+                         int src_stride_a,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height,
+                         int depth) {
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+
+  if (src_a == NULL) {
+    MergeARGB16To8PlaneOpaque(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                              src_stride_b, dst_argb, dst_stride_argb, width,
+                              height, depth);
+  } else {
+    MergeARGB16To8PlaneAlpha(src_r, src_stride_r, src_g, src_stride_g, src_b,
+                             src_stride_b, src_a, src_stride_a, dst_argb,
+                             dst_stride_argb, width, height, depth);
   }
 }
 
@@ -820,7 +1878,7 @@ int YUY2ToI422(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_YUY2TOYROW_MSA)
+#if defined(HAS_YUY2TOYROW_MSA) && defined(HAS_YUY2TOUV422ROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     YUY2ToYRow = YUY2ToYRow_Any_MSA;
     YUY2ToUV422Row = YUY2ToUV422Row_Any_MSA;
@@ -830,13 +1888,13 @@ int YUY2ToI422(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_YUY2TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    YUY2ToYRow = YUY2ToYRow_Any_MMI;
-    YUY2ToUV422Row = YUY2ToUV422Row_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      YUY2ToYRow = YUY2ToYRow_MMI;
-      YUY2ToUV422Row = YUY2ToUV422Row_MMI;
+#if defined(HAS_YUY2TOYROW_LASX) && defined(HAS_YUY2TOUV422ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    YUY2ToYRow = YUY2ToYRow_Any_LASX;
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToYRow = YUY2ToYRow_LASX;
+      YUY2ToUV422Row = YUY2ToUV422Row_LASX;
     }
   }
 #endif
@@ -916,7 +1974,7 @@ int UYVYToI422(const uint8_t* src_uyvy,
     }
   }
 #endif
-#if defined(HAS_UYVYTOYROW_MSA)
+#if defined(HAS_UYVYTOYROW_MSA) && defined(HAS_UYVYTOUV422ROW_MSA)
   if (TestCpuFlag(kCpuHasMSA)) {
     UYVYToYRow = UYVYToYRow_Any_MSA;
     UYVYToUV422Row = UYVYToUV422Row_Any_MSA;
@@ -926,13 +1984,13 @@ int UYVYToI422(const uint8_t* src_uyvy,
     }
   }
 #endif
-#if defined(HAS_UYVYTOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    UYVYToYRow = UYVYToYRow_Any_MMI;
-    UYVYToUV422Row = UYVYToUV422Row_Any_MMI;
-    if (IS_ALIGNED(width, 16)) {
-      UYVYToYRow = UYVYToYRow_MMI;
-      UYVYToUV422Row = UYVYToUV422Row_MMI;
+#if defined(HAS_UYVYTOYROW_LASX) && defined(HAS_UYVYTOUV422ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    UYVYToYRow = UYVYToYRow_Any_LASX;
+    UYVYToUV422Row = UYVYToUV422Row_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_LASX;
+      UYVYToUV422Row = UYVYToUV422Row_LASX;
     }
   }
 #endif
@@ -1006,23 +2064,214 @@ int YUY2ToY(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_YUY2TOYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    YUY2ToYRow = YUY2ToYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      YUY2ToYRow = YUY2ToYRow_MMI;
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Convert UYVY to Y.
+LIBYUV_API
+int UYVYToY(const uint8_t* src_uyvy,
+            int src_stride_uyvy,
+            uint8_t* dst_y,
+            int dst_stride_y,
+            int width,
+            int height) {
+  int y;
+  void (*UYVYToYRow)(const uint8_t* src_uyvy, uint8_t* dst_y, int width) =
+      UYVYToYRow_C;
+  if (!src_uyvy || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 && dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    UYVYToYRow = UYVYToYRow_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToYRow = UYVYToYRow_MSA;
     }
   }
 #endif
 
   for (y = 0; y < height; ++y) {
-    YUY2ToYRow(src_yuy2, dst_y, width);
-    src_yuy2 += src_stride_yuy2;
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
     dst_y += dst_stride_y;
   }
   return 0;
 }
 
+// Mirror a plane of data.
+// See Also I400Mirror
+LIBYUV_API
+void MirrorPlane(const uint8_t* src_y,
+                 int src_stride_y,
+                 uint8_t* dst_y,
+                 int dst_stride_y,
+                 int width,
+                 int height) {
+  int y;
+  void (*MirrorRow)(const uint8_t* src, uint8_t* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorRow = MirrorRow_Any_MSA;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorRow = MirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_LASX;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Mirror a plane of UV data.
+LIBYUV_API
+void MirrorUVPlane(const uint8_t* src_uv,
+                   int src_stride_uv,
+                   uint8_t* dst_uv,
+                   int dst_stride_uv,
+                   int width,
+                   int height) {
+  int y;
+  void (*MirrorUVRow)(const uint8_t* src, uint8_t* dst, int width) =
+      MirrorUVRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorUVRow = MirrorUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorUVRow = MirrorUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorUVRow = MirrorUVRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      MirrorUVRow = MirrorUVRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorUVRow = MirrorUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorUVRow = MirrorUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    MirrorUVRow = MirrorUVRow_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      MirrorUVRow = MirrorUVRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorUVRow = MirrorUVRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorUVRow = MirrorUVRow_LASX;
+    }
+  }
+#endif
+
+  // MirrorUV plane
+  for (y = 0; y < height; ++y) {
+    MirrorUVRow(src_uv, dst_uv, width);
+    src_uv += src_stride_uv;
+    dst_uv += dst_stride_uv;
+  }
+}
+
 // Mirror I400 with optional flipping
 LIBYUV_API
 int I400Mirror(const uint8_t* src_y,
@@ -1063,10 +2312,12 @@ int I420Mirror(const uint8_t* src_y,
                int height) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v || width <= 0 ||
+
+  if (!src_y || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
       height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -1087,6 +2338,43 @@ int I420Mirror(const uint8_t* src_y,
   return 0;
 }
 
+// NV12 mirror.
+LIBYUV_API
+int NV12Mirror(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_uv,
+               int src_stride_uv,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_uv,
+               int dst_stride_uv,
+               int width,
+               int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+
+  if (!src_y || !src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorUVPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, halfwidth,
+                halfheight);
+  return 0;
+}
+
 // ARGB mirror.
 LIBYUV_API
 int ARGBMirror(const uint8_t* src_argb,
@@ -1110,7 +2398,7 @@ int ARGBMirror(const uint8_t* src_argb,
 #if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 4)) {
+    if (IS_ALIGNED(width, 8)) {
       ARGBMirrorRow = ARGBMirrorRow_NEON;
     }
   }
@@ -1139,11 +2427,11 @@ int ARGBMirror(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBMIRRORROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBMirrorRow = ARGBMirrorRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBMirrorRow = ARGBMirrorRow_MMI;
+#if defined(HAS_ARGBMIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_LASX;
     }
   }
 #endif
@@ -1157,6 +2445,52 @@ int ARGBMirror(const uint8_t* src_argb,
   return 0;
 }
 
+// RGB24 mirror.
+LIBYUV_API
+int RGB24Mirror(const uint8_t* src_rgb24,
+                int src_stride_rgb24,
+                uint8_t* dst_rgb24,
+                int dst_stride_rgb24,
+                int width,
+                int height) {
+  int y;
+  void (*RGB24MirrorRow)(const uint8_t* src, uint8_t* dst, int width) =
+      RGB24MirrorRow_C;
+  if (!src_rgb24 || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+#if defined(HAS_RGB24MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24MirrorRow = RGB24MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24MirrorRow = RGB24MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGB24MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24MirrorRow = RGB24MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24MirrorRow = RGB24MirrorRow_SSSE3;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    RGB24MirrorRow(src_rgb24, dst_rgb24, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
 // Get a blender that optimized for the CPU and pixel count.
 // As there are 6 blenders to choose from, the caller should try to use
 // the same blend function for all pixels if possible.
@@ -1180,9 +2514,9 @@ ARGBBlendRow GetARGBBlend() {
     ARGBBlendRow = ARGBBlendRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBBLENDROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBBlendRow = ARGBBlendRow_MMI;
+#if defined(HAS_ARGBBLENDROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBBlendRow = ARGBBlendRow_LSX;
   }
 #endif
   return ARGBBlendRow;
@@ -1277,14 +2611,6 @@ int BlendPlane(const uint8_t* src_y0,
     }
   }
 #endif
-#if defined(HAS_BLENDPLANEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    BlendPlaneRow = BlendPlaneRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      BlendPlaneRow = BlendPlaneRow_MMI;
-    }
-  }
-#endif
 
   for (y = 0; y < height; ++y) {
     BlendPlaneRow(src_y0, src_y1, alpha, dst_y, width);
@@ -1329,6 +2655,7 @@ int I420Blend(const uint8_t* src_y0,
       BlendPlaneRow_C;
   void (*ScaleRowDown2)(const uint8_t* src_ptr, ptrdiff_t src_stride,
                         uint8_t* dst_ptr, int dst_width) = ScaleRowDown2Box_C;
+
   if (!src_y0 || !src_u0 || !src_v0 || !src_y1 || !src_u1 || !src_v1 ||
       !alpha || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
@@ -1361,14 +2688,6 @@ int I420Blend(const uint8_t* src_y0,
     }
   }
 #endif
-#if defined(HAS_BLENDPLANEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    BlendPlaneRow = BlendPlaneRow_Any_MMI;
-    if (IS_ALIGNED(halfwidth, 8)) {
-      BlendPlaneRow = BlendPlaneRow_MMI;
-    }
-  }
-#endif
   if (!IS_ALIGNED(width, 2)) {
     ScaleRowDown2 = ScaleRowDown2Box_Odd_C;
   }
@@ -1405,17 +2724,6 @@ int I420Blend(const uint8_t* src_y0,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN2_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ScaleRowDown2 = ScaleRowDown2Box_Odd_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ScaleRowDown2 = ScaleRowDown2Box_Any_MMI;
-      if (IS_ALIGNED(halfwidth, 8)) {
-        ScaleRowDown2 = ScaleRowDown2Box_MMI;
-      }
-    }
-  }
-#endif
 
   // Row buffer for intermediate alpha pixels.
   align_buffer_64(halfalpha, halfwidth);
@@ -1501,11 +2809,11 @@ int ARGBMultiply(const uint8_t* src_argb0,
     }
   }
 #endif
-#if defined(HAS_ARGBMULTIPLYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBMultiplyRow = ARGBMultiplyRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBMultiplyRow = ARGBMultiplyRow_MMI;
+#if defined(HAS_ARGBMULTIPLYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_LASX;
     }
   }
 #endif
@@ -1549,12 +2857,12 @@ int ARGBAdd(const uint8_t* src_argb0,
     height = 1;
     src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
   }
-#if defined(HAS_ARGBADDROW_SSE2) && (defined(_MSC_VER) && !defined(__clang__))
+#if defined(HAS_ARGBADDROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAddRow = ARGBAddRow_SSE2;
   }
 #endif
-#if defined(HAS_ARGBADDROW_SSE2) && !(defined(_MSC_VER) && !defined(__clang__))
+#if defined(HAS_ARGBADDROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBAddRow = ARGBAddRow_Any_SSE2;
     if (IS_ALIGNED(width, 4)) {
@@ -1586,11 +2894,11 @@ int ARGBAdd(const uint8_t* src_argb0,
     }
   }
 #endif
-#if defined(HAS_ARGBADDROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBAddRow = ARGBAddRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBAddRow = ARGBAddRow_MMI;
+#if defined(HAS_ARGBADDROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBAddRow = ARGBAddRow_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_LASX;
     }
   }
 #endif
@@ -1666,11 +2974,11 @@ int ARGBSubtract(const uint8_t* src_argb0,
     }
   }
 #endif
-#if defined(HAS_ARGBSUBTRACTROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBSubtractRow = ARGBSubtractRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBSubtractRow = ARGBSubtractRow_MMI;
+#if defined(HAS_ARGBSUBTRACTROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_LASX;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_LASX;
     }
   }
 #endif
@@ -1684,177 +2992,6 @@ int ARGBSubtract(const uint8_t* src_argb0,
   }
   return 0;
 }
-// Convert I422 to RGBA with matrix
-static int I422ToRGBAMatrix(const uint8_t* src_y,
-                            int src_stride_y,
-                            const uint8_t* src_u,
-                            int src_stride_u,
-                            const uint8_t* src_v,
-                            int src_stride_v,
-                            uint8_t* dst_rgba,
-                            int dst_stride_rgba,
-                            const struct YuvConstants* yuvconstants,
-                            int width,
-                            int height) {
-  int y;
-  void (*I422ToRGBARow)(const uint8_t* y_buf, const uint8_t* u_buf,
-                        const uint8_t* v_buf, uint8_t* rgb_buf,
-                        const struct YuvConstants* yuvconstants, int width) =
-      I422ToRGBARow_C;
-  if (!src_y || !src_u || !src_v || !dst_rgba || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
-    dst_stride_rgba = -dst_stride_rgba;
-  }
-#if defined(HAS_I422TORGBAROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      I422ToRGBARow = I422ToRGBARow_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    I422ToRGBARow = I422ToRGBARow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_NEON;
-    }
-  }
-#endif
-#if defined(HAS_I422TORGBAROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    I422ToRGBARow = I422ToRGBARow_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      I422ToRGBARow = I422ToRGBARow_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, yuvconstants, width);
-    dst_rgba += dst_stride_rgba;
-    src_y += src_stride_y;
-    src_u += src_stride_u;
-    src_v += src_stride_v;
-  }
-  return 0;
-}
-
-// Convert I422 to RGBA.
-LIBYUV_API
-int I422ToRGBA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_rgba,
-               int dst_stride_rgba,
-               int width,
-               int height) {
-  return I422ToRGBAMatrix(src_y, src_stride_y, src_u, src_stride_u, src_v,
-                          src_stride_v, dst_rgba, dst_stride_rgba,
-                          &kYuvI601Constants, width, height);
-}
-
-// Convert I422 to BGRA.
-LIBYUV_API
-int I422ToBGRA(const uint8_t* src_y,
-               int src_stride_y,
-               const uint8_t* src_u,
-               int src_stride_u,
-               const uint8_t* src_v,
-               int src_stride_v,
-               uint8_t* dst_bgra,
-               int dst_stride_bgra,
-               int width,
-               int height) {
-  return I422ToRGBAMatrix(src_y, src_stride_y, src_v,
-                          src_stride_v,  // Swap U and V
-                          src_u, src_stride_u, dst_bgra, dst_stride_bgra,
-                          &kYvuI601Constants,  // Use Yvu matrix
-                          width, height);
-}
-
-// Convert NV12 to RGB565.
-LIBYUV_API
-int NV12ToRGB565(const uint8_t* src_y,
-                 int src_stride_y,
-                 const uint8_t* src_uv,
-                 int src_stride_uv,
-                 uint8_t* dst_rgb565,
-                 int dst_stride_rgb565,
-                 int width,
-                 int height) {
-  int y;
-  void (*NV12ToRGB565Row)(
-      const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* rgb_buf,
-      const struct YuvConstants* yuvconstants, int width) = NV12ToRGB565Row_C;
-  if (!src_y || !src_uv || !dst_rgb565 || width <= 0 || height == 0) {
-    return -1;
-  }
-  // Negative height means invert the image.
-  if (height < 0) {
-    height = -height;
-    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
-    dst_stride_rgb565 = -dst_stride_rgb565;
-  }
-#if defined(HAS_NV12TORGB565ROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
-    }
-  }
-#endif
-#if defined(HAS_NV12TORGB565ROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
-    if (IS_ALIGNED(width, 16)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
-    }
-  }
-#endif
-#if defined(HAS_NV12TORGB565ROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
-    }
-  }
-#endif
-#if defined(HAS_NV12TORGB565ROW_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    NV12ToRGB565Row = NV12ToRGB565Row_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      NV12ToRGB565Row = NV12ToRGB565Row_MSA;
-    }
-  }
-#endif
-
-  for (y = 0; y < height; ++y) {
-    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, &kYuvI601Constants, width);
-    dst_rgb565 += dst_stride_rgb565;
-    src_y += src_stride_y;
-    if (y & 1) {
-      src_uv += src_stride_uv;
-    }
-  }
-  return 0;
-}
 
 // Convert RAW to RGB24.
 LIBYUV_API
@@ -1906,11 +3043,11 @@ int RAWToRGB24(const uint8_t* src_raw,
     }
   }
 #endif
-#if defined(HAS_RAWTORGB24ROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    RAWToRGB24Row = RAWToRGB24Row_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      RAWToRGB24Row = RAWToRGB24Row_MMI;
+#if defined(HAS_RAWTORGB24ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToRGB24Row = RAWToRGB24Row_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToRGB24Row = RAWToRGB24Row_LSX;
     }
   }
 #endif
@@ -1931,6 +3068,10 @@ void SetPlane(uint8_t* dst_y,
               uint32_t value) {
   int y;
   void (*SetRow)(uint8_t * dst, uint8_t value, int width) = SetRow_C;
+
+  if (width <= 0 || height == 0) {
+    return;
+  }
   if (height < 0) {
     height = -height;
     dst_y = dst_y + (height - 1) * dst_stride_y;
@@ -1968,6 +3109,14 @@ void SetPlane(uint8_t* dst_y,
     SetRow = SetRow_MSA;
   }
 #endif
+#if defined(HAS_SETROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SetRow = SetRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SetRow = SetRow_LSX;
+    }
+  }
+#endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
@@ -1996,6 +3145,7 @@ int I420Rect(uint8_t* dst_y,
   uint8_t* start_y = dst_y + y * dst_stride_y + x;
   uint8_t* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
   uint8_t* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+
   if (!dst_y || !dst_u || !dst_v || width <= 0 || height == 0 || x < 0 ||
       y < 0 || value_y < 0 || value_y > 255 || value_u < 0 || value_u > 255 ||
       value_v < 0 || value_v > 255) {
@@ -2057,6 +3207,14 @@ int ARGBRect(uint8_t* dst_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBSETROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBSetRow = ARGBSetRow_Any_LSX;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_LSX;
+    }
+  }
+#endif
 
   // Set plane
   for (y = 0; y < height; ++y) {
@@ -2135,11 +3293,11 @@ int ARGBAttenuate(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBATTENUATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBAttenuateRow = ARGBAttenuateRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBAttenuateRow = ARGBAttenuateRow_MMI;
+#if defined(HAS_ARGBATTENUATEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_LASX;
     }
   }
 #endif
@@ -2243,9 +3401,9 @@ int ARGBGrayTo(const uint8_t* src_argb,
     ARGBGrayRow = ARGBGrayRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBGRAYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 2)) {
-    ARGBGrayRow = ARGBGrayRow_MMI;
+#if defined(HAS_ARGBGRAYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBGrayRow = ARGBGrayRow_LASX;
   }
 #endif
 
@@ -2293,9 +3451,9 @@ int ARGBGray(uint8_t* dst_argb,
     ARGBGrayRow = ARGBGrayRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBGRAYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 2)) {
-    ARGBGrayRow = ARGBGrayRow_MMI;
+#if defined(HAS_ARGBGRAYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBGrayRow = ARGBGrayRow_LASX;
   }
 #endif
 
@@ -2341,9 +3499,9 @@ int ARGBSepia(uint8_t* dst_argb,
     ARGBSepiaRow = ARGBSepiaRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBSEPIAROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 2)) {
-    ARGBSepiaRow = ARGBSepiaRow_MMI;
+#if defined(HAS_ARGBSEPIAROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 16)) {
+    ARGBSepiaRow = ARGBSepiaRow_LASX;
   }
 #endif
 
@@ -2397,9 +3555,9 @@ int ARGBColorMatrix(const uint8_t* src_argb,
     ARGBColorMatrixRow = ARGBColorMatrixRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBCOLORMATRIXROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 2)) {
-    ARGBColorMatrixRow = ARGBColorMatrixRow_MMI;
+#if defined(HAS_ARGBCOLORMATRIXROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_LSX;
   }
 #endif
   for (y = 0; y < height; ++y) {
@@ -2567,6 +3725,11 @@ int ARGBQuantize(uint8_t* dst_argb,
     ARGBQuantizeRow = ARGBQuantizeRow_MSA;
   }
 #endif
+#if defined(HAS_ARGBQUANTIZEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_LSX;
+  }
+#endif
   for (y = 0; y < height; ++y) {
     ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
     dst += dst_stride_argb;
@@ -2596,11 +3759,6 @@ int ARGBComputeCumulativeSum(const uint8_t* src_argb,
     ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
   }
 #endif
-#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ComputeCumulativeSumRow = ComputeCumulativeSumRow_MMI;
-  }
-#endif
 
   memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
   for (y = 0; y < height; ++y) {
@@ -2651,7 +3809,7 @@ int ARGBBlur(const uint8_t* src_argb,
   if (radius > (width / 2 - 1)) {
     radius = width / 2 - 1;
   }
-  if (radius <= 0) {
+  if (radius <= 0 || height <= 1) {
     return -1;
   }
 #if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
@@ -2660,11 +3818,6 @@ int ARGBBlur(const uint8_t* src_argb,
     CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
   }
 #endif
-#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ComputeCumulativeSumRow = ComputeCumulativeSumRow_MMI;
-  }
-#endif
   // Compute enough CumulativeSum for first row to be blurred. After this
   // one row of CumulativeSum is updated at a time.
   ARGBComputeCumulativeSum(src_argb, src_stride_argb, dst_cumsum,
@@ -2771,9 +3924,9 @@ int ARGBShade(const uint8_t* src_argb,
     ARGBShadeRow = ARGBShadeRow_MSA;
   }
 #endif
-#if defined(HAS_ARGBSHADEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 2)) {
-    ARGBShadeRow = ARGBShadeRow_MMI;
+#if defined(HAS_ARGBSHADEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_LASX;
   }
 #endif
 
@@ -2847,11 +4000,11 @@ int InterpolatePlane(const uint8_t* src0,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -2865,6 +4018,86 @@ int InterpolatePlane(const uint8_t* src0,
   return 0;
 }
 
+// Interpolate 2 planes by specified amount (0 to 255).
+LIBYUV_API
+int InterpolatePlane_16(const uint16_t* src0,
+                        int src_stride0,
+                        const uint16_t* src1,
+                        int src_stride1,
+                        uint16_t* dst,
+                        int dst_stride,
+                        int width,
+                        int height,
+                        int interpolation) {
+  int y;
+  void (*InterpolateRow_16)(uint16_t * dst_ptr, const uint16_t* src_ptr,
+                            ptrdiff_t src_stride, int dst_width,
+                            int source_y_fraction) = InterpolateRow_16_C;
+  if (!src0 || !src1 || !dst || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst = dst + (height - 1) * dst_stride;
+    dst_stride = -dst_stride;
+  }
+  // Coalesce rows.
+  if (src_stride0 == width && src_stride1 == width && dst_stride == width) {
+    width *= height;
+    height = 1;
+    src_stride0 = src_stride1 = dst_stride = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow_16 = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow_16 = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_MSA;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow_16 = InterpolateRow_16_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow_16 = InterpolateRow_16_LSX;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow_16(dst, src0, src1 - src0, width, interpolation);
+    src0 += src_stride0;
+    src1 += src_stride1;
+    dst += dst_stride;
+  }
+  return 0;
+}
+
 // Interpolate 2 ARGB images by specified amount (0 to 255).
 LIBYUV_API
 int ARGBInterpolate(const uint8_t* src_argb0,
@@ -2906,10 +4139,12 @@ int I420Interpolate(const uint8_t* src0_y,
                     int interpolation) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
+
   if (!src0_y || !src0_u || !src0_v || !src1_y || !src1_u || !src1_v ||
       !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
     return -1;
   }
+
   InterpolatePlane(src0_y, src0_stride_y, src1_y, src1_stride_y, dst_y,
                    dst_stride_y, width, height, interpolation);
   InterpolatePlane(src0_u, src0_stride_u, src1_u, src1_stride_u, dst_u,
@@ -2978,11 +4213,11 @@ int ARGBShuffle(const uint8_t* src_bgra,
     }
   }
 #endif
-#if defined(HAS_ARGBSHUFFLEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBShuffleRow = ARGBShuffleRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBShuffleRow = ARGBShuffleRow_MMI;
+#if defined(HAS_ARGBSHUFFLEROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_LASX;
     }
   }
 #endif
@@ -2995,6 +4230,142 @@ int ARGBShuffle(const uint8_t* src_bgra,
   return 0;
 }
 
+// Shuffle AR64 channel order.  e.g. AR64 to AB64.
+LIBYUV_API
+int AR64Shuffle(const uint16_t* src_ar64,
+                int src_stride_ar64,
+                uint16_t* dst_ar64,
+                int dst_stride_ar64,
+                const uint8_t* shuffler,
+                int width,
+                int height) {
+  int y;
+  void (*AR64ShuffleRow)(const uint8_t* src_ar64, uint8_t* dst_ar64,
+                         const uint8_t* shuffler, int width) = AR64ShuffleRow_C;
+  if (!src_ar64 || !dst_ar64 || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_ar64 = src_ar64 + (height - 1) * src_stride_ar64;
+    src_stride_ar64 = -src_stride_ar64;
+  }
+  // Coalesce rows.
+  if (src_stride_ar64 == width * 4 && dst_stride_ar64 == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_ar64 = dst_stride_ar64 = 0;
+  }
+  // Assembly versions can be reused if it's implemented with shuffle.
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      AR64ShuffleRow = ARGBShuffleRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      AR64ShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    AR64ShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      AR64ShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    AR64ShuffleRow((uint8_t*)(src_ar64), (uint8_t*)(dst_ar64), shuffler,
+                   width * 2);
+    src_ar64 += src_stride_ar64;
+    dst_ar64 += dst_stride_ar64;
+  }
+  return 0;
+}
+
+// Gauss blur a float plane using Gaussian 5x5 filter with
+// coefficients of 1, 4, 6, 4, 1.
+// Each destination pixel is a blur of the 5x5
+// pixels from the source.
+// Source edges are clamped.
+// Edge is 2 pixels on each side, and interior is multiple of 4.
+LIBYUV_API
+int GaussPlane_F32(const float* src,
+                   int src_stride,
+                   float* dst,
+                   int dst_stride,
+                   int width,
+                   int height) {
+  int y;
+  void (*GaussCol_F32)(const float* src0, const float* src1, const float* src2,
+                       const float* src3, const float* src4, float* dst,
+                       int width) = GaussCol_F32_C;
+  void (*GaussRow_F32)(const float* src, float* dst, int width) =
+      GaussRow_F32_C;
+  if (!src || !dst || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+#if defined(HAS_GAUSSCOL_F32_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    GaussCol_F32 = GaussCol_F32_NEON;
+  }
+#endif
+#if defined(HAS_GAUSSROW_F32_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    GaussRow_F32 = GaussRow_F32_NEON;
+  }
+#endif
+  {
+    // 2 pixels on each side, but aligned out to 16 bytes.
+    align_buffer_64(rowbuf, (4 + width + 4) * 4);
+    memset(rowbuf, 0, 16);
+    memset(rowbuf + (4 + width) * 4, 0, 16);
+    float* row = (float*)(rowbuf + 16);
+    const float* src0 = src;
+    const float* src1 = src;
+    const float* src2 = src;
+    const float* src3 = src2 + ((height > 1) ? src_stride : 0);
+    const float* src4 = src3 + ((height > 2) ? src_stride : 0);
+
+    for (y = 0; y < height; ++y) {
+      GaussCol_F32(src0, src1, src2, src3, src4, row, width);
+
+      // Extrude edge by 2 floats
+      row[-2] = row[-1] = row[0];
+      row[width + 1] = row[width] = row[width - 1];
+
+      GaussRow_F32(row - 2, dst, width);
+
+      src0 = src1;
+      src1 = src2;
+      src2 = src3;
+      src3 = src4;
+      if ((y + 2) < (height - 1)) {
+        src4 += src_stride;
+      }
+      dst += dst_stride;
+    }
+    free_aligned_buffer_64(rowbuf);
+  }
+  return 0;
+}
+
 // Sobel ARGB effect.
 static int ARGBSobelize(const uint8_t* src_argb,
                         int src_stride_argb,
@@ -3044,7 +4415,7 @@ static int ARGBSobelize(const uint8_t* src_argb,
 #if defined(HAS_ARGBTOYJROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBToYJRow = ARGBToYJRow_Any_NEON;
-    if (IS_ALIGNED(width, 8)) {
+    if (IS_ALIGNED(width, 16)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
   }
@@ -3057,11 +4428,19 @@ static int ARGBSobelize(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBTOYJROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBToYJRow = ARGBToYJRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBToYJRow = ARGBToYJRow_MMI;
+#if defined(HAS_ARGBTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_LASX;
     }
   }
 #endif
@@ -3081,11 +4460,6 @@ static int ARGBSobelize(const uint8_t* src_argb,
     SobelYRow = SobelYRow_MSA;
   }
 #endif
-#if defined(HAS_SOBELYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SobelYRow = SobelYRow_MMI;
-  }
-#endif
 #if defined(HAS_SOBELXROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     SobelXRow = SobelXRow_SSE2;
@@ -3101,11 +4475,6 @@ static int ARGBSobelize(const uint8_t* src_argb,
     SobelXRow = SobelXRow_MSA;
   }
 #endif
-#if defined(HAS_SOBELXROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SobelXRow = SobelXRow_MMI;
-  }
-#endif
   {
     // 3 rows with edges before/after.
     const int kRowSize = (width + kEdge + 31) & ~31;
@@ -3188,11 +4557,11 @@ int ARGBSobel(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_SOBELROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SobelRow = SobelRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SobelRow = SobelRow_MMI;
+#if defined(HAS_SOBELROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelRow = SobelRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SobelRow = SobelRow_LSX;
     }
   }
 #endif
@@ -3234,11 +4603,11 @@ int ARGBSobelToPlane(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_SOBELTOPLANEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SobelToPlaneRow = SobelToPlaneRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SobelToPlaneRow = SobelToPlaneRow_MMI;
+#if defined(HAS_SOBELTOPLANEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SobelToPlaneRow = SobelToPlaneRow_LSX;
     }
   }
 #endif
@@ -3281,11 +4650,11 @@ int ARGBSobelXY(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_SOBELXYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SobelXYRow = SobelXYRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SobelXYRow = SobelXYRow_MMI;
+#if defined(HAS_SOBELXYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SobelXYRow = SobelXYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      SobelXYRow = SobelXYRow_LSX;
     }
   }
 #endif
@@ -3412,6 +4781,14 @@ int HalfFloatPlane(const uint16_t* src_y,
     }
   }
 #endif
+#if defined(HAS_HALFFLOATROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    HalfFloatRow = HalfFloatRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      HalfFloatRow = HalfFloatRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     HalfFloatRow(src_y, dst_y, scale, width);
@@ -3526,14 +4903,6 @@ int ARGBCopyAlpha(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBCOPYALPHAROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBCopyAlphaRow = ARGBCopyAlphaRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBCopyAlphaRow = ARGBCopyAlphaRow_MMI;
-    }
-  }
-#endif
 
   for (y = 0; y < height; ++y) {
     ARGBCopyAlphaRow(src_argb, dst_argb, width);
@@ -3592,10 +4961,10 @@ int ARGBExtractAlpha(const uint8_t* src_argb,
                                                 : ARGBExtractAlphaRow_Any_MSA;
   }
 #endif
-#if defined(HAS_ARGBEXTRACTALPHAROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBExtractAlphaRow = IS_ALIGNED(width, 8) ? ARGBExtractAlphaRow_MMI
-                                               : ARGBExtractAlphaRow_Any_MMI;
+#if defined(HAS_ARGBEXTRACTALPHAROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_LSX
+                                                : ARGBExtractAlphaRow_Any_LSX;
   }
 #endif
 
@@ -3649,14 +5018,6 @@ int ARGBCopyYToAlpha(const uint8_t* src_y,
     }
   }
 #endif
-#if defined(HAS_ARGBCOPYYTOALPHAROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_MMI;
-    }
-  }
-#endif
 
   for (y = 0; y < height; ++y) {
     ARGBCopyYToAlphaRow(src_y, dst_argb, width);
@@ -3685,9 +5046,11 @@ int YUY2ToNV12(const uint8_t* src_yuy2,
   void (*InterpolateRow)(uint8_t * dst_ptr, const uint8_t* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
+
   if (!src_yuy2 || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -3726,11 +5089,11 @@ int YUY2ToNV12(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SplitUVRow = SplitUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SplitUVRow = SplitUVRow_MMI;
+#if defined(HAS_SPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SplitUVRow = SplitUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_LSX;
     }
   }
 #endif
@@ -3766,11 +5129,11 @@ int YUY2ToNV12(const uint8_t* src_yuy2,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -3817,9 +5180,11 @@ int UYVYToNV12(const uint8_t* src_uyvy,
   void (*InterpolateRow)(uint8_t * dst_ptr, const uint8_t* src_ptr,
                          ptrdiff_t src_stride, int dst_width,
                          int source_y_fraction) = InterpolateRow_C;
+
   if (!src_uyvy || !dst_y || !dst_uv || width <= 0 || height == 0) {
     return -1;
   }
+
   // Negative height means invert the image.
   if (height < 0) {
     height = -height;
@@ -3858,11 +5223,11 @@ int UYVYToNV12(const uint8_t* src_uyvy,
     }
   }
 #endif
-#if defined(HAS_SPLITUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    SplitUVRow = SplitUVRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      SplitUVRow = SplitUVRow_MMI;
+#if defined(HAS_SPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    SplitUVRow = SplitUVRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_LSX;
     }
   }
 #endif
@@ -3898,11 +5263,11 @@ int UYVYToNV12(const uint8_t* src_uyvy,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -3933,6 +5298,56 @@ int UYVYToNV12(const uint8_t* src_uyvy,
   return 0;
 }
 
+// width and height are src size allowing odd size handling.
+LIBYUV_API
+void HalfMergeUVPlane(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int dst_stride_uv,
+                      int width,
+                      int height) {
+  int y;
+  void (*HalfMergeUVRow)(const uint8_t* src_u, int src_stride_u,
+                         const uint8_t* src_v, int src_stride_v,
+                         uint8_t* dst_uv, int width) = HalfMergeUVRow_C;
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+#if defined(HAS_HALFMERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    HalfMergeUVRow = HalfMergeUVRow_NEON;
+  }
+#endif
+#if defined(HAS_HALFMERGEUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
+    HalfMergeUVRow = HalfMergeUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_HALFMERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 32)) {
+    HalfMergeUVRow = HalfMergeUVRow_AVX2;
+  }
+#endif
+  for (y = 0; y < height - 1; y += 2) {
+    // Merge a row of U and V into a row of UV.
+    HalfMergeUVRow(src_u, src_stride_u, src_v, src_stride_v, dst_uv, width);
+    src_u += src_stride_u * 2;
+    src_v += src_stride_v * 2;
+    dst_uv += dst_stride_uv;
+  }
+  if (height & 1) {
+    HalfMergeUVRow(src_u, 0, src_v, 0, dst_uv, width);
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/rotate.cc b/files/source/rotate.cc
index d414186a..f1e83cbd 100644
--- a/files/source/rotate.cc
+++ b/files/source/rotate.cc
@@ -29,16 +29,20 @@ void TransposePlane(const uint8_t* src,
                     int width,
                     int height) {
   int i = height;
-#if defined(HAS_TRANSPOSEWX16_MSA)
+#if defined(HAS_TRANSPOSEWX16_MSA) || defined(HAS_TRANSPOSEWX16_LSX)
   void (*TransposeWx16)(const uint8_t* src, int src_stride, uint8_t* dst,
                         int dst_stride, int width) = TransposeWx16_C;
 #else
   void (*TransposeWx8)(const uint8_t* src, int src_stride, uint8_t* dst,
                        int dst_stride, int width) = TransposeWx8_C;
 #endif
+
 #if defined(HAS_TRANSPOSEWX8_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    TransposeWx8 = TransposeWx8_NEON;
+    TransposeWx8 = TransposeWx8_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeWx8 = TransposeWx8_NEON;
+    }
   }
 #endif
 #if defined(HAS_TRANSPOSEWX8_SSSE3)
@@ -49,11 +53,6 @@ void TransposePlane(const uint8_t* src,
     }
   }
 #endif
-#if defined(HAS_TRANSPOSEWX8_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    TransposeWx8 = TransposeWx8_MMI;
-  }
-#endif
 #if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     TransposeWx8 = TransposeWx8_Fast_Any_SSSE3;
@@ -70,8 +69,16 @@ void TransposePlane(const uint8_t* src,
     }
   }
 #endif
+#if defined(HAS_TRANSPOSEWX16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    TransposeWx16 = TransposeWx16_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx16 = TransposeWx16_LSX;
+    }
+  }
+#endif
 
-#if defined(HAS_TRANSPOSEWX16_MSA)
+#if defined(HAS_TRANSPOSEWX16_MSA) || defined(HAS_TRANSPOSEWX16_LSX)
   // Work across the source in 16x16 tiles
   while (i >= 16) {
     TransposeWx16(src, src_stride, dst, dst_stride, width);
@@ -142,7 +149,7 @@ void RotatePlane180(const uint8_t* src,
 #if defined(HAS_MIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     MirrorRow = MirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 16)) {
+    if (IS_ALIGNED(width, 32)) {
       MirrorRow = MirrorRow_NEON;
     }
   }
@@ -171,11 +178,11 @@ void RotatePlane180(const uint8_t* src,
     }
   }
 #endif
-#if defined(HAS_MIRRORROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    MirrorRow = MirrorRow_Any_MMI;
-    if (IS_ALIGNED(width, 8)) {
-      MirrorRow = MirrorRow_MMI;
+#if defined(HAS_MIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    MirrorRow = MirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 64)) {
+      MirrorRow = MirrorRow_LASX;
     }
   }
 #endif
@@ -199,19 +206,14 @@ void RotatePlane180(const uint8_t* src,
     CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
   }
 #endif
-#if defined(HAS_COPYROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    CopyRow = IS_ALIGNED(width, 8) ? CopyRow_MMI : CopyRow_Any_MMI;
-  }
-#endif
 
   // Odd height will harmlessly mirror the middle row twice.
   for (y = 0; y < half_height; ++y) {
-    MirrorRow(src, row, width);  // Mirror first row into a buffer
-    src += src_stride;
+    CopyRow(src, row, width);        // Copy first row into buffer
     MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    MirrorRow(row, dst_bot, width);  // Mirror buffer into last row
+    src += src_stride;
     dst += dst_stride;
-    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
     src_bot -= src_stride;
     dst_bot -= dst_stride;
   }
@@ -219,24 +221,44 @@ void RotatePlane180(const uint8_t* src,
 }
 
 LIBYUV_API
-void TransposeUV(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height) {
+void SplitTransposeUV(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   int i = height;
 #if defined(HAS_TRANSPOSEUVWX16_MSA)
   void (*TransposeUVWx16)(const uint8_t* src, int src_stride, uint8_t* dst_a,
                           int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
                           int width) = TransposeUVWx16_C;
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  void (*TransposeUVWx16)(const uint8_t* src, int src_stride, uint8_t* dst_a,
+                          int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
+                          int width) = TransposeUVWx16_C;
 #else
   void (*TransposeUVWx8)(const uint8_t* src, int src_stride, uint8_t* dst_a,
                          int dst_stride_a, uint8_t* dst_b, int dst_stride_b,
                          int width) = TransposeUVWx8_C;
 #endif
+
+#if defined(HAS_TRANSPOSEUVWX16_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    TransposeUVWx16 = TransposeUVWx16_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx16 = TransposeUVWx16_MSA;
+    }
+  }
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    TransposeUVWx16 = TransposeUVWx16_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeUVWx16 = TransposeUVWx16_LSX;
+    }
+  }
+#else
 #if defined(HAS_TRANSPOSEUVWX8_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     TransposeUVWx8 = TransposeUVWx8_NEON;
@@ -250,22 +272,7 @@ void TransposeUV(const uint8_t* src,
     }
   }
 #endif
-#if defined(HAS_TRANSPOSEUVWX8_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    TransposeUVWx8 = TransposeUVWx8_Any_MMI;
-    if (IS_ALIGNED(width, 4)) {
-      TransposeUVWx8 = TransposeUVWx8_MMI;
-    }
-  }
-#endif
-#if defined(HAS_TRANSPOSEUVWX16_MSA)
-  if (TestCpuFlag(kCpuHasMSA)) {
-    TransposeUVWx16 = TransposeUVWx16_Any_MSA;
-    if (IS_ALIGNED(width, 8)) {
-      TransposeUVWx16 = TransposeUVWx16_MSA;
-    }
-  }
-#endif
+#endif /* defined(HAS_TRANSPOSEUVWX16_MSA) */
 
 #if defined(HAS_TRANSPOSEUVWX16_MSA)
   // Work through the source in 8x8 tiles.
@@ -277,6 +284,16 @@ void TransposeUV(const uint8_t* src,
     dst_b += 16;             // Move over 8 columns.
     i -= 16;
   }
+#elif defined(HAS_TRANSPOSEUVWX16_LSX)
+  // Work through the source in 8x8 tiles.
+  while (i >= 16) {
+    TransposeUVWx16(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                    width);
+    src += 16 * src_stride;  // Go down 16 rows.
+    dst_a += 16;             // Move over 8 columns.
+    dst_b += 16;             // Move over 8 columns.
+    i -= 16;
+  }
 #else
   // Work through the source in 8x8 tiles.
   while (i >= 8) {
@@ -296,70 +313,70 @@ void TransposeUV(const uint8_t* src,
 }
 
 LIBYUV_API
-void RotateUV90(const uint8_t* src,
-                int src_stride,
-                uint8_t* dst_a,
-                int dst_stride_a,
-                uint8_t* dst_b,
-                int dst_stride_b,
-                int width,
-                int height) {
+void SplitRotateUV90(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst_a,
+                     int dst_stride_a,
+                     uint8_t* dst_b,
+                     int dst_stride_b,
+                     int width,
+                     int height) {
   src += src_stride * (height - 1);
   src_stride = -src_stride;
 
-  TransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, width,
-              height);
+  SplitTransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width, height);
 }
 
 LIBYUV_API
-void RotateUV270(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height) {
+void SplitRotateUV270(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   dst_a += dst_stride_a * (width - 1);
   dst_b += dst_stride_b * (width - 1);
   dst_stride_a = -dst_stride_a;
   dst_stride_b = -dst_stride_b;
 
-  TransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b, width,
-              height);
+  SplitTransposeUV(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width, height);
 }
 
 // Rotate 180 is a horizontal and vertical flip.
 LIBYUV_API
-void RotateUV180(const uint8_t* src,
-                 int src_stride,
-                 uint8_t* dst_a,
-                 int dst_stride_a,
-                 uint8_t* dst_b,
-                 int dst_stride_b,
-                 int width,
-                 int height) {
+void SplitRotateUV180(const uint8_t* src,
+                      int src_stride,
+                      uint8_t* dst_a,
+                      int dst_stride_a,
+                      uint8_t* dst_b,
+                      int dst_stride_b,
+                      int width,
+                      int height) {
   int i;
-  void (*MirrorUVRow)(const uint8_t* src, uint8_t* dst_u, uint8_t* dst_v,
-                      int width) = MirrorUVRow_C;
-#if defined(HAS_MIRRORUVROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
-    MirrorUVRow = MirrorUVRow_NEON;
+  void (*MirrorSplitUVRow)(const uint8_t* src, uint8_t* dst_u, uint8_t* dst_v,
+                           int width) = MirrorSplitUVRow_C;
+#if defined(HAS_MIRRORSPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
+    MirrorSplitUVRow = MirrorSplitUVRow_NEON;
   }
 #endif
-#if defined(HAS_MIRRORUVROW_SSSE3)
+#if defined(HAS_MIRRORSPLITUVROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
-    MirrorUVRow = MirrorUVRow_SSSE3;
+    MirrorSplitUVRow = MirrorSplitUVRow_SSSE3;
   }
 #endif
-#if defined(HAS_MIRRORUVROW_MSA)
+#if defined(HAS_MIRRORSPLITUVROW_MSA)
   if (TestCpuFlag(kCpuHasMSA) && IS_ALIGNED(width, 32)) {
-    MirrorUVRow = MirrorUVRow_MSA;
+    MirrorSplitUVRow = MirrorSplitUVRow_MSA;
   }
 #endif
-#if defined(HAS_MIRRORUVROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(width, 8)) {
-    MirrorUVRow = MirrorUVRow_MMI;
+#if defined(HAS_MIRRORSPLITUVROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && IS_ALIGNED(width, 32)) {
+    MirrorSplitUVRow = MirrorSplitUVRow_LSX;
   }
 #endif
 
@@ -367,13 +384,59 @@ void RotateUV180(const uint8_t* src,
   dst_b += dst_stride_b * (height - 1);
 
   for (i = 0; i < height; ++i) {
-    MirrorUVRow(src, dst_a, dst_b, width);
+    MirrorSplitUVRow(src, dst_a, dst_b, width);
     src += src_stride;
     dst_a -= dst_stride_a;
     dst_b -= dst_stride_b;
   }
 }
 
+// Rotate UV and split into planar.
+// width and height expected to be half size for NV12
+LIBYUV_API
+int SplitRotateUV(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_u,
+                  int dst_stride_u,
+                  uint8_t* dst_v,
+                  int dst_stride_v,
+                  int width,
+                  int height,
+                  enum RotationMode mode) {
+  if (!src_uv || width <= 0 || height == 0 || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      SplitUVPlane(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                   dst_stride_v, width, height);
+      return 0;
+    case kRotate90:
+      SplitRotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                      dst_stride_v, width, height);
+      return 0;
+    case kRotate270:
+      SplitRotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, width, height);
+      return 0;
+    case kRotate180:
+      SplitRotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
 LIBYUV_API
 int RotatePlane(const uint8_t* src,
                 int src_stride,
@@ -431,8 +494,8 @@ int I420Rotate(const uint8_t* src_y,
                enum RotationMode mode) {
   int halfwidth = (width + 1) >> 1;
   int halfheight = (height + 1) >> 1;
-  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
-      !dst_u || !dst_v) {
+  if ((!src_y && dst_y) || !src_u || !src_v || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
     return -1;
   }
 
@@ -482,6 +545,80 @@ int I420Rotate(const uint8_t* src_y,
 }
 
 LIBYUV_API
+int I422Rotate(const uint8_t* src_y,
+               int src_stride_y,
+               const uint8_t* src_u,
+               int src_stride_u,
+               const uint8_t* src_v,
+               int src_stride_v,
+               uint8_t* dst_y,
+               int dst_stride_y,
+               uint8_t* dst_u,
+               int dst_stride_u,
+               uint8_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+      CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+      return 0;
+    case kRotate90:
+      // We need to rotate and rescale, we use plane Y as temporal storage.
+      RotatePlane90(src_u, src_stride_u, dst_y, height, halfwidth, height);
+      ScalePlane(dst_y, height, height, halfwidth, dst_u, halfheight,
+                 halfheight, width, kFilterBilinear);
+      RotatePlane90(src_v, src_stride_v, dst_y, height, halfwidth, height);
+      ScalePlane(dst_y, height, height, halfwidth, dst_v, halfheight,
+                 halfheight, width, kFilterLinear);
+      RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      return 0;
+    case kRotate270:
+      // We need to rotate and rescale, we use plane Y as temporal storage.
+      RotatePlane270(src_u, src_stride_u, dst_y, height, halfwidth, height);
+      ScalePlane(dst_y, height, height, halfwidth, dst_u, halfheight,
+                 halfheight, width, kFilterBilinear);
+      RotatePlane270(src_v, src_stride_v, dst_y, height, halfwidth, height);
+      ScalePlane(dst_y, height, height, halfwidth, dst_v, halfheight,
+                 halfheight, width, kFilterLinear);
+      RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                     height);
+      RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                     height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
 int I444Rotate(const uint8_t* src_y,
                int src_stride_y,
                const uint8_t* src_u,
@@ -496,7 +633,7 @@ int I444Rotate(const uint8_t* src_y,
                int dst_stride_v,
                int width,
                int height,
-               enum libyuv::RotationMode mode) {
+               enum RotationMode mode) {
   if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
       !dst_u || !dst_v) {
     return -1;
@@ -514,23 +651,23 @@ int I444Rotate(const uint8_t* src_y,
   }
 
   switch (mode) {
-    case libyuv::kRotate0:
+    case kRotate0:
       // copy frame
       CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
       CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
       return 0;
-    case libyuv::kRotate90:
+    case kRotate90:
       RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       RotatePlane90(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
       RotatePlane90(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
       return 0;
-    case libyuv::kRotate270:
+    case kRotate270:
       RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       RotatePlane270(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
       RotatePlane270(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
       return 0;
-    case libyuv::kRotate180:
+    case kRotate180:
       RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       RotatePlane180(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
       RotatePlane180(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
@@ -580,18 +717,18 @@ int NV12ToI420Rotate(const uint8_t* src_y,
                         width, height);
     case kRotate90:
       RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-      RotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
-                 dst_stride_v, halfwidth, halfheight);
+      SplitRotateUV90(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                      dst_stride_v, halfwidth, halfheight);
       return 0;
     case kRotate270:
       RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-      RotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
-                  dst_stride_v, halfwidth, halfheight);
+      SplitRotateUV270(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, halfwidth, halfheight);
       return 0;
     case kRotate180:
       RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-      RotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
-                  dst_stride_v, halfwidth, halfheight);
+      SplitRotateUV180(src_uv, src_stride_uv, dst_u, dst_stride_u, dst_v,
+                       dst_stride_v, halfwidth, halfheight);
       return 0;
     default:
       break;
@@ -599,6 +736,98 @@ int NV12ToI420Rotate(const uint8_t* src_y,
   return -1;
 }
 
+static void SplitPixels(const uint8_t* src_u,
+                        int src_pixel_stride_uv,
+                        uint8_t* dst_u,
+                        int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst_u = *src_u;
+    ++dst_u;
+    src_u += src_pixel_stride_uv;
+  }
+}
+
+// Convert Android420 to I420 with Rotate
+LIBYUV_API
+int Android420ToI420Rotate(const uint8_t* src_y,
+                           int src_stride_y,
+                           const uint8_t* src_u,
+                           int src_stride_u,
+                           const uint8_t* src_v,
+                           int src_stride_v,
+                           int src_pixel_stride_uv,
+                           uint8_t* dst_y,
+                           int dst_stride_y,
+                           uint8_t* dst_u,
+                           int dst_stride_u,
+                           uint8_t* dst_v,
+                           int dst_stride_v,
+                           int width,
+                           int height,
+                           enum RotationMode rotation) {
+  int y;
+  const ptrdiff_t vu_off = src_v - src_u;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if ((!src_y && dst_y) || !src_u || !src_v || !dst_u || !dst_v || width <= 0 ||
+      height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    RotatePlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height,
+                rotation);
+  }
+
+  // Copy UV planes - I420
+  if (src_pixel_stride_uv == 1) {
+    RotatePlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight,
+                rotation);
+    RotatePlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight,
+                rotation);
+    return 0;
+  }
+  // Split UV planes - NV21
+  if (src_pixel_stride_uv == 2 && vu_off == -1 &&
+      src_stride_u == src_stride_v) {
+    SplitRotateUV(src_v, src_stride_v, dst_v, dst_stride_v, dst_u, dst_stride_u,
+                  halfwidth, halfheight, rotation);
+    return 0;
+  }
+  // Split UV planes - NV12
+  if (src_pixel_stride_uv == 2 && vu_off == 1 && src_stride_u == src_stride_v) {
+    SplitRotateUV(src_u, src_stride_u, dst_u, dst_stride_u, dst_v, dst_stride_v,
+                  halfwidth, halfheight, rotation);
+    return 0;
+  }
+
+  if (rotation == 0) {
+    for (y = 0; y < halfheight; ++y) {
+      SplitPixels(src_u, src_pixel_stride_uv, dst_u, halfwidth);
+      SplitPixels(src_v, src_pixel_stride_uv, dst_v, halfwidth);
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    return 0;
+  }
+  // unsupported type and/or rotation.
+  return -1;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/rotate_any.cc b/files/source/rotate_any.cc
index b3baf084..88ca7876 100644
--- a/files/source/rotate_any.cc
+++ b/files/source/rotate_any.cc
@@ -35,15 +35,15 @@ TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, 7)
 #ifdef HAS_TRANSPOSEWX8_SSSE3
 TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, 7)
 #endif
-#ifdef HAS_TRANSPOSEWX8_MMI
-TANY(TransposeWx8_Any_MMI, TransposeWx8_MMI, 7)
-#endif
 #ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
 TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, 15)
 #endif
 #ifdef HAS_TRANSPOSEWX16_MSA
 TANY(TransposeWx16_Any_MSA, TransposeWx16_MSA, 15)
 #endif
+#ifdef HAS_TRANSPOSEWX16_LSX
+TANY(TransposeWx16_Any_LSX, TransposeWx16_LSX, 15)
+#endif
 #undef TANY
 
 #define TUVANY(NAMEANY, TPOS_SIMD, MASK)                                       \
@@ -65,12 +65,12 @@ TUVANY(TransposeUVWx8_Any_NEON, TransposeUVWx8_NEON, 7)
 #ifdef HAS_TRANSPOSEUVWX8_SSE2
 TUVANY(TransposeUVWx8_Any_SSE2, TransposeUVWx8_SSE2, 7)
 #endif
-#ifdef HAS_TRANSPOSEUVWX8_MMI
-TUVANY(TransposeUVWx8_Any_MMI, TransposeUVWx8_MMI, 7)
-#endif
 #ifdef HAS_TRANSPOSEUVWX16_MSA
 TUVANY(TransposeUVWx16_Any_MSA, TransposeUVWx16_MSA, 7)
 #endif
+#ifdef HAS_TRANSPOSEUVWX16_LSX
+TUVANY(TransposeUVWx16_Any_LSX, TransposeUVWx16_LSX, 7)
+#endif
 #undef TUVANY
 
 #ifdef __cplusplus
diff --git a/files/source/rotate_argb.cc b/files/source/rotate_argb.cc
index a93fd55f..539cf98d 100644
--- a/files/source/rotate_argb.cc
+++ b/files/source/rotate_argb.cc
@@ -21,17 +21,21 @@ namespace libyuv {
 extern "C" {
 #endif
 
-static void ARGBTranspose(const uint8_t* src_argb,
-                          int src_stride_argb,
-                          uint8_t* dst_argb,
-                          int dst_stride_argb,
-                          int width,
-                          int height) {
+static int ARGBTranspose(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   int i;
   int src_pixel_step = src_stride_argb >> 2;
   void (*ScaleARGBRowDownEven)(
       const uint8_t* src_argb, ptrdiff_t src_stride_argb, int src_step,
       uint8_t* dst_argb, int dst_width) = ScaleARGBRowDownEven_C;
+  // Check stride is a multiple of 4.
+  if (src_stride_argb & 3) {
+    return -1;
+  }
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_SSE2;
@@ -56,11 +60,11 @@ static void ARGBTranspose(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBROWDOWNEVEN_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_MMI;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_Any_LSX;
     if (IS_ALIGNED(height, 4)) {  // Width of dest.
-      ScaleARGBRowDownEven = ScaleARGBRowDownEven_MMI;
+      ScaleARGBRowDownEven = ScaleARGBRowDownEven_LSX;
     }
   }
 #endif
@@ -70,44 +74,45 @@ static void ARGBTranspose(const uint8_t* src_argb,
     dst_argb += dst_stride_argb;
     src_argb += 4;
   }
+  return 0;
 }
 
-void ARGBRotate90(const uint8_t* src_argb,
-                  int src_stride_argb,
-                  uint8_t* dst_argb,
-                  int dst_stride_argb,
-                  int width,
-                  int height) {
+static int ARGBRotate90(const uint8_t* src_argb,
+                        int src_stride_argb,
+                        uint8_t* dst_argb,
+                        int dst_stride_argb,
+                        int width,
+                        int height) {
   // Rotate by 90 is a ARGBTranspose with the source read
   // from bottom to top. So set the source pointer to the end
   // of the buffer and flip the sign of the source stride.
   src_argb += src_stride_argb * (height - 1);
   src_stride_argb = -src_stride_argb;
-  ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
-                height);
+  return ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                       width, height);
 }
 
-void ARGBRotate270(const uint8_t* src_argb,
-                   int src_stride_argb,
-                   uint8_t* dst_argb,
-                   int dst_stride_argb,
-                   int width,
-                   int height) {
+static int ARGBRotate270(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   // Rotate by 270 is a ARGBTranspose with the destination written
   // from bottom to top. So set the destination pointer to the end
   // of the buffer and flip the sign of the destination stride.
   dst_argb += dst_stride_argb * (width - 1);
   dst_stride_argb = -dst_stride_argb;
-  ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
-                height);
+  return ARGBTranspose(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                       width, height);
 }
 
-void ARGBRotate180(const uint8_t* src_argb,
-                   int src_stride_argb,
-                   uint8_t* dst_argb,
-                   int dst_stride_argb,
-                   int width,
-                   int height) {
+static int ARGBRotate180(const uint8_t* src_argb,
+                         int src_stride_argb,
+                         uint8_t* dst_argb,
+                         int dst_stride_argb,
+                         int width,
+                         int height) {
   // Swap first and last row and mirror the content. Uses a temporary row.
   align_buffer_64(row, width * 4);
   const uint8_t* src_bot = src_argb + src_stride_argb * (height - 1);
@@ -121,7 +126,7 @@ void ARGBRotate180(const uint8_t* src_argb,
 #if defined(HAS_ARGBMIRRORROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
-    if (IS_ALIGNED(width, 4)) {
+    if (IS_ALIGNED(width, 8)) {
       ARGBMirrorRow = ARGBMirrorRow_NEON;
     }
   }
@@ -150,11 +155,11 @@ void ARGBRotate180(const uint8_t* src_argb,
     }
   }
 #endif
-#if defined(HAS_ARGBMIRRORROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ARGBMirrorRow = ARGBMirrorRow_Any_MMI;
-    if (IS_ALIGNED(width, 2)) {
-      ARGBMirrorRow = ARGBMirrorRow_MMI;
+#if defined(HAS_ARGBMIRRORROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBMirrorRow = ARGBMirrorRow_LASX;
     }
   }
 #endif
@@ -190,6 +195,7 @@ void ARGBRotate180(const uint8_t* src_argb,
     dst_bot -= dst_stride_argb;
   }
   free_aligned_buffer_64(row);
+  return 0;
 }
 
 LIBYUV_API
@@ -217,17 +223,14 @@ int ARGBRotate(const uint8_t* src_argb,
       return ARGBCopy(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
                       width, height);
     case kRotate90:
-      ARGBRotate90(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
-                   height);
-      return 0;
+      return ARGBRotate90(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                          width, height);
     case kRotate270:
-      ARGBRotate270(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
-                    height);
-      return 0;
+      return ARGBRotate270(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                           width, height);
     case kRotate180:
-      ARGBRotate180(src_argb, src_stride_argb, dst_argb, dst_stride_argb, width,
-                    height);
-      return 0;
+      return ARGBRotate180(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                           width, height);
     default:
       break;
   }
diff --git a/files/source/rotate_dspr2.cc b/files/source/rotate_dspr2.cc
deleted file mode 100644
index 5d2338de..00000000
--- a/files/source/rotate_dspr2.cc
+++ /dev/null
@@ -1,475 +0,0 @@
-/*
- *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/rotate_row.h"
-#include "libyuv/row.h"
-
-#include "libyuv/basic_types.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips_dsp) && \
-    (__mips_dsp_rev >= 2) && (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-void TransposeWx8_DSPR2(const uint8* src,
-                        int src_stride,
-                        uint8* dst,
-                        int dst_stride,
-                        int width) {
-  __asm__ __volatile__(
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "sll              $t2, %[src_stride], 0x1          \n"  // src_stride x 2
-      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
-      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
-      "addu             $t3, $t2, %[src_stride]          \n"
-      "addu             $t5, $t4, %[src_stride]          \n"
-      "addu             $t6, $t2, $t4                    \n"
-      "andi             $t0, %[dst], 0x3                 \n"
-      "andi             $t1, %[dst_stride], 0x3          \n"
-      "or               $t0, $t0, $t1                    \n"
-      "bnez             $t0, 11f                         \n"
-      " subu            $t7, $t9, %[src_stride]          \n"
-      // dst + dst_stride word aligned
-      "1:                                                \n"
-      "lbu              $t0, 0(%[src])                   \n"
-      "lbux             $t1, %[src_stride](%[src])       \n"
-      "lbux             $t8, $t2(%[src])                 \n"
-      "lbux             $t9, $t3(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s0, $t8, $t0                    \n"
-      "lbux             $t0, $t4(%[src])                 \n"
-      "lbux             $t1, $t5(%[src])                 \n"
-      "lbux             $t8, $t6(%[src])                 \n"
-      "lbux             $t9, $t7(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s1, $t8, $t0                    \n"
-      "sw               $s0, 0(%[dst])                   \n"
-      "addiu            %[width], -1                     \n"
-      "addiu            %[src], 1                        \n"
-      "sw               $s1, 4(%[dst])                   \n"
-      "bnez             %[width], 1b                     \n"
-      " addu            %[dst], %[dst], %[dst_stride]    \n"
-      "b                2f                               \n"
-      // dst + dst_stride unaligned
-      "11:                                               \n"
-      "lbu              $t0, 0(%[src])                   \n"
-      "lbux             $t1, %[src_stride](%[src])       \n"
-      "lbux             $t8, $t2(%[src])                 \n"
-      "lbux             $t9, $t3(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s0, $t8, $t0                    \n"
-      "lbux             $t0, $t4(%[src])                 \n"
-      "lbux             $t1, $t5(%[src])                 \n"
-      "lbux             $t8, $t6(%[src])                 \n"
-      "lbux             $t9, $t7(%[src])                 \n"
-      "sll              $t1, $t1, 16                     \n"
-      "sll              $t9, $t9, 16                     \n"
-      "or               $t0, $t0, $t1                    \n"
-      "or               $t8, $t8, $t9                    \n"
-      "precr.qb.ph      $s1, $t8, $t0                    \n"
-      "swr              $s0, 0(%[dst])                   \n"
-      "swl              $s0, 3(%[dst])                   \n"
-      "addiu            %[width], -1                     \n"
-      "addiu            %[src], 1                        \n"
-      "swr              $s1, 4(%[dst])                   \n"
-      "swl              $s1, 7(%[dst])                   \n"
-      "bnez             %[width], 11b                    \n"
-      "addu             %[dst], %[dst], %[dst_stride]    \n"
-      "2:                                                \n"
-      ".set pop                                          \n"
-      : [src] "+r"(src), [dst] "+r"(dst), [width] "+r"(width)
-      : [src_stride] "r"(src_stride), [dst_stride] "r"(dst_stride)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1");
-}
-
-void TransposeWx8_Fast_DSPR2(const uint8* src,
-                             int src_stride,
-                             uint8* dst,
-                             int dst_stride,
-                             int width) {
-  __asm__ __volatile__(
-      ".set noat                                         \n"
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "beqz             %[width], 2f                     \n"
-      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
-      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
-      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
-      "addu             $t3, $t2, %[src_stride]          \n"
-      "addu             $t5, $t4, %[src_stride]          \n"
-      "addu             $t6, $t2, $t4                    \n"
-
-      "srl              $AT, %[width], 0x2               \n"
-      "andi             $t0, %[dst], 0x3                 \n"
-      "andi             $t1, %[dst_stride], 0x3          \n"
-      "or               $t0, $t0, $t1                    \n"
-      "bnez             $t0, 11f                         \n"
-      " subu            $t7, $t9, %[src_stride]          \n"
-      // dst + dst_stride word aligned
-      "1:                                                \n"
-      "lw               $t0, 0(%[src])                   \n"
-      "lwx              $t1, %[src_stride](%[src])       \n"
-      "lwx              $t8, $t2(%[src])                 \n"
-      "lwx              $t9, $t3(%[src])                 \n"
-
-      // t0 = | 30 | 20 | 10 | 00 |
-      // t1 = | 31 | 21 | 11 | 01 |
-      // t8 = | 32 | 22 | 12 | 02 |
-      // t9 = | 33 | 23 | 13 | 03 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-      // s0 = | 21 | 01 | 20 | 00 |
-      // s1 = | 23 | 03 | 22 | 02 |
-      // s2 = | 31 | 11 | 30 | 10 |
-      // s3 = | 33 | 13 | 32 | 12 |
-
-      "precr.qb.ph     $s4, $s1, $s0                     \n"
-      "precrq.qb.ph    $s5, $s1, $s0                     \n"
-      "precr.qb.ph     $s6, $s3, $s2                     \n"
-      "precrq.qb.ph    $s7, $s3, $s2                     \n"
-
-      // s4 = | 03 | 02 | 01 | 00 |
-      // s5 = | 23 | 22 | 21 | 20 |
-      // s6 = | 13 | 12 | 11 | 10 |
-      // s7 = | 33 | 32 | 31 | 30 |
-
-      "lwx              $t0, $t4(%[src])                 \n"
-      "lwx              $t1, $t5(%[src])                 \n"
-      "lwx              $t8, $t6(%[src])                 \n"
-      "lwx              $t9, $t7(%[src])                 \n"
-
-      // t0 = | 34 | 24 | 14 | 04 |
-      // t1 = | 35 | 25 | 15 | 05 |
-      // t8 = | 36 | 26 | 16 | 06 |
-      // t9 = | 37 | 27 | 17 | 07 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-      // s0 = | 25 | 05 | 24 | 04 |
-      // s1 = | 27 | 07 | 26 | 06 |
-      // s2 = | 35 | 15 | 34 | 14 |
-      // s3 = | 37 | 17 | 36 | 16 |
-
-      "precr.qb.ph     $t0, $s1, $s0                     \n"
-      "precrq.qb.ph    $t1, $s1, $s0                     \n"
-      "precr.qb.ph     $t8, $s3, $s2                     \n"
-      "precrq.qb.ph    $t9, $s3, $s2                     \n"
-
-      // t0 = | 07 | 06 | 05 | 04 |
-      // t1 = | 27 | 26 | 25 | 24 |
-      // t8 = | 17 | 16 | 15 | 14 |
-      // t9 = | 37 | 36 | 35 | 34 |
-
-      "addu            $s0, %[dst], %[dst_stride]        \n"
-      "addu            $s1, $s0, %[dst_stride]           \n"
-      "addu            $s2, $s1, %[dst_stride]           \n"
-
-      "sw              $s4, 0(%[dst])                    \n"
-      "sw              $t0, 4(%[dst])                    \n"
-      "sw              $s6, 0($s0)                       \n"
-      "sw              $t8, 4($s0)                       \n"
-      "sw              $s5, 0($s1)                       \n"
-      "sw              $t1, 4($s1)                       \n"
-      "sw              $s7, 0($s2)                       \n"
-      "sw              $t9, 4($s2)                       \n"
-
-      "addiu            $AT, -1                          \n"
-      "addiu            %[src], 4                        \n"
-
-      "bnez             $AT, 1b                          \n"
-      " addu            %[dst], $s2, %[dst_stride]       \n"
-      "b                2f                               \n"
-      // dst + dst_stride unaligned
-      "11:                                               \n"
-      "lw               $t0, 0(%[src])                   \n"
-      "lwx              $t1, %[src_stride](%[src])       \n"
-      "lwx              $t8, $t2(%[src])                 \n"
-      "lwx              $t9, $t3(%[src])                 \n"
-
-      // t0 = | 30 | 20 | 10 | 00 |
-      // t1 = | 31 | 21 | 11 | 01 |
-      // t8 = | 32 | 22 | 12 | 02 |
-      // t9 = | 33 | 23 | 13 | 03 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-      // s0 = | 21 | 01 | 20 | 00 |
-      // s1 = | 23 | 03 | 22 | 02 |
-      // s2 = | 31 | 11 | 30 | 10 |
-      // s3 = | 33 | 13 | 32 | 12 |
-
-      "precr.qb.ph     $s4, $s1, $s0                     \n"
-      "precrq.qb.ph    $s5, $s1, $s0                     \n"
-      "precr.qb.ph     $s6, $s3, $s2                     \n"
-      "precrq.qb.ph    $s7, $s3, $s2                     \n"
-
-      // s4 = | 03 | 02 | 01 | 00 |
-      // s5 = | 23 | 22 | 21 | 20 |
-      // s6 = | 13 | 12 | 11 | 10 |
-      // s7 = | 33 | 32 | 31 | 30 |
-
-      "lwx              $t0, $t4(%[src])                 \n"
-      "lwx              $t1, $t5(%[src])                 \n"
-      "lwx              $t8, $t6(%[src])                 \n"
-      "lwx              $t9, $t7(%[src])                 \n"
-
-      // t0 = | 34 | 24 | 14 | 04 |
-      // t1 = | 35 | 25 | 15 | 05 |
-      // t8 = | 36 | 26 | 16 | 06 |
-      // t9 = | 37 | 27 | 17 | 07 |
-
-      "precr.qb.ph     $s0, $t1, $t0                     \n"
-      "precr.qb.ph     $s1, $t9, $t8                     \n"
-      "precrq.qb.ph    $s2, $t1, $t0                     \n"
-      "precrq.qb.ph    $s3, $t9, $t8                     \n"
-
-      // s0 = | 25 | 05 | 24 | 04 |
-      // s1 = | 27 | 07 | 26 | 06 |
-      // s2 = | 35 | 15 | 34 | 14 |
-      // s3 = | 37 | 17 | 36 | 16 |
-
-      "precr.qb.ph     $t0, $s1, $s0                     \n"
-      "precrq.qb.ph    $t1, $s1, $s0                     \n"
-      "precr.qb.ph     $t8, $s3, $s2                     \n"
-      "precrq.qb.ph    $t9, $s3, $s2                     \n"
-
-      // t0 = | 07 | 06 | 05 | 04 |
-      // t1 = | 27 | 26 | 25 | 24 |
-      // t8 = | 17 | 16 | 15 | 14 |
-      // t9 = | 37 | 36 | 35 | 34 |
-
-      "addu            $s0, %[dst], %[dst_stride]        \n"
-      "addu            $s1, $s0, %[dst_stride]           \n"
-      "addu            $s2, $s1, %[dst_stride]           \n"
-
-      "swr              $s4, 0(%[dst])                   \n"
-      "swl              $s4, 3(%[dst])                   \n"
-      "swr              $t0, 4(%[dst])                   \n"
-      "swl              $t0, 7(%[dst])                   \n"
-      "swr              $s6, 0($s0)                      \n"
-      "swl              $s6, 3($s0)                      \n"
-      "swr              $t8, 4($s0)                      \n"
-      "swl              $t8, 7($s0)                      \n"
-      "swr              $s5, 0($s1)                      \n"
-      "swl              $s5, 3($s1)                      \n"
-      "swr              $t1, 4($s1)                      \n"
-      "swl              $t1, 7($s1)                      \n"
-      "swr              $s7, 0($s2)                      \n"
-      "swl              $s7, 3($s2)                      \n"
-      "swr              $t9, 4($s2)                      \n"
-      "swl              $t9, 7($s2)                      \n"
-
-      "addiu            $AT, -1                          \n"
-      "addiu            %[src], 4                        \n"
-
-      "bnez             $AT, 11b                         \n"
-      " addu            %[dst], $s2, %[dst_stride]       \n"
-      "2:                                                \n"
-      ".set pop                                          \n"
-      ".set at                                           \n"
-      : [src] "+r"(src), [dst] "+r"(dst), [width] "+r"(width)
-      : [src_stride] "r"(src_stride), [dst_stride] "r"(dst_stride)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1",
-        "s2", "s3", "s4", "s5", "s6", "s7");
-}
-
-void TransposeUVWx8_DSPR2(const uint8* src,
-                          int src_stride,
-                          uint8* dst_a,
-                          int dst_stride_a,
-                          uint8* dst_b,
-                          int dst_stride_b,
-                          int width) {
-  __asm__ __volatile__(
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "beqz            %[width], 2f                      \n"
-      " sll            $t2, %[src_stride], 0x1           \n"  // src_stride x 2
-      "sll             $t4, %[src_stride], 0x2           \n"  // src_stride x 4
-      "sll             $t9, %[src_stride], 0x3           \n"  // src_stride x 8
-      "addu            $t3, $t2, %[src_stride]           \n"
-      "addu            $t5, $t4, %[src_stride]           \n"
-      "addu            $t6, $t2, $t4                     \n"
-      "subu            $t7, $t9, %[src_stride]           \n"
-      "srl             $t1, %[width], 1                  \n"
-
-      // check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
-      "andi            $t0, %[dst_a], 0x3                \n"
-      "andi            $t8, %[dst_b], 0x3                \n"
-      "or              $t0, $t0, $t8                     \n"
-      "andi            $t8, %[dst_stride_a], 0x3         \n"
-      "andi            $s5, %[dst_stride_b], 0x3         \n"
-      "or              $t8, $t8, $s5                     \n"
-      "or              $t0, $t0, $t8                     \n"
-      "bnez            $t0, 11f                          \n"
-      " nop                                              \n"
-      // dst + dst_stride word aligned (both, a & b dst addresses)
-      "1:                                                \n"
-      "lw              $t0, 0(%[src])                    \n"  // |B0|A0|b0|a0|
-      "lwx             $t8, %[src_stride](%[src])        \n"  // |B1|A1|b1|a1|
-      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
-      "lwx             $t9, $t2(%[src])                  \n"  // |B2|A2|b2|a2|
-      "lwx             $s0, $t3(%[src])                  \n"  // |B3|A3|b3|a3|
-      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B1|A1|B0|A0|
-      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B3|A3|B2|A2|
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A3|A2|A1|A0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B3|B2|B1|B0|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n"  // |b1|a1|b0|a0|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n"  // |b3|a3|b2|a2|
-
-      "sw              $s3, 0($s5)                       \n"
-      "sw              $s4, 0($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a3|a2|a1|a0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b3|b2|b1|b0|
-
-      "lwx             $t0, $t4(%[src])                  \n"  // |B4|A4|b4|a4|
-      "lwx             $t8, $t5(%[src])                  \n"  // |B5|A5|b5|a5|
-      "lwx             $t9, $t6(%[src])                  \n"  // |B6|A6|b6|a6|
-      "lwx             $s0, $t7(%[src])                  \n"  // |B7|A7|b7|a7|
-      "sw              $s3, 0(%[dst_a])                  \n"
-      "sw              $s4, 0(%[dst_b])                  \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B5|A5|B4|A4|
-      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B6|A6|B7|A7|
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A7|A6|A5|A4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B7|B6|B5|B4|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n"  // |b5|a5|b4|a4|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n"  // |b7|a7|b6|a6|
-      "sw              $s3, 4($s5)                       \n"
-      "sw              $s4, 4($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a7|a6|a5|a4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b7|b6|b5|b4|
-
-      "addiu           %[src], 4                         \n"
-      "addiu           $t1, -1                           \n"
-      "sll             $t0, %[dst_stride_a], 1           \n"
-      "sll             $t8, %[dst_stride_b], 1           \n"
-      "sw              $s3, 4(%[dst_a])                  \n"
-      "sw              $s4, 4(%[dst_b])                  \n"
-      "addu            %[dst_a], %[dst_a], $t0           \n"
-      "bnez            $t1, 1b                           \n"
-      " addu           %[dst_b], %[dst_b], $t8           \n"
-      "b               2f                                \n"
-      " nop                                              \n"
-
-      // dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
-      "11:                                               \n"
-      "lw              $t0, 0(%[src])                    \n"  // |B0|A0|b0|a0|
-      "lwx             $t8, %[src_stride](%[src])        \n"  // |B1|A1|b1|a1|
-      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
-      "lwx             $t9, $t2(%[src])                  \n"  // |B2|A2|b2|a2|
-      "lwx             $s0, $t3(%[src])                  \n"  // |B3|A3|b3|a3|
-      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B1|A1|B0|A0|
-      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B3|A3|B2|A2|
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A3|A2|A1|A0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B3|B2|B1|B0|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n"  // |b1|a1|b0|a0|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n"  // |b3|a3|b2|a2|
-
-      "swr             $s3, 0($s5)                       \n"
-      "swl             $s3, 3($s5)                       \n"
-      "swr             $s4, 0($s6)                       \n"
-      "swl             $s4, 3($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a3|a2|a1|a0|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b3|b2|b1|b0|
-
-      "lwx             $t0, $t4(%[src])                  \n"  // |B4|A4|b4|a4|
-      "lwx             $t8, $t5(%[src])                  \n"  // |B5|A5|b5|a5|
-      "lwx             $t9, $t6(%[src])                  \n"  // |B6|A6|b6|a6|
-      "lwx             $s0, $t7(%[src])                  \n"  // |B7|A7|b7|a7|
-      "swr             $s3, 0(%[dst_a])                  \n"
-      "swl             $s3, 3(%[dst_a])                  \n"
-      "swr             $s4, 0(%[dst_b])                  \n"
-      "swl             $s4, 3(%[dst_b])                  \n"
-
-      "precrq.ph.w     $s1, $t8, $t0                     \n"  // |B5|A5|B4|A4|
-      "precrq.ph.w     $s2, $s0, $t9                     \n"  // |B6|A6|B7|A7|
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |A7|A6|A5|A4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |B7|B6|B5|B4|
-
-      "sll             $t0, $t0, 16                      \n"
-      "packrl.ph       $s1, $t8, $t0                     \n"  // |b5|a5|b4|a4|
-      "sll             $t9, $t9, 16                      \n"
-      "packrl.ph       $s2, $s0, $t9                     \n"  // |b7|a7|b6|a6|
-
-      "swr             $s3, 4($s5)                       \n"
-      "swl             $s3, 7($s5)                       \n"
-      "swr             $s4, 4($s6)                       \n"
-      "swl             $s4, 7($s6)                       \n"
-
-      "precr.qb.ph     $s3, $s2, $s1                     \n"  // |a7|a6|a5|a4|
-      "precrq.qb.ph    $s4, $s2, $s1                     \n"  // |b7|b6|b5|b4|
-
-      "addiu           %[src], 4                         \n"
-      "addiu           $t1, -1                           \n"
-      "sll             $t0, %[dst_stride_a], 1           \n"
-      "sll             $t8, %[dst_stride_b], 1           \n"
-      "swr             $s3, 4(%[dst_a])                  \n"
-      "swl             $s3, 7(%[dst_a])                  \n"
-      "swr             $s4, 4(%[dst_b])                  \n"
-      "swl             $s4, 7(%[dst_b])                  \n"
-      "addu            %[dst_a], %[dst_a], $t0           \n"
-      "bnez            $t1, 11b                          \n"
-      " addu           %[dst_b], %[dst_b], $t8           \n"
-
-      "2:                                                \n"
-      ".set pop                                          \n"
-      : [src] "+r"(src), [dst_a] "+r"(dst_a), [dst_b] "+r"(dst_b),
-        [width] "+r"(width), [src_stride] "+r"(src_stride)
-      : [dst_stride_a] "r"(dst_stride_a), [dst_stride_b] "r"(dst_stride_b)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "s0", "s1",
-        "s2", "s3", "s4", "s5", "s6");
-}
-
-#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/files/source/rotate_gcc.cc b/files/source/rotate_gcc.cc
index 04e19e29..1a3f8cbb 100644
--- a/files/source/rotate_gcc.cc
+++ b/files/source/rotate_gcc.cc
@@ -17,8 +17,7 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 // Transpose 8x8. 32 or 64 bit, but not NaCL for 64 bit.
 #if defined(HAS_TRANSPOSEWX8_SSSE3)
@@ -31,75 +30,75 @@ void TransposeWx8_SSSE3(const uint8_t* src,
       // Read in the data from the source pointer.
       // First round of bit swap.
       LABELALIGN
-      "1:                                          \n"
-      "movq       (%0),%%xmm0                      \n"
-      "movq       (%0,%3),%%xmm1                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "punpcklbw  %%xmm1,%%xmm0                    \n"
-      "movq       (%0),%%xmm2                      \n"
-      "movdqa     %%xmm0,%%xmm1                    \n"
-      "palignr    $0x8,%%xmm1,%%xmm1               \n"
-      "movq       (%0,%3),%%xmm3                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "punpcklbw  %%xmm3,%%xmm2                    \n"
-      "movdqa     %%xmm2,%%xmm3                    \n"
-      "movq       (%0),%%xmm4                      \n"
-      "palignr    $0x8,%%xmm3,%%xmm3               \n"
-      "movq       (%0,%3),%%xmm5                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "punpcklbw  %%xmm5,%%xmm4                    \n"
-      "movdqa     %%xmm4,%%xmm5                    \n"
-      "movq       (%0),%%xmm6                      \n"
-      "palignr    $0x8,%%xmm5,%%xmm5               \n"
-      "movq       (%0,%3),%%xmm7                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "punpcklbw  %%xmm7,%%xmm6                    \n"
-      "neg        %3                               \n"
-      "movdqa     %%xmm6,%%xmm7                    \n"
-      "lea        0x8(%0,%3,8),%0                  \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
-      "neg        %3                               \n"
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        (%0,%3),%%xmm1                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "movq        (%0),%%xmm2                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm1            \n"
+      "movq        (%0,%3),%%xmm3                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movq        (%0),%%xmm4                   \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "movq        (%0,%3),%%xmm5                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "movq        (%0),%%xmm6                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        (%0,%3),%%xmm7                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "neg         %3                            \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "lea         0x8(%0,%3,8),%0               \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "neg         %3                            \n"
       // Second round of bit swap.
-      "punpcklwd  %%xmm2,%%xmm0                    \n"
-      "punpcklwd  %%xmm3,%%xmm1                    \n"
-      "movdqa     %%xmm0,%%xmm2                    \n"
-      "movdqa     %%xmm1,%%xmm3                    \n"
-      "palignr    $0x8,%%xmm2,%%xmm2               \n"
-      "palignr    $0x8,%%xmm3,%%xmm3               \n"
-      "punpcklwd  %%xmm6,%%xmm4                    \n"
-      "punpcklwd  %%xmm7,%%xmm5                    \n"
-      "movdqa     %%xmm4,%%xmm6                    \n"
-      "movdqa     %%xmm5,%%xmm7                    \n"
-      "palignr    $0x8,%%xmm6,%%xmm6               \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "palignr     $0x8,%%xmm2,%%xmm2            \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "movdqa      %%xmm5,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
       // Third round of bit swap.
       // Write to the destination pointer.
-      "punpckldq  %%xmm4,%%xmm0                    \n"
-      "movq       %%xmm0,(%1)                      \n"
-      "movdqa     %%xmm0,%%xmm4                    \n"
-      "palignr    $0x8,%%xmm4,%%xmm4               \n"
-      "movq       %%xmm4,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm6,%%xmm2                    \n"
-      "movdqa     %%xmm2,%%xmm6                    \n"
-      "movq       %%xmm2,(%1)                      \n"
-      "palignr    $0x8,%%xmm6,%%xmm6               \n"
-      "punpckldq  %%xmm5,%%xmm1                    \n"
-      "movq       %%xmm6,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "movdqa     %%xmm1,%%xmm5                    \n"
-      "movq       %%xmm1,(%1)                      \n"
-      "palignr    $0x8,%%xmm5,%%xmm5               \n"
-      "movq       %%xmm5,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm7,%%xmm3                    \n"
-      "movq       %%xmm3,(%1)                      \n"
-      "movdqa     %%xmm3,%%xmm7                    \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
-      "sub        $0x8,%2                          \n"
-      "movq       %%xmm7,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "jg         1b                               \n"
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "palignr     $0x8,%%xmm4,%%xmm4            \n"
+      "movq        %%xmm4,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "movq        %%xmm2,(%1)                   \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movq        %%xmm6,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "movq        %%xmm1,(%1)                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        %%xmm5,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movq        %%xmm3,(%1)                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "sub         $0x8,%2                       \n"
+      "movq        %%xmm7,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "jg          1b                            \n"
       : "+r"(src),                    // %0
         "+r"(dst),                    // %1
         "+r"(width)                   // %2
@@ -121,127 +120,127 @@ void TransposeWx8_Fast_SSSE3(const uint8_t* src,
       // Read in the data from the source pointer.
       // First round of bit swap.
       LABELALIGN
-      "1:                                          \n"
-      "movdqu     (%0),%%xmm0                      \n"
-      "movdqu     (%0,%3),%%xmm1                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "movdqa     %%xmm0,%%xmm8                    \n"
-      "punpcklbw  %%xmm1,%%xmm0                    \n"
-      "punpckhbw  %%xmm1,%%xmm8                    \n"
-      "movdqu     (%0),%%xmm2                      \n"
-      "movdqa     %%xmm0,%%xmm1                    \n"
-      "movdqa     %%xmm8,%%xmm9                    \n"
-      "palignr    $0x8,%%xmm1,%%xmm1               \n"
-      "palignr    $0x8,%%xmm9,%%xmm9               \n"
-      "movdqu     (%0,%3),%%xmm3                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "movdqa     %%xmm2,%%xmm10                   \n"
-      "punpcklbw  %%xmm3,%%xmm2                    \n"
-      "punpckhbw  %%xmm3,%%xmm10                   \n"
-      "movdqa     %%xmm2,%%xmm3                    \n"
-      "movdqa     %%xmm10,%%xmm11                  \n"
-      "movdqu     (%0),%%xmm4                      \n"
-      "palignr    $0x8,%%xmm3,%%xmm3               \n"
-      "palignr    $0x8,%%xmm11,%%xmm11             \n"
-      "movdqu     (%0,%3),%%xmm5                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "movdqa     %%xmm4,%%xmm12                   \n"
-      "punpcklbw  %%xmm5,%%xmm4                    \n"
-      "punpckhbw  %%xmm5,%%xmm12                   \n"
-      "movdqa     %%xmm4,%%xmm5                    \n"
-      "movdqa     %%xmm12,%%xmm13                  \n"
-      "movdqu     (%0),%%xmm6                      \n"
-      "palignr    $0x8,%%xmm5,%%xmm5               \n"
-      "palignr    $0x8,%%xmm13,%%xmm13             \n"
-      "movdqu     (%0,%3),%%xmm7                   \n"
-      "lea        (%0,%3,2),%0                     \n"
-      "movdqa     %%xmm6,%%xmm14                   \n"
-      "punpcklbw  %%xmm7,%%xmm6                    \n"
-      "punpckhbw  %%xmm7,%%xmm14                   \n"
-      "neg        %3                               \n"
-      "movdqa     %%xmm6,%%xmm7                    \n"
-      "movdqa     %%xmm14,%%xmm15                  \n"
-      "lea        0x10(%0,%3,8),%0                 \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
-      "palignr    $0x8,%%xmm15,%%xmm15             \n"
-      "neg        %3                               \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%0,%3),%%xmm1                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm8                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm8,%%xmm9                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm1            \n"
+      "palignr     $0x8,%%xmm9,%%xmm9            \n"
+      "movdqu      (%0,%3),%%xmm3                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm2,%%xmm10                \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "punpckhbw   %%xmm3,%%xmm10                \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm10,%%xmm11               \n"
+      "movdqu      (%0),%%xmm4                   \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "palignr     $0x8,%%xmm11,%%xmm11          \n"
+      "movdqu      (%0,%3),%%xmm5                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm4,%%xmm12                \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "punpckhbw   %%xmm5,%%xmm12                \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "movdqa      %%xmm12,%%xmm13               \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "palignr     $0x8,%%xmm13,%%xmm13          \n"
+      "movdqu      (%0,%3),%%xmm7                \n"
+      "lea         (%0,%3,2),%0                  \n"
+      "movdqa      %%xmm6,%%xmm14                \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "punpckhbw   %%xmm7,%%xmm14                \n"
+      "neg         %3                            \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "movdqa      %%xmm14,%%xmm15               \n"
+      "lea         0x10(%0,%3,8),%0              \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
+      "neg         %3                            \n"
       // Second round of bit swap.
-      "punpcklwd  %%xmm2,%%xmm0                    \n"
-      "punpcklwd  %%xmm3,%%xmm1                    \n"
-      "movdqa     %%xmm0,%%xmm2                    \n"
-      "movdqa     %%xmm1,%%xmm3                    \n"
-      "palignr    $0x8,%%xmm2,%%xmm2               \n"
-      "palignr    $0x8,%%xmm3,%%xmm3               \n"
-      "punpcklwd  %%xmm6,%%xmm4                    \n"
-      "punpcklwd  %%xmm7,%%xmm5                    \n"
-      "movdqa     %%xmm4,%%xmm6                    \n"
-      "movdqa     %%xmm5,%%xmm7                    \n"
-      "palignr    $0x8,%%xmm6,%%xmm6               \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
-      "punpcklwd  %%xmm10,%%xmm8                   \n"
-      "punpcklwd  %%xmm11,%%xmm9                   \n"
-      "movdqa     %%xmm8,%%xmm10                   \n"
-      "movdqa     %%xmm9,%%xmm11                   \n"
-      "palignr    $0x8,%%xmm10,%%xmm10             \n"
-      "palignr    $0x8,%%xmm11,%%xmm11             \n"
-      "punpcklwd  %%xmm14,%%xmm12                  \n"
-      "punpcklwd  %%xmm15,%%xmm13                  \n"
-      "movdqa     %%xmm12,%%xmm14                  \n"
-      "movdqa     %%xmm13,%%xmm15                  \n"
-      "palignr    $0x8,%%xmm14,%%xmm14             \n"
-      "palignr    $0x8,%%xmm15,%%xmm15             \n"
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "palignr     $0x8,%%xmm2,%%xmm2            \n"
+      "palignr     $0x8,%%xmm3,%%xmm3            \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "movdqa      %%xmm5,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "punpcklwd   %%xmm10,%%xmm8                \n"
+      "punpcklwd   %%xmm11,%%xmm9                \n"
+      "movdqa      %%xmm8,%%xmm10                \n"
+      "movdqa      %%xmm9,%%xmm11                \n"
+      "palignr     $0x8,%%xmm10,%%xmm10          \n"
+      "palignr     $0x8,%%xmm11,%%xmm11          \n"
+      "punpcklwd   %%xmm14,%%xmm12               \n"
+      "punpcklwd   %%xmm15,%%xmm13               \n"
+      "movdqa      %%xmm12,%%xmm14               \n"
+      "movdqa      %%xmm13,%%xmm15               \n"
+      "palignr     $0x8,%%xmm14,%%xmm14          \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
       // Third round of bit swap.
       // Write to the destination pointer.
-      "punpckldq  %%xmm4,%%xmm0                    \n"
-      "movq       %%xmm0,(%1)                      \n"
-      "movdqa     %%xmm0,%%xmm4                    \n"
-      "palignr    $0x8,%%xmm4,%%xmm4               \n"
-      "movq       %%xmm4,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm6,%%xmm2                    \n"
-      "movdqa     %%xmm2,%%xmm6                    \n"
-      "movq       %%xmm2,(%1)                      \n"
-      "palignr    $0x8,%%xmm6,%%xmm6               \n"
-      "punpckldq  %%xmm5,%%xmm1                    \n"
-      "movq       %%xmm6,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "movdqa     %%xmm1,%%xmm5                    \n"
-      "movq       %%xmm1,(%1)                      \n"
-      "palignr    $0x8,%%xmm5,%%xmm5               \n"
-      "movq       %%xmm5,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm7,%%xmm3                    \n"
-      "movq       %%xmm3,(%1)                      \n"
-      "movdqa     %%xmm3,%%xmm7                    \n"
-      "palignr    $0x8,%%xmm7,%%xmm7               \n"
-      "movq       %%xmm7,(%1,%4)                   \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm12,%%xmm8                   \n"
-      "movq       %%xmm8,(%1)                      \n"
-      "movdqa     %%xmm8,%%xmm12                   \n"
-      "palignr    $0x8,%%xmm12,%%xmm12             \n"
-      "movq       %%xmm12,(%1,%4)                  \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm14,%%xmm10                  \n"
-      "movdqa     %%xmm10,%%xmm14                  \n"
-      "movq       %%xmm10,(%1)                     \n"
-      "palignr    $0x8,%%xmm14,%%xmm14             \n"
-      "punpckldq  %%xmm13,%%xmm9                   \n"
-      "movq       %%xmm14,(%1,%4)                  \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "movdqa     %%xmm9,%%xmm13                   \n"
-      "movq       %%xmm9,(%1)                      \n"
-      "palignr    $0x8,%%xmm13,%%xmm13             \n"
-      "movq       %%xmm13,(%1,%4)                  \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "punpckldq  %%xmm15,%%xmm11                  \n"
-      "movq       %%xmm11,(%1)                     \n"
-      "movdqa     %%xmm11,%%xmm15                  \n"
-      "palignr    $0x8,%%xmm15,%%xmm15             \n"
-      "sub        $0x10,%2                         \n"
-      "movq       %%xmm15,(%1,%4)                  \n"
-      "lea        (%1,%4,2),%1                     \n"
-      "jg         1b                               \n"
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "palignr     $0x8,%%xmm4,%%xmm4            \n"
+      "movq        %%xmm4,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "movq        %%xmm2,(%1)                   \n"
+      "palignr     $0x8,%%xmm6,%%xmm6            \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movq        %%xmm6,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "movq        %%xmm1,(%1)                   \n"
+      "palignr     $0x8,%%xmm5,%%xmm5            \n"
+      "movq        %%xmm5,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movq        %%xmm3,(%1)                   \n"
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "palignr     $0x8,%%xmm7,%%xmm7            \n"
+      "movq        %%xmm7,(%1,%4)                \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm12,%%xmm8                \n"
+      "movq        %%xmm8,(%1)                   \n"
+      "movdqa      %%xmm8,%%xmm12                \n"
+      "palignr     $0x8,%%xmm12,%%xmm12          \n"
+      "movq        %%xmm12,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm14,%%xmm10               \n"
+      "movdqa      %%xmm10,%%xmm14               \n"
+      "movq        %%xmm10,(%1)                  \n"
+      "palignr     $0x8,%%xmm14,%%xmm14          \n"
+      "punpckldq   %%xmm13,%%xmm9                \n"
+      "movq        %%xmm14,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "movdqa      %%xmm9,%%xmm13                \n"
+      "movq        %%xmm9,(%1)                   \n"
+      "palignr     $0x8,%%xmm13,%%xmm13          \n"
+      "movq        %%xmm13,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "punpckldq   %%xmm15,%%xmm11               \n"
+      "movq        %%xmm11,(%1)                  \n"
+      "movdqa      %%xmm11,%%xmm15               \n"
+      "palignr     $0x8,%%xmm15,%%xmm15          \n"
+      "sub         $0x10,%2                      \n"
+      "movq        %%xmm15,(%1,%4)               \n"
+      "lea         (%1,%4,2),%1                  \n"
+      "jg          1b                            \n"
       : "+r"(src),                    // %0
         "+r"(dst),                    // %1
         "+r"(width)                   // %2
@@ -266,95 +265,95 @@ void TransposeUVWx8_SSE2(const uint8_t* src,
       // Read in the data from the source pointer.
       // First round of bit swap.
       LABELALIGN
-      "1:                                          \n"
-      "movdqu     (%0),%%xmm0                      \n"
-      "movdqu     (%0,%4),%%xmm1                   \n"
-      "lea        (%0,%4,2),%0                     \n"
-      "movdqa     %%xmm0,%%xmm8                    \n"
-      "punpcklbw  %%xmm1,%%xmm0                    \n"
-      "punpckhbw  %%xmm1,%%xmm8                    \n"
-      "movdqa     %%xmm8,%%xmm1                    \n"
-      "movdqu     (%0),%%xmm2                      \n"
-      "movdqu     (%0,%4),%%xmm3                   \n"
-      "lea        (%0,%4,2),%0                     \n"
-      "movdqa     %%xmm2,%%xmm8                    \n"
-      "punpcklbw  %%xmm3,%%xmm2                    \n"
-      "punpckhbw  %%xmm3,%%xmm8                    \n"
-      "movdqa     %%xmm8,%%xmm3                    \n"
-      "movdqu     (%0),%%xmm4                      \n"
-      "movdqu     (%0,%4),%%xmm5                   \n"
-      "lea        (%0,%4,2),%0                     \n"
-      "movdqa     %%xmm4,%%xmm8                    \n"
-      "punpcklbw  %%xmm5,%%xmm4                    \n"
-      "punpckhbw  %%xmm5,%%xmm8                    \n"
-      "movdqa     %%xmm8,%%xmm5                    \n"
-      "movdqu     (%0),%%xmm6                      \n"
-      "movdqu     (%0,%4),%%xmm7                   \n"
-      "lea        (%0,%4,2),%0                     \n"
-      "movdqa     %%xmm6,%%xmm8                    \n"
-      "punpcklbw  %%xmm7,%%xmm6                    \n"
-      "neg        %4                               \n"
-      "lea        0x10(%0,%4,8),%0                 \n"
-      "punpckhbw  %%xmm7,%%xmm8                    \n"
-      "movdqa     %%xmm8,%%xmm7                    \n"
-      "neg        %4                               \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%0,%4),%%xmm1                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      (%0,%4),%%xmm3                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm2,%%xmm8                 \n"
+      "punpcklbw   %%xmm3,%%xmm2                 \n"
+      "punpckhbw   %%xmm3,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm3                 \n"
+      "movdqu      (%0),%%xmm4                   \n"
+      "movdqu      (%0,%4),%%xmm5                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm4,%%xmm8                 \n"
+      "punpcklbw   %%xmm5,%%xmm4                 \n"
+      "punpckhbw   %%xmm5,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm5                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      (%0,%4),%%xmm7                \n"
+      "lea         (%0,%4,2),%0                  \n"
+      "movdqa      %%xmm6,%%xmm8                 \n"
+      "punpcklbw   %%xmm7,%%xmm6                 \n"
+      "neg         %4                            \n"
+      "lea         0x10(%0,%4,8),%0              \n"
+      "punpckhbw   %%xmm7,%%xmm8                 \n"
+      "movdqa      %%xmm8,%%xmm7                 \n"
+      "neg         %4                            \n"
       // Second round of bit swap.
-      "movdqa     %%xmm0,%%xmm8                    \n"
-      "movdqa     %%xmm1,%%xmm9                    \n"
-      "punpckhwd  %%xmm2,%%xmm8                    \n"
-      "punpckhwd  %%xmm3,%%xmm9                    \n"
-      "punpcklwd  %%xmm2,%%xmm0                    \n"
-      "punpcklwd  %%xmm3,%%xmm1                    \n"
-      "movdqa     %%xmm8,%%xmm2                    \n"
-      "movdqa     %%xmm9,%%xmm3                    \n"
-      "movdqa     %%xmm4,%%xmm8                    \n"
-      "movdqa     %%xmm5,%%xmm9                    \n"
-      "punpckhwd  %%xmm6,%%xmm8                    \n"
-      "punpckhwd  %%xmm7,%%xmm9                    \n"
-      "punpcklwd  %%xmm6,%%xmm4                    \n"
-      "punpcklwd  %%xmm7,%%xmm5                    \n"
-      "movdqa     %%xmm8,%%xmm6                    \n"
-      "movdqa     %%xmm9,%%xmm7                    \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "movdqa      %%xmm1,%%xmm9                 \n"
+      "punpckhwd   %%xmm2,%%xmm8                 \n"
+      "punpckhwd   %%xmm3,%%xmm9                 \n"
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpcklwd   %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm8,%%xmm2                 \n"
+      "movdqa      %%xmm9,%%xmm3                 \n"
+      "movdqa      %%xmm4,%%xmm8                 \n"
+      "movdqa      %%xmm5,%%xmm9                 \n"
+      "punpckhwd   %%xmm6,%%xmm8                 \n"
+      "punpckhwd   %%xmm7,%%xmm9                 \n"
+      "punpcklwd   %%xmm6,%%xmm4                 \n"
+      "punpcklwd   %%xmm7,%%xmm5                 \n"
+      "movdqa      %%xmm8,%%xmm6                 \n"
+      "movdqa      %%xmm9,%%xmm7                 \n"
       // Third round of bit swap.
       // Write to the destination pointer.
-      "movdqa     %%xmm0,%%xmm8                    \n"
-      "punpckldq  %%xmm4,%%xmm0                    \n"
-      "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
-      "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
-      "punpckhdq  %%xmm4,%%xmm8                    \n"
-      "movlpd     %%xmm8,(%1,%5)                   \n"
-      "lea        (%1,%5,2),%1                     \n"
-      "movhpd     %%xmm8,(%2,%6)                   \n"
-      "lea        (%2,%6,2),%2                     \n"
-      "movdqa     %%xmm2,%%xmm8                    \n"
-      "punpckldq  %%xmm6,%%xmm2                    \n"
-      "movlpd     %%xmm2,(%1)                      \n"
-      "movhpd     %%xmm2,(%2)                      \n"
-      "punpckhdq  %%xmm6,%%xmm8                    \n"
-      "movlpd     %%xmm8,(%1,%5)                   \n"
-      "lea        (%1,%5,2),%1                     \n"
-      "movhpd     %%xmm8,(%2,%6)                   \n"
-      "lea        (%2,%6,2),%2                     \n"
-      "movdqa     %%xmm1,%%xmm8                    \n"
-      "punpckldq  %%xmm5,%%xmm1                    \n"
-      "movlpd     %%xmm1,(%1)                      \n"
-      "movhpd     %%xmm1,(%2)                      \n"
-      "punpckhdq  %%xmm5,%%xmm8                    \n"
-      "movlpd     %%xmm8,(%1,%5)                   \n"
-      "lea        (%1,%5,2),%1                     \n"
-      "movhpd     %%xmm8,(%2,%6)                   \n"
-      "lea        (%2,%6,2),%2                     \n"
-      "movdqa     %%xmm3,%%xmm8                    \n"
-      "punpckldq  %%xmm7,%%xmm3                    \n"
-      "movlpd     %%xmm3,(%1)                      \n"
-      "movhpd     %%xmm3,(%2)                      \n"
-      "punpckhdq  %%xmm7,%%xmm8                    \n"
-      "sub        $0x8,%3                          \n"
-      "movlpd     %%xmm8,(%1,%5)                   \n"
-      "lea        (%1,%5,2),%1                     \n"
-      "movhpd     %%xmm8,(%2,%6)                   \n"
-      "lea        (%2,%6,2),%2                     \n"
-      "jg         1b                               \n"
+      "movdqa      %%xmm0,%%xmm8                 \n"
+      "punpckldq   %%xmm4,%%xmm0                 \n"
+      "movlpd      %%xmm0,(%1)                   \n"  // Write back U channel
+      "movhpd      %%xmm0,(%2)                   \n"  // Write back V channel
+      "punpckhdq   %%xmm4,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm2,%%xmm8                 \n"
+      "punpckldq   %%xmm6,%%xmm2                 \n"
+      "movlpd      %%xmm2,(%1)                   \n"
+      "movhpd      %%xmm2,(%2)                   \n"
+      "punpckhdq   %%xmm6,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm1,%%xmm8                 \n"
+      "punpckldq   %%xmm5,%%xmm1                 \n"
+      "movlpd      %%xmm1,(%1)                   \n"
+      "movhpd      %%xmm1,(%2)                   \n"
+      "punpckhdq   %%xmm5,%%xmm8                 \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "movdqa      %%xmm3,%%xmm8                 \n"
+      "punpckldq   %%xmm7,%%xmm3                 \n"
+      "movlpd      %%xmm3,(%1)                   \n"
+      "movhpd      %%xmm3,(%2)                   \n"
+      "punpckhdq   %%xmm7,%%xmm8                 \n"
+      "sub         $0x8,%3                       \n"
+      "movlpd      %%xmm8,(%1,%5)                \n"
+      "lea         (%1,%5,2),%1                  \n"
+      "movhpd      %%xmm8,(%2,%6)                \n"
+      "lea         (%2,%6,2),%2                  \n"
+      "jg          1b                            \n"
       : "+r"(src),                      // %0
         "+r"(dst_a),                    // %1
         "+r"(dst_b),                    // %2
diff --git a/files/source/rotate_lsx.cc b/files/source/rotate_lsx.cc
new file mode 100644
index 00000000..94a2b91c
--- /dev/null
+++ b/files/source/rotate_lsx.cc
@@ -0,0 +1,243 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate_row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ILVLH_B(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_b, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_b, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_H(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_h, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_h, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_W(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_w, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_w, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define ILVLH_D(in0, in1, in2, in3, out0, out1, out2, out3)   \
+  {                                                           \
+    DUP2_ARG2(__lsx_vilvl_d, in1, in0, in3, in2, out0, out2); \
+    DUP2_ARG2(__lsx_vilvh_d, in1, in0, in3, in2, out1, out3); \
+  }
+
+#define LSX_ST_4(_dst0, _dst1, _dst2, _dst3, _dst, _stride, _stride2, \
+                 _stride3, _stride4)                                  \
+  {                                                                   \
+    __lsx_vst(_dst0, _dst, 0);                                        \
+    __lsx_vstx(_dst1, _dst, _stride);                                 \
+    __lsx_vstx(_dst2, _dst, _stride2);                                \
+    __lsx_vstx(_dst3, _dst, _stride3);                                \
+    _dst += _stride4;                                                 \
+  }
+
+#define LSX_ST_2(_dst0, _dst1, _dst, _stride, _stride2) \
+  {                                                     \
+    __lsx_vst(_dst0, _dst, 0);                          \
+    __lsx_vstx(_dst1, _dst, _stride);                   \
+    _dst += _stride2;                                   \
+  }
+
+void TransposeWx16_C(const uint8_t* src,
+                     int src_stride,
+                     uint8_t* dst,
+                     int dst_stride,
+                     int width) {
+  TransposeWx8_C(src, src_stride, dst, dst_stride, width);
+  TransposeWx8_C((src + 8 * src_stride), src_stride, (dst + 8), dst_stride,
+                 width);
+}
+
+void TransposeUVWx16_C(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst_a,
+                       int dst_stride_a,
+                       uint8_t* dst_b,
+                       int dst_stride_b,
+                       int width) {
+  TransposeUVWx8_C(src, src_stride, dst_a, dst_stride_a, dst_b, dst_stride_b,
+                   width);
+  TransposeUVWx8_C((src + 8 * src_stride), src_stride, (dst_a + 8),
+                   dst_stride_a, (dst_b + 8), dst_stride_b, width);
+}
+
+void TransposeWx16_LSX(const uint8_t* src,
+                       int src_stride,
+                       uint8_t* dst,
+                       int dst_stride,
+                       int width) {
+  int x;
+  int len = width / 16;
+  uint8_t* s;
+  int src_stride2 = src_stride << 1;
+  int src_stride3 = src_stride + src_stride2;
+  int src_stride4 = src_stride2 << 1;
+  int dst_stride2 = dst_stride << 1;
+  int dst_stride3 = dst_stride + dst_stride2;
+  int dst_stride4 = dst_stride2 << 1;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < len; x++) {
+    s = (uint8_t*)src;
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    ILVLH_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVLH_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    res8 = __lsx_vilvl_w(reg4, reg0);
+    res9 = __lsx_vilvh_w(reg4, reg0);
+    ILVLH_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg5, reg1);
+    res9 = __lsx_vilvh_w(reg5, reg1);
+    ILVLH_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg6, reg2);
+    res9 = __lsx_vilvh_w(reg6, reg2);
+    ILVLH_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    res8 = __lsx_vilvl_w(reg7, reg3);
+    res9 = __lsx_vilvh_w(reg7, reg3);
+    ILVLH_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_4(dst0, dst1, dst2, dst3, dst, dst_stride, dst_stride2, dst_stride3,
+             dst_stride4);
+    src += 16;
+  }
+}
+
+void TransposeUVWx16_LSX(const uint8_t* src,
+                         int src_stride,
+                         uint8_t* dst_a,
+                         int dst_stride_a,
+                         uint8_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int x;
+  int len = width / 8;
+  uint8_t* s;
+  int src_stride2 = src_stride << 1;
+  int src_stride3 = src_stride + src_stride2;
+  int src_stride4 = src_stride2 << 1;
+  int dst_stride_a2 = dst_stride_a << 1;
+  int dst_stride_b2 = dst_stride_b << 1;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i res0, res1, res2, res3, res4, res5, res6, res7, res8, res9;
+
+  for (x = 0; x < len; x++) {
+    s = (uint8_t*)src;
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    ILVLH_W(reg0, reg4, reg1, reg5, res0, res1, res2, res3);
+    ILVLH_W(reg2, reg6, reg3, reg7, res4, res5, res6, res7);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg0, reg1, reg2, reg3);
+    src0 = __lsx_vld(s, 0);
+    src1 = __lsx_vldx(s, src_stride);
+    src2 = __lsx_vldx(s, src_stride2);
+    src3 = __lsx_vldx(s, src_stride3);
+    s += src_stride4;
+    ILVLH_B(src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3);
+    ILVLH_H(tmp0, tmp2, tmp1, tmp3, reg4, reg5, reg6, reg7);
+    res8 = __lsx_vilvl_w(reg4, reg0);
+    res9 = __lsx_vilvh_w(reg4, reg0);
+    ILVLH_D(res0, res8, res1, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg5, reg1);
+    res9 = __lsx_vilvh_w(reg5, reg1);
+    ILVLH_D(res2, res8, res3, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg6, reg2);
+    res9 = __lsx_vilvh_w(reg6, reg2);
+    ILVLH_D(res4, res8, res5, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    res8 = __lsx_vilvl_w(reg7, reg3);
+    res9 = __lsx_vilvh_w(reg7, reg3);
+    ILVLH_D(res6, res8, res7, res9, dst0, dst1, dst2, dst3);
+    LSX_ST_2(dst0, dst2, dst_a, dst_stride_a, dst_stride_a2);
+    LSX_ST_2(dst1, dst3, dst_b, dst_stride_b, dst_stride_b2);
+    src += 16;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/files/source/rotate_neon.cc b/files/source/rotate_neon.cc
index fdc0dd47..844df2bf 100644
--- a/files/source/rotate_neon.cc
+++ b/files/source/rotate_neon.cc
@@ -38,52 +38,52 @@ void TransposeWx8_NEON(const uint8_t* src,
 
       // handle 8x8 blocks. this should be the majority of the plane
       "1:                                        \n"
-      "mov         %0, %1                      \n"
-
-      "vld1.8      {d0}, [%0], %2              \n"
-      "vld1.8      {d1}, [%0], %2              \n"
-      "vld1.8      {d2}, [%0], %2              \n"
-      "vld1.8      {d3}, [%0], %2              \n"
-      "vld1.8      {d4}, [%0], %2              \n"
-      "vld1.8      {d5}, [%0], %2              \n"
-      "vld1.8      {d6}, [%0], %2              \n"
-      "vld1.8      {d7}, [%0]                  \n"
-
-      "vtrn.8      d1, d0                      \n"
-      "vtrn.8      d3, d2                      \n"
-      "vtrn.8      d5, d4                      \n"
-      "vtrn.8      d7, d6                      \n"
-
-      "vtrn.16     d1, d3                      \n"
-      "vtrn.16     d0, d2                      \n"
-      "vtrn.16     d5, d7                      \n"
-      "vtrn.16     d4, d6                      \n"
-
-      "vtrn.32     d1, d5                      \n"
-      "vtrn.32     d0, d4                      \n"
-      "vtrn.32     d3, d7                      \n"
-      "vtrn.32     d2, d6                      \n"
-
-      "vrev16.8    q0, q0                      \n"
-      "vrev16.8    q1, q1                      \n"
-      "vrev16.8    q2, q2                      \n"
-      "vrev16.8    q3, q3                      \n"
-
-      "mov         %0, %3                      \n"
-
-      "vst1.8      {d1}, [%0], %4              \n"
-      "vst1.8      {d0}, [%0], %4              \n"
-      "vst1.8      {d3}, [%0], %4              \n"
-      "vst1.8      {d2}, [%0], %4              \n"
-      "vst1.8      {d5}, [%0], %4              \n"
-      "vst1.8      {d4}, [%0], %4              \n"
-      "vst1.8      {d7}, [%0], %4              \n"
-      "vst1.8      {d6}, [%0]                  \n"
-
-      "add         %1, #8                      \n"  // src += 8
-      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
-      "subs        %5,  #8                     \n"  // w   -= 8
-      "bge         1b                          \n"
+      "mov         %0, %1                        \n"
+
+      "vld1.8      {d0}, [%0], %2                \n"
+      "vld1.8      {d1}, [%0], %2                \n"
+      "vld1.8      {d2}, [%0], %2                \n"
+      "vld1.8      {d3}, [%0], %2                \n"
+      "vld1.8      {d4}, [%0], %2                \n"
+      "vld1.8      {d5}, [%0], %2                \n"
+      "vld1.8      {d6}, [%0], %2                \n"
+      "vld1.8      {d7}, [%0]                    \n"
+
+      "vtrn.8      d1, d0                        \n"
+      "vtrn.8      d3, d2                        \n"
+      "vtrn.8      d5, d4                        \n"
+      "vtrn.8      d7, d6                        \n"
+
+      "vtrn.16     d1, d3                        \n"
+      "vtrn.16     d0, d2                        \n"
+      "vtrn.16     d5, d7                        \n"
+      "vtrn.16     d4, d6                        \n"
+
+      "vtrn.32     d1, d5                        \n"
+      "vtrn.32     d0, d4                        \n"
+      "vtrn.32     d3, d7                        \n"
+      "vtrn.32     d2, d6                        \n"
+
+      "vrev16.8    q0, q0                        \n"
+      "vrev16.8    q1, q1                        \n"
+      "vrev16.8    q2, q2                        \n"
+      "vrev16.8    q3, q3                        \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.8      {d1}, [%0], %4                \n"
+      "vst1.8      {d0}, [%0], %4                \n"
+      "vst1.8      {d3}, [%0], %4                \n"
+      "vst1.8      {d2}, [%0], %4                \n"
+      "vst1.8      {d5}, [%0], %4                \n"
+      "vst1.8      {d4}, [%0], %4                \n"
+      "vst1.8      {d7}, [%0], %4                \n"
+      "vst1.8      {d6}, [%0]                    \n"
+
+      "add         %1, #8                        \n"  // src += 8
+      "add         %3, %3, %4, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                       \n"  // w   -= 8
+      "bge         1b                            \n"
 
       // add 8 back to counter. if the result is 0 there are
       // no residuals.
@@ -208,68 +208,70 @@ void TransposeUVWx8_NEON(const uint8_t* src,
 
       // handle 8x8 blocks. this should be the majority of the plane
       "1:                                        \n"
-      "mov         %0, %1                      \n"
-
-      "vld2.8      {d0,  d1},  [%0], %2        \n"
-      "vld2.8      {d2,  d3},  [%0], %2        \n"
-      "vld2.8      {d4,  d5},  [%0], %2        \n"
-      "vld2.8      {d6,  d7},  [%0], %2        \n"
-      "vld2.8      {d16, d17}, [%0], %2        \n"
-      "vld2.8      {d18, d19}, [%0], %2        \n"
-      "vld2.8      {d20, d21}, [%0], %2        \n"
-      "vld2.8      {d22, d23}, [%0]            \n"
-
-      "vtrn.8      q1, q0                      \n"
-      "vtrn.8      q3, q2                      \n"
-      "vtrn.8      q9, q8                      \n"
-      "vtrn.8      q11, q10                    \n"
-
-      "vtrn.16     q1, q3                      \n"
-      "vtrn.16     q0, q2                      \n"
-      "vtrn.16     q9, q11                     \n"
-      "vtrn.16     q8, q10                     \n"
-
-      "vtrn.32     q1, q9                      \n"
-      "vtrn.32     q0, q8                      \n"
-      "vtrn.32     q3, q11                     \n"
-      "vtrn.32     q2, q10                     \n"
-
-      "vrev16.8    q0, q0                      \n"
-      "vrev16.8    q1, q1                      \n"
-      "vrev16.8    q2, q2                      \n"
-      "vrev16.8    q3, q3                      \n"
-      "vrev16.8    q8, q8                      \n"
-      "vrev16.8    q9, q9                      \n"
-      "vrev16.8    q10, q10                    \n"
-      "vrev16.8    q11, q11                    \n"
-
-      "mov         %0, %3                      \n"
-
-      "vst1.8      {d2},  [%0], %4             \n"
-      "vst1.8      {d0},  [%0], %4             \n"
-      "vst1.8      {d6},  [%0], %4             \n"
-      "vst1.8      {d4},  [%0], %4             \n"
-      "vst1.8      {d18}, [%0], %4             \n"
-      "vst1.8      {d16}, [%0], %4             \n"
-      "vst1.8      {d22}, [%0], %4             \n"
-      "vst1.8      {d20}, [%0]                 \n"
-
-      "mov         %0, %5                      \n"
-
-      "vst1.8      {d3},  [%0], %6             \n"
-      "vst1.8      {d1},  [%0], %6             \n"
-      "vst1.8      {d7},  [%0], %6             \n"
-      "vst1.8      {d5},  [%0], %6             \n"
-      "vst1.8      {d19}, [%0], %6             \n"
-      "vst1.8      {d17}, [%0], %6             \n"
-      "vst1.8      {d23}, [%0], %6             \n"
-      "vst1.8      {d21}, [%0]                 \n"
-
-      "add         %1, #8*2                    \n"  // src   += 8*2
-      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
-      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
-      "subs        %7,  #8                     \n"  // w     -= 8
-      "bge         1b                          \n"
+      "mov         %0, %1                        \n"
+
+      "vld2.8      {d0,  d1},  [%0], %2          \n"
+      "vld2.8      {d2,  d3},  [%0], %2          \n"
+      "vld2.8      {d4,  d5},  [%0], %2          \n"
+      "vld2.8      {d6,  d7},  [%0], %2          \n"
+      "vld2.8      {d16, d17}, [%0], %2          \n"
+      "vld2.8      {d18, d19}, [%0], %2          \n"
+      "vld2.8      {d20, d21}, [%0], %2          \n"
+      "vld2.8      {d22, d23}, [%0]              \n"
+
+      "vtrn.8      q1, q0                        \n"
+      "vtrn.8      q3, q2                        \n"
+      "vtrn.8      q9, q8                        \n"
+      "vtrn.8      q11, q10                      \n"
+
+      "vtrn.16     q1, q3                        \n"
+      "vtrn.16     q0, q2                        \n"
+      "vtrn.16     q9, q11                       \n"
+      "vtrn.16     q8, q10                       \n"
+
+      "vtrn.32     q1, q9                        \n"
+      "vtrn.32     q0, q8                        \n"
+      "vtrn.32     q3, q11                       \n"
+      "vtrn.32     q2, q10                       \n"
+
+      "vrev16.8    q0, q0                        \n"
+      "vrev16.8    q1, q1                        \n"
+      "vrev16.8    q2, q2                        \n"
+      "vrev16.8    q3, q3                        \n"
+      "vrev16.8    q8, q8                        \n"
+      "vrev16.8    q9, q9                        \n"
+      "vrev16.8    q10, q10                      \n"
+      "vrev16.8    q11, q11                      \n"
+
+      "mov         %0, %3                        \n"
+
+      "vst1.8      {d2},  [%0], %4               \n"
+      "vst1.8      {d0},  [%0], %4               \n"
+      "vst1.8      {d6},  [%0], %4               \n"
+      "vst1.8      {d4},  [%0], %4               \n"
+      "vst1.8      {d18}, [%0], %4               \n"
+      "vst1.8      {d16}, [%0], %4               \n"
+      "vst1.8      {d22}, [%0], %4               \n"
+      "vst1.8      {d20}, [%0]                   \n"
+
+      "mov         %0, %5                        \n"
+
+      "vst1.8      {d3},  [%0], %6               \n"
+      "vst1.8      {d1},  [%0], %6               \n"
+      "vst1.8      {d7},  [%0], %6               \n"
+      "vst1.8      {d5},  [%0], %6               \n"
+      "vst1.8      {d19}, [%0], %6               \n"
+      "vst1.8      {d17}, [%0], %6               \n"
+      "vst1.8      {d23}, [%0], %6               \n"
+      "vst1.8      {d21}, [%0]                   \n"
+
+      "add         %1, #8*2                      \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3            \n"  // dst_a += 8 *
+                                                      // dst_stride_a
+      "add         %5, %5, %6, lsl #3            \n"  // dst_b += 8 *
+                                                      // dst_stride_b
+      "subs        %7,  #8                       \n"  // w     -= 8
+      "bge         1b                            \n"
 
       // add 8 back to counter. if the result is 0 there are
       // no residuals.
diff --git a/files/source/rotate_neon64.cc b/files/source/rotate_neon64.cc
index f469baac..43c15817 100644
--- a/files/source/rotate_neon64.cc
+++ b/files/source/rotate_neon64.cc
@@ -34,58 +34,74 @@ void TransposeWx8_NEON(const uint8_t* src,
       // loops are on blocks of 8. loop will stop when
       // counter gets to or below 0. starting the counter
       // at w-8 allow for this
-      "sub         %w3, %w3, #8                     \n"
+      "sub         %w3, %w3, #8                  \n"
 
       // handle 8x8 blocks. this should be the majority of the plane
-      "1:                                          \n"
+      "1:                                        \n"
+      "mov         %0, %1                        \n"
+
+      "ld1         {v0.8b}, [%0], %5             \n"
+      "ld1         {v1.8b}, [%0], %5             \n"
+      "ld1         {v2.8b}, [%0], %5             \n"
+      "ld1         {v3.8b}, [%0], %5             \n"
+      "ld1         {v4.8b}, [%0], %5             \n"
+      "ld1         {v5.8b}, [%0], %5             \n"
+      "ld1         {v6.8b}, [%0], %5             \n"
+      "ld1         {v7.8b}, [%0]                 \n"
       "mov         %0, %1                        \n"
 
-      "ld1        {v0.8b}, [%0], %5              \n"
-      "ld1        {v1.8b}, [%0], %5              \n"
-      "ld1        {v2.8b}, [%0], %5              \n"
-      "ld1        {v3.8b}, [%0], %5              \n"
-      "ld1        {v4.8b}, [%0], %5              \n"
-      "ld1        {v5.8b}, [%0], %5              \n"
-      "ld1        {v6.8b}, [%0], %5              \n"
-      "ld1        {v7.8b}, [%0]                  \n"
-
-      "trn2     v16.8b, v0.8b, v1.8b             \n"
-      "trn1     v17.8b, v0.8b, v1.8b             \n"
-      "trn2     v18.8b, v2.8b, v3.8b             \n"
-      "trn1     v19.8b, v2.8b, v3.8b             \n"
-      "trn2     v20.8b, v4.8b, v5.8b             \n"
-      "trn1     v21.8b, v4.8b, v5.8b             \n"
-      "trn2     v22.8b, v6.8b, v7.8b             \n"
-      "trn1     v23.8b, v6.8b, v7.8b             \n"
-
-      "trn2     v3.4h, v17.4h, v19.4h            \n"
-      "trn1     v1.4h, v17.4h, v19.4h            \n"
-      "trn2     v2.4h, v16.4h, v18.4h            \n"
-      "trn1     v0.4h, v16.4h, v18.4h            \n"
-      "trn2     v7.4h, v21.4h, v23.4h            \n"
-      "trn1     v5.4h, v21.4h, v23.4h            \n"
-      "trn2     v6.4h, v20.4h, v22.4h            \n"
-      "trn1     v4.4h, v20.4h, v22.4h            \n"
-
-      "trn2     v21.2s, v1.2s, v5.2s             \n"
-      "trn1     v17.2s, v1.2s, v5.2s             \n"
-      "trn2     v20.2s, v0.2s, v4.2s             \n"
-      "trn1     v16.2s, v0.2s, v4.2s             \n"
-      "trn2     v23.2s, v3.2s, v7.2s             \n"
-      "trn1     v19.2s, v3.2s, v7.2s             \n"
-      "trn2     v22.2s, v2.2s, v6.2s             \n"
-      "trn1     v18.2s, v2.2s, v6.2s             \n"
+      "trn2        v16.8b, v0.8b, v1.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "trn1        v17.8b, v0.8b, v1.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v18.8b, v2.8b, v3.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 1
+      "trn1        v19.8b, v2.8b, v3.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v20.8b, v4.8b, v5.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 2
+      "trn1        v21.8b, v4.8b, v5.8b          \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v22.8b, v6.8b, v7.8b          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 3
+      "trn1        v23.8b, v6.8b, v7.8b          \n"
+      "add         %0, %0, %5                    \n"
+
+      "trn2        v3.4h, v17.4h, v19.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 4
+      "trn1        v1.4h, v17.4h, v19.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v2.4h, v16.4h, v18.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 5
+      "trn1        v0.4h, v16.4h, v18.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v7.4h, v21.4h, v23.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 6
+      "trn1        v5.4h, v21.4h, v23.4h         \n"
+      "add         %0, %0, %5                    \n"
+      "trn2        v6.4h, v20.4h, v22.4h         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // row 7
+      "trn1        v4.4h, v20.4h, v22.4h         \n"
+
+      "trn2        v21.2s, v1.2s, v5.2s          \n"
+      "trn1        v17.2s, v1.2s, v5.2s          \n"
+      "trn2        v20.2s, v0.2s, v4.2s          \n"
+      "trn1        v16.2s, v0.2s, v4.2s          \n"
+      "trn2        v23.2s, v3.2s, v7.2s          \n"
+      "trn1        v19.2s, v3.2s, v7.2s          \n"
+      "trn2        v22.2s, v2.2s, v6.2s          \n"
+      "trn1        v18.2s, v2.2s, v6.2s          \n"
 
       "mov         %0, %2                        \n"
 
-      "st1      {v17.8b}, [%0], %6               \n"
-      "st1      {v16.8b}, [%0], %6               \n"
-      "st1      {v19.8b}, [%0], %6               \n"
-      "st1      {v18.8b}, [%0], %6               \n"
-      "st1      {v21.8b}, [%0], %6               \n"
-      "st1      {v20.8b}, [%0], %6               \n"
-      "st1      {v23.8b}, [%0], %6               \n"
-      "st1      {v22.8b}, [%0]                   \n"
+      "st1         {v17.8b}, [%0], %6            \n"
+      "st1         {v16.8b}, [%0], %6            \n"
+      "st1         {v19.8b}, [%0], %6            \n"
+      "st1         {v18.8b}, [%0], %6            \n"
+      "st1         {v21.8b}, [%0], %6            \n"
+      "st1         {v20.8b}, [%0], %6            \n"
+      "st1         {v23.8b}, [%0], %6            \n"
+      "st1         {v22.8b}, [%0]                \n"
 
       "add         %1, %1, #8                    \n"  // src += 8
       "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
@@ -94,33 +110,33 @@ void TransposeWx8_NEON(const uint8_t* src,
 
       // add 8 back to counter. if the result is 0 there are
       // no residuals.
-      "adds        %w3, %w3, #8                    \n"
-      "b.eq        4f                              \n"
+      "adds        %w3, %w3, #8                  \n"
+      "b.eq        4f                            \n"
 
       // some residual, so between 1 and 7 lines left to transpose
-      "cmp         %w3, #2                          \n"
-      "b.lt        3f                              \n"
+      "cmp         %w3, #2                       \n"
+      "b.lt        3f                            \n"
 
-      "cmp         %w3, #4                          \n"
-      "b.lt        2f                              \n"
+      "cmp         %w3, #4                       \n"
+      "b.lt        2f                            \n"
 
       // 4x8 block
-      "mov         %0, %1                          \n"
-      "ld1     {v0.s}[0], [%0], %5                 \n"
-      "ld1     {v0.s}[1], [%0], %5                 \n"
-      "ld1     {v0.s}[2], [%0], %5                 \n"
-      "ld1     {v0.s}[3], [%0], %5                 \n"
-      "ld1     {v1.s}[0], [%0], %5                 \n"
-      "ld1     {v1.s}[1], [%0], %5                 \n"
-      "ld1     {v1.s}[2], [%0], %5                 \n"
-      "ld1     {v1.s}[3], [%0]                     \n"
+      "mov         %0, %1                        \n"
+      "ld1         {v0.s}[0], [%0], %5           \n"
+      "ld1         {v0.s}[1], [%0], %5           \n"
+      "ld1         {v0.s}[2], [%0], %5           \n"
+      "ld1         {v0.s}[3], [%0], %5           \n"
+      "ld1         {v1.s}[0], [%0], %5           \n"
+      "ld1         {v1.s}[1], [%0], %5           \n"
+      "ld1         {v1.s}[2], [%0], %5           \n"
+      "ld1         {v1.s}[3], [%0]               \n"
 
-      "mov         %0, %2                          \n"
+      "mov         %0, %2                        \n"
 
-      "ld1      {v2.16b}, [%4]                     \n"
+      "ld1         {v2.16b}, [%4]                \n"
 
-      "tbl      v3.16b, {v0.16b}, v2.16b           \n"
-      "tbl      v0.16b, {v1.16b}, v2.16b           \n"
+      "tbl         v3.16b, {v0.16b}, v2.16b      \n"
+      "tbl         v0.16b, {v1.16b}, v2.16b      \n"
 
       // TODO(frkoenig): Rework shuffle above to
       // write out with 4 instead of 8 writes.
@@ -212,89 +228,90 @@ void TransposeUVWx8_NEON(const uint8_t* src,
       // loops are on blocks of 8. loop will stop when
       // counter gets to or below 0. starting the counter
       // at w-8 allow for this
-      "sub       %w4, %w4, #8                    \n"
+      "sub         %w4, %w4, #8                  \n"
 
       // handle 8x8 blocks. this should be the majority of the plane
       "1:                                        \n"
-      "mov       %0, %1                          \n"
-
-      "ld1       {v0.16b}, [%0], %5              \n"
-      "ld1       {v1.16b}, [%0], %5              \n"
-      "ld1       {v2.16b}, [%0], %5              \n"
-      "ld1       {v3.16b}, [%0], %5              \n"
-      "ld1       {v4.16b}, [%0], %5              \n"
-      "ld1       {v5.16b}, [%0], %5              \n"
-      "ld1       {v6.16b}, [%0], %5              \n"
-      "ld1       {v7.16b}, [%0]                  \n"
-
-      "trn1      v16.16b, v0.16b, v1.16b         \n"
-      "trn2      v17.16b, v0.16b, v1.16b         \n"
-      "trn1      v18.16b, v2.16b, v3.16b         \n"
-      "trn2      v19.16b, v2.16b, v3.16b         \n"
-      "trn1      v20.16b, v4.16b, v5.16b         \n"
-      "trn2      v21.16b, v4.16b, v5.16b         \n"
-      "trn1      v22.16b, v6.16b, v7.16b         \n"
-      "trn2      v23.16b, v6.16b, v7.16b         \n"
-
-      "trn1      v0.8h, v16.8h, v18.8h           \n"
-      "trn2      v1.8h, v16.8h, v18.8h           \n"
-      "trn1      v2.8h, v20.8h, v22.8h           \n"
-      "trn2      v3.8h, v20.8h, v22.8h           \n"
-      "trn1      v4.8h, v17.8h, v19.8h           \n"
-      "trn2      v5.8h, v17.8h, v19.8h           \n"
-      "trn1      v6.8h, v21.8h, v23.8h           \n"
-      "trn2      v7.8h, v21.8h, v23.8h           \n"
-
-      "trn1      v16.4s, v0.4s, v2.4s            \n"
-      "trn2      v17.4s, v0.4s, v2.4s            \n"
-      "trn1      v18.4s, v1.4s, v3.4s            \n"
-      "trn2      v19.4s, v1.4s, v3.4s            \n"
-      "trn1      v20.4s, v4.4s, v6.4s            \n"
-      "trn2      v21.4s, v4.4s, v6.4s            \n"
-      "trn1      v22.4s, v5.4s, v7.4s            \n"
-      "trn2      v23.4s, v5.4s, v7.4s            \n"
+      "mov         %0, %1                        \n"
 
-      "mov       %0, %2                          \n"
+      "ld1         {v0.16b}, [%0], %5            \n"
+      "ld1         {v1.16b}, [%0], %5            \n"
+      "ld1         {v2.16b}, [%0], %5            \n"
+      "ld1         {v3.16b}, [%0], %5            \n"
+      "ld1         {v4.16b}, [%0], %5            \n"
+      "ld1         {v5.16b}, [%0], %5            \n"
+      "ld1         {v6.16b}, [%0], %5            \n"
+      "ld1         {v7.16b}, [%0]                \n"
+      "mov         %0, %1                        \n"
 
-      "st1       {v16.d}[0], [%0], %6            \n"
-      "st1       {v18.d}[0], [%0], %6            \n"
-      "st1       {v17.d}[0], [%0], %6            \n"
-      "st1       {v19.d}[0], [%0], %6            \n"
-      "st1       {v16.d}[1], [%0], %6            \n"
-      "st1       {v18.d}[1], [%0], %6            \n"
-      "st1       {v17.d}[1], [%0], %6            \n"
-      "st1       {v19.d}[1], [%0]                \n"
+      "trn1        v16.16b, v0.16b, v1.16b       \n"
+      "trn2        v17.16b, v0.16b, v1.16b       \n"
+      "trn1        v18.16b, v2.16b, v3.16b       \n"
+      "trn2        v19.16b, v2.16b, v3.16b       \n"
+      "trn1        v20.16b, v4.16b, v5.16b       \n"
+      "trn2        v21.16b, v4.16b, v5.16b       \n"
+      "trn1        v22.16b, v6.16b, v7.16b       \n"
+      "trn2        v23.16b, v6.16b, v7.16b       \n"
+
+      "trn1        v0.8h, v16.8h, v18.8h         \n"
+      "trn2        v1.8h, v16.8h, v18.8h         \n"
+      "trn1        v2.8h, v20.8h, v22.8h         \n"
+      "trn2        v3.8h, v20.8h, v22.8h         \n"
+      "trn1        v4.8h, v17.8h, v19.8h         \n"
+      "trn2        v5.8h, v17.8h, v19.8h         \n"
+      "trn1        v6.8h, v21.8h, v23.8h         \n"
+      "trn2        v7.8h, v21.8h, v23.8h         \n"
+
+      "trn1        v16.4s, v0.4s, v2.4s          \n"
+      "trn2        v17.4s, v0.4s, v2.4s          \n"
+      "trn1        v18.4s, v1.4s, v3.4s          \n"
+      "trn2        v19.4s, v1.4s, v3.4s          \n"
+      "trn1        v20.4s, v4.4s, v6.4s          \n"
+      "trn2        v21.4s, v4.4s, v6.4s          \n"
+      "trn1        v22.4s, v5.4s, v7.4s          \n"
+      "trn2        v23.4s, v5.4s, v7.4s          \n"
 
-      "mov       %0, %3                          \n"
+      "mov         %0, %2                        \n"
 
-      "st1       {v20.d}[0], [%0], %7            \n"
-      "st1       {v22.d}[0], [%0], %7            \n"
-      "st1       {v21.d}[0], [%0], %7            \n"
-      "st1       {v23.d}[0], [%0], %7            \n"
-      "st1       {v20.d}[1], [%0], %7            \n"
-      "st1       {v22.d}[1], [%0], %7            \n"
-      "st1       {v21.d}[1], [%0], %7            \n"
-      "st1       {v23.d}[1], [%0]                \n"
-
-      "add       %1, %1, #16                     \n"  // src   += 8*2
-      "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 *
+      "st1         {v16.d}[0], [%0], %6          \n"
+      "st1         {v18.d}[0], [%0], %6          \n"
+      "st1         {v17.d}[0], [%0], %6          \n"
+      "st1         {v19.d}[0], [%0], %6          \n"
+      "st1         {v16.d}[1], [%0], %6          \n"
+      "st1         {v18.d}[1], [%0], %6          \n"
+      "st1         {v17.d}[1], [%0], %6          \n"
+      "st1         {v19.d}[1], [%0]              \n"
+
+      "mov         %0, %3                        \n"
+
+      "st1         {v20.d}[0], [%0], %7          \n"
+      "st1         {v22.d}[0], [%0], %7          \n"
+      "st1         {v21.d}[0], [%0], %7          \n"
+      "st1         {v23.d}[0], [%0], %7          \n"
+      "st1         {v20.d}[1], [%0], %7          \n"
+      "st1         {v22.d}[1], [%0], %7          \n"
+      "st1         {v21.d}[1], [%0], %7          \n"
+      "st1         {v23.d}[1], [%0]              \n"
+
+      "add         %1, %1, #16                   \n"  // src   += 8*2
+      "add         %2, %2, %6, lsl #3            \n"  // dst_a += 8 *
                                                       // dst_stride_a
-      "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 *
+      "add         %3, %3, %7, lsl #3            \n"  // dst_b += 8 *
                                                       // dst_stride_b
-      "subs      %w4, %w4,  #8                   \n"  // w     -= 8
-      "b.ge      1b                              \n"
+      "subs        %w4, %w4,  #8                 \n"  // w     -= 8
+      "b.ge        1b                            \n"
 
       // add 8 back to counter. if the result is 0 there are
       // no residuals.
-      "adds      %w4, %w4, #8                    \n"
-      "b.eq      4f                              \n"
+      "adds        %w4, %w4, #8                  \n"
+      "b.eq        4f                            \n"
 
       // some residual, so between 1 and 7 lines left to transpose
-      "cmp       %w4, #2                         \n"
-      "b.lt      3f                              \n"
+      "cmp         %w4, #2                       \n"
+      "b.lt        3f                            \n"
 
-      "cmp       %w4, #4                         \n"
-      "b.lt      2f                              \n"
+      "cmp         %w4, #4                       \n"
+      "b.lt        2f                            \n"
 
       // TODO(frkoenig): Clean this up
       // 4x8 block
diff --git a/files/source/rotate_win.cc b/files/source/rotate_win.cc
index e887dd52..a78873f8 100644
--- a/files/source/rotate_win.cc
+++ b/files/source/rotate_win.cc
@@ -16,8 +16,9 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
 
 __declspec(naked) void TransposeWx8_SSSE3(const uint8_t* src,
                                           int src_stride,
diff --git a/files/source/row_any.cc b/files/source/row_any.cc
index 06ca723a..3781a9f2 100644
--- a/files/source/row_any.cc
+++ b/files/source/row_any.cc
@@ -30,6 +30,39 @@ extern "C" {
 // Subsampled source needs to be increase by 1 of not even.
 #define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift))
 
+// Any 4 planes to 1
+#define ANY41(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)               \
+  void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,                   \
+               const uint8_t* v_buf, const uint8_t* a_buf, uint8_t* dst_ptr, \
+               int width) {                                                  \
+    SIMD_ALIGNED(uint8_t temp[64 * 5]);                                      \
+    memset(temp, 0, 64 * 4); /* for msan */                                  \
+    int r = width & MASK;                                                    \
+    int n = width & ~MASK;                                                   \
+    if (n > 0) {                                                             \
+      ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, n);                      \
+    }                                                                        \
+    memcpy(temp, y_buf + n, r);                                              \
+    memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
+    memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
+    memcpy(temp + 192, a_buf + n, r);                                        \
+    ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, temp + 256, MASK + 1); \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 256,                      \
+           SS(r, DUVSHIFT) * BPP);                                           \
+  }
+
+#ifdef HAS_MERGEARGBROW_SSE2
+ANY41(MergeARGBRow_Any_SSE2, MergeARGBRow_SSE2, 0, 0, 4, 7)
+#endif
+#ifdef HAS_MERGEARGBROW_AVX2
+ANY41(MergeARGBRow_Any_AVX2, MergeARGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_MERGEARGBROW_NEON
+ANY41(MergeARGBRow_Any_NEON, MergeARGBRow_NEON, 0, 0, 4, 15)
+#endif
+
+// Note that odd width replication includes 444 due to implementation
+// on arm that subsamples 444 to 422 internally.
 // Any 4 planes to 1 with yuvconstants
 #define ANY41C(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)              \
   void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,                   \
@@ -46,26 +79,163 @@ extern "C" {
     memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
     memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
     memcpy(temp + 192, a_buf + n, r);                                        \
+    if (width & 1) {                                                         \
+      temp[64 + SS(r, UVSHIFT)] = temp[64 + SS(r, UVSHIFT) - 1];             \
+      temp[128 + SS(r, UVSHIFT)] = temp[128 + SS(r, UVSHIFT) - 1];           \
+    }                                                                        \
     ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, temp + 256,            \
              yuvconstants, MASK + 1);                                        \
     memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 256,                      \
            SS(r, DUVSHIFT) * BPP);                                           \
   }
 
+#ifdef HAS_I444ALPHATOARGBROW_SSSE3
+ANY41C(I444AlphaToARGBRow_Any_SSSE3, I444AlphaToARGBRow_SSSE3, 0, 0, 4, 7)
+#endif
+#ifdef HAS_I444ALPHATOARGBROW_AVX2
+ANY41C(I444AlphaToARGBRow_Any_AVX2, I444AlphaToARGBRow_AVX2, 0, 0, 4, 15)
+#endif
 #ifdef HAS_I422ALPHATOARGBROW_SSSE3
 ANY41C(I422AlphaToARGBRow_Any_SSSE3, I422AlphaToARGBRow_SSSE3, 1, 0, 4, 7)
 #endif
 #ifdef HAS_I422ALPHATOARGBROW_AVX2
 ANY41C(I422AlphaToARGBRow_Any_AVX2, I422AlphaToARGBRow_AVX2, 1, 0, 4, 15)
 #endif
+#ifdef HAS_I444ALPHATOARGBROW_NEON
+ANY41C(I444AlphaToARGBRow_Any_NEON, I444AlphaToARGBRow_NEON, 0, 0, 4, 7)
+#endif
 #ifdef HAS_I422ALPHATOARGBROW_NEON
 ANY41C(I422AlphaToARGBRow_Any_NEON, I422AlphaToARGBRow_NEON, 1, 0, 4, 7)
 #endif
+#ifdef HAS_I444ALPHATOARGBROW_MSA
+ANY41C(I444AlphaToARGBRow_Any_MSA, I444AlphaToARGBRow_MSA, 0, 0, 4, 7)
+#endif
 #ifdef HAS_I422ALPHATOARGBROW_MSA
 ANY41C(I422AlphaToARGBRow_Any_MSA, I422AlphaToARGBRow_MSA, 1, 0, 4, 7)
 #endif
+#ifdef HAS_I422ALPHATOARGBROW_LASX
+ANY41C(I422AlphaToARGBRow_Any_LASX, I422AlphaToARGBRow_LASX, 1, 0, 4, 15)
+#endif
 #undef ANY41C
 
+// Any 4 planes to 1 plane of 8 bit with yuvconstants
+#define ANY41CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
+  void NAMEANY(const T* y_buf, const T* u_buf, const T* v_buf, const T* a_buf, \
+               uint8_t* dst_ptr, const struct YuvConstants* yuvconstants,      \
+               int width) {                                                    \
+    SIMD_ALIGNED(T temp[16 * 4]);                                              \
+    SIMD_ALIGNED(uint8_t out[64]);                                             \
+    memset(temp, 0, 16 * 4 * SBPP); /* for YUY2 and msan */                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(y_buf, u_buf, v_buf, a_buf, dst_ptr, yuvconstants, n);          \
+    }                                                                          \
+    memcpy(temp, y_buf + n, r * SBPP);                                         \
+    memcpy(temp + 16, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);          \
+    memcpy(temp + 32, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP);          \
+    memcpy(temp + 48, a_buf + n, r * SBPP);                                    \
+    ANY_SIMD(temp, temp + 16, temp + 32, temp + 48, out, yuvconstants,         \
+             MASK + 1);                                                        \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, out, SS(r, DUVSHIFT) * BPP);       \
+  }
+
+#ifdef HAS_I210ALPHATOARGBROW_SSSE3
+ANY41CT(I210AlphaToARGBRow_Any_SSSE3,
+        I210AlphaToARGBRow_SSSE3,
+        1,
+        0,
+        uint16_t,
+        2,
+        4,
+        7)
+#endif
+
+#ifdef HAS_I210ALPHATOARGBROW_AVX2
+ANY41CT(I210AlphaToARGBRow_Any_AVX2,
+        I210AlphaToARGBRow_AVX2,
+        1,
+        0,
+        uint16_t,
+        2,
+        4,
+        15)
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_SSSE3
+ANY41CT(I410AlphaToARGBRow_Any_SSSE3,
+        I410AlphaToARGBRow_SSSE3,
+        0,
+        0,
+        uint16_t,
+        2,
+        4,
+        7)
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_AVX2
+ANY41CT(I410AlphaToARGBRow_Any_AVX2,
+        I410AlphaToARGBRow_AVX2,
+        0,
+        0,
+        uint16_t,
+        2,
+        4,
+        15)
+#endif
+
+#undef ANY41CT
+
+// Any 4 planes to 1 plane with parameter
+#define ANY41PT(NAMEANY, ANY_SIMD, STYPE, SBPP, DTYPE, BPP, MASK)          \
+  void NAMEANY(const STYPE* r_buf, const STYPE* g_buf, const STYPE* b_buf, \
+               const STYPE* a_buf, DTYPE* dst_ptr, int depth, int width) { \
+    SIMD_ALIGNED(STYPE temp[16 * 4]);                                      \
+    SIMD_ALIGNED(DTYPE out[64]);                                           \
+    memset(temp, 0, 16 * 4 * SBPP); /* for YUY2 and msan */                \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(r_buf, g_buf, b_buf, a_buf, dst_ptr, depth, n);             \
+    }                                                                      \
+    memcpy(temp, r_buf + n, r * SBPP);                                     \
+    memcpy(temp + 16, g_buf + n, r * SBPP);                                \
+    memcpy(temp + 32, b_buf + n, r * SBPP);                                \
+    memcpy(temp + 48, a_buf + n, r * SBPP);                                \
+    ANY_SIMD(temp, temp + 16, temp + 32, temp + 48, out, depth, MASK + 1); \
+    memcpy((uint8_t*)dst_ptr + n * BPP, out, r * BPP);                     \
+  }
+
+#ifdef HAS_MERGEAR64ROW_AVX2
+ANY41PT(MergeAR64Row_Any_AVX2, MergeAR64Row_AVX2, uint16_t, 2, uint16_t, 8, 15)
+#endif
+
+#ifdef HAS_MERGEAR64ROW_NEON
+ANY41PT(MergeAR64Row_Any_NEON, MergeAR64Row_NEON, uint16_t, 2, uint16_t, 8, 7)
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_AVX2
+ANY41PT(MergeARGB16To8Row_Any_AVX2,
+        MergeARGB16To8Row_AVX2,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        15)
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_NEON
+ANY41PT(MergeARGB16To8Row_Any_NEON,
+        MergeARGB16To8Row_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        7)
+#endif
+
+#undef ANY41PT
+
 // Any 3 planes to 1.
 #define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)      \
   void NAMEANY(const uint8_t* y_buf, const uint8_t* u_buf,          \
@@ -92,8 +262,14 @@ ANY31(MergeRGBRow_Any_SSSE3, MergeRGBRow_SSSE3, 0, 0, 3, 15)
 #ifdef HAS_MERGERGBROW_NEON
 ANY31(MergeRGBRow_Any_NEON, MergeRGBRow_NEON, 0, 0, 3, 15)
 #endif
-#ifdef HAS_MERGERGBROW_MMI
-ANY31(MergeRGBRow_Any_MMI, MergeRGBRow_MMI, 0, 0, 3, 7)
+#ifdef HAS_MERGEXRGBROW_SSE2
+ANY31(MergeXRGBRow_Any_SSE2, MergeXRGBRow_SSE2, 0, 0, 4, 7)
+#endif
+#ifdef HAS_MERGEXRGBROW_AVX2
+ANY31(MergeXRGBRow_Any_AVX2, MergeXRGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_MERGEXRGBROW_NEON
+ANY31(MergeXRGBRow_Any_NEON, MergeXRGBRow_NEON, 0, 0, 4, 15)
 #endif
 #ifdef HAS_I422TOYUY2ROW_SSE2
 ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15)
@@ -109,8 +285,8 @@ ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15)
 #ifdef HAS_I422TOYUY2ROW_MSA
 ANY31(I422ToYUY2Row_Any_MSA, I422ToYUY2Row_MSA, 1, 1, 4, 31)
 #endif
-#ifdef HAS_I422TOYUY2ROW_MMI
-ANY31(I422ToYUY2Row_Any_MMI, I422ToYUY2Row_MMI, 1, 1, 4, 7)
+#ifdef HAS_I422TOYUY2ROW_LASX
+ANY31(I422ToYUY2Row_Any_LASX, I422ToYUY2Row_LASX, 1, 1, 4, 31)
 #endif
 #ifdef HAS_I422TOUYVYROW_NEON
 ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15)
@@ -118,8 +294,8 @@ ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15)
 #ifdef HAS_I422TOUYVYROW_MSA
 ANY31(I422ToUYVYRow_Any_MSA, I422ToUYVYRow_MSA, 1, 1, 4, 31)
 #endif
-#ifdef HAS_I422TOUYVYROW_MMI
-ANY31(I422ToUYVYRow_Any_MMI, I422ToUYVYRow_MMI, 1, 1, 4, 7)
+#ifdef HAS_I422TOUYVYROW_LASX
+ANY31(I422ToUYVYRow_Any_LASX, I422ToUYVYRow_LASX, 1, 1, 4, 31)
 #endif
 #ifdef HAS_BLENDPLANEROW_AVX2
 ANY31(BlendPlaneRow_Any_AVX2, BlendPlaneRow_AVX2, 0, 0, 1, 31)
@@ -127,9 +303,6 @@ ANY31(BlendPlaneRow_Any_AVX2, BlendPlaneRow_AVX2, 0, 0, 1, 31)
 #ifdef HAS_BLENDPLANEROW_SSSE3
 ANY31(BlendPlaneRow_Any_SSSE3, BlendPlaneRow_SSSE3, 0, 0, 1, 7)
 #endif
-#ifdef HAS_BLENDPLANEROW_MMI
-ANY31(BlendPlaneRow_Any_MMI, BlendPlaneRow_MMI, 0, 0, 1, 7)
-#endif
 #undef ANY31
 
 // Note that odd width replication includes 444 due to implementation
@@ -162,6 +335,21 @@ ANY31(BlendPlaneRow_Any_MMI, BlendPlaneRow_MMI, 0, 0, 1, 7)
 #ifdef HAS_I422TOARGBROW_SSSE3
 ANY31C(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, 1, 0, 4, 7)
 #endif
+#ifdef HAS_I422TORGBAROW_SSSE3
+ANY31C(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I422TOARGB4444ROW_SSSE3
+ANY31C(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGB1555ROW_SSSE3
+ANY31C(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TORGB565ROW_SSSE3
+ANY31C(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TORGB24ROW_SSSE3
+ANY31C(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 15)
+#endif
 #ifdef HAS_I422TOAR30ROW_SSSE3
 ANY31C(I422ToAR30Row_Any_SSSE3, I422ToAR30Row_SSSE3, 1, 0, 4, 7)
 #endif
@@ -170,18 +358,16 @@ ANY31C(I422ToAR30Row_Any_AVX2, I422ToAR30Row_AVX2, 1, 0, 4, 15)
 #endif
 #ifdef HAS_I444TOARGBROW_SSSE3
 ANY31C(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7)
-ANY31C(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7)
-ANY31C(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7)
-ANY31C(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7)
-ANY31C(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7)
-ANY31C(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 15)
-#endif  // HAS_I444TOARGBROW_SSSE3
+#endif
 #ifdef HAS_I422TORGB24ROW_AVX2
 ANY31C(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 31)
 #endif
 #ifdef HAS_I422TOARGBROW_AVX2
 ANY31C(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, 1, 0, 4, 15)
 #endif
+#ifdef HAS_I422TOARGBROW_AVX512BW
+ANY31C(I422ToARGBRow_Any_AVX512BW, I422ToARGBRow_AVX512BW, 1, 0, 4, 31)
+#endif
 #ifdef HAS_I422TORGBAROW_AVX2
 ANY31C(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, 1, 0, 4, 15)
 #endif
@@ -215,6 +401,17 @@ ANY31C(I422ToARGB4444Row_Any_MSA, I422ToARGB4444Row_MSA, 1, 0, 2, 7)
 ANY31C(I422ToARGB1555Row_Any_MSA, I422ToARGB1555Row_MSA, 1, 0, 2, 7)
 ANY31C(I422ToRGB565Row_Any_MSA, I422ToRGB565Row_MSA, 1, 0, 2, 7)
 #endif
+#ifdef HAS_I422TOARGBROW_LASX
+ANY31C(I422ToARGBRow_Any_LASX, I422ToARGBRow_LASX, 1, 0, 4, 31)
+ANY31C(I422ToRGBARow_Any_LASX, I422ToRGBARow_LASX, 1, 0, 4, 31)
+ANY31C(I422ToRGB24Row_Any_LASX, I422ToRGB24Row_LASX, 1, 0, 3, 31)
+ANY31C(I422ToRGB565Row_Any_LASX, I422ToRGB565Row_LASX, 1, 0, 2, 31)
+ANY31C(I422ToARGB4444Row_Any_LASX, I422ToARGB4444Row_LASX, 1, 0, 2, 31)
+ANY31C(I422ToARGB1555Row_Any_LASX, I422ToARGB1555Row_LASX, 1, 0, 2, 31)
+#endif
+#ifdef HAS_I444TOARGBROW_LSX
+ANY31C(I444ToARGBRow_Any_LSX, I444ToARGBRow_LSX, 0, 0, 4, 15)
+#endif
 #undef ANY31C
 
 // Any 3 planes of 16 bit to 1 with yuvconstants
@@ -250,24 +447,112 @@ ANY31CT(I210ToARGBRow_Any_AVX2, I210ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
 #ifdef HAS_I210TOAR30ROW_AVX2
 ANY31CT(I210ToAR30Row_Any_AVX2, I210ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
 #endif
+#ifdef HAS_I410TOAR30ROW_SSSE3
+ANY31CT(I410ToAR30Row_Any_SSSE3, I410ToAR30Row_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I410TOARGBROW_SSSE3
+ANY31CT(I410ToARGBRow_Any_SSSE3, I410ToARGBRow_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I410TOARGBROW_AVX2
+ANY31CT(I410ToARGBRow_Any_AVX2, I410ToARGBRow_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I410TOAR30ROW_AVX2
+ANY31CT(I410ToAR30Row_Any_AVX2, I410ToAR30Row_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I212TOAR30ROW_SSSE3
+ANY31CT(I212ToAR30Row_Any_SSSE3, I212ToAR30Row_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I212TOARGBROW_SSSE3
+ANY31CT(I212ToARGBRow_Any_SSSE3, I212ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_I212TOARGBROW_AVX2
+ANY31CT(I212ToARGBRow_Any_AVX2, I212ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_I212TOAR30ROW_AVX2
+ANY31CT(I212ToAR30Row_Any_AVX2, I212ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
 #undef ANY31CT
 
+// Any 3 planes to 1 plane with parameter
+#define ANY31PT(NAMEANY, ANY_SIMD, STYPE, SBPP, DTYPE, BPP, MASK)          \
+  void NAMEANY(const STYPE* r_buf, const STYPE* g_buf, const STYPE* b_buf, \
+               DTYPE* dst_ptr, int depth, int width) {                     \
+    SIMD_ALIGNED(STYPE temp[16 * 3]);                                      \
+    SIMD_ALIGNED(DTYPE out[64]);                                           \
+    memset(temp, 0, 16 * 3 * SBPP); /* for YUY2 and msan */                \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(r_buf, g_buf, b_buf, dst_ptr, depth, n);                    \
+    }                                                                      \
+    memcpy(temp, r_buf + n, r * SBPP);                                     \
+    memcpy(temp + 16, g_buf + n, r * SBPP);                                \
+    memcpy(temp + 32, b_buf + n, r * SBPP);                                \
+    ANY_SIMD(temp, temp + 16, temp + 32, out, depth, MASK + 1);            \
+    memcpy((uint8_t*)dst_ptr + n * BPP, out, r * BPP);                     \
+  }
+
+#ifdef HAS_MERGEXR30ROW_AVX2
+ANY31PT(MergeXR30Row_Any_AVX2, MergeXR30Row_AVX2, uint16_t, 2, uint8_t, 4, 15)
+#endif
+
+#ifdef HAS_MERGEXR30ROW_NEON
+ANY31PT(MergeXR30Row_Any_NEON, MergeXR30Row_NEON, uint16_t, 2, uint8_t, 4, 3)
+ANY31PT(MergeXR30Row_10_Any_NEON,
+        MergeXR30Row_10_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        3)
+#endif
+
+#ifdef HAS_MERGEXR64ROW_AVX2
+ANY31PT(MergeXR64Row_Any_AVX2, MergeXR64Row_AVX2, uint16_t, 2, uint16_t, 8, 15)
+#endif
+
+#ifdef HAS_MERGEXR64ROW_NEON
+ANY31PT(MergeXR64Row_Any_NEON, MergeXR64Row_NEON, uint16_t, 2, uint16_t, 8, 7)
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_AVX2
+ANY31PT(MergeXRGB16To8Row_Any_AVX2,
+        MergeXRGB16To8Row_AVX2,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        15)
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_NEON
+ANY31PT(MergeXRGB16To8Row_Any_NEON,
+        MergeXRGB16To8Row_NEON,
+        uint16_t,
+        2,
+        uint8_t,
+        4,
+        7)
+#endif
+
+#undef ANY31PT
+
 // Any 2 planes to 1.
 #define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)             \
   void NAMEANY(const uint8_t* y_buf, const uint8_t* uv_buf, uint8_t* dst_ptr, \
                int width) {                                                   \
-    SIMD_ALIGNED(uint8_t temp[64 * 3]);                                       \
-    memset(temp, 0, 64 * 2); /* for msan */                                   \
+    SIMD_ALIGNED(uint8_t temp[128 * 3]);                                      \
+    memset(temp, 0, 128 * 2); /* for msan */                                  \
     int r = width & MASK;                                                     \
     int n = width & ~MASK;                                                    \
     if (n > 0) {                                                              \
       ANY_SIMD(y_buf, uv_buf, dst_ptr, n);                                    \
     }                                                                         \
     memcpy(temp, y_buf + n * SBPP, r * SBPP);                                 \
-    memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2,                        \
+    memcpy(temp + 128, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
            SS(r, UVSHIFT) * SBPP2);                                           \
-    ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1);                          \
-    memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                           \
+    ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                         \
+    memcpy(dst_ptr + n * BPP, temp + 256, r * BPP);                           \
   }
 
 // Merge functions.
@@ -283,12 +568,15 @@ ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15)
 #ifdef HAS_MERGEUVROW_MSA
 ANY21(MergeUVRow_Any_MSA, MergeUVRow_MSA, 0, 1, 1, 2, 15)
 #endif
-#ifdef HAS_MERGEUVROW_MMI
-ANY21(MergeUVRow_Any_MMI, MergeUVRow_MMI, 0, 1, 1, 2, 7)
+#ifdef HAS_MERGEUVROW_LSX
+ANY21(MergeUVRow_Any_LSX, MergeUVRow_LSX, 0, 1, 1, 2, 15)
 #endif
 #ifdef HAS_NV21TOYUV24ROW_NEON
 ANY21(NV21ToYUV24Row_Any_NEON, NV21ToYUV24Row_NEON, 1, 1, 2, 3, 15)
 #endif
+#ifdef HAS_NV21TOYUV24ROW_SSSE3
+ANY21(NV21ToYUV24Row_Any_SSSE3, NV21ToYUV24Row_SSSE3, 1, 1, 2, 3, 15)
+#endif
 #ifdef HAS_NV21TOYUV24ROW_AVX2
 ANY21(NV21ToYUV24Row_Any_AVX2, NV21ToYUV24Row_AVX2, 1, 1, 2, 3, 31)
 #endif
@@ -323,20 +611,20 @@ ANY21(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, 0, 4, 4, 4, 7)
 #ifdef HAS_ARGBMULTIPLYROW_MSA
 ANY21(ARGBMultiplyRow_Any_MSA, ARGBMultiplyRow_MSA, 0, 4, 4, 4, 3)
 #endif
-#ifdef HAS_ARGBMULTIPLYROW_MMI
-ANY21(ARGBMultiplyRow_Any_MMI, ARGBMultiplyRow_MMI, 0, 4, 4, 4, 1)
+#ifdef HAS_ARGBMULTIPLYROW_LASX
+ANY21(ARGBMultiplyRow_Any_LASX, ARGBMultiplyRow_LASX, 0, 4, 4, 4, 7)
 #endif
 #ifdef HAS_ARGBADDROW_MSA
 ANY21(ARGBAddRow_Any_MSA, ARGBAddRow_MSA, 0, 4, 4, 4, 7)
 #endif
-#ifdef HAS_ARGBADDROW_MMI
-ANY21(ARGBAddRow_Any_MMI, ARGBAddRow_MMI, 0, 4, 4, 4, 1)
+#ifdef HAS_ARGBADDROW_LASX
+ANY21(ARGBAddRow_Any_LASX, ARGBAddRow_LASX, 0, 4, 4, 4, 7)
 #endif
 #ifdef HAS_ARGBSUBTRACTROW_MSA
 ANY21(ARGBSubtractRow_Any_MSA, ARGBSubtractRow_MSA, 0, 4, 4, 4, 7)
 #endif
-#ifdef HAS_ARGBSUBTRACTROW_MMI
-ANY21(ARGBSubtractRow_Any_MMI, ARGBSubtractRow_MMI, 0, 4, 4, 4, 1)
+#ifdef HAS_ARGBSUBTRACTROW_LASX
+ANY21(ARGBSubtractRow_Any_LASX, ARGBSubtractRow_LASX, 0, 4, 4, 4, 7)
 #endif
 #ifdef HAS_SOBELROW_SSE2
 ANY21(SobelRow_Any_SSE2, SobelRow_SSE2, 0, 1, 1, 4, 15)
@@ -347,8 +635,8 @@ ANY21(SobelRow_Any_NEON, SobelRow_NEON, 0, 1, 1, 4, 7)
 #ifdef HAS_SOBELROW_MSA
 ANY21(SobelRow_Any_MSA, SobelRow_MSA, 0, 1, 1, 4, 15)
 #endif
-#ifdef HAS_SOBELROW_MMI
-ANY21(SobelRow_Any_MMI, SobelRow_MMI, 0, 1, 1, 4, 7)
+#ifdef HAS_SOBELROW_LSX
+ANY21(SobelRow_Any_LSX, SobelRow_LSX, 0, 1, 1, 4, 15)
 #endif
 #ifdef HAS_SOBELTOPLANEROW_SSE2
 ANY21(SobelToPlaneRow_Any_SSE2, SobelToPlaneRow_SSE2, 0, 1, 1, 1, 15)
@@ -359,8 +647,8 @@ ANY21(SobelToPlaneRow_Any_NEON, SobelToPlaneRow_NEON, 0, 1, 1, 1, 15)
 #ifdef HAS_SOBELTOPLANEROW_MSA
 ANY21(SobelToPlaneRow_Any_MSA, SobelToPlaneRow_MSA, 0, 1, 1, 1, 31)
 #endif
-#ifdef HAS_SOBELTOPLANEROW_MMI
-ANY21(SobelToPlaneRow_Any_MMI, SobelToPlaneRow_MMI, 0, 1, 1, 1, 7)
+#ifdef HAS_SOBELTOPLANEROW_LSX
+ANY21(SobelToPlaneRow_Any_LSX, SobelToPlaneRow_LSX, 0, 1, 1, 1, 31)
 #endif
 #ifdef HAS_SOBELXYROW_SSE2
 ANY21(SobelXYRow_Any_SSE2, SobelXYRow_SSE2, 0, 1, 1, 4, 15)
@@ -371,8 +659,8 @@ ANY21(SobelXYRow_Any_NEON, SobelXYRow_NEON, 0, 1, 1, 4, 7)
 #ifdef HAS_SOBELXYROW_MSA
 ANY21(SobelXYRow_Any_MSA, SobelXYRow_MSA, 0, 1, 1, 4, 15)
 #endif
-#ifdef HAS_SOBELXYROW_MMI
-ANY21(SobelXYRow_Any_MMI, SobelXYRow_MMI, 0, 1, 1, 4, 7)
+#ifdef HAS_SOBELXYROW_LSX
+ANY21(SobelXYRow_Any_LSX, SobelXYRow_LSX, 0, 1, 1, 4, 15)
 #endif
 #undef ANY21
 
@@ -407,6 +695,12 @@ ANY21C(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, 1, 1, 2, 4, 7)
 #ifdef HAS_NV12TOARGBROW_MSA
 ANY21C(NV12ToARGBRow_Any_MSA, NV12ToARGBRow_MSA, 1, 1, 2, 4, 7)
 #endif
+#ifdef HAS_NV12TOARGBROW_LSX
+ANY21C(NV12ToARGBRow_Any_LSX, NV12ToARGBRow_LSX, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_LASX
+ANY21C(NV12ToARGBRow_Any_LASX, NV12ToARGBRow_LASX, 1, 1, 2, 4, 15)
+#endif
 #ifdef HAS_NV21TOARGBROW_SSSE3
 ANY21C(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
 #endif
@@ -419,6 +713,12 @@ ANY21C(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, 1, 1, 2, 4, 7)
 #ifdef HAS_NV21TOARGBROW_MSA
 ANY21C(NV21ToARGBRow_Any_MSA, NV21ToARGBRow_MSA, 1, 1, 2, 4, 7)
 #endif
+#ifdef HAS_NV21TOARGBROW_LSX
+ANY21C(NV21ToARGBRow_Any_LSX, NV21ToARGBRow_LSX, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV21TOARGBROW_LASX
+ANY21C(NV21ToARGBRow_Any_LASX, NV21ToARGBRow_LASX, 1, 1, 2, 4, 15)
+#endif
 #ifdef HAS_NV12TORGB24ROW_NEON
 ANY21C(NV12ToRGB24Row_Any_NEON, NV12ToRGB24Row_NEON, 1, 1, 2, 3, 7)
 #endif
@@ -449,8 +749,85 @@ ANY21C(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, 1, 1, 2, 2, 7)
 #ifdef HAS_NV12TORGB565ROW_MSA
 ANY21C(NV12ToRGB565Row_Any_MSA, NV12ToRGB565Row_MSA, 1, 1, 2, 2, 7)
 #endif
+#ifdef HAS_NV12TORGB565ROW_LSX
+ANY21C(NV12ToRGB565Row_Any_LSX, NV12ToRGB565Row_LSX, 1, 1, 2, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_LASX
+ANY21C(NV12ToRGB565Row_Any_LASX, NV12ToRGB565Row_LASX, 1, 1, 2, 2, 15)
+#endif
 #undef ANY21C
 
+// Any 2 planes of 16 bit to 1 with yuvconstants
+#define ANY21CT(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, T, SBPP, BPP, MASK)      \
+  void NAMEANY(const T* y_buf, const T* uv_buf, uint8_t* dst_ptr,              \
+               const struct YuvConstants* yuvconstants, int width) {           \
+    SIMD_ALIGNED(T temp[16 * 3]);                                              \
+    SIMD_ALIGNED(uint8_t out[64]);                                             \
+    memset(temp, 0, 16 * 3 * SBPP); /* for YUY2 and msan */                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(y_buf, uv_buf, dst_ptr, yuvconstants, n);                       \
+    }                                                                          \
+    memcpy(temp, y_buf + n, r * SBPP);                                         \
+    memcpy(temp + 16, uv_buf + 2 * (n >> UVSHIFT), SS(r, UVSHIFT) * SBPP * 2); \
+    ANY_SIMD(temp, temp + 16, out, yuvconstants, MASK + 1);                    \
+    memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, out, SS(r, DUVSHIFT) * BPP);       \
+  }
+
+#ifdef HAS_P210TOAR30ROW_SSSE3
+ANY21CT(P210ToAR30Row_Any_SSSE3, P210ToAR30Row_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P210TOARGBROW_SSSE3
+ANY21CT(P210ToARGBRow_Any_SSSE3, P210ToARGBRow_SSSE3, 1, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P210TOARGBROW_AVX2
+ANY21CT(P210ToARGBRow_Any_AVX2, P210ToARGBRow_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P210TOAR30ROW_AVX2
+ANY21CT(P210ToAR30Row_Any_AVX2, P210ToAR30Row_AVX2, 1, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P410TOAR30ROW_SSSE3
+ANY21CT(P410ToAR30Row_Any_SSSE3, P410ToAR30Row_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P410TOARGBROW_SSSE3
+ANY21CT(P410ToARGBRow_Any_SSSE3, P410ToARGBRow_SSSE3, 0, 0, uint16_t, 2, 4, 7)
+#endif
+#ifdef HAS_P410TOARGBROW_AVX2
+ANY21CT(P410ToARGBRow_Any_AVX2, P410ToARGBRow_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+#ifdef HAS_P410TOAR30ROW_AVX2
+ANY21CT(P410ToAR30Row_Any_AVX2, P410ToAR30Row_AVX2, 0, 0, uint16_t, 2, 4, 15)
+#endif
+
+#undef ANY21CT
+
+// Any 2 16 bit planes with parameter to 1
+#define ANY21PT(NAMEANY, ANY_SIMD, T, BPP, MASK)                     \
+  void NAMEANY(const T* src_u, const T* src_v, T* dst_uv, int depth, \
+               int width) {                                          \
+    SIMD_ALIGNED(T temp[16 * 4]);                                    \
+    memset(temp, 0, 16 * 4 * BPP); /* for msan */                    \
+    int r = width & MASK;                                            \
+    int n = width & ~MASK;                                           \
+    if (n > 0) {                                                     \
+      ANY_SIMD(src_u, src_v, dst_uv, depth, n);                      \
+    }                                                                \
+    memcpy(temp, src_u + n, r * BPP);                                \
+    memcpy(temp + 16, src_v + n, r * BPP);                           \
+    ANY_SIMD(temp, temp + 16, temp + 32, depth, MASK + 1);           \
+    memcpy(dst_uv + n * 2, temp + 32, r * BPP * 2);                  \
+  }
+
+#ifdef HAS_MERGEUVROW_16_AVX2
+ANY21PT(MergeUVRow_16_Any_AVX2, MergeUVRow_16_AVX2, uint16_t, 2, 15)
+#endif
+#ifdef HAS_MERGEUVROW_16_NEON
+ANY21PT(MergeUVRow_16_Any_NEON, MergeUVRow_16_NEON, uint16_t, 2, 7)
+#endif
+
+#undef ANY21CT
+
 // Any 1 to 1.
 #define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                \
   void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_ptr, int width) {     \
@@ -516,12 +893,6 @@ ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7)
 #if defined(HAS_J400TOARGBROW_AVX2)
 ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15)
 #endif
-#if defined(HAS_I400TOARGBROW_SSE2)
-ANY11(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, 0, 1, 4, 7)
-#endif
-#if defined(HAS_I400TOARGBROW_AVX2)
-ANY11(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, 0, 1, 4, 15)
-#endif
 #if defined(HAS_RGB24TOARGBROW_SSSE3)
 ANY11(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, 0, 3, 4, 15)
 ANY11(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, 0, 3, 4, 15)
@@ -529,6 +900,9 @@ ANY11(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, 0, 2, 4, 7)
 ANY11(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, 0, 2, 4, 7)
 ANY11(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, 0, 2, 4, 7)
 #endif
+#if defined(HAS_RAWTORGBAROW_SSSE3)
+ANY11(RAWToRGBARow_Any_SSSE3, RAWToRGBARow_SSSE3, 0, 3, 4, 15)
+#endif
 #if defined(HAS_RAWTORGB24ROW_SSSE3)
 ANY11(RAWToRGB24Row_Any_SSSE3, RAWToRGB24Row_SSSE3, 0, 3, 3, 7)
 #endif
@@ -542,13 +916,12 @@ ANY11(ARGB1555ToARGBRow_Any_AVX2, ARGB1555ToARGBRow_AVX2, 0, 2, 4, 15)
 ANY11(ARGB4444ToARGBRow_Any_AVX2, ARGB4444ToARGBRow_AVX2, 0, 2, 4, 15)
 #endif
 #if defined(HAS_ARGBTORGB24ROW_NEON)
-ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 15)
 ANY11(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, 0, 4, 3, 7)
 ANY11(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, 0, 4, 2, 7)
 ANY11(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, 0, 4, 2, 7)
 ANY11(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, 0, 4, 2, 7)
 ANY11(J400ToARGBRow_Any_NEON, J400ToARGBRow_NEON, 0, 1, 4, 7)
-ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7)
 #endif
 #if defined(HAS_ARGBTORGB24ROW_MSA)
 ANY11(ARGBToRGB24Row_Any_MSA, ARGBToRGB24Row_MSA, 0, 4, 3, 15)
@@ -557,16 +930,16 @@ ANY11(ARGBToRGB565Row_Any_MSA, ARGBToRGB565Row_MSA, 0, 4, 2, 7)
 ANY11(ARGBToARGB1555Row_Any_MSA, ARGBToARGB1555Row_MSA, 0, 4, 2, 7)
 ANY11(ARGBToARGB4444Row_Any_MSA, ARGBToARGB4444Row_MSA, 0, 4, 2, 7)
 ANY11(J400ToARGBRow_Any_MSA, J400ToARGBRow_MSA, 0, 1, 4, 15)
-ANY11(I400ToARGBRow_Any_MSA, I400ToARGBRow_MSA, 0, 1, 4, 15)
 #endif
-#if defined(HAS_ARGBTORGB24ROW_MMI)
-ANY11(ARGBToRGB24Row_Any_MMI, ARGBToRGB24Row_MMI, 0, 4, 3, 3)
-ANY11(ARGBToRAWRow_Any_MMI, ARGBToRAWRow_MMI, 0, 4, 3, 3)
-ANY11(ARGBToRGB565Row_Any_MMI, ARGBToRGB565Row_MMI, 0, 4, 2, 3)
-ANY11(ARGBToARGB1555Row_Any_MMI, ARGBToARGB1555Row_MMI, 0, 4, 2, 3)
-ANY11(ARGBToARGB4444Row_Any_MMI, ARGBToARGB4444Row_MMI, 0, 4, 2, 3)
-ANY11(J400ToARGBRow_Any_MMI, J400ToARGBRow_MMI, 0, 1, 4, 3)
-ANY11(I400ToARGBRow_Any_MMI, I400ToARGBRow_MMI, 0, 1, 4, 7)
+#if defined(HAS_ARGBTORGB24ROW_LASX)
+ANY11(ARGBToRGB24Row_Any_LASX, ARGBToRGB24Row_LASX, 0, 4, 3, 31)
+ANY11(ARGBToRAWRow_Any_LASX, ARGBToRAWRow_LASX, 0, 4, 3, 31)
+ANY11(ARGBToRGB565Row_Any_LASX, ARGBToRGB565Row_LASX, 0, 4, 2, 15)
+ANY11(ARGBToARGB1555Row_Any_LASX, ARGBToARGB1555Row_LASX, 0, 4, 2, 15)
+ANY11(ARGBToARGB4444Row_Any_LASX, ARGBToARGB4444Row_LASX, 0, 4, 2, 15)
+#endif
+#if defined(HAS_J400TOARGBROW_LSX)
+ANY11(J400ToARGBRow_Any_LSX, J400ToARGBRow_LSX, 0, 1, 4, 15)
 #endif
 #if defined(HAS_RAWTORGB24ROW_NEON)
 ANY11(RAWToRGB24Row_Any_NEON, RAWToRGB24Row_NEON, 0, 3, 3, 7)
@@ -574,15 +947,21 @@ ANY11(RAWToRGB24Row_Any_NEON, RAWToRGB24Row_NEON, 0, 3, 3, 7)
 #if defined(HAS_RAWTORGB24ROW_MSA)
 ANY11(RAWToRGB24Row_Any_MSA, RAWToRGB24Row_MSA, 0, 3, 3, 15)
 #endif
-#if defined(HAS_RAWTORGB24ROW_MMI)
-ANY11(RAWToRGB24Row_Any_MMI, RAWToRGB24Row_MMI, 0, 3, 3, 3)
+#if defined(HAS_RAWTORGB24ROW_LSX)
+ANY11(RAWToRGB24Row_Any_LSX, RAWToRGB24Row_LSX, 0, 3, 3, 15)
 #endif
 #ifdef HAS_ARGBTOYROW_AVX2
 ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31)
 #endif
+#ifdef HAS_ABGRTOYROW_AVX2
+ANY11(ABGRToYRow_Any_AVX2, ABGRToYRow_AVX2, 0, 4, 1, 31)
+#endif
 #ifdef HAS_ARGBTOYJROW_AVX2
 ANY11(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 0, 4, 1, 31)
 #endif
+#ifdef HAS_RGBATOYJROW_AVX2
+ANY11(RGBAToYJRow_Any_AVX2, RGBAToYJRow_AVX2, 0, 4, 1, 31)
+#endif
 #ifdef HAS_UYVYTOYROW_AVX2
 ANY11(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 0, 2, 1, 31)
 #endif
@@ -596,74 +975,109 @@ ANY11(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 0, 4, 1, 15)
 ANY11(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOYROW_SSE2
 ANY11(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 1, 4, 1, 15)
 ANY11(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 1, 4, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYJROW_SSSE3
 ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15)
 #endif
+#ifdef HAS_RGBATOYJROW_SSSE3
+ANY11(RGBAToYJRow_Any_SSSE3, RGBAToYJRow_SSSE3, 0, 4, 1, 15)
+#endif
 #ifdef HAS_ARGBTOYROW_NEON
-ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYROW_MSA
 ANY11(ARGBToYRow_Any_MSA, ARGBToYRow_MSA, 0, 4, 1, 15)
 #endif
-#ifdef HAS_ARGBTOYROW_MMI
-ANY11(ARGBToYRow_Any_MMI, ARGBToYRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_ARGBTOYROW_LASX
+ANY11(ARGBToYRow_Any_LASX, ARGBToYRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_ARGBTOYJROW_NEON
-ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 15)
+#endif
+#ifdef HAS_RGBATOYJROW_NEON
+ANY11(RGBAToYJRow_Any_NEON, RGBAToYJRow_NEON, 0, 4, 1, 15)
 #endif
 #ifdef HAS_ARGBTOYJROW_MSA
 ANY11(ARGBToYJRow_Any_MSA, ARGBToYJRow_MSA, 0, 4, 1, 15)
 #endif
-#ifdef HAS_ARGBTOYJROW_MMI
-ANY11(ARGBToYJRow_Any_MMI, ARGBToYJRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_ARGBTOYJROW_LSX
+ANY11(ARGBToYJRow_Any_LSX, ARGBToYJRow_LSX, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYJROW_LASX
+ANY11(ARGBToYJRow_Any_LASX, ARGBToYJRow_LASX, 0, 4, 1, 31)
 #endif
 #ifdef HAS_BGRATOYROW_NEON
-ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7)
+ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 15)
 #endif
 #ifdef HAS_BGRATOYROW_MSA
 ANY11(BGRAToYRow_Any_MSA, BGRAToYRow_MSA, 0, 4, 1, 15)
 #endif
-#ifdef HAS_BGRATOYROW_MMI
-ANY11(BGRAToYRow_Any_MMI, BGRAToYRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_BGRATOYROW_LSX
+ANY11(BGRAToYRow_Any_LSX, BGRAToYRow_LSX, 0, 4, 1, 15)
 #endif
 #ifdef HAS_ABGRTOYROW_NEON
-ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7)
+ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 15)
 #endif
 #ifdef HAS_ABGRTOYROW_MSA
 ANY11(ABGRToYRow_Any_MSA, ABGRToYRow_MSA, 0, 4, 1, 7)
 #endif
-#ifdef HAS_ABGRTOYROW_MMI
-ANY11(ABGRToYRow_Any_MMI, ABGRToYRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_ABGRTOYROW_LSX
+ANY11(ABGRToYRow_Any_LSX, ABGRToYRow_LSX, 0, 4, 1, 15)
 #endif
 #ifdef HAS_RGBATOYROW_NEON
-ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7)
+ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 15)
 #endif
 #ifdef HAS_RGBATOYROW_MSA
 ANY11(RGBAToYRow_Any_MSA, RGBAToYRow_MSA, 0, 4, 1, 15)
 #endif
-#ifdef HAS_RGBATOYROW_MMI
-ANY11(RGBAToYRow_Any_MMI, RGBAToYRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_RGBATOYROW_LSX
+ANY11(RGBAToYRow_Any_LSX, RGBAToYRow_LSX, 0, 4, 1, 15)
 #endif
 #ifdef HAS_RGB24TOYROW_NEON
-ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
+ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_AVX2
+ANY11(RGB24ToYJRow_Any_AVX2, RGB24ToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RGB24TOYJROW_SSSE3
+ANY11(RGB24ToYJRow_Any_SSSE3, RGB24ToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYJROW_NEON
+ANY11(RGB24ToYJRow_Any_NEON, RGB24ToYJRow_NEON, 0, 3, 1, 15)
 #endif
 #ifdef HAS_RGB24TOYROW_MSA
 ANY11(RGB24ToYRow_Any_MSA, RGB24ToYRow_MSA, 0, 3, 1, 15)
 #endif
-#ifdef HAS_RGB24TOYROW_MMI
-ANY11(RGB24ToYRow_Any_MMI, RGB24ToYRow_MMI, 0, 3, 1, 7)
+#ifdef HAS_RGB24TOYROW_LSX
+ANY11(RGB24ToYRow_Any_LSX, RGB24ToYRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RGB24TOYROW_LASX
+ANY11(RGB24ToYRow_Any_LASX, RGB24ToYRow_LASX, 0, 3, 1, 31)
 #endif
 #ifdef HAS_RAWTOYROW_NEON
-ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
+ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYJROW_AVX2
+ANY11(RAWToYJRow_Any_AVX2, RAWToYJRow_AVX2, 0, 3, 1, 31)
+#endif
+#ifdef HAS_RAWTOYJROW_SSSE3
+ANY11(RAWToYJRow_Any_SSSE3, RAWToYJRow_SSSE3, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYJROW_NEON
+ANY11(RAWToYJRow_Any_NEON, RAWToYJRow_NEON, 0, 3, 1, 15)
 #endif
 #ifdef HAS_RAWTOYROW_MSA
 ANY11(RAWToYRow_Any_MSA, RAWToYRow_MSA, 0, 3, 1, 15)
 #endif
-#ifdef HAS_RAWTOYROW_MMI
-ANY11(RAWToYRow_Any_MMI, RAWToYRow_MMI, 0, 3, 1, 7)
+#ifdef HAS_RAWTOYROW_LSX
+ANY11(RAWToYRow_Any_LSX, RAWToYRow_LSX, 0, 3, 1, 15)
+#endif
+#ifdef HAS_RAWTOYROW_LASX
+ANY11(RAWToYRow_Any_LASX, RAWToYRow_LASX, 0, 3, 1, 31)
 #endif
 #ifdef HAS_RGB565TOYROW_NEON
 ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7)
@@ -671,8 +1085,11 @@ ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7)
 #ifdef HAS_RGB565TOYROW_MSA
 ANY11(RGB565ToYRow_Any_MSA, RGB565ToYRow_MSA, 0, 2, 1, 15)
 #endif
-#ifdef HAS_RGB565TOYROW_MMI
-ANY11(RGB565ToYRow_Any_MMI, RGB565ToYRow_MMI, 0, 2, 1, 7)
+#ifdef HAS_RGB565TOYROW_LSX
+ANY11(RGB565ToYRow_Any_LSX, RGB565ToYRow_LSX, 0, 2, 1, 15)
+#endif
+#ifdef HAS_RGB565TOYROW_LASX
+ANY11(RGB565ToYRow_Any_LASX, RGB565ToYRow_LASX, 0, 2, 1, 31)
 #endif
 #ifdef HAS_ARGB1555TOYROW_NEON
 ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7)
@@ -680,15 +1097,15 @@ ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7)
 #ifdef HAS_ARGB1555TOYROW_MSA
 ANY11(ARGB1555ToYRow_Any_MSA, ARGB1555ToYRow_MSA, 0, 2, 1, 15)
 #endif
-#ifdef HAS_ARGB1555TOYROW_MMI
-ANY11(ARGB1555ToYRow_Any_MMI, ARGB1555ToYRow_MMI, 0, 2, 1, 7)
+#ifdef HAS_ARGB1555TOYROW_LSX
+ANY11(ARGB1555ToYRow_Any_LSX, ARGB1555ToYRow_LSX, 0, 2, 1, 15)
+#endif
+#ifdef HAS_ARGB1555TOYROW_LASX
+ANY11(ARGB1555ToYRow_Any_LASX, ARGB1555ToYRow_LASX, 0, 2, 1, 31)
 #endif
 #ifdef HAS_ARGB4444TOYROW_NEON
 ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7)
 #endif
-#ifdef HAS_ARGB4444TOYROW_MMI
-ANY11(ARGB4444ToYRow_Any_MMI, ARGB4444ToYRow_MMI, 0, 2, 1, 7)
-#endif
 #ifdef HAS_YUY2TOYROW_NEON
 ANY11(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 1, 4, 1, 15)
 #endif
@@ -698,20 +1115,26 @@ ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 1, 4, 1, 15)
 #ifdef HAS_YUY2TOYROW_MSA
 ANY11(YUY2ToYRow_Any_MSA, YUY2ToYRow_MSA, 1, 4, 1, 31)
 #endif
-#ifdef HAS_YUY2TOYROW_MMI
-ANY11(YUY2ToYRow_Any_MMI, YUY2ToYRow_MMI, 1, 4, 1, 7)
+#ifdef HAS_YUY2TOYROW_LASX
+ANY11(YUY2ToYRow_Any_LASX, YUY2ToYRow_LASX, 1, 4, 1, 31)
 #endif
 #ifdef HAS_UYVYTOYROW_MSA
 ANY11(UYVYToYRow_Any_MSA, UYVYToYRow_MSA, 1, 4, 1, 31)
 #endif
-#ifdef HAS_UYVYTOYROW_MMI
-ANY11(UYVYToYRow_Any_MMI, UYVYToYRow_MMI, 1, 4, 1, 15)
+#ifdef HAS_UYVYTOYROW_LASX
+ANY11(UYVYToYRow_Any_LASX, UYVYToYRow_LASX, 1, 4, 1, 31)
 #endif
 #ifdef HAS_AYUVTOYROW_NEON
 ANY11(AYUVToYRow_Any_NEON, AYUVToYRow_NEON, 0, 4, 1, 15)
 #endif
-#ifdef HAS_AYUVTOYROW_NEON
-ANY11(UVToVURow_Any_NEON, UVToVURow_NEON, 0, 2, 2, 15)
+#ifdef HAS_SWAPUVROW_SSSE3
+ANY11(SwapUVRow_Any_SSSE3, SwapUVRow_SSSE3, 0, 2, 2, 15)
+#endif
+#ifdef HAS_SWAPUVROW_AVX2
+ANY11(SwapUVRow_Any_AVX2, SwapUVRow_AVX2, 0, 2, 2, 31)
+#endif
+#ifdef HAS_SWAPUVROW_NEON
+ANY11(SwapUVRow_Any_NEON, SwapUVRow_NEON, 0, 2, 2, 15)
 #endif
 #ifdef HAS_RGB24TOARGBROW_NEON
 ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7)
@@ -719,17 +1142,26 @@ ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7)
 #ifdef HAS_RGB24TOARGBROW_MSA
 ANY11(RGB24ToARGBRow_Any_MSA, RGB24ToARGBRow_MSA, 0, 3, 4, 15)
 #endif
-#ifdef HAS_RGB24TOARGBROW_MMI
-ANY11(RGB24ToARGBRow_Any_MMI, RGB24ToARGBRow_MMI, 0, 3, 4, 3)
+#ifdef HAS_RGB24TOARGBROW_LSX
+ANY11(RGB24ToARGBRow_Any_LSX, RGB24ToARGBRow_LSX, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RGB24TOARGBROW_LASX
+ANY11(RGB24ToARGBRow_Any_LASX, RGB24ToARGBRow_LASX, 0, 3, 4, 31)
 #endif
 #ifdef HAS_RAWTOARGBROW_NEON
 ANY11(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 0, 3, 4, 7)
 #endif
+#ifdef HAS_RAWTORGBAROW_NEON
+ANY11(RAWToRGBARow_Any_NEON, RAWToRGBARow_NEON, 0, 3, 4, 7)
+#endif
 #ifdef HAS_RAWTOARGBROW_MSA
 ANY11(RAWToARGBRow_Any_MSA, RAWToARGBRow_MSA, 0, 3, 4, 15)
 #endif
-#ifdef HAS_RAWTOARGBROW_MMI
-ANY11(RAWToARGBRow_Any_MMI, RAWToARGBRow_MMI, 0, 3, 4, 3)
+#ifdef HAS_RAWTOARGBROW_LSX
+ANY11(RAWToARGBRow_Any_LSX, RAWToARGBRow_LSX, 0, 3, 4, 15)
+#endif
+#ifdef HAS_RAWTOARGBROW_LASX
+ANY11(RAWToARGBRow_Any_LASX, RAWToARGBRow_LASX, 0, 3, 4, 31)
 #endif
 #ifdef HAS_RGB565TOARGBROW_NEON
 ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7)
@@ -737,8 +1169,11 @@ ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7)
 #ifdef HAS_RGB565TOARGBROW_MSA
 ANY11(RGB565ToARGBRow_Any_MSA, RGB565ToARGBRow_MSA, 0, 2, 4, 15)
 #endif
-#ifdef HAS_RGB565TOARGBROW_MMI
-ANY11(RGB565ToARGBRow_Any_MMI, RGB565ToARGBRow_MMI, 0, 2, 4, 3)
+#ifdef HAS_RGB565TOARGBROW_LSX
+ANY11(RGB565ToARGBRow_Any_LSX, RGB565ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_RGB565TOARGBROW_LASX
+ANY11(RGB565ToARGBRow_Any_LASX, RGB565ToARGBRow_LASX, 0, 2, 4, 31)
 #endif
 #ifdef HAS_ARGB1555TOARGBROW_NEON
 ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7)
@@ -746,8 +1181,11 @@ ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7)
 #ifdef HAS_ARGB1555TOARGBROW_MSA
 ANY11(ARGB1555ToARGBRow_Any_MSA, ARGB1555ToARGBRow_MSA, 0, 2, 4, 15)
 #endif
-#ifdef HAS_ARGB1555TOARGBROW_MMI
-ANY11(ARGB1555ToARGBRow_Any_MMI, ARGB1555ToARGBRow_MMI, 0, 2, 4, 3)
+#ifdef HAS_ARGB1555TOARGBROW_LSX
+ANY11(ARGB1555ToARGBRow_Any_LSX, ARGB1555ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB1555TOARGBROW_LASX
+ANY11(ARGB1555ToARGBRow_Any_LASX, ARGB1555ToARGBRow_LASX, 0, 2, 4, 31)
 #endif
 #ifdef HAS_ARGB4444TOARGBROW_NEON
 ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
@@ -755,8 +1193,11 @@ ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
 #ifdef HAS_ARGB4444TOARGBROW_MSA
 ANY11(ARGB4444ToARGBRow_Any_MSA, ARGB4444ToARGBRow_MSA, 0, 2, 4, 15)
 #endif
-#ifdef HAS_ARGB4444TOARGBROW_MMI
-ANY11(ARGB4444ToARGBRow_Any_MMI, ARGB4444ToARGBRow_MMI, 0, 2, 4, 3)
+#ifdef HAS_ARGB4444TOARGBROW_LSX
+ANY11(ARGB4444ToARGBRow_Any_LSX, ARGB4444ToARGBRow_LSX, 0, 2, 4, 15)
+#endif
+#ifdef HAS_ARGB4444TOARGBROW_LASX
+ANY11(ARGB4444ToARGBRow_Any_LASX, ARGB4444ToARGBRow_LASX, 0, 2, 4, 31)
 #endif
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
 ANY11(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, 0, 4, 4, 3)
@@ -776,8 +1217,8 @@ ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
 #ifdef HAS_ARGBATTENUATEROW_MSA
 ANY11(ARGBAttenuateRow_Any_MSA, ARGBAttenuateRow_MSA, 0, 4, 4, 7)
 #endif
-#ifdef HAS_ARGBATTENUATEROW_MMI
-ANY11(ARGBAttenuateRow_Any_MMI, ARGBAttenuateRow_MMI, 0, 4, 4, 1)
+#ifdef HAS_ARGBATTENUATEROW_LASX
+ANY11(ARGBAttenuateRow_Any_LASX, ARGBAttenuateRow_LASX, 0, 4, 4, 15)
 #endif
 #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2
 ANY11(ARGBExtractAlphaRow_Any_SSE2, ARGBExtractAlphaRow_SSE2, 0, 4, 1, 7)
@@ -791,8 +1232,8 @@ ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 15)
 #ifdef HAS_ARGBEXTRACTALPHAROW_MSA
 ANY11(ARGBExtractAlphaRow_Any_MSA, ARGBExtractAlphaRow_MSA, 0, 4, 1, 15)
 #endif
-#ifdef HAS_ARGBEXTRACTALPHAROW_MMI
-ANY11(ARGBExtractAlphaRow_Any_MMI, ARGBExtractAlphaRow_MMI, 0, 4, 1, 7)
+#ifdef HAS_ARGBEXTRACTALPHAROW_LSX
+ANY11(ARGBExtractAlphaRow_Any_LSX, ARGBExtractAlphaRow_LSX, 0, 4, 1, 15)
 #endif
 #undef ANY11
 
@@ -818,18 +1259,12 @@ ANY11B(ARGBCopyAlphaRow_Any_AVX2, ARGBCopyAlphaRow_AVX2, 0, 4, 4, 15)
 #ifdef HAS_ARGBCOPYALPHAROW_SSE2
 ANY11B(ARGBCopyAlphaRow_Any_SSE2, ARGBCopyAlphaRow_SSE2, 0, 4, 4, 7)
 #endif
-#ifdef HAS_ARGBCOPYALPHAROW_MMI
-ANY11B(ARGBCopyAlphaRow_Any_MMI, ARGBCopyAlphaRow_MMI, 0, 4, 4, 1)
-#endif
 #ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
 ANY11B(ARGBCopyYToAlphaRow_Any_AVX2, ARGBCopyYToAlphaRow_AVX2, 0, 1, 4, 15)
 #endif
 #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
 ANY11B(ARGBCopyYToAlphaRow_Any_SSE2, ARGBCopyYToAlphaRow_SSE2, 0, 1, 4, 7)
 #endif
-#ifdef HAS_ARGBCOPYYTOALPHAROW_MMI
-ANY11B(ARGBCopyYToAlphaRow_Any_MMI, ARGBCopyYToAlphaRow_MMI, 0, 1, 4, 7)
-#endif
 #undef ANY11B
 
 // Any 1 to 1 with parameter.
@@ -847,6 +1282,47 @@ ANY11B(ARGBCopyYToAlphaRow_Any_MMI, ARGBCopyYToAlphaRow_MMI, 0, 1, 4, 7)
     memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                             \
   }
 
+#if defined(HAS_I400TOARGBROW_SSE2)
+ANY11P(I400ToARGBRow_Any_SSE2,
+       I400ToARGBRow_SSE2,
+       const struct YuvConstants*,
+       1,
+       4,
+       7)
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+ANY11P(I400ToARGBRow_Any_AVX2,
+       I400ToARGBRow_AVX2,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
+#if defined(HAS_I400TOARGBROW_NEON)
+ANY11P(I400ToARGBRow_Any_NEON,
+       I400ToARGBRow_NEON,
+       const struct YuvConstants*,
+       1,
+       4,
+       7)
+#endif
+#if defined(HAS_I400TOARGBROW_MSA)
+ANY11P(I400ToARGBRow_Any_MSA,
+       I400ToARGBRow_MSA,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
+#if defined(HAS_I400TOARGBROW_LSX)
+ANY11P(I400ToARGBRow_Any_LSX,
+       I400ToARGBRow_LSX,
+       const struct YuvConstants*,
+       1,
+       4,
+       15)
+#endif
+
 #if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
 ANY11P(ARGBToRGB565DitherRow_Any_SSE2,
        ARGBToRGB565DitherRow_SSE2,
@@ -879,13 +1355,13 @@ ANY11P(ARGBToRGB565DitherRow_Any_MSA,
        2,
        7)
 #endif
-#if defined(HAS_ARGBTORGB565DITHERROW_MMI)
-ANY11P(ARGBToRGB565DitherRow_Any_MMI,
-       ARGBToRGB565DitherRow_MMI,
+#if defined(HAS_ARGBTORGB565DITHERROW_LASX)
+ANY11P(ARGBToRGB565DitherRow_Any_LASX,
+       ARGBToRGB565DitherRow_LASX,
        const uint32_t,
        4,
        2,
-       3)
+       15)
 #endif
 #ifdef HAS_ARGBSHUFFLEROW_SSSE3
 ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8_t*, 4, 4, 7)
@@ -899,12 +1375,78 @@ ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8_t*, 4, 4, 3)
 #ifdef HAS_ARGBSHUFFLEROW_MSA
 ANY11P(ARGBShuffleRow_Any_MSA, ARGBShuffleRow_MSA, const uint8_t*, 4, 4, 7)
 #endif
-#ifdef HAS_ARGBSHUFFLEROW_MMI
-ANY11P(ARGBShuffleRow_Any_MMI, ARGBShuffleRow_MMI, const uint8_t*, 4, 4, 1)
+#ifdef HAS_ARGBSHUFFLEROW_LASX
+ANY11P(ARGBShuffleRow_Any_LASX, ARGBShuffleRow_LASX, const uint8_t*, 4, 4, 15)
 #endif
 #undef ANY11P
 #undef ANY11P
 
+// Any 1 to 1 with type
+#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, STYPE, DTYPE, MASK)  \
+  void NAMEANY(const STYPE* src_ptr, DTYPE* dst_ptr, int width) { \
+    SIMD_ALIGNED(uint8_t temp[(MASK + 1) * SBPP]);                \
+    SIMD_ALIGNED(uint8_t out[(MASK + 1) * BPP]);                  \
+    memset(temp, 0, (MASK + 1) * SBPP); /* for msan */            \
+    int r = width & MASK;                                         \
+    int n = width & ~MASK;                                        \
+    if (n > 0) {                                                  \
+      ANY_SIMD(src_ptr, dst_ptr, n);                              \
+    }                                                             \
+    memcpy(temp, (uint8_t*)(src_ptr) + n * SBPP, r * SBPP);       \
+    ANY_SIMD((STYPE*)temp, (DTYPE*)out, MASK + 1);                \
+    memcpy((uint8_t*)(dst_ptr) + n * BPP, out, r * BPP);          \
+  }
+
+#ifdef HAS_ARGBTOAR64ROW_SSSE3
+ANY11T(ARGBToAR64Row_Any_SSSE3, ARGBToAR64Row_SSSE3, 4, 8, uint8_t, uint16_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_SSSE3
+ANY11T(ARGBToAB64Row_Any_SSSE3, ARGBToAB64Row_SSSE3, 4, 8, uint8_t, uint16_t, 3)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_SSSE3
+ANY11T(AR64ToARGBRow_Any_SSSE3, AR64ToARGBRow_SSSE3, 8, 4, uint16_t, uint8_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_SSSE3
+ANY11T(AB64ToARGBRow_Any_SSSE3, AB64ToARGBRow_SSSE3, 8, 4, uint16_t, uint8_t, 3)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+ANY11T(ARGBToAR64Row_Any_AVX2, ARGBToAR64Row_AVX2, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_AVX2
+ANY11T(ARGBToAB64Row_Any_AVX2, ARGBToAB64Row_AVX2, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_AVX2
+ANY11T(AR64ToARGBRow_Any_AVX2, AR64ToARGBRow_AVX2, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+ANY11T(AB64ToARGBRow_Any_AVX2, AB64ToARGBRow_AVX2, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_NEON
+ANY11T(ARGBToAR64Row_Any_NEON, ARGBToAR64Row_NEON, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_NEON
+ANY11T(ARGBToAB64Row_Any_NEON, ARGBToAB64Row_NEON, 4, 8, uint8_t, uint16_t, 7)
+#endif
+
+#ifdef HAS_AR64TOARGBROW_NEON
+ANY11T(AR64ToARGBRow_Any_NEON, AR64ToARGBRow_NEON, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#ifdef HAS_ARGBTOAR64ROW_NEON
+ANY11T(AB64ToARGBRow_Any_NEON, AB64ToARGBRow_NEON, 8, 4, uint16_t, uint8_t, 7)
+#endif
+
+#undef ANY11T
+
 // Any 1 to 1 with parameter and shorts.  BPP measures in shorts.
 #define ANY11C(NAMEANY, ANY_SIMD, SBPP, BPP, STYPE, DTYPE, MASK)             \
   void NAMEANY(const STYPE* src_ptr, DTYPE* dst_ptr, int scale, int width) { \
@@ -939,6 +1481,15 @@ ANY11C(Convert16To8Row_Any_AVX2,
        uint8_t,
        31)
 #endif
+#ifdef HAS_CONVERT16TO8ROW_NEON
+ANY11C(Convert16To8Row_Any_NEON,
+       Convert16To8Row_NEON,
+       2,
+       1,
+       uint16_t,
+       uint8_t,
+       15)
+#endif
 #ifdef HAS_CONVERT8TO16ROW_SSE2
 ANY11C(Convert8To16Row_Any_SSE2,
        Convert8To16Row_SSE2,
@@ -957,6 +1508,30 @@ ANY11C(Convert8To16Row_Any_AVX2,
        uint16_t,
        31)
 #endif
+#ifdef HAS_MULTIPLYROW_16_AVX2
+ANY11C(MultiplyRow_16_Any_AVX2,
+       MultiplyRow_16_AVX2,
+       2,
+       2,
+       uint16_t,
+       uint16_t,
+       31)
+#endif
+#ifdef HAS_MULTIPLYROW_16_NEON
+ANY11C(MultiplyRow_16_Any_NEON,
+       MultiplyRow_16_NEON,
+       2,
+       2,
+       uint16_t,
+       uint16_t,
+       15)
+#endif
+#ifdef HAS_DIVIDEROW_16_AVX2
+ANY11C(DivideRow_16_Any_AVX2, DivideRow_16_AVX2, 2, 2, uint16_t, uint16_t, 31)
+#endif
+#ifdef HAS_DIVIDEROW_16_NEON
+ANY11C(DivideRow_16_Any_NEON, DivideRow_16_NEON, 2, 2, uint16_t, uint16_t, 15)
+#endif
 #undef ANY11C
 
 // Any 1 to 1 with parameter and shorts to byte.  BPP measures in shorts.
@@ -1007,6 +1582,9 @@ ANY11P16(HalfFloatRow_Any_MSA, HalfFloatRow_MSA, uint16_t, uint16_t, 2, 2, 31)
 #ifdef HAS_BYTETOFLOATROW_NEON
 ANY11P16(ByteToFloatRow_Any_NEON, ByteToFloatRow_NEON, uint8_t, float, 1, 3, 7)
 #endif
+#ifdef HAS_HALFFLOATROW_LSX
+ANY11P16(HalfFloatRow_Any_LSX, HalfFloatRow_LSX, uint16_t, uint16_t, 2, 2, 31)
+#endif
 #undef ANY11P16
 
 // Any 1 to 1 with yuvconstants
@@ -1040,41 +1618,107 @@ ANY11C(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, 1, 4, 4, 7)
 ANY11C(YUY2ToARGBRow_Any_MSA, YUY2ToARGBRow_MSA, 1, 4, 4, 7)
 ANY11C(UYVYToARGBRow_Any_MSA, UYVYToARGBRow_MSA, 1, 4, 4, 7)
 #endif
+#if defined(HAS_YUY2TOARGBROW_LSX)
+ANY11C(YUY2ToARGBRow_Any_LSX, YUY2ToARGBRow_LSX, 1, 4, 4, 7)
+ANY11C(UYVYToARGBRow_Any_LSX, UYVYToARGBRow_LSX, 1, 4, 4, 7)
+#endif
 #undef ANY11C
 
 // Any 1 to 1 interpolate.  Takes 2 rows of source via stride.
-#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                           \
-  void NAMEANY(uint8_t* dst_ptr, const uint8_t* src_ptr,                     \
-               ptrdiff_t src_stride_ptr, int width, int source_y_fraction) { \
-    SIMD_ALIGNED(uint8_t temp[64 * 3]);                                      \
-    memset(temp, 0, 64 * 2); /* for msan */                                  \
-    int r = width & MASK;                                                    \
-    int n = width & ~MASK;                                                   \
-    if (n > 0) {                                                             \
-      ANY_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction);      \
-    }                                                                        \
-    memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
-    memcpy(temp + 64, src_ptr + src_stride_ptr + n * SBPP, r * SBPP);        \
-    ANY_SIMD(temp + 128, temp, 64, MASK + 1, source_y_fraction);             \
-    memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+#define ANY11I(NAMEANY, ANY_SIMD, TD, TS, SBPP, BPP, MASK)           \
+  void NAMEANY(TD* dst_ptr, const TS* src_ptr, ptrdiff_t src_stride, \
+               int width, int source_y_fraction) {                   \
+    SIMD_ALIGNED(TS temps[64 * 2]);                                  \
+    SIMD_ALIGNED(TD tempd[64]);                                      \
+    memset(temps, 0, sizeof(temps)); /* for msan */                  \
+    int r = width & MASK;                                            \
+    int n = width & ~MASK;                                           \
+    if (n > 0) {                                                     \
+      ANY_SIMD(dst_ptr, src_ptr, src_stride, n, source_y_fraction);  \
+    }                                                                \
+    memcpy(temps, src_ptr + n * SBPP, r * SBPP * sizeof(TS));        \
+    if (source_y_fraction) {                                         \
+      memcpy(temps + 64, src_ptr + src_stride + n * SBPP,            \
+             r * SBPP * sizeof(TS));                                 \
+    }                                                                \
+    ANY_SIMD(tempd, temps, 64, MASK + 1, source_y_fraction);         \
+    memcpy(dst_ptr + n * BPP, tempd, r * BPP * sizeof(TD));          \
   }
 
 #ifdef HAS_INTERPOLATEROW_AVX2
-ANY11T(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, 1, 1, 31)
+ANY11I(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, uint8_t, uint8_t, 1, 1, 31)
 #endif
 #ifdef HAS_INTERPOLATEROW_SSSE3
-ANY11T(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, 1, 1, 15)
+ANY11I(InterpolateRow_Any_SSSE3,
+       InterpolateRow_SSSE3,
+       uint8_t,
+       uint8_t,
+       1,
+       1,
+       15)
 #endif
 #ifdef HAS_INTERPOLATEROW_NEON
-ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15)
+ANY11I(InterpolateRow_Any_NEON, InterpolateRow_NEON, uint8_t, uint8_t, 1, 1, 15)
 #endif
 #ifdef HAS_INTERPOLATEROW_MSA
-ANY11T(InterpolateRow_Any_MSA, InterpolateRow_MSA, 1, 1, 31)
+ANY11I(InterpolateRow_Any_MSA, InterpolateRow_MSA, uint8_t, uint8_t, 1, 1, 31)
 #endif
-#ifdef HAS_INTERPOLATEROW_MMI
-ANY11T(InterpolateRow_Any_MMI, InterpolateRow_MMI, 1, 1, 7)
+#ifdef HAS_INTERPOLATEROW_LSX
+ANY11I(InterpolateRow_Any_LSX, InterpolateRow_LSX, uint8_t, uint8_t, 1, 1, 31)
 #endif
-#undef ANY11T
+
+#ifdef HAS_INTERPOLATEROW_16_NEON
+ANY11I(InterpolateRow_16_Any_NEON,
+       InterpolateRow_16_NEON,
+       uint16_t,
+       uint16_t,
+       1,
+       1,
+       7)
+#endif
+#undef ANY11I
+
+// Any 1 to 1 interpolate with scale param
+#define ANY11IS(NAMEANY, ANY_SIMD, TD, TS, SBPP, BPP, MASK)                \
+  void NAMEANY(TD* dst_ptr, const TS* src_ptr, ptrdiff_t src_stride,       \
+               int scale, int width, int source_y_fraction) {              \
+    SIMD_ALIGNED(TS temps[64 * 2]);                                        \
+    SIMD_ALIGNED(TD tempd[64]);                                            \
+    memset(temps, 0, sizeof(temps)); /* for msan */                        \
+    int r = width & MASK;                                                  \
+    int n = width & ~MASK;                                                 \
+    if (n > 0) {                                                           \
+      ANY_SIMD(dst_ptr, src_ptr, src_stride, scale, n, source_y_fraction); \
+    }                                                                      \
+    memcpy(temps, src_ptr + n * SBPP, r * SBPP * sizeof(TS));              \
+    if (source_y_fraction) {                                               \
+      memcpy(temps + 64, src_ptr + src_stride + n * SBPP,                  \
+             r * SBPP * sizeof(TS));                                       \
+    }                                                                      \
+    ANY_SIMD(tempd, temps, 64, scale, MASK + 1, source_y_fraction);        \
+    memcpy(dst_ptr + n * BPP, tempd, r * BPP * sizeof(TD));                \
+  }
+
+#ifdef HAS_INTERPOLATEROW_16TO8_NEON
+ANY11IS(InterpolateRow_16To8_Any_NEON,
+        InterpolateRow_16To8_NEON,
+        uint8_t,
+        uint16_t,
+        1,
+        1,
+        7)
+#endif
+#ifdef HAS_INTERPOLATEROW_16TO8_AVX2
+ANY11IS(InterpolateRow_16To8_Any_AVX2,
+        InterpolateRow_16To8_AVX2,
+        uint8_t,
+        uint16_t,
+        1,
+        1,
+        31)
+#endif
+
+#undef ANY11IS
 
 // Any 1 to 1 mirror.
 #define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK)                              \
@@ -1098,13 +1742,28 @@ ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
 ANY11M(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, 1, 15)
 #endif
 #ifdef HAS_MIRRORROW_NEON
-ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 15)
+ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 31)
 #endif
 #ifdef HAS_MIRRORROW_MSA
 ANY11M(MirrorRow_Any_MSA, MirrorRow_MSA, 1, 63)
 #endif
-#ifdef HAS_MIRRORROW_MMI
-ANY11M(MirrorRow_Any_MMI, MirrorRow_MMI, 1, 7)
+#ifdef HAS_MIRRORROW_LASX
+ANY11M(MirrorRow_Any_LASX, MirrorRow_LASX, 1, 63)
+#endif
+#ifdef HAS_MIRRORUVROW_AVX2
+ANY11M(MirrorUVRow_Any_AVX2, MirrorUVRow_AVX2, 2, 15)
+#endif
+#ifdef HAS_MIRRORUVROW_SSSE3
+ANY11M(MirrorUVRow_Any_SSSE3, MirrorUVRow_SSSE3, 2, 7)
+#endif
+#ifdef HAS_MIRRORUVROW_NEON
+ANY11M(MirrorUVRow_Any_NEON, MirrorUVRow_NEON, 2, 31)
+#endif
+#ifdef HAS_MIRRORUVROW_MSA
+ANY11M(MirrorUVRow_Any_MSA, MirrorUVRow_MSA, 2, 7)
+#endif
+#ifdef HAS_MIRRORUVROW_LASX
+ANY11M(MirrorUVRow_Any_LASX, MirrorUVRow_LASX, 2, 15)
 #endif
 #ifdef HAS_ARGBMIRRORROW_AVX2
 ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
@@ -1113,13 +1772,19 @@ ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
 ANY11M(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, 4, 3)
 #endif
 #ifdef HAS_ARGBMIRRORROW_NEON
-ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 3)
+ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 7)
 #endif
 #ifdef HAS_ARGBMIRRORROW_MSA
 ANY11M(ARGBMirrorRow_Any_MSA, ARGBMirrorRow_MSA, 4, 15)
 #endif
-#ifdef HAS_ARGBMIRRORROW_MMI
-ANY11M(ARGBMirrorRow_Any_MMI, ARGBMirrorRow_MMI, 4, 1)
+#ifdef HAS_ARGBMIRRORROW_LASX
+ANY11M(ARGBMirrorRow_Any_LASX, ARGBMirrorRow_LASX, 4, 15)
+#endif
+#ifdef HAS_RGB24MIRRORROW_SSSE3
+ANY11M(RGB24MirrorRow_Any_SSSE3, RGB24MirrorRow_SSSE3, 3, 15)
+#endif
+#ifdef HAS_RGB24MIRRORROW_NEON
+ANY11M(RGB24MirrorRow_Any_NEON, RGB24MirrorRow_NEON, 3, 15)
 #endif
 #undef ANY11M
 
@@ -1127,6 +1792,7 @@ ANY11M(ARGBMirrorRow_Any_MMI, ARGBMirrorRow_MMI, 4, 1)
 #define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK)        \
   void NAMEANY(uint8_t* dst_ptr, T v32, int width) { \
     SIMD_ALIGNED(uint8_t temp[64]);                  \
+    memset(temp, 0, 64); /* for msan */              \
     int r = width & MASK;                            \
     int n = width & ~MASK;                           \
     if (n > 0) {                                     \
@@ -1142,12 +1808,18 @@ ANY1(SetRow_Any_X86, SetRow_X86, uint8_t, 1, 3)
 #ifdef HAS_SETROW_NEON
 ANY1(SetRow_Any_NEON, SetRow_NEON, uint8_t, 1, 15)
 #endif
+#ifdef HAS_SETROW_LSX
+ANY1(SetRow_Any_LSX, SetRow_LSX, uint8_t, 1, 15)
+#endif
 #ifdef HAS_ARGBSETROW_NEON
 ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32_t, 4, 3)
 #endif
 #ifdef HAS_ARGBSETROW_MSA
 ANY1(ARGBSetRow_Any_MSA, ARGBSetRow_MSA, uint32_t, 4, 3)
 #endif
+#ifdef HAS_ARGBSETROW_LSX
+ANY1(ARGBSetRow_Any_LSX, ARGBSetRow_LSX, uint32_t, 4, 3)
+#endif
 #undef ANY1
 
 // Any 1 to 2.  Outputs UV planes.
@@ -1179,8 +1851,8 @@ ANY12(SplitUVRow_Any_NEON, SplitUVRow_NEON, 0, 2, 0, 15)
 #ifdef HAS_SPLITUVROW_MSA
 ANY12(SplitUVRow_Any_MSA, SplitUVRow_MSA, 0, 2, 0, 31)
 #endif
-#ifdef HAS_SPLITUVROW_MMI
-ANY12(SplitUVRow_Any_MMI, SplitUVRow_MMI, 0, 2, 0, 7)
+#ifdef HAS_SPLITUVROW_LSX
+ANY12(SplitUVRow_Any_LSX, SplitUVRow_LSX, 0, 2, 0, 31)
 #endif
 #ifdef HAS_ARGBTOUV444ROW_SSSE3
 ANY12(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, 0, 4, 0, 15)
@@ -1203,13 +1875,39 @@ ANY12(ARGBToUV444Row_Any_MSA, ARGBToUV444Row_MSA, 0, 4, 0, 15)
 ANY12(YUY2ToUV422Row_Any_MSA, YUY2ToUV422Row_MSA, 1, 4, 1, 31)
 ANY12(UYVYToUV422Row_Any_MSA, UYVYToUV422Row_MSA, 1, 4, 1, 31)
 #endif
-#ifdef HAS_YUY2TOUV422ROW_MMI
-ANY12(ARGBToUV444Row_Any_MMI, ARGBToUV444Row_MMI, 0, 4, 0, 7)
-ANY12(UYVYToUV422Row_Any_MMI, UYVYToUV422Row_MMI, 1, 4, 1, 15)
-ANY12(YUY2ToUV422Row_Any_MMI, YUY2ToUV422Row_MMI, 1, 4, 1, 15)
+#ifdef HAS_YUY2TOUV422ROW_LASX
+ANY12(ARGBToUV444Row_Any_LASX, ARGBToUV444Row_LASX, 0, 4, 0, 31)
+ANY12(YUY2ToUV422Row_Any_LASX, YUY2ToUV422Row_LASX, 1, 4, 1, 31)
+ANY12(UYVYToUV422Row_Any_LASX, UYVYToUV422Row_LASX, 1, 4, 1, 31)
 #endif
 #undef ANY12
 
+// Any 2 16 bit planes with parameter to 1
+#define ANY12PT(NAMEANY, ANY_SIMD, T, BPP, MASK)                            \
+  void NAMEANY(const T* src_uv, T* dst_u, T* dst_v, int depth, int width) { \
+    SIMD_ALIGNED(T temp[16 * 4]);                                           \
+    memset(temp, 0, 16 * 4 * BPP); /* for msan */                           \
+    int r = width & MASK;                                                   \
+    int n = width & ~MASK;                                                  \
+    if (n > 0) {                                                            \
+      ANY_SIMD(src_uv, dst_u, dst_v, depth, n);                             \
+    }                                                                       \
+    memcpy(temp, src_uv + n * 2, r * BPP * 2);                              \
+    ANY_SIMD(temp, temp + 32, temp + 48, depth, MASK + 1);                  \
+    memcpy(dst_u + n, temp + 32, r * BPP);                                  \
+    memcpy(dst_v + n, temp + 48, r * BPP);                                  \
+  }
+
+#ifdef HAS_SPLITUVROW_16_AVX2
+ANY12PT(SplitUVRow_16_Any_AVX2, SplitUVRow_16_AVX2, uint16_t, 2, 15)
+#endif
+
+#ifdef HAS_SPLITUVROW_16_NEON
+ANY12PT(SplitUVRow_16_Any_NEON, SplitUVRow_16_NEON, uint16_t, 2, 7)
+#endif
+
+#undef ANY21CT
+
 // Any 1 to 3.  Outputs RGB planes.
 #define ANY13(NAMEANY, ANY_SIMD, BPP, MASK)                                \
   void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_r, uint8_t* dst_g,     \
@@ -1234,24 +1932,66 @@ ANY13(SplitRGBRow_Any_SSSE3, SplitRGBRow_SSSE3, 3, 15)
 #ifdef HAS_SPLITRGBROW_NEON
 ANY13(SplitRGBRow_Any_NEON, SplitRGBRow_NEON, 3, 15)
 #endif
-#ifdef HAS_SPLITRGBROW_MMI
-ANY13(SplitRGBRow_Any_MMI, SplitRGBRow_MMI, 3, 3)
+#ifdef HAS_SPLITXRGBROW_SSE2
+ANY13(SplitXRGBRow_Any_SSE2, SplitXRGBRow_SSE2, 4, 7)
+#endif
+#ifdef HAS_SPLITXRGBROW_SSSE3
+ANY13(SplitXRGBRow_Any_SSSE3, SplitXRGBRow_SSSE3, 4, 7)
+#endif
+#ifdef HAS_SPLITXRGBROW_AVX2
+ANY13(SplitXRGBRow_Any_AVX2, SplitXRGBRow_AVX2, 4, 15)
+#endif
+#ifdef HAS_SPLITXRGBROW_NEON
+ANY13(SplitXRGBRow_Any_NEON, SplitXRGBRow_NEON, 4, 15)
+#endif
+
+// Any 1 to 4.  Outputs ARGB planes.
+#define ANY14(NAMEANY, ANY_SIMD, BPP, MASK)                                    \
+  void NAMEANY(const uint8_t* src_ptr, uint8_t* dst_r, uint8_t* dst_g,         \
+               uint8_t* dst_b, uint8_t* dst_a, int width) {                    \
+    SIMD_ALIGNED(uint8_t temp[16 * 8]);                                        \
+    memset(temp, 0, 16 * 4); /* for msan */                                    \
+    int r = width & MASK;                                                      \
+    int n = width & ~MASK;                                                     \
+    if (n > 0) {                                                               \
+      ANY_SIMD(src_ptr, dst_r, dst_g, dst_b, dst_a, n);                        \
+    }                                                                          \
+    memcpy(temp, src_ptr + n * BPP, r * BPP);                                  \
+    ANY_SIMD(temp, temp + 16 * 4, temp + 16 * 5, temp + 16 * 6, temp + 16 * 7, \
+             MASK + 1);                                                        \
+    memcpy(dst_r + n, temp + 16 * 4, r);                                       \
+    memcpy(dst_g + n, temp + 16 * 5, r);                                       \
+    memcpy(dst_b + n, temp + 16 * 6, r);                                       \
+    memcpy(dst_a + n, temp + 16 * 7, r);                                       \
+  }
+
+#ifdef HAS_SPLITARGBROW_SSE2
+ANY14(SplitARGBRow_Any_SSE2, SplitARGBRow_SSE2, 4, 7)
+#endif
+#ifdef HAS_SPLITARGBROW_SSSE3
+ANY14(SplitARGBRow_Any_SSSE3, SplitARGBRow_SSSE3, 4, 7)
+#endif
+#ifdef HAS_SPLITARGBROW_AVX2
+ANY14(SplitARGBRow_Any_AVX2, SplitARGBRow_AVX2, 4, 15)
+#endif
+#ifdef HAS_SPLITARGBROW_NEON
+ANY14(SplitARGBRow_Any_NEON, SplitARGBRow_NEON, 4, 15)
 #endif
 
 // Any 1 to 2 with source stride (2 rows of source).  Outputs UV planes.
 // 128 byte row allows for 32 avx ARGB pixels.
 #define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                        \
-  void NAMEANY(const uint8_t* src_ptr, int src_stride_ptr, uint8_t* dst_u,   \
+  void NAMEANY(const uint8_t* src_ptr, int src_stride, uint8_t* dst_u,       \
                uint8_t* dst_v, int width) {                                  \
     SIMD_ALIGNED(uint8_t temp[128 * 4]);                                     \
     memset(temp, 0, 128 * 2); /* for msan */                                 \
     int r = width & MASK;                                                    \
     int n = width & ~MASK;                                                   \
     if (n > 0) {                                                             \
-      ANY_SIMD(src_ptr, src_stride_ptr, dst_u, dst_v, n);                    \
+      ANY_SIMD(src_ptr, src_stride, dst_u, dst_v, n);                        \
     }                                                                        \
     memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);      \
-    memcpy(temp + 128, src_ptr + src_stride_ptr + (n >> UVSHIFT) * BPP,      \
+    memcpy(temp + 128, src_ptr + src_stride + (n >> UVSHIFT) * BPP,          \
            SS(r, UVSHIFT) * BPP);                                            \
     if ((width & 1) && UVSHIFT == 0) { /* repeat last pixel for subsample */ \
       memcpy(temp + SS(r, UVSHIFT) * BPP, temp + SS(r, UVSHIFT) * BPP - BPP, \
@@ -1267,6 +2007,9 @@ ANY13(SplitRGBRow_Any_MMI, SplitRGBRow_MMI, 3, 3)
 #ifdef HAS_ARGBTOUVROW_AVX2
 ANY12S(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, 0, 4, 31)
 #endif
+#ifdef HAS_ABGRTOUVROW_AVX2
+ANY12S(ABGRToUVRow_Any_AVX2, ABGRToUVRow_AVX2, 0, 4, 31)
+#endif
 #ifdef HAS_ARGBTOUVJROW_AVX2
 ANY12S(ARGBToUVJRow_Any_AVX2, ARGBToUVJRow_AVX2, 0, 4, 31)
 #endif
@@ -1291,8 +2034,8 @@ ANY12S(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, 0, 4, 15)
 #ifdef HAS_ARGBTOUVROW_MSA
 ANY12S(ARGBToUVRow_Any_MSA, ARGBToUVRow_MSA, 0, 4, 31)
 #endif
-#ifdef HAS_ARGBTOUVROW_MMI
-ANY12S(ARGBToUVRow_Any_MMI, ARGBToUVRow_MMI, 0, 4, 15)
+#ifdef HAS_ARGBTOUVROW_LASX
+ANY12S(ARGBToUVRow_Any_LASX, ARGBToUVRow_LASX, 0, 4, 31)
 #endif
 #ifdef HAS_ARGBTOUVJROW_NEON
 ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15)
@@ -1300,53 +2043,68 @@ ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15)
 #ifdef HAS_ARGBTOUVJROW_MSA
 ANY12S(ARGBToUVJRow_Any_MSA, ARGBToUVJRow_MSA, 0, 4, 31)
 #endif
-#ifdef HAS_ARGBTOUVJROW_MMI
-ANY12S(ARGBToUVJRow_Any_MMI, ARGBToUVJRow_MMI, 0, 4, 15)
+#ifdef HAS_ARGBTOUVJROW_LSX
+ANY12S(ARGBToUVJRow_Any_LSX, ARGBToUVJRow_LSX, 0, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_LASX
+ANY12S(ARGBToUVJRow_Any_LASX, ARGBToUVJRow_LASX, 0, 4, 31)
 #endif
 #ifdef HAS_BGRATOUVROW_NEON
 ANY12S(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, 0, 4, 15)
 #endif
 #ifdef HAS_BGRATOUVROW_MSA
-ANY12S(BGRAToUVRow_Any_MSA, BGRAToUVRow_MSA, 0, 4, 31)
+ANY12S(BGRAToUVRow_Any_MSA, BGRAToUVRow_MSA, 0, 4, 15)
 #endif
-#ifdef HAS_BGRATOUVROW_MMI
-ANY12S(BGRAToUVRow_Any_MMI, BGRAToUVRow_MMI, 0, 4, 15)
+#ifdef HAS_BGRATOUVROW_LSX
+ANY12S(BGRAToUVRow_Any_LSX, BGRAToUVRow_LSX, 0, 4, 15)
 #endif
 #ifdef HAS_ABGRTOUVROW_NEON
 ANY12S(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, 0, 4, 15)
 #endif
 #ifdef HAS_ABGRTOUVROW_MSA
-ANY12S(ABGRToUVRow_Any_MSA, ABGRToUVRow_MSA, 0, 4, 31)
+ANY12S(ABGRToUVRow_Any_MSA, ABGRToUVRow_MSA, 0, 4, 15)
 #endif
-#ifdef HAS_ABGRTOUVROW_MMI
-ANY12S(ABGRToUVRow_Any_MMI, ABGRToUVRow_MMI, 0, 4, 15)
+#ifdef HAS_ABGRTOUVROW_LSX
+ANY12S(ABGRToUVRow_Any_LSX, ABGRToUVRow_LSX, 0, 4, 15)
 #endif
 #ifdef HAS_RGBATOUVROW_NEON
 ANY12S(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, 0, 4, 15)
 #endif
 #ifdef HAS_RGBATOUVROW_MSA
-ANY12S(RGBAToUVRow_Any_MSA, RGBAToUVRow_MSA, 0, 4, 31)
+ANY12S(RGBAToUVRow_Any_MSA, RGBAToUVRow_MSA, 0, 4, 15)
 #endif
-#ifdef HAS_RGBATOUVROW_MMI
-ANY12S(RGBAToUVRow_Any_MMI, RGBAToUVRow_MMI, 0, 4, 15)
+#ifdef HAS_RGBATOUVROW_LSX
+ANY12S(RGBAToUVRow_Any_LSX, RGBAToUVRow_LSX, 0, 4, 15)
 #endif
 #ifdef HAS_RGB24TOUVROW_NEON
 ANY12S(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, 0, 3, 15)
 #endif
+#ifdef HAS_RGB24TOUVJROW_NEON
+ANY12S(RGB24ToUVJRow_Any_NEON, RGB24ToUVJRow_NEON, 0, 3, 15)
+#endif
 #ifdef HAS_RGB24TOUVROW_MSA
 ANY12S(RGB24ToUVRow_Any_MSA, RGB24ToUVRow_MSA, 0, 3, 15)
 #endif
-#ifdef HAS_RGB24TOUVROW_MMI
-ANY12S(RGB24ToUVRow_Any_MMI, RGB24ToUVRow_MMI, 0, 3, 15)
+#ifdef HAS_RGB24TOUVROW_LSX
+ANY12S(RGB24ToUVRow_Any_LSX, RGB24ToUVRow_LSX, 0, 3, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_LASX
+ANY12S(RGB24ToUVRow_Any_LASX, RGB24ToUVRow_LASX, 0, 3, 31)
 #endif
 #ifdef HAS_RAWTOUVROW_NEON
 ANY12S(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, 0, 3, 15)
 #endif
+#ifdef HAS_RAWTOUVJROW_NEON
+ANY12S(RAWToUVJRow_Any_NEON, RAWToUVJRow_NEON, 0, 3, 15)
+#endif
 #ifdef HAS_RAWTOUVROW_MSA
 ANY12S(RAWToUVRow_Any_MSA, RAWToUVRow_MSA, 0, 3, 15)
 #endif
-#ifdef HAS_RAWTOUVROW_MMI
-ANY12S(RAWToUVRow_Any_MMI, RAWToUVRow_MMI, 0, 3, 15)
+#ifdef HAS_RAWTOUVROW_LSX
+ANY12S(RAWToUVRow_Any_LSX, RAWToUVRow_LSX, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_LASX
+ANY12S(RAWToUVRow_Any_LASX, RAWToUVRow_LASX, 0, 3, 31)
 #endif
 #ifdef HAS_RGB565TOUVROW_NEON
 ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15)
@@ -1354,8 +2112,11 @@ ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15)
 #ifdef HAS_RGB565TOUVROW_MSA
 ANY12S(RGB565ToUVRow_Any_MSA, RGB565ToUVRow_MSA, 0, 2, 15)
 #endif
-#ifdef HAS_RGB565TOUVROW_MMI
-ANY12S(RGB565ToUVRow_Any_MMI, RGB565ToUVRow_MMI, 0, 2, 15)
+#ifdef HAS_RGB565TOUVROW_LSX
+ANY12S(RGB565ToUVRow_Any_LSX, RGB565ToUVRow_LSX, 0, 2, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_LASX
+ANY12S(RGB565ToUVRow_Any_LASX, RGB565ToUVRow_LASX, 0, 2, 31)
 #endif
 #ifdef HAS_ARGB1555TOUVROW_NEON
 ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15)
@@ -1363,15 +2124,15 @@ ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15)
 #ifdef HAS_ARGB1555TOUVROW_MSA
 ANY12S(ARGB1555ToUVRow_Any_MSA, ARGB1555ToUVRow_MSA, 0, 2, 15)
 #endif
-#ifdef HAS_ARGB1555TOUVROW_MMI
-ANY12S(ARGB1555ToUVRow_Any_MMI, ARGB1555ToUVRow_MMI, 0, 2, 15)
+#ifdef HAS_ARGB1555TOUVROW_LSX
+ANY12S(ARGB1555ToUVRow_Any_LSX, ARGB1555ToUVRow_LSX, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_LASX
+ANY12S(ARGB1555ToUVRow_Any_LASX, ARGB1555ToUVRow_LASX, 0, 2, 31)
 #endif
 #ifdef HAS_ARGB4444TOUVROW_NEON
 ANY12S(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, 0, 2, 15)
 #endif
-#ifdef HAS_ARGB4444TOUVROW_MMI
-ANY12S(ARGB4444ToUVRow_Any_MMI, ARGB4444ToUVRow_MMI, 0, 2, 15)
-#endif
 #ifdef HAS_YUY2TOUVROW_NEON
 ANY12S(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, 1, 4, 15)
 #endif
@@ -1381,31 +2142,31 @@ ANY12S(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, 1, 4, 15)
 #ifdef HAS_YUY2TOUVROW_MSA
 ANY12S(YUY2ToUVRow_Any_MSA, YUY2ToUVRow_MSA, 1, 4, 31)
 #endif
-#ifdef HAS_YUY2TOUVROW_MMI
-ANY12S(YUY2ToUVRow_Any_MMI, YUY2ToUVRow_MMI, 1, 4, 15)
+#ifdef HAS_YUY2TOUVROW_LASX
+ANY12S(YUY2ToUVRow_Any_LASX, YUY2ToUVRow_LASX, 1, 4, 31)
 #endif
 #ifdef HAS_UYVYTOUVROW_MSA
 ANY12S(UYVYToUVRow_Any_MSA, UYVYToUVRow_MSA, 1, 4, 31)
 #endif
-#ifdef HAS_UYVYTOUVROW_MMI
-ANY12S(UYVYToUVRow_Any_MMI, UYVYToUVRow_MMI, 1, 4, 15)
+#ifdef HAS_UYVYTOUVROW_LASX
+ANY12S(UYVYToUVRow_Any_LASX, UYVYToUVRow_LASX, 1, 4, 31)
 #endif
 #undef ANY12S
 
 // Any 1 to 1 with source stride (2 rows of source).  Outputs UV plane.
 // 128 byte row allows for 32 avx ARGB pixels.
 #define ANY11S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                        \
-  void NAMEANY(const uint8_t* src_ptr, int src_stride_ptr, uint8_t* dst_vu,  \
+  void NAMEANY(const uint8_t* src_ptr, int src_stride, uint8_t* dst_vu,      \
                int width) {                                                  \
     SIMD_ALIGNED(uint8_t temp[128 * 3]);                                     \
     memset(temp, 0, 128 * 2); /* for msan */                                 \
     int r = width & MASK;                                                    \
     int n = width & ~MASK;                                                   \
     if (n > 0) {                                                             \
-      ANY_SIMD(src_ptr, src_stride_ptr, dst_vu, n);                          \
+      ANY_SIMD(src_ptr, src_stride, dst_vu, n);                              \
     }                                                                        \
     memcpy(temp, src_ptr + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);      \
-    memcpy(temp + 128, src_ptr + src_stride_ptr + (n >> UVSHIFT) * BPP,      \
+    memcpy(temp + 128, src_ptr + src_stride + (n >> UVSHIFT) * BPP,          \
            SS(r, UVSHIFT) * BPP);                                            \
     if ((width & 1) && UVSHIFT == 0) { /* repeat last pixel for subsample */ \
       memcpy(temp + SS(r, UVSHIFT) * BPP, temp + SS(r, UVSHIFT) * BPP - BPP, \
@@ -1423,6 +2184,51 @@ ANY11S(AYUVToVURow_Any_NEON, AYUVToVURow_NEON, 0, 4, 15)
 #endif
 #undef ANY11S
 
+#define ANYDETILE(NAMEANY, ANY_SIMD, MASK)                                  \
+  void NAMEANY(const uint8_t* src, ptrdiff_t src_tile_stride, uint8_t* dst, \
+               int width) {                                                 \
+    SIMD_ALIGNED(uint8_t temp[16 * 2]);                                     \
+    memset(temp, 0, 16); /* for msan */                                     \
+    int r = width & MASK;                                                   \
+    int n = width & ~MASK;                                                  \
+    if (n > 0) {                                                            \
+      ANY_SIMD(src, src_tile_stride, dst, n);                               \
+    }                                                                       \
+    memcpy(temp, src + (n / 16) * src_tile_stride, r);                      \
+    ANY_SIMD(temp, src_tile_stride, temp + 16, MASK + 1);                   \
+    memcpy(dst + n, temp + 16, r);                                          \
+  }
+
+#ifdef HAS_DETILEROW_NEON
+ANYDETILE(DetileRow_Any_NEON, DetileRow_NEON, 15)
+#endif
+#ifdef HAS_DETILEROW_SSE2
+ANYDETILE(DetileRow_Any_SSE2, DetileRow_SSE2, 15)
+#endif
+
+#define ANYDETILESPLITUV(NAMEANY, ANY_SIMD, MASK)                \
+  void NAMEANY(const uint8_t* src_uv, ptrdiff_t src_tile_stride, \
+               uint8_t* dst_u, uint8_t* dst_v, int width) {      \
+    SIMD_ALIGNED(uint8_t temp[16 * 2]);                          \
+    memset(temp, 0, 16 * 2); /* for msan */                      \
+    int r = width & MASK;                                        \
+    int n = width & ~MASK;                                       \
+    if (n > 0) {                                                 \
+      ANY_SIMD(src_uv, src_tile_stride, dst_u, dst_v, n);        \
+    }                                                            \
+    memcpy(temp, src_uv + (n / 16) * src_tile_stride, r);        \
+    ANY_SIMD(temp, src_tile_stride, temp + 16, temp + 24, r);    \
+    memcpy(dst_u + n / 2, temp + 16, (r + 1) / 2);               \
+    memcpy(dst_v + n / 2, temp + 24, (r + 1) / 2);               \
+  }
+
+#ifdef HAS_DETILESPLITUVROW_NEON
+ANYDETILESPLITUV(DetileSplitUVRow_Any_NEON, DetileSplitUVRow_NEON, 15)
+#endif
+#ifdef HAS_DETILESPLITUVROW_SSSE3
+ANYDETILESPLITUV(DetileSplitUVRow_Any_SSSE3, DetileSplitUVRow_SSSE3, 15)
+#endif
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/row_common.cc b/files/source/row_common.cc
index 8951d003..83442496 100644
--- a/files/source/row_common.cc
+++ b/files/source/row_common.cc
@@ -10,34 +10,67 @@
 
 #include "libyuv/row.h"
 
-#include <stdio.h>
+#include <assert.h>
 #include <string.h>  // For memcpy and memset.
 
 #include "libyuv/basic_types.h"
+#include "libyuv/convert_argb.h"  // For kYuvI601Constants
 
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
+// This macro controls YUV to RGB using unsigned math to extend range of
+// YUV to RGB coefficients to 0 to 4 instead of 0 to 2 for more accuracy on B:
+// LIBYUV_UNLIMITED_DATA
+
+// Macros to enable unlimited data for each colorspace
+// LIBYUV_UNLIMITED_BT601
+// LIBYUV_UNLIMITED_BT709
+// LIBYUV_UNLIMITED_BT2020
+
+// The following macro from row_win makes the C code match the row_win code,
+// which is 7 bit fixed point for ARGBToI420:
+#if !defined(LIBYUV_BIT_EXACT) && !defined(LIBYUV_DISABLE_X86) && \
+    defined(_MSC_VER) && !defined(__clang__) &&                   \
+    (defined(_M_IX86) || defined(_M_X64))
+#define LIBYUV_RGB7 1
+#endif
+
+#if !defined(LIBYUV_BIT_EXACT) && (defined(__x86_64__) || defined(_M_X64) || \
+                                   defined(__i386__) || defined(_M_IX86))
+#define LIBYUV_ARGBTOUV_PAVGB 1
+#define LIBYUV_RGBTOU_TRUNCATE 1
+#define LIBYUV_ATTENUATE_DUP 1
+#endif
+#if defined(LIBYUV_BIT_EXACT)
+#define LIBYUV_UNATTENUATE_DUP 1
+#endif
+
 // llvm x86 is poor at ternary operator, so use branchless min/max.
 
 #define USE_BRANCHLESS 1
 #if USE_BRANCHLESS
 static __inline int32_t clamp0(int32_t v) {
-  return ((-(v) >> 31) & (v));
+  return -(v >= 0) & v;
 }
-
+// TODO(fbarchard): make clamp255 preserve negative values.
 static __inline int32_t clamp255(int32_t v) {
-  return (((255 - (v)) >> 31) | (v)) & 255;
+  return (-(v >= 255) | v) & 255;
 }
 
 static __inline int32_t clamp1023(int32_t v) {
-  return (((1023 - (v)) >> 31) | (v)) & 1023;
+  return (-(v >= 1023) | v) & 1023;
+}
+
+// clamp to max
+static __inline int32_t ClampMax(int32_t v, int32_t max) {
+  return (-(v >= max) | v) & max;
 }
 
 static __inline uint32_t Abs(int32_t v) {
-  int m = v >> 31;
+  int m = -(v < 0);
   return (v + m) ^ m;
 }
 #else   // USE_BRANCHLESS
@@ -53,6 +86,10 @@ static __inline int32_t clamp1023(int32_t v) {
   return (v > 1023) ? 1023 : v;
 }
 
+static __inline int32_t ClampMax(int32_t v, int32_t max) {
+  return (v > max) ? max : v;
+}
+
 static __inline uint32_t Abs(int32_t v) {
   return (v < 0) ? -v : v;
 }
@@ -111,6 +148,21 @@ void RAWToARGBRow_C(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   }
 }
 
+void RAWToRGBARow_C(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8_t r = src_raw[0];
+    uint8_t g = src_raw[1];
+    uint8_t b = src_raw[2];
+    dst_rgba[0] = 255u;
+    dst_rgba[1] = b;
+    dst_rgba[2] = g;
+    dst_rgba[3] = r;
+    dst_rgba += 4;
+    src_raw += 3;
+  }
+}
+
 void RAWToRGB24Row_C(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
   int x;
   for (x = 0; x < width; ++x) {
@@ -181,7 +233,8 @@ void ARGB4444ToARGBRow_C(const uint8_t* src_argb4444,
 void AR30ToARGBRow_C(const uint8_t* src_ar30, uint8_t* dst_argb, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint32_t ar30 = *(const uint32_t*)src_ar30;
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
     uint32_t b = (ar30 >> 2) & 0xff;
     uint32_t g = (ar30 >> 12) & 0xff;
     uint32_t r = (ar30 >> 22) & 0xff;
@@ -195,7 +248,8 @@ void AR30ToARGBRow_C(const uint8_t* src_ar30, uint8_t* dst_argb, int width) {
 void AR30ToABGRRow_C(const uint8_t* src_ar30, uint8_t* dst_abgr, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint32_t ar30 = *(const uint32_t*)src_ar30;
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
     uint32_t b = (ar30 >> 2) & 0xff;
     uint32_t g = (ar30 >> 12) & 0xff;
     uint32_t r = (ar30 >> 22) & 0xff;
@@ -209,7 +263,8 @@ void AR30ToABGRRow_C(const uint8_t* src_ar30, uint8_t* dst_abgr, int width) {
 void AR30ToAB30Row_C(const uint8_t* src_ar30, uint8_t* dst_ab30, int width) {
   int x;
   for (x = 0; x < width; ++x) {
-    uint32_t ar30 = *(const uint32_t*)src_ar30;
+    uint32_t ar30;
+    memcpy(&ar30, src_ar30, sizeof ar30);
     uint32_t b = ar30 & 0x3ff;
     uint32_t ga = ar30 & 0xc00ffc00;
     uint32_t r = (ar30 >> 20) & 0x3ff;
@@ -291,8 +346,8 @@ void ARGBToRGB565DitherRow_C(const uint8_t* src_argb,
     uint8_t b1 = clamp255(src_argb[4] + dither1) >> 3;
     uint8_t g1 = clamp255(src_argb[5] + dither1) >> 2;
     uint8_t r1 = clamp255(src_argb[6] + dither1) >> 3;
-    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) | (b1 << 16) | (g1 << 21) |
-                           (r1 << 27));
+    *(uint16_t*)(dst_rgb + 0) = b0 | (g0 << 5) | (r0 << 11);
+    *(uint16_t*)(dst_rgb + 2) = b1 | (g1 << 5) | (r1 << 11);
     dst_rgb += 4;
     src_argb += 8;
   }
@@ -316,8 +371,8 @@ void ARGBToARGB1555Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
     uint8_t g1 = src_argb[5] >> 3;
     uint8_t r1 = src_argb[6] >> 3;
     uint8_t a1 = src_argb[7] >> 7;
-    *(uint32_t*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
-                            (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    *(uint16_t*)(dst_rgb + 0) = b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+    *(uint16_t*)(dst_rgb + 2) = b1 | (g1 << 5) | (r1 << 10) | (a1 << 15);
     dst_rgb += 4;
     src_argb += 8;
   }
@@ -341,8 +396,8 @@ void ARGBToARGB4444Row_C(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
     uint8_t g1 = src_argb[5] >> 4;
     uint8_t r1 = src_argb[6] >> 4;
     uint8_t a1 = src_argb[7] >> 4;
-    *(uint32_t*)(dst_rgb) = b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
-                            (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    *(uint16_t*)(dst_rgb + 0) = b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+    *(uint16_t*)(dst_rgb + 2) = b1 | (g1 << 4) | (r1 << 8) | (a1 << 12);
     dst_rgb += 4;
     src_argb += 8;
   }
@@ -381,56 +436,208 @@ void ARGBToAR30Row_C(const uint8_t* src_argb, uint8_t* dst_ar30, int width) {
   }
 }
 
+void ARGBToAR64Row_C(const uint8_t* src_argb, uint16_t* dst_ar64, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_ar64[0] = src_argb[0] * 0x0101;
+    dst_ar64[1] = src_argb[1] * 0x0101;
+    dst_ar64[2] = src_argb[2] * 0x0101;
+    dst_ar64[3] = src_argb[3] * 0x0101;
+    dst_ar64 += 4;
+    src_argb += 4;
+  }
+}
+
+void ARGBToAB64Row_C(const uint8_t* src_argb, uint16_t* dst_ab64, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_ab64[0] = src_argb[2] * 0x0101;
+    dst_ab64[1] = src_argb[1] * 0x0101;
+    dst_ab64[2] = src_argb[0] * 0x0101;
+    dst_ab64[3] = src_argb[3] * 0x0101;
+    dst_ab64 += 4;
+    src_argb += 4;
+  }
+}
+
+void AR64ToARGBRow_C(const uint16_t* src_ar64, uint8_t* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_ar64[0] >> 8;
+    dst_argb[1] = src_ar64[1] >> 8;
+    dst_argb[2] = src_ar64[2] >> 8;
+    dst_argb[3] = src_ar64[3] >> 8;
+    dst_argb += 4;
+    src_ar64 += 4;
+  }
+}
+
+void AB64ToARGBRow_C(const uint16_t* src_ab64, uint8_t* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_ab64[2] >> 8;
+    dst_argb[1] = src_ab64[1] >> 8;
+    dst_argb[2] = src_ab64[0] >> 8;
+    dst_argb[3] = src_ab64[3] >> 8;
+    dst_argb += 4;
+    src_ab64 += 4;
+  }
+}
+
+// TODO(fbarchard): Make shuffle compatible with SIMD versions
+void AR64ShuffleRow_C(const uint8_t* src_ar64,
+                      uint8_t* dst_ar64,
+                      const uint8_t* shuffler,
+                      int width) {
+  const uint16_t* src_ar64_16 = (const uint16_t*)src_ar64;
+  uint16_t* dst_ar64_16 = (uint16_t*)dst_ar64;
+  int index0 = shuffler[0] / 2;
+  int index1 = shuffler[2] / 2;
+  int index2 = shuffler[4] / 2;
+  int index3 = shuffler[6] / 2;
+  // Shuffle a row of AR64.
+  int x;
+  for (x = 0; x < width / 2; ++x) {
+    // To support in-place conversion.
+    uint16_t b = src_ar64_16[index0];
+    uint16_t g = src_ar64_16[index1];
+    uint16_t r = src_ar64_16[index2];
+    uint16_t a = src_ar64_16[index3];
+    dst_ar64_16[0] = b;
+    dst_ar64_16[1] = g;
+    dst_ar64_16[2] = r;
+    dst_ar64_16[3] = a;
+    src_ar64_16 += 4;
+    dst_ar64_16 += 4;
+  }
+}
+
+#ifdef LIBYUV_RGB7
+// Old 7 bit math for compatibility on unsupported platforms.
+static __inline int RGBToY(uint8_t r, uint8_t g, uint8_t b) {
+  return ((33 * r + 65 * g + 13 * b) >> 7) + 16;
+}
+#else
+// 8 bit
+// Intel SSE/AVX uses the following equivalent formula
+// 0x7e80 = (66 + 129 + 25) * -128 + 0x1000 (for +16) and 0x0080 for round.
+//  return (66 * ((int)r - 128) + 129 * ((int)g - 128) + 25 * ((int)b - 128) +
+//  0x7e80) >> 8;
+
 static __inline int RGBToY(uint8_t r, uint8_t g, uint8_t b) {
   return (66 * r + 129 * g + 25 * b + 0x1080) >> 8;
 }
+#endif
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
 
+// LIBYUV_RGBTOU_TRUNCATE mimics x86 code that does not round.
+#ifdef LIBYUV_RGBTOU_TRUNCATE
+static __inline int RGBToU(uint8_t r, uint8_t g, uint8_t b) {
+  return (112 * b - 74 * g - 38 * r + 0x8000) >> 8;
+}
+static __inline int RGBToV(uint8_t r, uint8_t g, uint8_t b) {
+  return (112 * r - 94 * g - 18 * b + 0x8000) >> 8;
+}
+#else
+// TODO(fbarchard): Add rounding to x86 SIMD and use this
 static __inline int RGBToU(uint8_t r, uint8_t g, uint8_t b) {
   return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
 }
 static __inline int RGBToV(uint8_t r, uint8_t g, uint8_t b) {
   return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
 }
+#endif
+
+// LIBYUV_ARGBTOUV_PAVGB mimics x86 code that subsamples with 2 pavgb.
+#if !defined(LIBYUV_ARGBTOUV_PAVGB)
+static __inline int RGB2xToU(uint16_t r, uint16_t g, uint16_t b) {
+  return ((112 / 2) * b - (74 / 2) * g - (38 / 2) * r + 0x8080) >> 8;
+}
+static __inline int RGB2xToV(uint16_t r, uint16_t g, uint16_t b) {
+  return ((112 / 2) * r - (94 / 2) * g - (18 / 2) * b + 0x8080) >> 8;
+}
+#endif
 
 // ARGBToY_C and ARGBToUV_C
-#define MAKEROWY(NAME, R, G, B, BPP)                                         \
-  void NAME##ToYRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width) { \
-    int x;                                                                   \
-    for (x = 0; x < width; ++x) {                                            \
-      dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);           \
-      src_argb0 += BPP;                                                      \
-      dst_y += 1;                                                            \
-    }                                                                        \
-  }                                                                          \
-  void NAME##ToUVRow_C(const uint8_t* src_rgb0, int src_stride_rgb,          \
-                       uint8_t* dst_u, uint8_t* dst_v, int width) {          \
-    const uint8_t* src_rgb1 = src_rgb0 + src_stride_rgb;                     \
-    int x;                                                                   \
-    for (x = 0; x < width - 1; x += 2) {                                     \
-      uint8_t ab = (src_rgb0[B] + src_rgb0[B + BPP] + src_rgb1[B] +          \
-                    src_rgb1[B + BPP]) >>                                    \
-                   2;                                                        \
-      uint8_t ag = (src_rgb0[G] + src_rgb0[G + BPP] + src_rgb1[G] +          \
-                    src_rgb1[G + BPP]) >>                                    \
-                   2;                                                        \
-      uint8_t ar = (src_rgb0[R] + src_rgb0[R + BPP] + src_rgb1[R] +          \
-                    src_rgb1[R + BPP]) >>                                    \
-                   2;                                                        \
-      dst_u[0] = RGBToU(ar, ag, ab);                                         \
-      dst_v[0] = RGBToV(ar, ag, ab);                                         \
-      src_rgb0 += BPP * 2;                                                   \
-      src_rgb1 += BPP * 2;                                                   \
-      dst_u += 1;                                                            \
-      dst_v += 1;                                                            \
-    }                                                                        \
-    if (width & 1) {                                                         \
-      uint8_t ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                         \
-      uint8_t ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                         \
-      uint8_t ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                         \
-      dst_u[0] = RGBToU(ar, ag, ab);                                         \
-      dst_v[0] = RGBToV(ar, ag, ab);                                         \
-    }                                                                        \
+// Intel version mimic SSE/AVX which does 2 pavgb
+#if LIBYUV_ARGBTOUV_PAVGB
+#define MAKEROWY(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                 \
+    for (x = 0; x < width; ++x) {                                          \
+      dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                      \
+      dst_y += 1;                                                          \
+    }                                                                      \
+  }                                                                        \
+  void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                       uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                    \
+    int x;                                                                 \
+    for (x = 0; x < width - 1; x += 2) {                                   \
+      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                     \
+                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));        \
+      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                     \
+                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));        \
+      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                     \
+                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));        \
+      dst_u[0] = RGBToU(ar, ag, ab);                                       \
+      dst_v[0] = RGBToV(ar, ag, ab);                                       \
+      src_rgb += BPP * 2;                                                  \
+      src_rgb1 += BPP * 2;                                                 \
+      dst_u += 1;                                                          \
+      dst_v += 1;                                                          \
+    }                                                                      \
+    if (width & 1) {                                                       \
+      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                          \
+      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                          \
+      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                          \
+      dst_u[0] = RGBToU(ar, ag, ab);                                       \
+      dst_v[0] = RGBToV(ar, ag, ab);                                       \
+    }                                                                      \
   }
+#else
+// ARM version does sum / 2 then multiply by 2x smaller coefficients
+#define MAKEROWY(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                 \
+    for (x = 0; x < width; ++x) {                                          \
+      dst_y[0] = RGBToY(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                      \
+      dst_y += 1;                                                          \
+    }                                                                      \
+  }                                                                        \
+  void NAME##ToUVRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                       uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                    \
+    int x;                                                                 \
+    for (x = 0; x < width - 1; x += 2) {                                   \
+      uint16_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] +         \
+                     src_rgb1[B + BPP] + 1) >>                             \
+                    1;                                                     \
+      uint16_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] +         \
+                     src_rgb1[G + BPP] + 1) >>                             \
+                    1;                                                     \
+      uint16_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] +         \
+                     src_rgb1[R + BPP] + 1) >>                             \
+                    1;                                                     \
+      dst_u[0] = RGB2xToU(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToV(ar, ag, ab);                                     \
+      src_rgb += BPP * 2;                                                  \
+      src_rgb1 += BPP * 2;                                                 \
+      dst_u += 1;                                                          \
+      dst_v += 1;                                                          \
+    }                                                                      \
+    if (width & 1) {                                                       \
+      uint16_t ab = src_rgb[B] + src_rgb1[B];                              \
+      uint16_t ag = src_rgb[G] + src_rgb1[G];                              \
+      uint16_t ar = src_rgb[R] + src_rgb1[R];                              \
+      dst_u[0] = RGB2xToU(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToV(ar, ag, ab);                                     \
+    }                                                                      \
+  }
+#endif
 
 MAKEROWY(ARGB, 2, 1, 0, 4)
 MAKEROWY(BGRA, 1, 2, 3, 4)
@@ -448,14 +655,14 @@ MAKEROWY(RAW, 0, 1, 2, 3)
 // b 0.1016 * 255 = 25.908 = 25
 // g 0.5078 * 255 = 129.489 = 129
 // r 0.2578 * 255 = 65.739 = 66
-// JPeg 8 bit Y (not used):
-// b 0.11400 * 256 = 29.184 = 29
-// g 0.58700 * 256 = 150.272 = 150
-// r 0.29900 * 256 = 76.544 = 77
-// JPeg 7 bit Y:
+// JPeg 7 bit Y (deprecated)
 // b 0.11400 * 128 = 14.592 = 15
 // g 0.58700 * 128 = 75.136 = 75
 // r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit Y:
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
 // JPeg 8 bit U:
 // b  0.50000 * 255 = 127.5 = 127
 // g -0.33126 * 255 = -84.4713 = -84
@@ -465,57 +672,119 @@ MAKEROWY(RAW, 0, 1, 2, 3)
 // g -0.41869 * 255 = -106.76595 = -107
 // r  0.50000 * 255 = 127.5 = 127
 
+#ifdef LIBYUV_RGB7
+// Old 7 bit math for compatibility on unsupported platforms.
 static __inline int RGBToYJ(uint8_t r, uint8_t g, uint8_t b) {
   return (38 * r + 75 * g + 15 * b + 64) >> 7;
 }
+#else
+// 8 bit
+static __inline int RGBToYJ(uint8_t r, uint8_t g, uint8_t b) {
+  return (77 * r + 150 * g + 29 * b + 128) >> 8;
+}
+#endif
 
+#if defined(LIBYUV_ARGBTOUV_PAVGB)
 static __inline int RGBToUJ(uint8_t r, uint8_t g, uint8_t b) {
   return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
 }
 static __inline int RGBToVJ(uint8_t r, uint8_t g, uint8_t b) {
   return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
 }
-
-#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+#else
+static __inline int RGB2xToUJ(uint16_t r, uint16_t g, uint16_t b) {
+  return ((127 / 2) * b - (84 / 2) * g - (43 / 2) * r + 0x8080) >> 8;
+}
+static __inline int RGB2xToVJ(uint16_t r, uint16_t g, uint16_t b) {
+  return ((127 / 2) * r - (107 / 2) * g - (20 / 2) * b + 0x8080) >> 8;
+}
+#endif
 
 // ARGBToYJ_C and ARGBToUVJ_C
-#define MAKEROWYJ(NAME, R, G, B, BPP)                                         \
-  void NAME##ToYJRow_C(const uint8_t* src_argb0, uint8_t* dst_y, int width) { \
-    int x;                                                                    \
-    for (x = 0; x < width; ++x) {                                             \
-      dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);           \
-      src_argb0 += BPP;                                                       \
-      dst_y += 1;                                                             \
-    }                                                                         \
-  }                                                                           \
-  void NAME##ToUVJRow_C(const uint8_t* src_rgb0, int src_stride_rgb,          \
-                        uint8_t* dst_u, uint8_t* dst_v, int width) {          \
-    const uint8_t* src_rgb1 = src_rgb0 + src_stride_rgb;                      \
-    int x;                                                                    \
-    for (x = 0; x < width - 1; x += 2) {                                      \
-      uint8_t ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                       \
-                        AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));          \
-      uint8_t ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                       \
-                        AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));          \
-      uint8_t ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                       \
-                        AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));          \
-      dst_u[0] = RGBToUJ(ar, ag, ab);                                         \
-      dst_v[0] = RGBToVJ(ar, ag, ab);                                         \
-      src_rgb0 += BPP * 2;                                                    \
-      src_rgb1 += BPP * 2;                                                    \
-      dst_u += 1;                                                             \
-      dst_v += 1;                                                             \
-    }                                                                         \
-    if (width & 1) {                                                          \
-      uint8_t ab = AVGB(src_rgb0[B], src_rgb1[B]);                            \
-      uint8_t ag = AVGB(src_rgb0[G], src_rgb1[G]);                            \
-      uint8_t ar = AVGB(src_rgb0[R], src_rgb1[R]);                            \
-      dst_u[0] = RGBToUJ(ar, ag, ab);                                         \
-      dst_v[0] = RGBToVJ(ar, ag, ab);                                         \
-    }                                                                         \
+// Intel version mimic SSE/AVX which does 2 pavgb
+#if LIBYUV_ARGBTOUV_PAVGB
+#define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                  \
+    for (x = 0; x < width; ++x) {                                           \
+      dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                       \
+      dst_y += 1;                                                           \
+    }                                                                       \
+  }                                                                         \
+  void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                        uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                     \
+    int x;                                                                  \
+    for (x = 0; x < width - 1; x += 2) {                                    \
+      uint8_t ab = AVGB(AVGB(src_rgb[B], src_rgb1[B]),                      \
+                        AVGB(src_rgb[B + BPP], src_rgb1[B + BPP]));         \
+      uint8_t ag = AVGB(AVGB(src_rgb[G], src_rgb1[G]),                      \
+                        AVGB(src_rgb[G + BPP], src_rgb1[G + BPP]));         \
+      uint8_t ar = AVGB(AVGB(src_rgb[R], src_rgb1[R]),                      \
+                        AVGB(src_rgb[R + BPP], src_rgb1[R + BPP]));         \
+      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
+      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
+      src_rgb += BPP * 2;                                                   \
+      src_rgb1 += BPP * 2;                                                  \
+      dst_u += 1;                                                           \
+      dst_v += 1;                                                           \
+    }                                                                       \
+    if (width & 1) {                                                        \
+      uint8_t ab = AVGB(src_rgb[B], src_rgb1[B]);                           \
+      uint8_t ag = AVGB(src_rgb[G], src_rgb1[G]);                           \
+      uint8_t ar = AVGB(src_rgb[R], src_rgb1[R]);                           \
+      dst_u[0] = RGBToUJ(ar, ag, ab);                                       \
+      dst_v[0] = RGBToVJ(ar, ag, ab);                                       \
+    }                                                                       \
   }
+#else
+// ARM version does sum / 2 then multiply by 2x smaller coefficients
+#define MAKEROWYJ(NAME, R, G, B, BPP)                                       \
+  void NAME##ToYJRow_C(const uint8_t* src_rgb, uint8_t* dst_y, int width) { \
+    int x;                                                                  \
+    for (x = 0; x < width; ++x) {                                           \
+      dst_y[0] = RGBToYJ(src_rgb[R], src_rgb[G], src_rgb[B]);               \
+      src_rgb += BPP;                                                       \
+      dst_y += 1;                                                           \
+    }                                                                       \
+  }                                                                         \
+  void NAME##ToUVJRow_C(const uint8_t* src_rgb, int src_stride_rgb,         \
+                        uint8_t* dst_u, uint8_t* dst_v, int width) {        \
+    const uint8_t* src_rgb1 = src_rgb + src_stride_rgb;                     \
+    int x;                                                                  \
+    for (x = 0; x < width - 1; x += 2) {                                    \
+      uint16_t ab = (src_rgb[B] + src_rgb[B + BPP] + src_rgb1[B] +          \
+                     src_rgb1[B + BPP] + 1) >>                              \
+                    1;                                                      \
+      uint16_t ag = (src_rgb[G] + src_rgb[G + BPP] + src_rgb1[G] +          \
+                     src_rgb1[G + BPP] + 1) >>                              \
+                    1;                                                      \
+      uint16_t ar = (src_rgb[R] + src_rgb[R + BPP] + src_rgb1[R] +          \
+                     src_rgb1[R + BPP] + 1) >>                              \
+                    1;                                                      \
+      dst_u[0] = RGB2xToUJ(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToVJ(ar, ag, ab);                                     \
+      src_rgb += BPP * 2;                                                   \
+      src_rgb1 += BPP * 2;                                                  \
+      dst_u += 1;                                                           \
+      dst_v += 1;                                                           \
+    }                                                                       \
+    if (width & 1) {                                                        \
+      uint16_t ab = (src_rgb[B] + src_rgb1[B]);                             \
+      uint16_t ag = (src_rgb[G] + src_rgb1[G]);                             \
+      uint16_t ar = (src_rgb[R] + src_rgb1[R]);                             \
+      dst_u[0] = RGB2xToUJ(ar, ag, ab);                                     \
+      dst_v[0] = RGB2xToVJ(ar, ag, ab);                                     \
+    }                                                                       \
+  }
+
+#endif
 
 MAKEROWYJ(ARGB, 2, 1, 0, 4)
+MAKEROWYJ(RGBA, 3, 2, 1, 4)
+MAKEROWYJ(RGB24, 2, 1, 0, 3)
+MAKEROWYJ(RAW, 0, 1, 2, 3)
 #undef MAKEROWYJ
 
 void RGB565ToYRow_C(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
@@ -583,13 +852,34 @@ void RGB565ToUVRow_C(const uint8_t* src_rgb565,
     uint8_t b3 = next_rgb565[2] & 0x1f;
     uint8_t g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
     uint8_t r3 = next_rgb565[3] >> 3;
-    uint8_t b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
-    uint8_t g = (g0 + g1 + g2 + g3);
-    uint8_t r = (r0 + r1 + r2 + r3);
-    b = (b << 1) | (b >> 6);  // 787 -> 888.
-    r = (r << 1) | (r >> 6);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+
+    b0 = (b0 << 3) | (b0 >> 2);
+    g0 = (g0 << 2) | (g0 >> 4);
+    r0 = (r0 << 3) | (r0 >> 2);
+    b1 = (b1 << 3) | (b1 >> 2);
+    g1 = (g1 << 2) | (g1 >> 4);
+    r1 = (r1 << 3) | (r1 >> 2);
+    b2 = (b2 << 3) | (b2 >> 2);
+    g2 = (g2 << 2) | (g2 >> 4);
+    r2 = (r2 << 3) | (r2 >> 2);
+    b3 = (b3 << 3) | (b3 >> 2);
+    g3 = (g3 << 2) | (g3 >> 4);
+    r3 = (r3 << 3) | (r3 >> 2);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_rgb565 += 4;
     next_rgb565 += 4;
     dst_u += 1;
@@ -602,14 +892,27 @@ void RGB565ToUVRow_C(const uint8_t* src_rgb565,
     uint8_t b2 = next_rgb565[0] & 0x1f;
     uint8_t g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
     uint8_t r2 = next_rgb565[1] >> 3;
-    uint8_t b = (b0 + b2);  // 565 * 2 = 676.
-    uint8_t g = (g0 + g2);
-    uint8_t r = (r0 + r2);
-    b = (b << 2) | (b >> 4);  // 676 -> 888
-    g = (g << 1) | (g >> 6);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+
+    b0 = (b0 << 3) | (b0 >> 2);
+    g0 = (g0 << 2) | (g0 >> 4);
+    r0 = (r0 << 3) | (r0 >> 2);
+    b2 = (b2 << 3) | (b2 >> 2);
+    g2 = (g2 << 2) | (g2 >> 4);
+    r2 = (r2 << 3) | (r2 >> 2);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
   }
 }
 
@@ -633,14 +936,34 @@ void ARGB1555ToUVRow_C(const uint8_t* src_argb1555,
     uint8_t b3 = next_argb1555[2] & 0x1f;
     uint8_t g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
     uint8_t r3 = (next_argb1555[3] & 0x7c) >> 2;
-    uint8_t b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
-    uint8_t g = (g0 + g1 + g2 + g3);
-    uint8_t r = (r0 + r1 + r2 + r3);
-    b = (b << 1) | (b >> 6);  // 777 -> 888.
-    g = (g << 1) | (g >> 6);
-    r = (r << 1) | (r >> 6);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+
+    b0 = (b0 << 3) | (b0 >> 2);
+    g0 = (g0 << 3) | (g0 >> 2);
+    r0 = (r0 << 3) | (r0 >> 2);
+    b1 = (b1 << 3) | (b1 >> 2);
+    g1 = (g1 << 3) | (g1 >> 2);
+    r1 = (r1 << 3) | (r1 >> 2);
+    b2 = (b2 << 3) | (b2 >> 2);
+    g2 = (g2 << 3) | (g2 >> 2);
+    r2 = (r2 << 3) | (r2 >> 2);
+    b3 = (b3 << 3) | (b3 >> 2);
+    g3 = (g3 << 3) | (g3 >> 2);
+    r3 = (r3 << 3) | (r3 >> 2);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_argb1555 += 4;
     next_argb1555 += 4;
     dst_u += 1;
@@ -652,15 +975,28 @@ void ARGB1555ToUVRow_C(const uint8_t* src_argb1555,
     uint8_t r0 = (src_argb1555[1] & 0x7c) >> 2;
     uint8_t b2 = next_argb1555[0] & 0x1f;
     uint8_t g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
-    uint8_t r2 = next_argb1555[1] >> 3;
-    uint8_t b = (b0 + b2);  // 555 * 2 = 666.
-    uint8_t g = (g0 + g2);
-    uint8_t r = (r0 + r2);
-    b = (b << 2) | (b >> 4);  // 666 -> 888.
-    g = (g << 2) | (g >> 4);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+    uint8_t r2 = (next_argb1555[1] & 0x7c) >> 2;
+
+    b0 = (b0 << 3) | (b0 >> 2);
+    g0 = (g0 << 3) | (g0 >> 2);
+    r0 = (r0 << 3) | (r0 >> 2);
+    b2 = (b2 << 3) | (b2 >> 2);
+    g2 = (g2 << 3) | (g2 >> 2);
+    r2 = (r2 << 3) | (r2 >> 2);
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
   }
 }
 
@@ -684,14 +1020,34 @@ void ARGB4444ToUVRow_C(const uint8_t* src_argb4444,
     uint8_t b3 = next_argb4444[2] & 0x0f;
     uint8_t g3 = next_argb4444[2] >> 4;
     uint8_t r3 = next_argb4444[3] & 0x0f;
-    uint8_t b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
-    uint8_t g = (g0 + g1 + g2 + g3);
-    uint8_t r = (r0 + r1 + r2 + r3);
-    b = (b << 2) | (b >> 4);  // 666 -> 888.
-    g = (g << 2) | (g >> 4);
-    r = (r << 2) | (r >> 4);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+
+    b0 = (b0 << 4) | b0;
+    g0 = (g0 << 4) | g0;
+    r0 = (r0 << 4) | r0;
+    b1 = (b1 << 4) | b1;
+    g1 = (g1 << 4) | g1;
+    r1 = (r1 << 4) | r1;
+    b2 = (b2 << 4) | b2;
+    g2 = (g2 << 4) | g2;
+    r2 = (r2 << 4) | r2;
+    b3 = (b3 << 4) | b3;
+    g3 = (g3 << 4) | g3;
+    r3 = (r3 << 4) | r3;
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(AVGB(b0, b2), AVGB(b1, b3));
+    uint8_t ag = AVGB(AVGB(g0, g2), AVGB(g1, g3));
+    uint8_t ar = AVGB(AVGB(r0, r2), AVGB(r1, r3));
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = (b0 + b1 + b2 + b3 + 1) >> 1;
+    uint16_t g = (g0 + g1 + g2 + g3 + 1) >> 1;
+    uint16_t r = (r0 + r1 + r2 + r3 + 1) >> 1;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
+
     src_argb4444 += 4;
     next_argb4444 += 4;
     dst_u += 1;
@@ -704,14 +1060,27 @@ void ARGB4444ToUVRow_C(const uint8_t* src_argb4444,
     uint8_t b2 = next_argb4444[0] & 0x0f;
     uint8_t g2 = next_argb4444[0] >> 4;
     uint8_t r2 = next_argb4444[1] & 0x0f;
-    uint8_t b = (b0 + b2);  // 444 * 2 = 555.
-    uint8_t g = (g0 + g2);
-    uint8_t r = (r0 + r2);
-    b = (b << 3) | (b >> 2);  // 555 -> 888.
-    g = (g << 3) | (g >> 2);
-    r = (r << 3) | (r >> 2);
-    dst_u[0] = RGBToU(r, g, b);
-    dst_v[0] = RGBToV(r, g, b);
+
+    b0 = (b0 << 4) | b0;
+    g0 = (g0 << 4) | g0;
+    r0 = (r0 << 4) | r0;
+    b2 = (b2 << 4) | b2;
+    g2 = (g2 << 4) | g2;
+    r2 = (r2 << 4) | r2;
+
+#if LIBYUV_ARGBTOUV_PAVGB
+    uint8_t ab = AVGB(b0, b2);
+    uint8_t ag = AVGB(g0, g2);
+    uint8_t ar = AVGB(r0, r2);
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+#else
+    uint16_t b = b0 + b2;
+    uint16_t g = g0 + g2;
+    uint16_t r = r0 + r2;
+    dst_u[0] = RGB2xToU(r, g, b);
+    dst_v[0] = RGB2xToV(r, g, b);
+#endif
   }
 }
 
@@ -877,16 +1246,16 @@ void ARGBShadeRow_C(const uint8_t* src_argb,
 #define REPEAT8(v) (v) | ((v) << 8)
 #define SHADE(f, v) v* f >> 16
 
-void ARGBMultiplyRow_C(const uint8_t* src_argb0,
+void ARGBMultiplyRow_C(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const uint32_t b = REPEAT8(src_argb0[0]);
-    const uint32_t g = REPEAT8(src_argb0[1]);
-    const uint32_t r = REPEAT8(src_argb0[2]);
-    const uint32_t a = REPEAT8(src_argb0[3]);
+    const uint32_t b = REPEAT8(src_argb[0]);
+    const uint32_t g = REPEAT8(src_argb[1]);
+    const uint32_t r = REPEAT8(src_argb[2]);
+    const uint32_t a = REPEAT8(src_argb[3]);
     const uint32_t b_scale = src_argb1[0];
     const uint32_t g_scale = src_argb1[1];
     const uint32_t r_scale = src_argb1[2];
@@ -895,7 +1264,7 @@ void ARGBMultiplyRow_C(const uint8_t* src_argb0,
     dst_argb[1] = SHADE(g, g_scale);
     dst_argb[2] = SHADE(r, r_scale);
     dst_argb[3] = SHADE(a, a_scale);
-    src_argb0 += 4;
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -905,16 +1274,16 @@ void ARGBMultiplyRow_C(const uint8_t* src_argb0,
 
 #define SHADE(f, v) clamp255(v + f)
 
-void ARGBAddRow_C(const uint8_t* src_argb0,
+void ARGBAddRow_C(const uint8_t* src_argb,
                   const uint8_t* src_argb1,
                   uint8_t* dst_argb,
                   int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const int b = src_argb0[0];
-    const int g = src_argb0[1];
-    const int r = src_argb0[2];
-    const int a = src_argb0[3];
+    const int b = src_argb[0];
+    const int g = src_argb[1];
+    const int r = src_argb[2];
+    const int a = src_argb[3];
     const int b_add = src_argb1[0];
     const int g_add = src_argb1[1];
     const int r_add = src_argb1[2];
@@ -923,7 +1292,7 @@ void ARGBAddRow_C(const uint8_t* src_argb0,
     dst_argb[1] = SHADE(g, g_add);
     dst_argb[2] = SHADE(r, r_add);
     dst_argb[3] = SHADE(a, a_add);
-    src_argb0 += 4;
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -932,16 +1301,16 @@ void ARGBAddRow_C(const uint8_t* src_argb0,
 
 #define SHADE(f, v) clamp0(f - v)
 
-void ARGBSubtractRow_C(const uint8_t* src_argb0,
+void ARGBSubtractRow_C(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width) {
   int i;
   for (i = 0; i < width; ++i) {
-    const int b = src_argb0[0];
-    const int g = src_argb0[1];
-    const int r = src_argb0[2];
-    const int a = src_argb0[3];
+    const int b = src_argb[0];
+    const int g = src_argb[1];
+    const int r = src_argb[2];
+    const int a = src_argb[3];
     const int b_sub = src_argb1[0];
     const int g_sub = src_argb1[1];
     const int r_sub = src_argb1[2];
@@ -950,7 +1319,7 @@ void ARGBSubtractRow_C(const uint8_t* src_argb0,
     dst_argb[1] = SHADE(g, g_sub);
     dst_argb[2] = SHADE(r, r_sub);
     dst_argb[3] = SHADE(a, a_sub);
-    src_argb0 += 4;
+    src_argb += 4;
     src_argb1 += 4;
     dst_argb += 4;
   }
@@ -1058,257 +1427,244 @@ void J400ToARGBRow_C(const uint8_t* src_y, uint8_t* dst_argb, int width) {
   }
 }
 
-// TODO(fbarchard): Unify these structures to be platform independent.
-// TODO(fbarchard): Generate SIMD structures from float matrix.
+// Macros to create SIMD specific yuv to rgb conversion constants.
 
-// BT.601 YUV to RGB reference
-//  R = (Y - 16) * 1.164              - V * -1.596
-//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
-//  B = (Y - 16) * 1.164 - U * -2.018
+// clang-format off
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997  /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+#if defined(__aarch64__) || defined(__arm__)
+// Bias values include subtract 128 from U and V, bias from Y and rounding.
+// For B and R bias is negative. For G bias is positive.
+#define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR)                             \
+  {{UB, VR, UG, VG, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},                     \
+   {YG, (UB * 128 - YB), (UG * 128 + VG * 128 + YB), (VR * 128 - YB), YB, 0, \
+    0, 0}}
+#else
+#define YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR)                     \
+  {{UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,          \
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0},         \
+   {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,  \
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG}, \
+   {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,          \
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR},         \
+   {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}, \
+   {YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB, YB}}
+#endif
+
+// clang-format on
+
+#define MAKEYUVCONSTANTS(name, YG, YB, UB, UG, VG, VR)            \
+  const struct YuvConstants SIMD_ALIGNED(kYuv##name##Constants) = \
+      YUVCONSTANTSBODY(YG, YB, UB, UG, VG, VR);                   \
+  const struct YuvConstants SIMD_ALIGNED(kYvu##name##Constants) = \
+      YUVCONSTANTSBODY(YG, YB, VR, VG, UG, UB);
+
+// TODO(fbarchard): Generate SIMD structures from float matrix.
+
+// BT.601 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164             + V * 1.596
+//  G = (Y - 16) * 1.164 - U * 0.391 - V * 0.813
+//  B = (Y - 16) * 1.164 + U * 2.018
+// KR = 0.299; KB = 0.114
 
 // U and V contributions to R,G,B.
-#define UB -128 /* max(-128, round(-2.018 * 64)) */
-#define UG 25   /* round(0.391 * 64) */
-#define VG 52   /* round(0.813 * 64) */
-#define VR -102 /* round(-1.596 * 64) */
-
-// Bias values to subtract 16 from Y and 128 from U and V.
-#define BB (UB * 128 + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR (VR * 128 + YGB)
-
-#if defined(__aarch64__)  // 64 bit arm
-const struct YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-#elif defined(__arm__)  // 32 bit arm
-const struct YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-    {-UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0},
-    {UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-    {-VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0},
-    {VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT601)
+#define UB 129 /* round(2.018 * 64) */
 #else
-const struct YuvConstants SIMD_ALIGNED(kYuvI601Constants) = {
-    {UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-     UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0},
-    {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-     UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG},
-    {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-     0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
-const struct YuvConstants SIMD_ALIGNED(kYvuI601Constants) = {
-    {VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-     VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0},
-    {VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-     VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG},
-    {0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-     0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
+#define UB 128 /* max(128, round(2.018 * 64)) */
 #endif
+#define UG 25  /* round(0.391 * 64) */
+#define VG 52  /* round(0.813 * 64) */
+#define VR 102 /* round(1.596 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(I601, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
-#undef YG
 
-// JPEG YUV to RGB reference
-// *  R = Y                - V * -1.40200
-// *  G = Y - U *  0.34414 - V *  0.71414
-// *  B = Y - U * -1.77200
+// BT.601 full range YUV to RGB reference (aka JPEG)
+// *  R = Y               + V * 1.40200
+// *  G = Y - U * 0.34414 - V * 0.71414
+// *  B = Y + U * 1.77200
+// KR = 0.299; KB = 0.114
+
+// U and V contributions to R,G,B.
+#define UB 113 /* round(1.77200 * 64) */
+#define UG 22  /* round(0.34414 * 64) */
+#define VG 46  /* round(0.71414 * 64) */
+#define VR 90  /* round(1.40200 * 64) */
 
 // Y contribution to R,G,B.  Scale and bias.
 #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
-#define YGB 32   /* 64 / 2 */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(JPEG, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+// BT.709 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164             + V * 1.793
+//  G = (Y - 16) * 1.164 - U * 0.213 - V * 0.533
+//  B = (Y - 16) * 1.164 + U * 2.112
+//  KR = 0.2126, KB = 0.0722
 
 // U and V contributions to R,G,B.
-#define UB -113 /* round(-1.77200 * 64) */
-#define UG 22   /* round(0.34414 * 64) */
-#define VG 46   /* round(0.71414  * 64) */
-#define VR -90  /* round(-1.40200 * 64) */
-
-// Bias values to round, and subtract 128 from U and V.
-#define BB (UB * 128 + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR (VR * 128 + YGB)
-
-#if defined(__aarch64__)
-const struct YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-#elif defined(__arm__)
-const struct YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-    {-UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0},
-    {UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-    {-VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0},
-    {VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT709)
+#define UB 135 /* round(2.112 * 64) */
 #else
-const struct YuvConstants SIMD_ALIGNED(kYuvJPEGConstants) = {
-    {UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-     UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0},
-    {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-     UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG},
-    {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-     0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
-const struct YuvConstants SIMD_ALIGNED(kYvuJPEGConstants) = {
-    {VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-     VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0},
-    {VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-     VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG},
-    {0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-     0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
+#define UB 128 /* max(128, round(2.112 * 64)) */
 #endif
+#define UG 14  /* round(0.213 * 64) */
+#define VG 34  /* round(0.533 * 64) */
+#define VR 115 /* round(1.793 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(H709, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
-#undef YG
 
-// BT.709 YUV to RGB reference
-//  R = (Y - 16) * 1.164              - V * -1.793
-//  G = (Y - 16) * 1.164 - U *  0.213 - V *  0.533
-//  B = (Y - 16) * 1.164 - U * -2.112
-// See also http://www.equasys.de/colorconversion.html
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126, KB = 0.0722
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997  /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+// U and V contributions to R,G,B.
+#define UB 119 /* round(1.8556 * 64) */
+#define UG 12  /* round(0.18732 * 64) */
+#define VG 30  /* round(0.46812 * 64) */
+#define VR 101 /* round(1.5748 * 64) */
+
+// Y contribution to R,G,B.  Scale and bias.  (same as jpeg)
+#define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(F709, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+// BT.2020 limited range YUV to RGB reference
+//  R = (Y - 16) * 1.164384                + V * 1.67867
+//  G = (Y - 16) * 1.164384 - U * 0.187326 - V * 0.65042
+//  B = (Y - 16) * 1.164384 + U * 2.14177
+// KR = 0.2627; KB = 0.0593
 
-// TODO(fbarchard): Find way to express 2.112 instead of 2.0.
 // U and V contributions to R,G,B.
-#define UB -128 /* max(-128, round(-2.112 * 64)) */
-#define UG 14   /* round(0.213 * 64) */
-#define VG 34   /* round(0.533  * 64) */
-#define VR -115 /* round(-1.793 * 64) */
-
-// Bias values to round, and subtract 128 from U and V.
-#define BB (UB * 128 + YGB)
-#define BG (UG * 128 + VG * 128 + YGB)
-#define BR (VR * 128 + YGB)
-
-#if defined(__aarch64__)
-const struct YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {-UB, -VR, -UB, -VR, -UB, -VR, -UB, -VR},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {UG, VG, UG, VG, UG, VG, UG, VG},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {-VR, -UB, -VR, -UB, -VR, -UB, -VR, -UB},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {VG, UG, VG, UG, VG, UG, VG, UG},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-#elif defined(__arm__)
-const struct YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-    {-UB, -UB, -UB, -UB, -VR, -VR, -VR, -VR, 0, 0, 0, 0, 0, 0, 0, 0},
-    {UG, UG, UG, UG, VG, VG, VG, VG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BB, BG, BR, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
-const struct YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-    {-VR, -VR, -VR, -VR, -UB, -UB, -UB, -UB, 0, 0, 0, 0, 0, 0, 0, 0},
-    {VG, VG, VG, VG, UG, UG, UG, UG, 0, 0, 0, 0, 0, 0, 0, 0},
-    {BR, BG, BB, 0, 0, 0, 0, 0},
-    {0x0101 * YG, 0, 0, 0}};
+#if defined(LIBYUV_UNLIMITED_DATA) || defined(LIBYUV_UNLIMITED_BT2020)
+#define UB 137 /* round(2.142 * 64) */
 #else
-const struct YuvConstants SIMD_ALIGNED(kYuvH709Constants) = {
-    {UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
-     UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0},
-    {UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
-     UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG},
-    {0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
-     0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
-const struct YuvConstants SIMD_ALIGNED(kYvuH709Constants) = {
-    {VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
-     VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0},
-    {VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
-     VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG},
-    {0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
-     0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB},
-    {BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR},
-    {BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG},
-    {BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB},
-    {YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG}};
+#define UB 128 /* max(128, round(2.142 * 64)) */
 #endif
+#define UG 12  /* round(0.187326 * 64) */
+#define VG 42  /* round(0.65042 * 64) */
+#define VR 107 /* round(1.67867 * 64) */
 
-#undef BB
-#undef BG
-#undef BR
-#undef YGB
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 19003 /* round(1.164384 * 64 * 256 * 256 / 257) */
+#define YB -1160 /* 1.164384 * 64 * -16 + 64 / 2 */
+
+MAKEYUVCONSTANTS(2020, YG, YB, UB, UG, VG, VR)
+
+#undef YG
+#undef YB
 #undef UB
 #undef UG
 #undef VG
 #undef VR
+
+// BT.2020 full range YUV to RGB reference
+//  R = Y                + V * 1.474600
+//  G = Y - U * 0.164553 - V * 0.571353
+//  B = Y + U * 1.881400
+// KR = 0.2627; KB = 0.0593
+
+#define UB 120 /* round(1.881400 * 64) */
+#define UG 11  /* round(0.164553 * 64) */
+#define VG 37  /* round(0.571353 * 64) */
+#define VR 94  /* round(1.474600 * 64) */
+
+// Y contribution to R,G,B.  Scale and bias.  (same as jpeg)
+#define YG 16320 /* round(1 * 64 * 256 * 256 / 257) */
+#define YB 32    /* 64 / 2 */
+
+MAKEYUVCONSTANTS(V2020, YG, YB, UB, UG, VG, VR)
+
 #undef YG
+#undef YB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+
+#undef BB
+#undef BG
+#undef BR
+
+#undef MAKEYUVCONSTANTS
+
+#if defined(__aarch64__) || defined(__arm__)
+#define LOAD_YUV_CONSTANTS                 \
+  int ub = yuvconstants->kUVCoeff[0];      \
+  int vr = yuvconstants->kUVCoeff[1];      \
+  int ug = yuvconstants->kUVCoeff[2];      \
+  int vg = yuvconstants->kUVCoeff[3];      \
+  int yg = yuvconstants->kRGBCoeffBias[0]; \
+  int bb = yuvconstants->kRGBCoeffBias[1]; \
+  int bg = yuvconstants->kRGBCoeffBias[2]; \
+  int br = yuvconstants->kRGBCoeffBias[3]
+
+#define CALC_RGB16                         \
+  int32_t y1 = (uint32_t)(y32 * yg) >> 16; \
+  int b16 = y1 + (u * ub) - bb;            \
+  int g16 = y1 + bg - (u * ug + v * vg);   \
+  int r16 = y1 + (v * vr) - br
+#else
+#define LOAD_YUV_CONSTANTS           \
+  int ub = yuvconstants->kUVToB[0];  \
+  int ug = yuvconstants->kUVToG[0];  \
+  int vg = yuvconstants->kUVToG[1];  \
+  int vr = yuvconstants->kUVToR[1];  \
+  int yg = yuvconstants->kYToRgb[0]; \
+  int yb = yuvconstants->kYBiasToRgb[0]
+
+#define CALC_RGB16                                \
+  int32_t y1 = ((uint32_t)(y32 * yg) >> 16) + yb; \
+  int8_t ui = (int8_t)u;                          \
+  int8_t vi = (int8_t)v;                          \
+  ui -= 0x80;                                     \
+  vi -= 0x80;                                     \
+  int b16 = y1 + (ui * ub);                       \
+  int g16 = y1 - (ui * ug + vi * vg);             \
+  int r16 = y1 + (vi * vr)
+#endif
 
 // C reference code that mimics the YUV assembly.
 // Reads 8 bit YUV and leaves result as 16 bit.
-
 static __inline void YuvPixel(uint8_t y,
                               uint8_t u,
                               uint8_t v,
@@ -1316,39 +1672,12 @@ static __inline void YuvPixel(uint8_t y,
                               uint8_t* g,
                               uint8_t* r,
                               const struct YuvConstants* yuvconstants) {
-#if defined(__aarch64__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = -yuvconstants->kUVToRB[1];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#elif defined(__arm__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[4];
-  int vr = -yuvconstants->kUVToRB[4];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#else
-  int ub = yuvconstants->kUVToB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = yuvconstants->kUVToR[1];
-  int bb = yuvconstants->kUVBiasB[0];
-  int bg = yuvconstants->kUVBiasG[0];
-  int br = yuvconstants->kUVBiasR[0];
-  int yg = yuvconstants->kYToRgb[0];
-#endif
-
-  uint32_t y1 = (uint32_t)(y * 0x0101 * yg) >> 16;
-  *b = Clamp((int32_t)(-(u * ub) + y1 + bb) >> 6);
-  *g = Clamp((int32_t)(-(u * ug + v * vg) + y1 + bg) >> 6);
-  *r = Clamp((int32_t)(-(v * vr) + y1 + br) >> 6);
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y * 0x0101;
+  CALC_RGB16;
+  *b = Clamp((int32_t)(b16) >> 6);
+  *g = Clamp((int32_t)(g16) >> 6);
+  *r = Clamp((int32_t)(r16) >> 6);
 }
 
 // Reads 8 bit YUV and leaves result as 16 bit.
@@ -1359,85 +1688,50 @@ static __inline void YuvPixel8_16(uint8_t y,
                                   int* g,
                                   int* r,
                                   const struct YuvConstants* yuvconstants) {
-#if defined(__aarch64__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = -yuvconstants->kUVToRB[1];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#elif defined(__arm__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[4];
-  int vr = -yuvconstants->kUVToRB[4];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#else
-  int ub = yuvconstants->kUVToB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = yuvconstants->kUVToR[1];
-  int bb = yuvconstants->kUVBiasB[0];
-  int bg = yuvconstants->kUVBiasG[0];
-  int br = yuvconstants->kUVBiasR[0];
-  int yg = yuvconstants->kYToRgb[0];
-#endif
-
-  uint32_t y1 = (uint32_t)(y * 0x0101 * yg) >> 16;
-  *b = (int)(-(u * ub) + y1 + bb);
-  *g = (int)(-(u * ug + v * vg) + y1 + bg);
-  *r = (int)(-(v * vr) + y1 + br);
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y * 0x0101;
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
 }
 
 // C reference code that mimics the YUV 16 bit assembly.
 // Reads 10 bit YUV and leaves result as 16 bit.
-static __inline void YuvPixel16(int16_t y,
-                                int16_t u,
-                                int16_t v,
-                                int* b,
-                                int* g,
-                                int* r,
-                                const struct YuvConstants* yuvconstants) {
-#if defined(__aarch64__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = -yuvconstants->kUVToRB[1];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#elif defined(__arm__)
-  int ub = -yuvconstants->kUVToRB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[4];
-  int vr = -yuvconstants->kUVToRB[4];
-  int bb = yuvconstants->kUVBiasBGR[0];
-  int bg = yuvconstants->kUVBiasBGR[1];
-  int br = yuvconstants->kUVBiasBGR[2];
-  int yg = yuvconstants->kYToRgb[0] / 0x0101;
-#else
-  int ub = yuvconstants->kUVToB[0];
-  int ug = yuvconstants->kUVToG[0];
-  int vg = yuvconstants->kUVToG[1];
-  int vr = yuvconstants->kUVToR[1];
-  int bb = yuvconstants->kUVBiasB[0];
-  int bg = yuvconstants->kUVBiasG[0];
-  int br = yuvconstants->kUVBiasR[0];
-  int yg = yuvconstants->kYToRgb[0];
-#endif
-
-  uint32_t y1 = (uint32_t)((y << 6) * yg) >> 16;
+static __inline void YuvPixel10_16(uint16_t y,
+                                   uint16_t u,
+                                   uint16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y << 6;
   u = clamp255(u >> 2);
   v = clamp255(v >> 2);
-  *b = (int)(-(u * ub) + y1 + bb);
-  *g = (int)(-(u * ug + v * vg) + y1 + bg);
-  *r = (int)(-(v * vr) + y1 + br);
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
+}
+
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 12 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel12_16(int16_t y,
+                                   int16_t u,
+                                   int16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y << 4;
+  u = clamp255(u >> 4);
+  v = clamp255(v >> 4);
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
 }
 
 // C reference code that mimics the YUV 10 bit assembly.
@@ -1452,59 +1746,88 @@ static __inline void YuvPixel10(uint16_t y,
   int b16;
   int g16;
   int r16;
-  YuvPixel16(y, u, v, &b16, &g16, &r16, yuvconstants);
+  YuvPixel10_16(y, u, v, &b16, &g16, &r16, yuvconstants);
   *b = Clamp(b16 >> 6);
   *g = Clamp(g16 >> 6);
   *r = Clamp(r16 >> 6);
 }
 
-// Y contribution to R,G,B.  Scale and bias.
-#define YG 18997  /* round(1.164 * 64 * 256 * 256 / 257) */
-#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
-
-// C reference code that mimics the YUV assembly.
-static __inline void YPixel(uint8_t y, uint8_t* b, uint8_t* g, uint8_t* r) {
-  uint32_t y1 = (uint32_t)(y * 0x0101 * YG) >> 16;
-  *b = Clamp((int32_t)(y1 + YGB) >> 6);
-  *g = Clamp((int32_t)(y1 + YGB) >> 6);
-  *r = Clamp((int32_t)(y1 + YGB) >> 6);
+// C reference code that mimics the YUV 12 bit assembly.
+// Reads 12 bit YUV and clamps down to 8 bit RGB.
+static __inline void YuvPixel12(uint16_t y,
+                                uint16_t u,
+                                uint16_t v,
+                                uint8_t* b,
+                                uint8_t* g,
+                                uint8_t* r,
+                                const struct YuvConstants* yuvconstants) {
+  int b16;
+  int g16;
+  int r16;
+  YuvPixel12_16(y, u, v, &b16, &g16, &r16, yuvconstants);
+  *b = Clamp(b16 >> 6);
+  *g = Clamp(g16 >> 6);
+  *r = Clamp(r16 >> 6);
 }
 
-#undef YG
-#undef YGB
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 16 bit YUV and leaves result as 8 bit.
+static __inline void YuvPixel16_8(uint16_t y,
+                                  uint16_t u,
+                                  uint16_t v,
+                                  uint8_t* b,
+                                  uint8_t* g,
+                                  uint8_t* r,
+                                  const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y;
+  u = clamp255(u >> 8);
+  v = clamp255(v >> 8);
+  CALC_RGB16;
+  *b = Clamp((int32_t)(b16) >> 6);
+  *g = Clamp((int32_t)(g16) >> 6);
+  *r = Clamp((int32_t)(r16) >> 6);
+}
 
-#if !defined(LIBYUV_DISABLE_NEON) && \
-    (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON))
-// C mimic assembly.
-// TODO(fbarchard): Remove subsampling from Neon.
-void I444ToARGBRow_C(const uint8_t* src_y,
-                     const uint8_t* src_u,
-                     const uint8_t* src_v,
-                     uint8_t* rgb_buf,
-                     const struct YuvConstants* yuvconstants,
-                     int width) {
-  int x;
-  for (x = 0; x < width - 1; x += 2) {
-    uint8_t u = (src_u[0] + src_u[1] + 1) >> 1;
-    uint8_t v = (src_v[0] + src_v[1] + 1) >> 1;
-    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2,
-             yuvconstants);
-    rgb_buf[3] = 255;
-    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6,
-             yuvconstants);
-    rgb_buf[7] = 255;
-    src_y += 2;
-    src_u += 2;
-    src_v += 2;
-    rgb_buf += 8;  // Advance 2 pixels.
-  }
-  if (width & 1) {
-    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
-             rgb_buf + 2, yuvconstants);
-    rgb_buf[3] = 255;
-  }
+// C reference code that mimics the YUV 16 bit assembly.
+// Reads 16 bit YUV and leaves result as 16 bit.
+static __inline void YuvPixel16_16(uint16_t y,
+                                   uint16_t u,
+                                   uint16_t v,
+                                   int* b,
+                                   int* g,
+                                   int* r,
+                                   const struct YuvConstants* yuvconstants) {
+  LOAD_YUV_CONSTANTS;
+  uint32_t y32 = y;
+  u = clamp255(u >> 8);
+  v = clamp255(v >> 8);
+  CALC_RGB16;
+  *b = b16;
+  *g = g16;
+  *r = r16;
 }
+
+// C reference code that mimics the YUV assembly.
+// Reads 8 bit YUV and leaves result as 8 bit.
+static __inline void YPixel(uint8_t y,
+                            uint8_t* b,
+                            uint8_t* g,
+                            uint8_t* r,
+                            const struct YuvConstants* yuvconstants) {
+#if defined(__aarch64__) || defined(__arm__)
+  int yg = yuvconstants->kRGBCoeffBias[0];
+  int ygb = yuvconstants->kRGBCoeffBias[4];
 #else
+  int ygb = yuvconstants->kYBiasToRgb[0];
+  int yg = yuvconstants->kYToRgb[0];
+#endif
+  uint32_t y1 = (uint32_t)(y * 0x0101 * yg) >> 16;
+  *b = Clamp(((int32_t)(y1) + ygb) >> 6);
+  *g = Clamp(((int32_t)(y1) + ygb) >> 6);
+  *r = Clamp(((int32_t)(y1) + ygb) >> 6);
+}
+
 void I444ToARGBRow_C(const uint8_t* src_y,
                      const uint8_t* src_u,
                      const uint8_t* src_v,
@@ -1522,7 +1845,6 @@ void I444ToARGBRow_C(const uint8_t* src_y,
     rgb_buf += 4;  // Advance 1 pixel.
   }
 }
-#endif
 
 // Also used for 420
 void I422ToARGBRow_C(const uint8_t* src_y,
@@ -1578,9 +1900,102 @@ void I210ToARGBRow_C(const uint16_t* src_y,
   }
 }
 
+void I410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixels.
+  }
+}
+
+void I210AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = clamp255(src_a[0] >> 2);
+    YuvPixel10(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+               rgb_buf + 6, yuvconstants);
+    rgb_buf[7] = clamp255(src_a[1] >> 2);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    src_a += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = clamp255(src_a[0] >> 2);
+  }
+}
+
+void I410AlphaToARGBRow_C(const uint16_t* src_y,
+                          const uint16_t* src_u,
+                          const uint16_t* src_v,
+                          const uint16_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel10(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = clamp255(src_a[0] >> 2);
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    src_a += 1;
+    rgb_buf += 4;  // Advance 1 pixels.
+  }
+}
+
+// 12 bit YUV to ARGB
+void I212ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+    YuvPixel12(src_y[1], src_u[0], src_v[0], rgb_buf + 4, rgb_buf + 5,
+               rgb_buf + 6, yuvconstants);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel12(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+               rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = 255;
+  }
+}
+
 static void StoreAR30(uint8_t* rgb_buf, int b, int g, int r) {
   uint32_t ar30;
-  b = b >> 4;  // convert 10.6 to 10 bit.
+  b = b >> 4;  // convert 8 bit 10.6 to 10 bit.
   g = g >> 4;
   r = r >> 4;
   b = Clamp10(b);
@@ -1602,9 +2017,9 @@ void I210ToAR30Row_C(const uint16_t* src_y,
   int g;
   int r;
   for (x = 0; x < width - 1; x += 2) {
-    YuvPixel16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
     StoreAR30(rgb_buf, b, g, r);
-    YuvPixel16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    YuvPixel10_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
     StoreAR30(rgb_buf + 4, b, g, r);
     src_y += 2;
     src_u += 1;
@@ -1612,11 +2027,141 @@ void I210ToAR30Row_C(const uint16_t* src_y,
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YuvPixel16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
     StoreAR30(rgb_buf, b, g, r);
   }
 }
 
+// 12 bit YUV to 10 bit AR30
+void I212ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    YuvPixel12_16(src_y[1], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf + 4, b, g, r);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel12_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+  }
+}
+
+void I410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_u,
+                     const uint16_t* src_v,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width; ++x) {
+    YuvPixel10_16(src_y[0], src_u[0], src_v[0], &b, &g, &r, yuvconstants);
+    StoreAR30(rgb_buf, b, g, r);
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+
+// P210 has 10 bits in msb of 16 bit NV12 style layout.
+void P210ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+    YuvPixel16_8(src_y[1], src_uv[0], src_uv[1], dst_argb + 4, dst_argb + 5,
+                 dst_argb + 6, yuvconstants);
+    dst_argb[7] = 255;
+    src_y += 2;
+    src_uv += 2;
+    dst_argb += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+  }
+}
+
+void P410ToARGBRow_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_argb,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel16_8(src_y[0], src_uv[0], src_uv[1], dst_argb + 0, dst_argb + 1,
+                 dst_argb + 2, yuvconstants);
+    dst_argb[3] = 255;
+    src_y += 1;
+    src_uv += 2;
+    dst_argb += 4;  // Advance 1 pixels.
+  }
+}
+
+void P210ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+    YuvPixel16_16(src_y[1], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30 + 4, b, g, r);
+    src_y += 2;
+    src_uv += 2;
+    dst_ar30 += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+  }
+}
+
+void P410ToAR30Row_C(const uint16_t* src_y,
+                     const uint16_t* src_uv,
+                     uint8_t* dst_ar30,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
+  int x;
+  int b;
+  int g;
+  int r;
+  for (x = 0; x < width; ++x) {
+    YuvPixel16_16(src_y[0], src_uv[0], src_uv[1], &b, &g, &r, yuvconstants);
+    StoreAR30(dst_ar30, b, g, r);
+    src_y += 1;
+    src_uv += 2;
+    dst_ar30 += 4;  // Advance 1 pixel.
+  }
+}
+
 // 8 bit YUV to 10 bit AR30
 // Uses same code as 10 bit YUV bit shifts the 8 bit values up to 10 bits.
 void I422ToAR30Row_C(const uint8_t* src_y,
@@ -1645,6 +2190,26 @@ void I422ToAR30Row_C(const uint8_t* src_y,
   }
 }
 
+void I444AlphaToARGBRow_C(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          const uint8_t* src_a,
+                          uint8_t* rgb_buf,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], rgb_buf + 0, rgb_buf + 1,
+             rgb_buf + 2, yuvconstants);
+    rgb_buf[3] = src_a[0];
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    src_a += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+
 void I422AlphaToARGBRow_C(const uint8_t* src_y,
                           const uint8_t* src_u,
                           const uint8_t* src_v,
@@ -1718,8 +2283,8 @@ void I422ToARGB4444Row_C(const uint8_t* src_y,
     b1 = b1 >> 4;
     g1 = g1 >> 4;
     r1 = r1 >> 4;
-    *(uint32_t*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) | (b1 << 16) |
-                                 (g1 << 20) | (r1 << 24) | 0xf000f000;
+    *(uint16_t*)(dst_argb4444 + 0) = b0 | (g0 << 4) | (r0 << 8) | 0xf000;
+    *(uint16_t*)(dst_argb4444 + 2) = b1 | (g1 << 4) | (r1 << 8) | 0xf000;
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1756,8 +2321,8 @@ void I422ToARGB1555Row_C(const uint8_t* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 3;
     r1 = r1 >> 3;
-    *(uint32_t*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) | (b1 << 16) |
-                                 (g1 << 21) | (r1 << 26) | 0x80008000;
+    *(uint16_t*)(dst_argb1555 + 0) = b0 | (g0 << 5) | (r0 << 10) | 0x8000;
+    *(uint16_t*)(dst_argb1555 + 2) = b1 | (g1 << 5) | (r1 << 10) | 0x8000;
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1794,8 +2359,8 @@ void I422ToRGB565Row_C(const uint8_t* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 2;
     r1 = r1 >> 3;
-    *(uint32_t*)(dst_rgb565) =
-        b0 | (g0 << 5) | (r0 << 11) | (b1 << 16) | (g1 << 21) | (r1 << 27);
+    *(uint16_t*)(dst_rgb565 + 0) = b0 | (g0 << 5) | (r0 << 11);  // for ubsan
+    *(uint16_t*)(dst_rgb565 + 2) = b1 | (g1 << 5) | (r1 << 11);
     src_y += 2;
     src_u += 1;
     src_v += 1;
@@ -1921,8 +2486,8 @@ void NV12ToRGB565Row_C(const uint8_t* src_y,
     b1 = b1 >> 3;
     g1 = g1 >> 2;
     r1 = r1 >> 3;
-    *(uint32_t*)(dst_rgb565) =
-        b0 | (g0 << 5) | (r0 << 11) | (b1 << 16) | (g1 << 21) | (r1 << 27);
+    *(uint16_t*)(dst_rgb565 + 0) = b0 | (g0 << 5) | (r0 << 11);
+    *(uint16_t*)(dst_rgb565 + 2) = b1 | (g1 << 5) | (r1 << 11);
     src_y += 2;
     src_uv += 2;
     dst_rgb565 += 4;  // Advance 2 pixels.
@@ -2006,18 +2571,21 @@ void I422ToRGBARow_C(const uint8_t* src_y,
   }
 }
 
-void I400ToARGBRow_C(const uint8_t* src_y, uint8_t* rgb_buf, int width) {
+void I400ToARGBRow_C(const uint8_t* src_y,
+                     uint8_t* rgb_buf,
+                     const struct YuvConstants* yuvconstants,
+                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
-    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6, yuvconstants);
     rgb_buf[7] = 255;
     src_y += 2;
     rgb_buf += 8;  // Advance 2 pixels.
   }
   if (width & 1) {
-    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2, yuvconstants);
     rgb_buf[3] = 255;
   }
 }
@@ -2035,10 +2603,21 @@ void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width) {
   }
 }
 
-void MirrorUVRow_C(const uint8_t* src_uv,
-                   uint8_t* dst_u,
-                   uint8_t* dst_v,
-                   int width) {
+void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width; ++x) {
+    dst_uv[0] = src_uv[0];
+    dst_uv[1] = src_uv[1];
+    src_uv -= 2;
+    dst_uv += 2;
+  }
+}
+
+void MirrorSplitUVRow_C(const uint8_t* src_uv,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
   int x;
   src_uv += (width - 1) << 1;
   for (x = 0; x < width - 1; x += 2) {
@@ -2069,6 +2648,21 @@ void ARGBMirrorRow_C(const uint8_t* src, uint8_t* dst, int width) {
   }
 }
 
+void RGB24MirrorRow_C(const uint8_t* src_rgb24, uint8_t* dst_rgb24, int width) {
+  int x;
+  src_rgb24 += width * 3 - 3;
+  for (x = 0; x < width; ++x) {
+    uint8_t b = src_rgb24[0];
+    uint8_t g = src_rgb24[1];
+    uint8_t r = src_rgb24[2];
+    dst_rgb24[0] = b;
+    dst_rgb24[1] = g;
+    dst_rgb24[2] = r;
+    src_rgb24 -= 3;
+    dst_rgb24 += 3;
+  }
+}
+
 void SplitUVRow_C(const uint8_t* src_uv,
                   uint8_t* dst_u,
                   uint8_t* dst_v,
@@ -2105,6 +2699,38 @@ void MergeUVRow_C(const uint8_t* src_u,
   }
 }
 
+void DetileRow_C(const uint8_t* src,
+                 ptrdiff_t src_tile_stride,
+                 uint8_t* dst,
+                 int width) {
+  int x;
+  for (x = 0; x < width - 15; x += 16) {
+    memcpy(dst, src, 16);
+    dst += 16;
+    src += src_tile_stride;
+  }
+  if (width & 15) {
+    memcpy(dst, src, width & 15);
+  }
+}
+
+void DetileSplitUVRow_C(const uint8_t* src_uv,
+                        ptrdiff_t src_tile_stride,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  for (x = 0; x < width - 15; x += 16) {
+    SplitUVRow_C(src_uv, dst_u, dst_v, 8);
+    dst_u += 8;
+    dst_v += 8;
+    src_uv += src_tile_stride;
+  }
+  if (width & 15) {
+    SplitUVRow_C(src_uv, dst_u, dst_v, ((width & 15) + 1) / 2);
+  }
+}
+
 void SplitRGBRow_C(const uint8_t* src_rgb,
                    uint8_t* dst_r,
                    uint8_t* dst_g,
@@ -2133,27 +2759,197 @@ void MergeRGBRow_C(const uint8_t* src_r,
   }
 }
 
-// Use scale to convert lsb formats to msb, depending how many bits there are:
-// 128 = 9 bits
-// 64 = 10 bits
-// 16 = 12 bits
-// 1 = 16 bits
+void SplitARGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    uint8_t* dst_a,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_b[x] = src_argb[0];
+    dst_g[x] = src_argb[1];
+    dst_r[x] = src_argb[2];
+    dst_a[x] = src_argb[3];
+    src_argb += 4;
+  }
+}
+
+void MergeARGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    const uint8_t* src_a,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_b[x];
+    dst_argb[1] = src_g[x];
+    dst_argb[2] = src_r[x];
+    dst_argb[3] = src_a[x];
+    dst_argb += 4;
+  }
+}
+
+void MergeXR30Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint8_t* dst_ar30,
+                    int depth,
+                    int width) {
+  assert(depth >= 10);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 10;
+  uint32_t* dst_ar30_32 = (uint32_t*)dst_ar30;
+  for (x = 0; x < width; ++x) {
+    uint32_t r = clamp1023(src_r[x] >> shift);
+    uint32_t g = clamp1023(src_g[x] >> shift);
+    uint32_t b = clamp1023(src_b[x] >> shift);
+    dst_ar30_32[x] = b | (g << 10) | (r << 20) | 0xc0000000;
+  }
+}
+
+void MergeAR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    const uint16_t* src_a,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width) {
+  assert(depth >= 1);
+  assert(depth <= 16);
+  int x;
+  int shift = 16 - depth;
+  int max = (1 << depth) - 1;
+  for (x = 0; x < width; ++x) {
+    dst_ar64[0] = ClampMax(src_b[x], max) << shift;
+    dst_ar64[1] = ClampMax(src_g[x], max) << shift;
+    dst_ar64[2] = ClampMax(src_r[x], max) << shift;
+    dst_ar64[3] = ClampMax(src_a[x], max) << shift;
+    dst_ar64 += 4;
+  }
+}
+
+void MergeARGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         const uint16_t* src_a,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width) {
+  assert(depth >= 8);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 8;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = clamp255(src_b[x] >> shift);
+    dst_argb[1] = clamp255(src_g[x] >> shift);
+    dst_argb[2] = clamp255(src_r[x] >> shift);
+    dst_argb[3] = clamp255(src_a[x] >> shift);
+    dst_argb += 4;
+  }
+}
+
+void MergeXR64Row_C(const uint16_t* src_r,
+                    const uint16_t* src_g,
+                    const uint16_t* src_b,
+                    uint16_t* dst_ar64,
+                    int depth,
+                    int width) {
+  assert(depth >= 1);
+  assert(depth <= 16);
+  int x;
+  int shift = 16 - depth;
+  int max = (1 << depth) - 1;
+  for (x = 0; x < width; ++x) {
+    dst_ar64[0] = ClampMax(src_b[x], max) << shift;
+    dst_ar64[1] = ClampMax(src_g[x], max) << shift;
+    dst_ar64[2] = ClampMax(src_r[x], max) << shift;
+    dst_ar64[3] = 0xffff;
+    dst_ar64 += 4;
+  }
+}
+
+void MergeXRGB16To8Row_C(const uint16_t* src_r,
+                         const uint16_t* src_g,
+                         const uint16_t* src_b,
+                         uint8_t* dst_argb,
+                         int depth,
+                         int width) {
+  assert(depth >= 8);
+  assert(depth <= 16);
+  int x;
+  int shift = depth - 8;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = clamp255(src_b[x] >> shift);
+    dst_argb[1] = clamp255(src_g[x] >> shift);
+    dst_argb[2] = clamp255(src_r[x] >> shift);
+    dst_argb[3] = 0xff;
+    dst_argb += 4;
+  }
+}
+
+void SplitXRGBRow_C(const uint8_t* src_argb,
+                    uint8_t* dst_r,
+                    uint8_t* dst_g,
+                    uint8_t* dst_b,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_b[x] = src_argb[0];
+    dst_g[x] = src_argb[1];
+    dst_r[x] = src_argb[2];
+    src_argb += 4;
+  }
+}
+
+void MergeXRGBRow_C(const uint8_t* src_r,
+                    const uint8_t* src_g,
+                    const uint8_t* src_b,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_argb[0] = src_b[x];
+    dst_argb[1] = src_g[x];
+    dst_argb[2] = src_r[x];
+    dst_argb[3] = 255;
+    dst_argb += 4;
+  }
+}
+
+// Convert lsb formats to msb, depending on sample depth.
 void MergeUVRow_16_C(const uint16_t* src_u,
                      const uint16_t* src_v,
                      uint16_t* dst_uv,
-                     int scale,
+                     int depth,
                      int width) {
+  int shift = 16 - depth;
+  assert(depth >= 8);
+  assert(depth <= 16);
   int x;
-  for (x = 0; x < width - 1; x += 2) {
-    dst_uv[0] = src_u[x] * scale;
-    dst_uv[1] = src_v[x] * scale;
-    dst_uv[2] = src_u[x + 1] * scale;
-    dst_uv[3] = src_v[x + 1] * scale;
-    dst_uv += 4;
+  for (x = 0; x < width; ++x) {
+    dst_uv[0] = src_u[x] << shift;
+    dst_uv[1] = src_v[x] << shift;
+    dst_uv += 2;
   }
-  if (width & 1) {
-    dst_uv[0] = src_u[width - 1] * scale;
-    dst_uv[1] = src_v[width - 1] * scale;
+}
+
+// Convert msb formats to lsb, depending on sample depth.
+void SplitUVRow_16_C(const uint16_t* src_uv,
+                     uint16_t* dst_u,
+                     uint16_t* dst_v,
+                     int depth,
+                     int width) {
+  int shift = 16 - depth;
+  int x;
+  assert(depth >= 8);
+  assert(depth <= 16);
+  for (x = 0; x < width; ++x) {
+    dst_u[x] = src_uv[0] >> shift;
+    dst_v[x] = src_uv[1] >> shift;
+    src_uv += 2;
   }
 }
 
@@ -2167,18 +2963,34 @@ void MultiplyRow_16_C(const uint16_t* src_y,
   }
 }
 
+void DivideRow_16_C(const uint16_t* src_y,
+                    uint16_t* dst_y,
+                    int scale,
+                    int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_y[x] = (src_y[x] * scale) >> 16;
+  }
+}
+
 // Use scale to convert lsb formats to msb, depending how many bits there are:
 // 32768 = 9 bits
 // 16384 = 10 bits
 // 4096 = 12 bits
 // 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+#define C16TO8(v, scale) clamp255(((v) * (scale)) >> 16)
+
 void Convert16To8Row_C(const uint16_t* src_y,
                        uint8_t* dst_y,
                        int scale,
                        int width) {
   int x;
+  assert(scale >= 256);
+  assert(scale <= 32768);
+
   for (x = 0; x < width; ++x) {
-    dst_y[x] = clamp255((src_y[x] * scale) >> 16);
+    dst_y[x] = C16TO8(src_y[x], scale);
   }
 }
 
@@ -2208,10 +3020,9 @@ void SetRow_C(uint8_t* dst, uint8_t v8, int width) {
 }
 
 void ARGBSetRow_C(uint8_t* dst_argb, uint32_t v32, int width) {
-  uint32_t* d = (uint32_t*)(dst_argb);
   int x;
   for (x = 0; x < width; ++x) {
-    d[x] = v32;
+    memcpy(dst_argb + x * sizeof v32, &v32, sizeof v32);
   }
 }
 
@@ -2309,21 +3120,21 @@ void UYVYToYRow_C(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
   }
 }
 
-#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+#define BLEND(f, b, a) clamp255((((256 - a) * b) >> 8) + f)
 
-// Blend src_argb0 over src_argb1 and store to dst_argb.
-// dst_argb may be src_argb0 or src_argb1.
+// Blend src_argb over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb or src_argb1.
 // This code mimics the SSSE3 version for better testability.
-void ARGBBlendRow_C(const uint8_t* src_argb0,
+void ARGBBlendRow_C(const uint8_t* src_argb,
                     const uint8_t* src_argb1,
                     uint8_t* dst_argb,
                     int width) {
   int x;
   for (x = 0; x < width - 1; x += 2) {
-    uint32_t fb = src_argb0[0];
-    uint32_t fg = src_argb0[1];
-    uint32_t fr = src_argb0[2];
-    uint32_t a = src_argb0[3];
+    uint32_t fb = src_argb[0];
+    uint32_t fg = src_argb[1];
+    uint32_t fr = src_argb[2];
+    uint32_t a = src_argb[3];
     uint32_t bb = src_argb1[0];
     uint32_t bg = src_argb1[1];
     uint32_t br = src_argb1[2];
@@ -2332,10 +3143,10 @@ void ARGBBlendRow_C(const uint8_t* src_argb0,
     dst_argb[2] = BLEND(fr, br, a);
     dst_argb[3] = 255u;
 
-    fb = src_argb0[4 + 0];
-    fg = src_argb0[4 + 1];
-    fr = src_argb0[4 + 2];
-    a = src_argb0[4 + 3];
+    fb = src_argb[4 + 0];
+    fg = src_argb[4 + 1];
+    fr = src_argb[4 + 2];
+    a = src_argb[4 + 3];
     bb = src_argb1[4 + 0];
     bg = src_argb1[4 + 1];
     br = src_argb1[4 + 2];
@@ -2343,16 +3154,16 @@ void ARGBBlendRow_C(const uint8_t* src_argb0,
     dst_argb[4 + 1] = BLEND(fg, bg, a);
     dst_argb[4 + 2] = BLEND(fr, br, a);
     dst_argb[4 + 3] = 255u;
-    src_argb0 += 8;
+    src_argb += 8;
     src_argb1 += 8;
     dst_argb += 8;
   }
 
   if (width & 1) {
-    uint32_t fb = src_argb0[0];
-    uint32_t fg = src_argb0[1];
-    uint32_t fr = src_argb0[2];
-    uint32_t a = src_argb0[3];
+    uint32_t fb = src_argb[0];
+    uint32_t fg = src_argb[1];
+    uint32_t fr = src_argb[2];
+    uint32_t a = src_argb[3];
     uint32_t bb = src_argb1[0];
     uint32_t bg = src_argb1[1];
     uint32_t br = src_argb1[2];
@@ -2385,10 +3196,14 @@ void BlendPlaneRow_C(const uint8_t* src0,
 }
 #undef UBLEND
 
+#if LIBYUV_ATTENUATE_DUP
+// This code mimics the SSSE3 version for better testability.
 #define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+#else
+#define ATTENUATE(f, a) (f * a + 128) >> 8
+#endif
 
 // Multiply source RGB by alpha and store to destination.
-// This code mimics the SSSE3 version for better testability.
 void ARGBAttenuateRow_C(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   int i;
   for (i = 0; i < width - 1; i += 2) {
@@ -2472,6 +3287,14 @@ const uint32_t fixed_invtbl8[256] = {
     T(0xfc),    T(0xfd),    T(0xfe), 0x01000100};
 #undef T
 
+#if LIBYUV_UNATTENUATE_DUP
+// This code mimics the Intel SIMD version for better testability.
+#define UNATTENUATE(f, ia) clamp255(((f | (f << 8)) * ia) >> 16)
+#else
+#define UNATTENUATE(f, ia) clamp255((f * ia) >> 8)
+#endif
+
+// mimics the Intel SIMD code for exactness.
 void ARGBUnattenuateRow_C(const uint8_t* src_argb,
                           uint8_t* dst_argb,
                           int width) {
@@ -2482,13 +3305,11 @@ void ARGBUnattenuateRow_C(const uint8_t* src_argb,
     uint32_t r = src_argb[2];
     const uint32_t a = src_argb[3];
     const uint32_t ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
-    b = (b * ia) >> 8;
-    g = (g * ia) >> 8;
-    r = (r * ia) >> 8;
+
     // Clamping should not be necessary but is free in assembly.
-    dst_argb[0] = clamp255(b);
-    dst_argb[1] = clamp255(g);
-    dst_argb[2] = clamp255(r);
+    dst_argb[0] = UNATTENUATE(b, ia);
+    dst_argb[1] = UNATTENUATE(g, ia);
+    dst_argb[2] = UNATTENUATE(r, ia);
     dst_argb[3] = a;
     src_argb += 4;
     dst_argb += 4;
@@ -2519,8 +3340,11 @@ void CumulativeSumToAverageRow_C(const int32_t* tl,
                                  int area,
                                  uint8_t* dst,
                                  int count) {
-  float ooa = 1.0f / area;
+  float ooa;
   int i;
+  assert(area != 0);
+
+  ooa = 1.0f / area;
   for (i = 0; i < count; ++i) {
     dst[0] = (uint8_t)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
     dst[1] = (uint8_t)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
@@ -2576,6 +3400,17 @@ static void HalfRow_16_C(const uint16_t* src_uv,
   }
 }
 
+static void HalfRow_16To8_C(const uint16_t* src_uv,
+                            ptrdiff_t src_uv_stride,
+                            uint8_t* dst_uv,
+                            int scale,
+                            int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    dst_uv[x] = C16TO8((src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1, scale);
+  }
+}
+
 // C version 2x2 -> 2x1.
 void InterpolateRow_C(uint8_t* dst_ptr,
                       const uint8_t* src_ptr,
@@ -2586,6 +3421,9 @@ void InterpolateRow_C(uint8_t* dst_ptr,
   int y0_fraction = 256 - y1_fraction;
   const uint8_t* src_ptr1 = src_ptr + src_stride;
   int x;
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
   if (y1_fraction == 0) {
     memcpy(dst_ptr, src_ptr, width);
     return;
@@ -2594,21 +3432,16 @@ void InterpolateRow_C(uint8_t* dst_ptr,
     HalfRow_C(src_ptr, src_stride, dst_ptr, width);
     return;
   }
-  for (x = 0; x < width - 1; x += 2) {
-    dst_ptr[0] =
-        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8;
-    dst_ptr[1] =
-        (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction + 128) >> 8;
-    src_ptr += 2;
-    src_ptr1 += 2;
-    dst_ptr += 2;
-  }
-  if (width & 1) {
+  for (x = 0; x < width; ++x) {
     dst_ptr[0] =
         (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8;
+    ++src_ptr;
+    ++src_ptr1;
+    ++dst_ptr;
   }
 }
 
+// C version 2x2 -> 2x1.
 void InterpolateRow_16_C(uint16_t* dst_ptr,
                          const uint16_t* src_ptr,
                          ptrdiff_t src_stride,
@@ -2618,23 +3451,62 @@ void InterpolateRow_16_C(uint16_t* dst_ptr,
   int y0_fraction = 256 - y1_fraction;
   const uint16_t* src_ptr1 = src_ptr + src_stride;
   int x;
-  if (source_y_fraction == 0) {
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
+  if (y1_fraction == 0) {
     memcpy(dst_ptr, src_ptr, width * 2);
     return;
   }
-  if (source_y_fraction == 128) {
+  if (y1_fraction == 128) {
     HalfRow_16_C(src_ptr, src_stride, dst_ptr, width);
     return;
   }
-  for (x = 0; x < width - 1; x += 2) {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
-    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
-    src_ptr += 2;
-    src_ptr1 += 2;
-    dst_ptr += 2;
+  for (x = 0; x < width; ++x) {
+    dst_ptr[0] =
+        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8;
+    ++src_ptr;
+    ++src_ptr1;
+    ++dst_ptr;
   }
-  if (width & 1) {
-    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+}
+
+// C version 2x2 16 bit-> 2x1 8 bit.
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+
+void InterpolateRow_16To8_C(uint8_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int scale,
+                            int width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  int x;
+  assert(source_y_fraction >= 0);
+  assert(source_y_fraction < 256);
+
+  if (source_y_fraction == 0) {
+    Convert16To8Row_C(src_ptr, dst_ptr, scale, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16To8_C(src_ptr, src_stride, dst_ptr, scale, width);
+    return;
+  }
+  for (x = 0; x < width; ++x) {
+    dst_ptr[0] = C16TO8(
+        (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction + 128) >> 8,
+        scale);
+    src_ptr += 1;
+    src_ptr1 += 1;
+    dst_ptr += 1;
   }
 }
 
@@ -2873,7 +3745,7 @@ void ARGBCopyYToAlphaRow_C(const uint8_t* src, uint8_t* dst, int width) {
 // Maximum temporary width for wrappers to process at a time, in pixels.
 #define MAXTWIDTH 2048
 
-#if !(defined(_MSC_VER) && defined(_M_IX86)) && \
+#if !(defined(_MSC_VER) && !defined(__clang__) && defined(_M_IX86)) && \
     defined(HAS_I422TORGB565ROW_SSSE3)
 // row_win.cc has asm version, but GCC uses 2 step wrapper.
 void I422ToRGB565Row_SSSE3(const uint8_t* src_y,
@@ -3175,12 +4047,93 @@ void NV12ToRGB565Row_AVX2(const uint8_t* src_y,
 }
 #endif
 
+#ifdef HAS_RGB24TOYJROW_AVX2
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_AVX2(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RGB24TOYJROW_AVX2
+
+#ifdef HAS_RAWTOYJROW_AVX2
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_AVX2(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_AVX2(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RAWTOYJROW_AVX2
+
+#ifdef HAS_RGB24TOYJROW_SSSE3
+// Convert 16 RGB24 pixels (64 bytes) to 16 YJ values.
+void RGB24ToYJRow_SSSE3(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RGB24ToARGBRow_SSSE3(src_rgb24, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_rgb24 += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RGB24TOYJROW_SSSE3
+
+#ifdef HAS_RAWTOYJROW_SSSE3
+// Convert 16 RAW pixels (64 bytes) to 16 YJ values.
+void RAWToYJRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8_t row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    RAWToARGBRow_SSSE3(src_raw, row, twidth);
+    ARGBToYJRow_SSSE3(row, dst_yj, twidth);
+    src_raw += twidth * 3;
+    dst_yj += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_RAWTOYJROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_16TO8_AVX2
+void InterpolateRow_16To8_AVX2(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int width,
+                               int source_y_fraction) {
+  // Row buffer for intermediate 16 bit pixels.
+  SIMD_ALIGNED(uint16_t row[MAXTWIDTH]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    InterpolateRow_16_C(row, src_ptr, src_stride, twidth, source_y_fraction);
+    Convert16To8Row_AVX2(row, dst_ptr, scale, twidth);
+    src_ptr += twidth;
+    dst_ptr += twidth;
+    width -= twidth;
+  }
+}
+#endif  // HAS_INTERPOLATEROW_16TO8_AVX2
+
 float ScaleSumSamples_C(const float* src, float* dst, float scale, int width) {
   float fsum = 0.f;
   int i;
-#if defined(__clang__)
-#pragma clang loop vectorize_width(4)
-#endif
   for (i = 0; i < width; ++i) {
     float v = *src++;
     fsum += v * v;
@@ -3231,6 +4184,29 @@ void GaussCol_C(const uint16_t* src0,
   }
 }
 
+void GaussRow_F32_C(const float* src, float* dst, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = (src[0] + src[1] * 4 + src[2] * 6 + src[3] * 4 + src[4]) *
+             (1.0f / 256.0f);
+    ++src;
+  }
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_F32_C(const float* src0,
+                    const float* src1,
+                    const float* src2,
+                    const float* src3,
+                    const float* src4,
+                    float* dst,
+                    int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    *dst++ = *src0++ + *src1++ * 4 + *src2++ * 6 + *src3++ * 4 + *src4++;
+  }
+}
+
 // Convert biplanar NV21 to packed YUV24
 void NV21ToYUV24Row_C(const uint8_t* src_y,
                       const uint8_t* src_vu,
@@ -3256,13 +4232,14 @@ void NV21ToYUV24Row_C(const uint8_t* src_y,
 }
 
 // Filter 2 rows of AYUV UV's (444) into UV (420).
+// AYUV is VUYA in memory.  UV for NV12 is UV order in memory.
 void AYUVToUVRow_C(const uint8_t* src_ayuv,
                    int src_stride_ayuv,
                    uint8_t* dst_uv,
                    int width) {
   // Output a row of UV values, filtering 2x2 rows of AYUV.
   int x;
-  for (x = 0; x < width; x += 2) {
+  for (x = 0; x < width - 1; x += 2) {
     dst_uv[0] = (src_ayuv[1] + src_ayuv[5] + src_ayuv[src_stride_ayuv + 1] +
                  src_ayuv[src_stride_ayuv + 5] + 2) >>
                 2;
@@ -3273,12 +4250,8 @@ void AYUVToUVRow_C(const uint8_t* src_ayuv,
     dst_uv += 2;
   }
   if (width & 1) {
-    dst_uv[0] = (src_ayuv[0] + src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] +
-                 src_ayuv[src_stride_ayuv + 0] + 2) >>
-                2;
-    dst_uv[1] = (src_ayuv[1] + src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] +
-                 src_ayuv[src_stride_ayuv + 1] + 2) >>
-                2;
+    dst_uv[0] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1;
+    dst_uv[1] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1;
   }
 }
 
@@ -3289,7 +4262,7 @@ void AYUVToVURow_C(const uint8_t* src_ayuv,
                    int width) {
   // Output a row of VU values, filtering 2x2 rows of AYUV.
   int x;
-  for (x = 0; x < width; x += 2) {
+  for (x = 0; x < width - 1; x += 2) {
     dst_vu[0] = (src_ayuv[0] + src_ayuv[4] + src_ayuv[src_stride_ayuv + 0] +
                  src_ayuv[src_stride_ayuv + 4] + 2) >>
                 2;
@@ -3300,12 +4273,8 @@ void AYUVToVURow_C(const uint8_t* src_ayuv,
     dst_vu += 2;
   }
   if (width & 1) {
-    dst_vu[0] = (src_ayuv[0] + src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] +
-                 src_ayuv[src_stride_ayuv + 0] + 2) >>
-                2;
-    dst_vu[1] = (src_ayuv[1] + src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] +
-                 src_ayuv[src_stride_ayuv + 1] + 2) >>
-                2;
+    dst_vu[0] = (src_ayuv[0] + src_ayuv[src_stride_ayuv + 0] + 1) >> 1;
+    dst_vu[1] = (src_ayuv[1] + src_ayuv[src_stride_ayuv + 1] + 1) >> 1;
   }
 }
 
@@ -3319,7 +4288,8 @@ void AYUVToYRow_C(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
   }
 }
 
-void UVToVURow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
   int x;
   for (x = 0; x < width; ++x) {
     uint8_t u = src_uv[0];
@@ -3331,16 +4301,27 @@ void UVToVURow_C(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
   }
 }
 
-// divide values by weights and provide mask to indicate weight of 0.
-void FloatDivToByteRow_C(const float* src_weights,
-                         const float* src_values,
-                         uint8_t* dst_out,
-                         uint8_t* dst_mask,
-                         int width) {
+void HalfMergeUVRow_C(const uint8_t* src_u,
+                      int src_stride_u,
+                      const uint8_t* src_v,
+                      int src_stride_v,
+                      uint8_t* dst_uv,
+                      int width) {
   int x;
-  for (x = 0; x < width; ++x) {
-    dst_out[x] = Clamp(src_values[x] / src_weights[x]);
-    dst_mask[x] = src_weights[x] > 0 ? 0 : 0xff;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = (src_u[0] + src_u[1] + src_u[src_stride_u] +
+                 src_u[src_stride_u + 1] + 2) >>
+                2;
+    dst_uv[1] = (src_v[0] + src_v[1] + src_v[src_stride_v] +
+                 src_v[src_stride_v + 1] + 2) >>
+                2;
+    src_u += 2;
+    src_v += 2;
+    dst_uv += 2;
+  }
+  if (width & 1) {
+    dst_uv[0] = (src_u[0] + src_u[src_stride_u] + 1) >> 1;
+    dst_uv[1] = (src_v[0] + src_v[src_stride_v] + 1) >> 1;
   }
 }
 
diff --git a/files/source/row_dspr2.cc b/files/source/row_dspr2.cc
deleted file mode 100644
index 11f78e0d..00000000
--- a/files/source/row_dspr2.cc
+++ /dev/null
@@ -1,1721 +0,0 @@
-/*
- *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// The following are available on Mips platforms:
-#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips__) && \
-    (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-#ifdef HAS_COPYROW_MIPS
-void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
-  __asm__ __volatile__(
-      ".set      noreorder                         \n"
-      ".set      noat                              \n"
-      "slti      $at, %[count], 8                  \n"
-      "bne       $at ,$zero, $last8                \n"
-      "xor       $t8, %[src], %[dst]               \n"
-      "andi      $t8, $t8, 0x3                     \n"
-
-      "bne       $t8, $zero, unaligned             \n"
-      "negu      $a3, %[dst]                       \n"
-      // make dst/src aligned
-      "andi      $a3, $a3, 0x3                     \n"
-      "beq       $a3, $zero, $chk16w               \n"
-      // word-aligned now count is the remining bytes count
-      "subu     %[count], %[count], $a3            \n"
-
-      "lwr       $t8, 0(%[src])                    \n"
-      "addu      %[src], %[src], $a3               \n"
-      "swr       $t8, 0(%[dst])                    \n"
-      "addu      %[dst], %[dst], $a3               \n"
-
-      // Now the dst/src are mutually word-aligned with word-aligned addresses
-      "$chk16w:                                    \n"
-      "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
-      // t8 is the byte count after 64-byte chunks
-      "beq       %[count], $t8, chk8w              \n"
-      // There will be at most 1 32-byte chunk after it
-      "subu      $a3, %[count], $t8                \n"  // the reminder
-      // Here a3 counts bytes in 16w chunks
-      "addu      $a3, %[dst], $a3                  \n"
-      // Now a3 is the final dst after 64-byte chunks
-      "addu      $t0, %[dst], %[count]             \n"
-      // t0 is the "past the end" address
-
-      // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be
-      // past
-      // the "t0-32" address
-      // This means: for x=128 the last "safe" a1 address is "t0-160"
-      // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
-      // we will use "pref 30,128(a1)", so "t0-160" is the limit
-      "subu      $t9, $t0, 160                     \n"
-      // t9 is the "last safe pref 30,128(a1)" address
-      "pref      0, 0(%[src])                      \n"  // first line of src
-      "pref      0, 32(%[src])                     \n"  // second line of src
-      "pref      0, 64(%[src])                     \n"
-      "pref      30, 32(%[dst])                    \n"
-      // In case the a1 > t9 don't use "pref 30" at all
-      "sltu      $v1, $t9, %[dst]                  \n"
-      "bgtz      $v1, $loop16w                     \n"
-      "nop                                         \n"
-      // otherwise, start with using pref30
-      "pref      30, 64(%[dst])                    \n"
-      "$loop16w:                                    \n"
-      "pref      0, 96(%[src])                     \n"
-      "lw        $t0, 0(%[src])                    \n"
-      "bgtz      $v1, $skip_pref30_96              \n"  // skip
-      "lw        $t1, 4(%[src])                    \n"
-      "pref      30, 96(%[dst])                    \n"  // continue
-      "$skip_pref30_96:                            \n"
-      "lw        $t2, 8(%[src])                    \n"
-      "lw        $t3, 12(%[src])                   \n"
-      "lw        $t4, 16(%[src])                   \n"
-      "lw        $t5, 20(%[src])                   \n"
-      "lw        $t6, 24(%[src])                   \n"
-      "lw        $t7, 28(%[src])                   \n"
-      "pref      0, 128(%[src])                    \n"
-      //  bring the next lines of src, addr 128
-      "sw        $t0, 0(%[dst])                    \n"
-      "sw        $t1, 4(%[dst])                    \n"
-      "sw        $t2, 8(%[dst])                    \n"
-      "sw        $t3, 12(%[dst])                   \n"
-      "sw        $t4, 16(%[dst])                   \n"
-      "sw        $t5, 20(%[dst])                   \n"
-      "sw        $t6, 24(%[dst])                   \n"
-      "sw        $t7, 28(%[dst])                   \n"
-      "lw        $t0, 32(%[src])                   \n"
-      "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
-      "lw        $t1, 36(%[src])                   \n"
-      "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
-      "$skip_pref30_128:                           \n"
-      "lw        $t2, 40(%[src])                   \n"
-      "lw        $t3, 44(%[src])                   \n"
-      "lw        $t4, 48(%[src])                   \n"
-      "lw        $t5, 52(%[src])                   \n"
-      "lw        $t6, 56(%[src])                   \n"
-      "lw        $t7, 60(%[src])                   \n"
-      "pref      0, 160(%[src])                    \n"
-      // bring the next lines of src, addr 160
-      "sw        $t0, 32(%[dst])                   \n"
-      "sw        $t1, 36(%[dst])                   \n"
-      "sw        $t2, 40(%[dst])                   \n"
-      "sw        $t3, 44(%[dst])                   \n"
-      "sw        $t4, 48(%[dst])                   \n"
-      "sw        $t5, 52(%[dst])                   \n"
-      "sw        $t6, 56(%[dst])                   \n"
-      "sw        $t7, 60(%[dst])                   \n"
-
-      "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
-      "sltu      $v1, $t9, %[dst]                  \n"
-      "bne       %[dst], $a3, $loop16w             \n"
-      " addiu    %[src], %[src], 64                \n"  // adding 64 to src
-      "move      %[count], $t8                     \n"
-
-      // Here we have src and dest word-aligned but less than 64-bytes to go
-
-      "chk8w:                                      \n"
-      "pref      0, 0x0(%[src])                    \n"
-      "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
-      // the t8 is the reminder count past 32-bytes
-      "beq       %[count], $t8, chk1w              \n"
-      // count=t8,no 32-byte chunk
-      " nop                                        \n"
-
-      "lw        $t0, 0(%[src])                    \n"
-      "lw        $t1, 4(%[src])                    \n"
-      "lw        $t2, 8(%[src])                    \n"
-      "lw        $t3, 12(%[src])                   \n"
-      "lw        $t4, 16(%[src])                   \n"
-      "lw        $t5, 20(%[src])                   \n"
-      "lw        $t6, 24(%[src])                   \n"
-      "lw        $t7, 28(%[src])                   \n"
-      "addiu     %[src], %[src], 32                \n"
-
-      "sw        $t0, 0(%[dst])                    \n"
-      "sw        $t1, 4(%[dst])                    \n"
-      "sw        $t2, 8(%[dst])                    \n"
-      "sw        $t3, 12(%[dst])                   \n"
-      "sw        $t4, 16(%[dst])                   \n"
-      "sw        $t5, 20(%[dst])                   \n"
-      "sw        $t6, 24(%[dst])                   \n"
-      "sw        $t7, 28(%[dst])                   \n"
-      "addiu     %[dst], %[dst], 32                \n"
-
-      "chk1w:                                      \n"
-      "andi      %[count], $t8, 0x3                \n"
-      // now count is the reminder past 1w chunks
-      "beq       %[count], $t8, $last8             \n"
-      " subu     $a3, $t8, %[count]                \n"
-      // a3 is count of bytes in 1w chunks
-      "addu      $a3, %[dst], $a3                  \n"
-      // now a3 is the dst address past the 1w chunks
-      // copying in words (4-byte chunks)
-      "$wordCopy_loop:                             \n"
-      "lw        $t3, 0(%[src])                    \n"
-      // the first t3 may be equal t0 ... optimize?
-      "addiu     %[src], %[src],4                  \n"
-      "addiu     %[dst], %[dst],4                  \n"
-      "bne       %[dst], $a3,$wordCopy_loop        \n"
-      " sw       $t3, -4(%[dst])                   \n"
-
-      // For the last (<8) bytes
-      "$last8:                                     \n"
-      "blez      %[count], leave                   \n"
-      " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
-      "$last8loop:                                 \n"
-      "lb        $v1, 0(%[src])                    \n"
-      "addiu     %[src], %[src], 1                 \n"
-      "addiu     %[dst], %[dst], 1                 \n"
-      "bne       %[dst], $a3, $last8loop           \n"
-      " sb       $v1, -1(%[dst])                   \n"
-
-      "leave:                                      \n"
-      "  j       $ra                               \n"
-      "  nop                                       \n"
-
-      //
-      // UNALIGNED case
-      //
-
-      "unaligned:                                  \n"
-      // got here with a3="negu a1"
-      "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
-      "beqz      $a3, $ua_chk16w                   \n"
-      " subu     %[count], %[count], $a3           \n"
-      // bytes left after initial a3 bytes
-      "lwr       $v1, 0(%[src])                    \n"
-      "lwl       $v1, 3(%[src])                    \n"
-      "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
-      "swr       $v1, 0(%[dst])                    \n"
-      "addu      %[dst], %[dst], $a3               \n"
-      // below the dst will be word aligned (NOTE1)
-      "$ua_chk16w:                                 \n"
-      "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
-      // t8 is the byte count after 64-byte chunks
-      "beq       %[count], $t8, ua_chk8w           \n"
-      // if a2==t8, no 64-byte chunks
-      // There will be at most 1 32-byte chunk after it
-      "subu      $a3, %[count], $t8                \n"  // the reminder
-      // Here a3 counts bytes in 16w chunks
-      "addu      $a3, %[dst], $a3                  \n"
-      // Now a3 is the final dst after 64-byte chunks
-      "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
-      "subu      $t9, $t0, 160                     \n"
-      // t9 is the "last safe pref 30,128(a1)" address
-      "pref      0, 0(%[src])                      \n"  // first line of src
-      "pref      0, 32(%[src])                     \n"  // second line  addr 32
-      "pref      0, 64(%[src])                     \n"
-      "pref      30, 32(%[dst])                    \n"
-      // safe, as we have at least 64 bytes ahead
-      // In case the a1 > t9 don't use "pref 30" at all
-      "sltu      $v1, $t9, %[dst]                  \n"
-      "bgtz      $v1, $ua_loop16w                  \n"
-      // skip "pref 30,64(a1)" for too short arrays
-      " nop                                        \n"
-      // otherwise, start with using pref30
-      "pref      30, 64(%[dst])                    \n"
-      "$ua_loop16w:                                \n"
-      "pref      0, 96(%[src])                     \n"
-      "lwr       $t0, 0(%[src])                    \n"
-      "lwl       $t0, 3(%[src])                    \n"
-      "lwr       $t1, 4(%[src])                    \n"
-      "bgtz      $v1, $ua_skip_pref30_96           \n"
-      " lwl      $t1, 7(%[src])                    \n"
-      "pref      30, 96(%[dst])                    \n"
-      // continue setting up the dest, addr 96
-      "$ua_skip_pref30_96:                         \n"
-      "lwr       $t2, 8(%[src])                    \n"
-      "lwl       $t2, 11(%[src])                   \n"
-      "lwr       $t3, 12(%[src])                   \n"
-      "lwl       $t3, 15(%[src])                   \n"
-      "lwr       $t4, 16(%[src])                   \n"
-      "lwl       $t4, 19(%[src])                   \n"
-      "lwr       $t5, 20(%[src])                   \n"
-      "lwl       $t5, 23(%[src])                   \n"
-      "lwr       $t6, 24(%[src])                   \n"
-      "lwl       $t6, 27(%[src])                   \n"
-      "lwr       $t7, 28(%[src])                   \n"
-      "lwl       $t7, 31(%[src])                   \n"
-      "pref      0, 128(%[src])                    \n"
-      // bring the next lines of src, addr 128
-      "sw        $t0, 0(%[dst])                    \n"
-      "sw        $t1, 4(%[dst])                    \n"
-      "sw        $t2, 8(%[dst])                    \n"
-      "sw        $t3, 12(%[dst])                   \n"
-      "sw        $t4, 16(%[dst])                   \n"
-      "sw        $t5, 20(%[dst])                   \n"
-      "sw        $t6, 24(%[dst])                   \n"
-      "sw        $t7, 28(%[dst])                   \n"
-      "lwr       $t0, 32(%[src])                   \n"
-      "lwl       $t0, 35(%[src])                   \n"
-      "lwr       $t1, 36(%[src])                   \n"
-      "bgtz      $v1, ua_skip_pref30_128           \n"
-      " lwl      $t1, 39(%[src])                   \n"
-      "pref      30, 128(%[dst])                   \n"
-      // continue setting up the dest, addr 128
-      "ua_skip_pref30_128:                         \n"
-
-      "lwr       $t2, 40(%[src])                   \n"
-      "lwl       $t2, 43(%[src])                   \n"
-      "lwr       $t3, 44(%[src])                   \n"
-      "lwl       $t3, 47(%[src])                   \n"
-      "lwr       $t4, 48(%[src])                   \n"
-      "lwl       $t4, 51(%[src])                   \n"
-      "lwr       $t5, 52(%[src])                   \n"
-      "lwl       $t5, 55(%[src])                   \n"
-      "lwr       $t6, 56(%[src])                   \n"
-      "lwl       $t6, 59(%[src])                   \n"
-      "lwr       $t7, 60(%[src])                   \n"
-      "lwl       $t7, 63(%[src])                   \n"
-      "pref      0, 160(%[src])                    \n"
-      // bring the next lines of src, addr 160
-      "sw        $t0, 32(%[dst])                   \n"
-      "sw        $t1, 36(%[dst])                   \n"
-      "sw        $t2, 40(%[dst])                   \n"
-      "sw        $t3, 44(%[dst])                   \n"
-      "sw        $t4, 48(%[dst])                   \n"
-      "sw        $t5, 52(%[dst])                   \n"
-      "sw        $t6, 56(%[dst])                   \n"
-      "sw        $t7, 60(%[dst])                   \n"
-
-      "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
-      "sltu      $v1,$t9,%[dst]                    \n"
-      "bne       %[dst],$a3,$ua_loop16w            \n"
-      " addiu    %[src],%[src],64                  \n"  // adding 64 to src
-      "move      %[count],$t8                      \n"
-
-      // Here we have src and dest word-aligned but less than 64-bytes to go
-
-      "ua_chk8w:                                   \n"
-      "pref      0, 0x0(%[src])                    \n"
-      "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
-      // the t8 is the reminder count
-      "beq       %[count], $t8, $ua_chk1w          \n"
-      // when count==t8, no 32-byte chunk
-
-      "lwr       $t0, 0(%[src])                    \n"
-      "lwl       $t0, 3(%[src])                    \n"
-      "lwr       $t1, 4(%[src])                    \n"
-      "lwl       $t1, 7(%[src])                    \n"
-      "lwr       $t2, 8(%[src])                    \n"
-      "lwl       $t2, 11(%[src])                   \n"
-      "lwr       $t3, 12(%[src])                   \n"
-      "lwl       $t3, 15(%[src])                   \n"
-      "lwr       $t4, 16(%[src])                   \n"
-      "lwl       $t4, 19(%[src])                   \n"
-      "lwr       $t5, 20(%[src])                   \n"
-      "lwl       $t5, 23(%[src])                   \n"
-      "lwr       $t6, 24(%[src])                   \n"
-      "lwl       $t6, 27(%[src])                   \n"
-      "lwr       $t7, 28(%[src])                   \n"
-      "lwl       $t7, 31(%[src])                   \n"
-      "addiu     %[src], %[src], 32                \n"
-
-      "sw        $t0, 0(%[dst])                    \n"
-      "sw        $t1, 4(%[dst])                    \n"
-      "sw        $t2, 8(%[dst])                    \n"
-      "sw        $t3, 12(%[dst])                   \n"
-      "sw        $t4, 16(%[dst])                   \n"
-      "sw        $t5, 20(%[dst])                   \n"
-      "sw        $t6, 24(%[dst])                   \n"
-      "sw        $t7, 28(%[dst])                   \n"
-      "addiu     %[dst], %[dst], 32                \n"
-
-      "$ua_chk1w:                                  \n"
-      "andi      %[count], $t8, 0x3                \n"
-      // now count is the reminder past 1w chunks
-      "beq       %[count], $t8, ua_smallCopy       \n"
-      "subu      $a3, $t8, %[count]                \n"
-      // a3 is count of bytes in 1w chunks
-      "addu      $a3, %[dst], $a3                  \n"
-      // now a3 is the dst address past the 1w chunks
-
-      // copying in words (4-byte chunks)
-      "$ua_wordCopy_loop:                          \n"
-      "lwr       $v1, 0(%[src])                    \n"
-      "lwl       $v1, 3(%[src])                    \n"
-      "addiu     %[src], %[src], 4                 \n"
-      "addiu     %[dst], %[dst], 4                 \n"
-      // note: dst=a1 is word aligned here, see NOTE1
-      "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
-      " sw       $v1,-4(%[dst])                    \n"
-
-      // Now less than 4 bytes (value in count) left to copy
-      "ua_smallCopy:                               \n"
-      "beqz      %[count], leave                   \n"
-      " addu     $a3, %[dst], %[count]             \n"  // a3 = last dst address
-      "$ua_smallCopy_loop:                         \n"
-      "lb        $v1, 0(%[src])                    \n"
-      "addiu     %[src], %[src], 1                 \n"
-      "addiu     %[dst], %[dst], 1                 \n"
-      "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
-      " sb       $v1, -1(%[dst])                   \n"
-
-      "j         $ra                               \n"
-      " nop                                        \n"
-      ".set      at                                \n"
-      ".set      reorder                           \n"
-      : [dst] "+r"(dst), [src] "+r"(src)
-      : [count] "r"(count)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9", "a3", "v1",
-        "at");
-}
-#endif  // HAS_COPYROW_MIPS
-
-// DSPR2 functions
-#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips_dsp) &&   \
-    (__mips_dsp_rev >= 2) && (_MIPS_SIM == _MIPS_SIM_ABI32) && \
-    (__mips_isa_rev < 6)
-
-void SplitUVRow_DSPR2(const uint8* src_uv,
-                      uint8* dst_u,
-                      uint8* dst_v,
-                      int width) {
-  __asm__ __volatile__(
-      ".set push                                     \n"
-      ".set noreorder                                \n"
-      "srl             $t4, %[width], 4              \n"  // multiplies of 16
-      "blez            $t4, 2f                       \n"
-      " andi           %[width], %[width], 0xf       \n"  // residual
-
-      "1:                                            \n"
-      "addiu           $t4, $t4, -1                  \n"
-      "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
-      "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
-      "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
-      "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
-      "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
-      "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 |
-                                                          // U10
-      "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 |
-                                                          // U12
-      "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 |
-                                                          // U14
-      "addiu           %[src_uv], %[src_uv], 32      \n"
-      "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
-      "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
-      "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
-      "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
-      "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
-      "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
-      "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 |
-                                                          // V12
-      "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 |
-                                                          // U12
-      "sw              $t9, 0(%[dst_v])              \n"
-      "sw              $t0, 0(%[dst_u])              \n"
-      "sw              $t1, 4(%[dst_v])              \n"
-      "sw              $t2, 4(%[dst_u])              \n"
-      "sw              $t3, 8(%[dst_v])              \n"
-      "sw              $t5, 8(%[dst_u])              \n"
-      "sw              $t6, 12(%[dst_v])             \n"
-      "sw              $t7, 12(%[dst_u])             \n"
-      "addiu           %[dst_v], %[dst_v], 16        \n"
-      "bgtz            $t4, 1b                       \n"
-      " addiu          %[dst_u], %[dst_u], 16        \n"
-
-      "beqz            %[width], 3f                  \n"
-      " nop                                          \n"
-
-      "2:                                              \n"
-      "lbu             $t0, 0(%[src_uv])             \n"
-      "lbu             $t1, 1(%[src_uv])             \n"
-      "addiu           %[src_uv], %[src_uv], 2       \n"
-      "addiu           %[width], %[width], -1        \n"
-      "sb              $t0, 0(%[dst_u])              \n"
-      "sb              $t1, 0(%[dst_v])              \n"
-      "addiu           %[dst_u], %[dst_u], 1         \n"
-      "bgtz            %[width], 2b                  \n"
-      " addiu          %[dst_v], %[dst_v], 1         \n"
-
-      "3:                                              \n"
-      ".set pop                                      \n"
-      : [src_uv] "+r"(src_uv), [width] "+r"(width), [dst_u] "+r"(dst_u),
-        [dst_v] "+r"(dst_v)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void MirrorRow_DSPR2(const uint8* src, uint8* dst, int width) {
-  __asm__ __volatile__(
-      ".set push                             \n"
-      ".set noreorder                        \n"
-
-      "srl       $t4, %[width], 4            \n"  // multiplies of 16
-      "andi      $t5, %[width], 0xf          \n"
-      "blez      $t4, 2f                     \n"
-      " addu     %[src], %[src], %[width]    \n"  // src += width
-
-      "1:                                     \n"
-      "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
-      "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
-      "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
-      "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
-      "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
-      "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
-      "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
-      "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
-      "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
-      "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
-      "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
-      "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
-      "addiu     %[src], %[src], -16         \n"
-      "addiu     $t4, $t4, -1                \n"
-      "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
-      "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
-      "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
-      "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
-      "bgtz      $t4, 1b                     \n"
-      " addiu    %[dst], %[dst], 16          \n"
-      "beqz      $t5, 3f                     \n"
-      " nop                                  \n"
-
-      "2:                                     \n"
-      "lbu       $t0, -1(%[src])             \n"
-      "addiu     $t5, $t5, -1                \n"
-      "addiu     %[src], %[src], -1          \n"
-      "sb        $t0, 0(%[dst])              \n"
-      "bgez      $t5, 2b                     \n"
-      " addiu    %[dst], %[dst], 1           \n"
-
-      "3:                                     \n"
-      ".set pop                              \n"
-      : [src] "+r"(src), [dst] "+r"(dst)
-      : [width] "r"(width)
-      : "t0", "t1", "t2", "t3", "t4", "t5");
-}
-
-void MirrorUVRow_DSPR2(const uint8* src_uv,
-                       uint8* dst_u,
-                       uint8* dst_v,
-                       int width) {
-  int x;
-  int y;
-  __asm__ __volatile__(
-      ".set push                                    \n"
-      ".set noreorder                               \n"
-
-      "addu            $t4, %[width], %[width]      \n"
-      "srl             %[x], %[width], 4            \n"
-      "andi            %[y], %[width], 0xf          \n"
-      "blez            %[x], 2f                     \n"
-      " addu           %[src_uv], %[src_uv], $t4    \n"
-
-      "1:                                          \n"
-      "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
-      "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
-      "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
-      "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
-      "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
-      "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
-      "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
-      "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
-
-      "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
-      "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
-      "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
-      "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
-      "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
-      "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
-      "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
-      "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
-      "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
-      "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
-      "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
-      "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
-      "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
-      "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
-      "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
-      "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
-      "addiu           %[src_uv], %[src_uv], -32    \n"
-      "addiu           %[x], %[x], -1               \n"
-      "swr             $t4, 0(%[dst_u])             \n"
-      "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
-      "swr             $t6, 0(%[dst_v])             \n"
-      "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
-      "swr             $t2, 4(%[dst_u])             \n"
-      "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
-      "swr             $t3, 4(%[dst_v])             \n"
-      "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
-      "swr             $t0, 8(%[dst_u])             \n"
-      "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
-      "swr             $t1, 8(%[dst_v])             \n"
-      "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
-      "swr             $t9, 12(%[dst_u])            \n"
-      "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
-      "swr             $t5, 12(%[dst_v])            \n"
-      "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
-      "addiu           %[dst_v], %[dst_v], 16       \n"
-      "bgtz            %[x], 1b                     \n"
-      " addiu          %[dst_u], %[dst_u], 16       \n"
-      "beqz            %[y], 3f                     \n"
-      " nop                                         \n"
-      "b               2f                           \n"
-      " nop                                         \n"
-
-      "2:                                            \n"
-      "lbu             $t0, -2(%[src_uv])           \n"
-      "lbu             $t1, -1(%[src_uv])           \n"
-      "addiu           %[src_uv], %[src_uv], -2     \n"
-      "addiu           %[y], %[y], -1               \n"
-      "sb              $t0, 0(%[dst_u])             \n"
-      "sb              $t1, 0(%[dst_v])             \n"
-      "addiu           %[dst_u], %[dst_u], 1        \n"
-      "bgtz            %[y], 2b                     \n"
-      " addiu          %[dst_v], %[dst_v], 1        \n"
-
-      "3:                                            \n"
-      ".set pop                                     \n"
-      : [src_uv] "+r"(src_uv), [dst_u] "+r"(dst_u), [dst_v] "+r"(dst_v),
-        [x] "=&r"(x), [y] "=&r"(y)
-      : [width] "r"(width)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t7", "t8", "t9");
-}
-
-void I422ToARGBRow_DSPR2(const uint8* src_y,
-                         const uint8* src_u,
-                         const uint8* src_v,
-                         uint8* rgb_buf,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
-  int x;
-  uint32 tmp_ub = yuvconstants->kUVToB[0];
-  uint32 tmp_ug = yuvconstants->kUVToG[0];
-  uint32 tmp_vg = yuvconstants->kUVToG[1];
-  uint32 tmp_vr = yuvconstants->kUVToR[1];
-  uint32 tmp_bb = yuvconstants->kUVBiasB[0];
-  uint32 tmp_bg = yuvconstants->kUVBiasG[0];
-  uint32 tmp_br = yuvconstants->kUVBiasR[0];
-  uint32 yg = yuvconstants->kYToRgb[0];
-  uint32 tmp_yg;
-  uint32 tmp_mask = 0x7fff7fff;
-  tmp_bb = ((uint)(tmp_bb & 0xffff) << 16) | (tmp_bb & 0xffff);
-  tmp_bg = ((uint)(tmp_bg & 0xffff) << 16) | (tmp_bg & 0xffff);
-  tmp_br = ((uint)(tmp_br & 0xffff) << 16) | (tmp_br & 0xffff);
-  tmp_yg = ((uint)(yg & 0xffff) << 16) | (yg & 0xffff);
-  tmp_ub = ~(((uint)(tmp_ub & 0xffff) << 16) | (tmp_ub & 0xffff)) + 0x00010001;
-  tmp_ug = ((uint)(tmp_ug & 0xffff) << 16) | (tmp_ug & 0xffff);
-  tmp_vg = ((uint)(tmp_vg & 0xffff) << 16) | (tmp_vg & 0xffff);
-  tmp_vr = ~(((uint)(tmp_vr & 0xffff) << 16) | (tmp_vr & 0xffff)) + 0x00010001;
-  yg = yg * 0x0101;
-
-  for (x = 0; x < width - 1; x += 2) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    uint32 tmp_t6, tmp_t7, tmp_t8, tmp_t9;
-    __asm__ __volatile__(
-        ".set push                                             \n"
-        ".set noreorder                                        \n"
-        "lbu              %[tmp_t7], 0(%[src_y])               \n"
-        "lbu              %[tmp_t1], 1(%[src_y])               \n"
-        "mul              %[tmp_t7], %[tmp_t7],     %[yg]      \n"
-        "mul              %[tmp_t1], %[tmp_t1],     %[yg]      \n"
-        "lbu              %[tmp_t2], 0(%[src_u])               \n"
-        "lbu              %[tmp_t3], 0(%[src_v])               \n"
-        "replv.ph         %[tmp_t2], %[tmp_t2]                 \n"
-        "replv.ph         %[tmp_t3], %[tmp_t3]                 \n"
-        "mul.ph           %[tmp_t4], %[tmp_t2],     %[tmp_ub]  \n"
-        "mul.ph           %[tmp_t5], %[tmp_t2],     %[tmp_ug]  \n"
-        "mul.ph           %[tmp_t6], %[tmp_t3],     %[tmp_vr]  \n"
-        "mul.ph           %[tmp_t3], %[tmp_t3],     %[tmp_vg]  \n"
-        "srl              %[tmp_t7], %[tmp_t7],     16         \n"
-        "ins              %[tmp_t1], %[tmp_t7],     0,      16 \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t1],     %[tmp_bb]  \n"
-        "addq_s.ph        %[tmp_t8], %[tmp_t1],     %[tmp_bg]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t1],     %[tmp_br]  \n"
-        "addq_s.ph        %[tmp_t5], %[tmp_t5],     %[tmp_t3]  \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t7],     %[tmp_t4]  \n"
-        "subq_s.ph        %[tmp_t8], %[tmp_t8],     %[tmp_t5]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t9],     %[tmp_t6]  \n"
-        "shra.ph          %[tmp_t7], %[tmp_t7],     6          \n"
-        "shra.ph          %[tmp_t8], %[tmp_t8],     6          \n"
-        "shra.ph          %[tmp_t9], %[tmp_t9],     6          \n"
-        "shll_s.ph        %[tmp_t7], %[tmp_t7],     7          \n"
-        "shll_s.ph        %[tmp_t8], %[tmp_t8],     7          \n"
-        "shll_s.ph        %[tmp_t9], %[tmp_t9],     7          \n"
-        "precrqu_s.qb.ph  %[tmp_t8], %[tmp_mask],   %[tmp_t8]  \n"
-        "precrqu_s.qb.ph  %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.ph.w      %[tmp_t9], %[tmp_t8],     %[tmp_t7]  \n"
-        "ins              %[tmp_t7], %[tmp_t8],     16,     16 \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.qb.ph     %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "sw               %[tmp_t8], 0(%[rgb_buf])             \n"
-        "sw               %[tmp_t7], 4(%[rgb_buf])             \n"
-        ".set pop                                              \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [tmp_t9] "=&r"(tmp_t9)
-        : [src_y] "r"(src_y), [src_u] "r"(src_u), [src_v] "r"(src_v),
-          [tmp_ub] "r"(tmp_ub), [tmp_ug] "r"(tmp_ug), [yg] "r"(yg),
-          [tmp_vg] "r"(tmp_vg), [tmp_vr] "r"(tmp_vr), [tmp_bb] "r"(tmp_bb),
-          [tmp_bg] "r"(tmp_bg), [tmp_br] "r"(tmp_br), [tmp_yg] "r"(tmp_yg),
-          [rgb_buf] "r"(rgb_buf), [tmp_mask] "r"(tmp_mask));
-    src_y += 2;
-    src_u += 1;
-    src_v += 1;
-    rgb_buf += 8;  // Advance 4 pixels.
-  }
-}
-
-// Bilinear filter 8x2 -> 8x1
-void InterpolateRow_DSPR2(uint8* dst_ptr,
-                          const uint8* src_ptr,
-                          ptrdiff_t src_stride,
-                          int dst_width,
-                          int source_y_fraction) {
-  int y0_fraction = 256 - source_y_fraction;
-  const uint8* src_ptr1 = src_ptr + src_stride;
-
-  __asm__ __volatile__(
-      ".set push                                           \n"
-      ".set noreorder                                      \n"
-
-      "replv.ph          $t0, %[y0_fraction]               \n"
-      "replv.ph          $t1, %[source_y_fraction]         \n"
-
-      "1:                                                    \n"
-      "lw                $t2, 0(%[src_ptr])                \n"
-      "lw                $t3, 0(%[src_ptr1])               \n"
-      "lw                $t4, 4(%[src_ptr])                \n"
-      "lw                $t5, 4(%[src_ptr1])               \n"
-      "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
-      "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
-      "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
-      "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
-      "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
-      "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
-      "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
-      "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
-      "addq.ph           $t6, $t6, $t8                     \n"
-      "addq.ph           $t7, $t7, $t9                     \n"
-      "addq.ph           $t2, $t2, $t4                     \n"
-      "addq.ph           $t3, $t3, $t5                     \n"
-      "shra_r.ph         $t6, $t6, 8                       \n"
-      "shra_r.ph         $t7, $t7, 8                       \n"
-      "shra_r.ph         $t2, $t2, 8                       \n"
-      "shra_r.ph         $t3, $t3, 8                       \n"
-      "precr.qb.ph       $t6, $t6, $t7                     \n"
-      "precr.qb.ph       $t2, $t2, $t3                     \n"
-      "addiu             %[src_ptr], %[src_ptr], 8         \n"
-      "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
-      "addiu             %[dst_width], %[dst_width], -8    \n"
-      "sw                $t6, 0(%[dst_ptr])                \n"
-      "sw                $t2, 4(%[dst_ptr])                \n"
-      "bgtz              %[dst_width], 1b                  \n"
-      " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
-
-      ".set pop                                            \n"
-      : [dst_ptr] "+r"(dst_ptr), [src_ptr1] "+r"(src_ptr1),
-        [src_ptr] "+r"(src_ptr), [dst_width] "+r"(dst_width)
-      : [source_y_fraction] "r"(source_y_fraction),
-        [y0_fraction] "r"(y0_fraction), [src_stride] "r"(src_stride)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-#include <stdio.h>
-void RGB24ToARGBRow_DSPR2(const uint8* src_rgb24, uint8* dst_argb, int width) {
-  int x;
-  uint32 tmp_mask = 0xff;
-  uint32 tmp_t1;
-  for (x = 0; x < (width - 1); ++x) {
-    __asm__ __volatile__(
-        ".set push                                                  \n"
-        ".set noreorder                                             \n"
-        "ulw             %[tmp_t1],    0(%[src_rgb24])              \n"
-        "addiu           %[dst_argb],  %[dst_argb],     4           \n"
-        "addiu           %[src_rgb24], %[src_rgb24],    3           \n"
-        "ins             %[tmp_t1],    %[tmp_mask],     24,    8    \n"
-        "sw              %[tmp_t1],    -4(%[dst_argb])              \n"
-        ".set pop                                                   \n"
-        : [src_rgb24] "+r"(src_rgb24), [dst_argb] "+r"(dst_argb),
-          [tmp_t1] "=&r"(tmp_t1)
-        : [tmp_mask] "r"(tmp_mask)
-        : "memory");
-  }
-  uint8 b = src_rgb24[0];
-  uint8 g = src_rgb24[1];
-  uint8 r = src_rgb24[2];
-  dst_argb[0] = b;
-  dst_argb[1] = g;
-  dst_argb[2] = r;
-  dst_argb[3] = 255u;
-}
-
-void RAWToARGBRow_DSPR2(const uint8* src_raw, uint8* dst_argb, int width) {
-  int x;
-  uint32 tmp_mask = 0xff;
-  uint32 tmp_t1, tmp_t2;
-  for (x = 0; x < (width - 1); ++x) {
-    __asm__ __volatile__(
-        ".set push                                               \n"
-        ".set noreorder                                          \n"
-        "ulw               %[tmp_t1],   0(%[src_raw])            \n"
-        "addiu             %[dst_argb], %[dst_argb],      4      \n"
-        "addiu             %[src_raw],  %[src_raw],       3      \n"
-        "srl               %[tmp_t2],   %[tmp_t1],        16     \n"
-        "ins               %[tmp_t1],   %[tmp_mask],      24, 8  \n"
-        "ins               %[tmp_t1],   %[tmp_t1],        16, 8  \n"
-        "ins               %[tmp_t1],   %[tmp_t2],        0,  8  \n"
-        "sw                %[tmp_t1],   -4(%[dst_argb])          \n"
-        ".set pop                                                \n"
-        : [src_raw] "+r"(src_raw), [dst_argb] "+r"(dst_argb),
-          [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2)
-        : [tmp_mask] "r"(tmp_mask)
-        : "memory");
-  }
-  uint8 r = src_raw[0];
-  uint8 g = src_raw[1];
-  uint8 b = src_raw[2];
-  dst_argb[0] = b;
-  dst_argb[1] = g;
-  dst_argb[2] = r;
-  dst_argb[3] = 255u;
-}
-
-void RGB565ToARGBRow_DSPR2(const uint8* src_rgb565,
-                           uint8* dst_argb,
-                           int width) {
-  int x;
-  uint32 tmp_mask = 0xff;
-  uint32 tmp_t1, tmp_t2, tmp_t3;
-  for (x = 0; x < width; ++x) {
-    __asm__ __volatile__(
-        ".set push                                                   \n"
-        ".set noreorder                                              \n"
-        "lhu               %[tmp_t1],     0(%[src_rgb565])           \n"
-        "addiu             %[dst_argb],   %[dst_argb],      4        \n"
-        "addiu             %[src_rgb565], %[src_rgb565],    2        \n"
-        "sll               %[tmp_t2],     %[tmp_t1],        8        \n"
-        "ins               %[tmp_t2],     %[tmp_mask],      24,8     \n"
-        "ins               %[tmp_t2],     %[tmp_t1],        3, 16    \n"
-        "ins               %[tmp_t2],     %[tmp_t1],        5, 11    \n"
-        "srl               %[tmp_t3],     %[tmp_t1],        9        \n"
-        "ins               %[tmp_t2],     %[tmp_t3],        8, 2     \n"
-        "ins               %[tmp_t2],     %[tmp_t1],        3, 5     \n"
-        "srl               %[tmp_t3],     %[tmp_t1],        2        \n"
-        "ins               %[tmp_t2],     %[tmp_t3],        0, 3     \n"
-        "sw                %[tmp_t2],     -4(%[dst_argb])            \n"
-        ".set pop                                                    \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [src_rgb565] "+r"(src_rgb565),
-          [dst_argb] "+r"(dst_argb)
-        : [tmp_mask] "r"(tmp_mask));
-  }
-}
-
-void ARGB1555ToARGBRow_DSPR2(const uint8* src_argb1555,
-                             uint8* dst_argb,
-                             int width) {
-  int x;
-  uint32 tmp_t1, tmp_t2, tmp_t3;
-  for (x = 0; x < width; ++x) {
-    __asm__ __volatile__(
-        ".set push                                                   \n"
-        ".set noreorder                                              \n"
-        "lh                %[tmp_t1],       0(%[src_argb1555])       \n"
-        "addiu             %[dst_argb],     %[dst_argb],      4      \n"
-        "addiu             %[src_argb1555], %[src_argb1555],  2      \n"
-        "sll               %[tmp_t2],       %[tmp_t1],        9      \n"
-        "ins               %[tmp_t2],       %[tmp_t1],        4, 15  \n"
-        "ins               %[tmp_t2],       %[tmp_t1],        6, 10  \n"
-        "srl               %[tmp_t3],       %[tmp_t1],        7      \n"
-        "ins               %[tmp_t2],       %[tmp_t3],        8, 3   \n"
-        "ins               %[tmp_t2],       %[tmp_t1],        3, 5   \n"
-        "srl               %[tmp_t3],       %[tmp_t1],        2      \n"
-        "ins               %[tmp_t2],       %[tmp_t3],        0, 3   \n"
-        "sw                %[tmp_t2],       -4(%[dst_argb])          \n"
-        ".set pop                                                    \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [src_argb1555] "+r"(src_argb1555),
-          [dst_argb] "+r"(dst_argb)
-        :);
-  }
-}
-
-void ARGB4444ToARGBRow_DSPR2(const uint8* src_argb4444,
-                             uint8* dst_argb,
-                             int width) {
-  int x;
-  uint32 tmp_t1;
-  for (x = 0; x < width; ++x) {
-    __asm__ __volatile__(
-        ".set push                                                    \n"
-        ".set noreorder                                               \n"
-        "lh                %[tmp_t1],       0(%[src_argb4444])        \n"
-        "addiu             %[dst_argb],     %[dst_argb],       4      \n"
-        "addiu             %[src_argb4444], %[src_argb4444],   2      \n"
-        "ins               %[tmp_t1],       %[tmp_t1],         16, 16 \n"
-        "ins               %[tmp_t1],       %[tmp_t1],         12, 16 \n"
-        "ins               %[tmp_t1],       %[tmp_t1],         8,  12 \n"
-        "ins               %[tmp_t1],       %[tmp_t1],         4,  8  \n"
-        "sw                %[tmp_t1],       -4(%[dst_argb])           \n"
-        ".set pop                                                     \n"
-        : [src_argb4444] "+r"(src_argb4444), [dst_argb] "+r"(dst_argb),
-          [tmp_t1] "=&r"(tmp_t1));
-  }
-}
-
-void I444ToARGBRow_DSPR2(const uint8* y_buf,
-                         const uint8* u_buf,
-                         const uint8* v_buf,
-                         uint8* rgb_buf,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
-  int x;
-  uint32 tmp_ub = yuvconstants->kUVToB[0];
-  uint32 tmp_ug = yuvconstants->kUVToG[0];
-  uint32 tmp_vg = yuvconstants->kUVToG[1];
-  uint32 tmp_vr = yuvconstants->kUVToR[1];
-  uint32 tmp_bb = yuvconstants->kUVBiasB[0];
-  uint32 tmp_bg = yuvconstants->kUVBiasG[0];
-  uint32 tmp_br = yuvconstants->kUVBiasR[0];
-  uint32 yg = yuvconstants->kYToRgb[0];
-  uint32 tmp_mask = 0x7fff7fff;
-  uint32 tmp_yg;
-
-  tmp_bb = ((uint)(tmp_bb & 0xffff) << 16) | (tmp_bb & 0xffff);
-  tmp_bg = ((uint)(tmp_bg & 0xffff) << 16) | (tmp_bg & 0xffff);
-  tmp_br = ((uint)(tmp_br & 0xffff) << 16) | (tmp_br & 0xffff);
-  tmp_yg = ((uint)(yg & 0xffff) << 16) | (yg & 0xffff);
-  tmp_ub = ~(((uint)(tmp_ub & 0xffff) << 16) | (tmp_ub & 0xffff)) + 0x00010001;
-  tmp_ug = ((uint)(tmp_ug & 0xffff) << 16) | (tmp_ug & 0xffff);
-  tmp_vg = ((uint)(tmp_vg & 0xffff) << 16) | (tmp_vg & 0xffff);
-  tmp_vr = ~(((uint)(tmp_vr & 0xffff) << 16) | (tmp_vr & 0xffff)) + 0x00010001;
-  yg = yg * 0x0101;
-
-  for (x = 0; x < width - 1; x += 2) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    uint32 tmp_t6, tmp_t7, tmp_t8, tmp_t9;
-    __asm__ __volatile__(
-        ".set push                                              \n"
-        ".set noreorder                                         \n"
-        "lbu              %[tmp_t7], 0(%[y_buf])               \n"
-        "lbu              %[tmp_t1], 1(%[y_buf])               \n"
-        "mul              %[tmp_t7], %[tmp_t7],     %[yg]      \n"
-        "mul              %[tmp_t1], %[tmp_t1],     %[yg]      \n"
-        "lh               %[tmp_t2], 0(%[u_buf])               \n"
-        "lh               %[tmp_t3], 0(%[v_buf])               \n"
-        "preceu.ph.qbr    %[tmp_t2], %[tmp_t2]                 \n"
-        "preceu.ph.qbr    %[tmp_t3], %[tmp_t3]                 \n"
-        "mul.ph           %[tmp_t4], %[tmp_t2],     %[tmp_ub]  \n"
-        "mul.ph           %[tmp_t5], %[tmp_t2],     %[tmp_ug]  \n"
-        "mul.ph           %[tmp_t6], %[tmp_t3],     %[tmp_vr]  \n"
-        "mul.ph           %[tmp_t3], %[tmp_t3],     %[tmp_vg]  \n"
-        "srl              %[tmp_t7], %[tmp_t7],     16         \n"
-        "ins              %[tmp_t1], %[tmp_t7],     0,      16 \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t1],     %[tmp_bb]  \n"
-        "addq_s.ph        %[tmp_t8], %[tmp_t1],     %[tmp_bg]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t1],     %[tmp_br]  \n"
-        "addq_s.ph        %[tmp_t5], %[tmp_t5],     %[tmp_t3]  \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t7],     %[tmp_t4]  \n"
-        "subq_s.ph        %[tmp_t8], %[tmp_t8],     %[tmp_t5]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t9],     %[tmp_t6]  \n"
-        "shra.ph          %[tmp_t7], %[tmp_t7],     6          \n"
-        "shra.ph          %[tmp_t8], %[tmp_t8],     6          \n"
-        "shra.ph          %[tmp_t9], %[tmp_t9],     6          \n"
-        "shll_s.ph        %[tmp_t7], %[tmp_t7],     7          \n"
-        "shll_s.ph        %[tmp_t8], %[tmp_t8],     7          \n"
-        "shll_s.ph        %[tmp_t9], %[tmp_t9],     7          \n"
-        "precrqu_s.qb.ph  %[tmp_t8], %[tmp_mask],   %[tmp_t8]  \n"
-        "precrqu_s.qb.ph  %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.ph.w      %[tmp_t2], %[tmp_t8],     %[tmp_t7]  \n"
-        "ins              %[tmp_t7], %[tmp_t8],     16,     16 \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t2],     %[tmp_t7]  \n"
-        "precrq.qb.ph     %[tmp_t7], %[tmp_t2],     %[tmp_t7]  \n"
-        "sw               %[tmp_t8], 0(%[rgb_buf])             \n"
-        "sw               %[tmp_t7], 4(%[rgb_buf])             \n"
-        ".set pop                                              \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [tmp_t9] "=&r"(tmp_t9)
-        : [y_buf] "r"(y_buf), [yg] "r"(yg), [u_buf] "r"(u_buf),
-          [v_buf] "r"(v_buf), [tmp_ub] "r"(tmp_ub), [tmp_ug] "r"(tmp_ug),
-          [tmp_vg] "r"(tmp_vg), [tmp_vr] "r"(tmp_vr), [tmp_bb] "r"(tmp_bb),
-          [tmp_bg] "r"(tmp_bg), [tmp_br] "r"(tmp_br), [tmp_yg] "r"(tmp_yg),
-          [rgb_buf] "r"(rgb_buf), [tmp_mask] "r"(tmp_mask));
-    y_buf += 2;
-    u_buf += 2;
-    v_buf += 2;
-    rgb_buf += 8;  // Advance 1 pixel.
-  }
-}
-
-void I422ToARGB4444Row_DSPR2(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb4444,
-                             const struct YuvConstants* yuvconstants,
-                             int width) {
-  int x;
-  uint32 tmp_ub = yuvconstants->kUVToB[0];
-  uint32 tmp_ug = yuvconstants->kUVToG[0];
-  uint32 tmp_vg = yuvconstants->kUVToG[1];
-  uint32 tmp_vr = yuvconstants->kUVToR[1];
-  uint32 tmp_bb = yuvconstants->kUVBiasB[0];
-  uint32 tmp_bg = yuvconstants->kUVBiasG[0];
-  uint32 tmp_br = yuvconstants->kUVBiasR[0];
-  uint32 yg = yuvconstants->kYToRgb[0];
-  uint32 tmp_yg;
-  uint32 tmp_mask = 0x7fff7fff;
-  tmp_bb = ((uint)(tmp_bb & 0xffff) << 16) | (tmp_bb & 0xffff);
-  tmp_bg = ((uint)(tmp_bg & 0xffff) << 16) | (tmp_bg & 0xffff);
-  tmp_br = ((uint)(tmp_br & 0xffff) << 16) | (tmp_br & 0xffff);
-  tmp_yg = ((uint)(yg & 0xffff) << 16) | (yg & 0xffff);
-  tmp_ub = ~(((uint)(tmp_ub & 0xffff) << 16) | (tmp_ub & 0xffff)) + 0x00010001;
-  tmp_ug = ((uint)(tmp_ug & 0xffff) << 16) | (tmp_ug & 0xffff);
-  tmp_vg = ((uint)(tmp_vg & 0xffff) << 16) | (tmp_vg & 0xffff);
-  tmp_vr = ~(((uint)(tmp_vr & 0xffff) << 16) | (tmp_vr & 0xffff)) + 0x00010001;
-  yg = yg * 0x0101;
-
-  for (x = 0; x < width - 1; x += 2) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    uint32 tmp_t6, tmp_t7, tmp_t8, tmp_t9;
-    __asm__ __volatile__(
-        ".set push                                             \n"
-        ".set noreorder                                        \n"
-        "lbu              %[tmp_t7], 0(%[src_y])               \n"
-        "lbu              %[tmp_t1], 1(%[src_y])               \n"
-        "mul              %[tmp_t7], %[tmp_t7],     %[yg]      \n"
-        "mul              %[tmp_t1], %[tmp_t1],     %[yg]      \n"
-        "lbu              %[tmp_t2], 0(%[src_u])               \n"
-        "lbu              %[tmp_t3], 0(%[src_v])               \n"
-        "replv.ph         %[tmp_t2], %[tmp_t2]                 \n"
-        "replv.ph         %[tmp_t3], %[tmp_t3]                 \n"
-        "mul.ph           %[tmp_t4], %[tmp_t2],     %[tmp_ub]  \n"
-        "mul.ph           %[tmp_t5], %[tmp_t2],     %[tmp_ug]  \n"
-        "mul.ph           %[tmp_t6], %[tmp_t3],     %[tmp_vr]  \n"
-        "mul.ph           %[tmp_t3], %[tmp_t3],     %[tmp_vg]  \n"
-        "srl              %[tmp_t7], %[tmp_t7],     16         \n"
-        "ins              %[tmp_t1], %[tmp_t7],     0,      16 \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t1],     %[tmp_bb]  \n"
-        "addq_s.ph        %[tmp_t8], %[tmp_t1],     %[tmp_bg]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t1],     %[tmp_br]  \n"
-        "addq_s.ph        %[tmp_t5], %[tmp_t5],     %[tmp_t3]  \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t7],     %[tmp_t4]  \n"
-        "subq_s.ph        %[tmp_t8], %[tmp_t8],     %[tmp_t5]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t9],     %[tmp_t6]  \n"
-        "shra.ph          %[tmp_t7], %[tmp_t7],     6          \n"
-        "shra.ph          %[tmp_t8], %[tmp_t8],     6          \n"
-        "shra.ph          %[tmp_t9], %[tmp_t9],     6          \n"
-        "shll_s.ph        %[tmp_t7], %[tmp_t7],     7          \n"
-        "shll_s.ph        %[tmp_t8], %[tmp_t8],     7          \n"
-        "shll_s.ph        %[tmp_t9], %[tmp_t9],     7          \n"
-        "precrqu_s.qb.ph  %[tmp_t8], %[tmp_mask],   %[tmp_t8]  \n"
-        "precrqu_s.qb.ph  %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.ph.w      %[tmp_t2], %[tmp_t8],     %[tmp_t7]  \n"
-        "ins              %[tmp_t7], %[tmp_t8],     16,     16 \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t2],     %[tmp_t7]  \n"
-        "precrq.qb.ph     %[tmp_t7], %[tmp_t2],     %[tmp_t7]  \n"
-        "shrl.qb          %[tmp_t1], %[tmp_t8],     4          \n"
-        "shrl.qb          %[tmp_t2], %[tmp_t7],     4          \n"
-        "shrl.ph          %[tmp_t8], %[tmp_t1],     4          \n"
-        "shrl.ph          %[tmp_t7], %[tmp_t2],     4          \n"
-        "or               %[tmp_t8], %[tmp_t8],     %[tmp_t1]  \n"
-        "or               %[tmp_t7], %[tmp_t7],     %[tmp_t2]  \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t7],     %[tmp_t8]  \n"
-        "sw               %[tmp_t8], 0(%[dst_argb4444])        \n"
-        ".set pop                                              \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [tmp_t9] "=&r"(tmp_t9)
-        : [dst_argb4444] "r"(dst_argb4444), [yg] "r"(yg), [src_u] "r"(src_u),
-          [src_v] "r"(src_v), [src_y] "r"(src_y), [tmp_ub] "r"(tmp_ub),
-          [tmp_ug] "r"(tmp_ug), [tmp_vg] "r"(tmp_vg), [tmp_vr] "r"(tmp_vr),
-          [tmp_bb] "r"(tmp_bb), [tmp_bg] "r"(tmp_bg), [tmp_br] "r"(tmp_br),
-          [tmp_yg] "r"(tmp_yg), [tmp_mask] "r"(tmp_mask));
-    src_y += 2;
-    src_u += 1;
-    src_v += 1;
-    dst_argb4444 += 4;  // Advance 2 pixels.
-  }
-}
-
-void I422ToARGB1555Row_DSPR2(const uint8* src_y,
-                             const uint8* src_u,
-                             const uint8* src_v,
-                             uint8* dst_argb1555,
-                             const struct YuvConstants* yuvconstants,
-                             int width) {
-  int x;
-  uint32 tmp_ub = yuvconstants->kUVToB[0];
-  uint32 tmp_ug = yuvconstants->kUVToG[0];
-  uint32 tmp_vg = yuvconstants->kUVToG[1];
-  uint32 tmp_vr = yuvconstants->kUVToR[1];
-  uint32 tmp_bb = yuvconstants->kUVBiasB[0];
-  uint32 tmp_bg = yuvconstants->kUVBiasG[0];
-  uint32 tmp_br = yuvconstants->kUVBiasR[0];
-  uint32 yg = yuvconstants->kYToRgb[0];
-  uint32 tmp_yg;
-  uint32 tmp_mask = 0x80008000;
-  tmp_bb = ((uint)(tmp_bb & 0xffff) << 16) | (tmp_bb & 0xffff);
-  tmp_bg = ((uint)(tmp_bg & 0xffff) << 16) | (tmp_bg & 0xffff);
-  tmp_br = ((uint)(tmp_br & 0xffff) << 16) | (tmp_br & 0xffff);
-  tmp_yg = ((uint)(yg & 0xffff) << 16) | (yg & 0xffff);
-  tmp_ub = ~(((uint)(tmp_ub & 0xffff) << 16) | (tmp_ub & 0xffff)) + 0x00010001;
-  tmp_ug = ((uint)(tmp_ug & 0xffff) << 16) | (tmp_ug & 0xffff);
-  tmp_vg = ((uint)(tmp_vg & 0xffff) << 16) | (tmp_vg & 0xffff);
-  tmp_vr = ~(((uint)(tmp_vr & 0xffff) << 16) | (tmp_vr & 0xffff)) + 0x00010001;
-  yg = yg * 0x0101;
-
-  for (x = 0; x < width - 1; x += 2) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    uint32 tmp_t6, tmp_t7, tmp_t8, tmp_t9;
-    __asm__ __volatile__(
-        ".set push                                             \n"
-        ".set noreorder                                        \n"
-        "lbu              %[tmp_t7], 0(%[src_y])               \n"
-        "lbu              %[tmp_t1], 1(%[src_y])               \n"
-        "mul              %[tmp_t7], %[tmp_t7],     %[yg]      \n"
-        "mul              %[tmp_t1], %[tmp_t1],     %[yg]      \n"
-        "lbu              %[tmp_t2], 0(%[src_u])               \n"
-        "lbu              %[tmp_t3], 0(%[src_v])               \n"
-        "replv.ph         %[tmp_t2], %[tmp_t2]                 \n"
-        "replv.ph         %[tmp_t3], %[tmp_t3]                 \n"
-        "mul.ph           %[tmp_t4], %[tmp_t2],     %[tmp_ub]  \n"
-        "mul.ph           %[tmp_t5], %[tmp_t2],     %[tmp_ug]  \n"
-        "mul.ph           %[tmp_t6], %[tmp_t3],     %[tmp_vr]  \n"
-        "mul.ph           %[tmp_t3], %[tmp_t3],     %[tmp_vg]  \n"
-        "srl              %[tmp_t7], %[tmp_t7],     16         \n"
-        "ins              %[tmp_t1], %[tmp_t7],     0,      16 \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t1],     %[tmp_bb]  \n"
-        "addq_s.ph        %[tmp_t8], %[tmp_t1],     %[tmp_bg]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t1],     %[tmp_br]  \n"
-        "addq_s.ph        %[tmp_t5], %[tmp_t5],     %[tmp_t3]  \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t7],     %[tmp_t4]  \n"
-        "subq_s.ph        %[tmp_t8], %[tmp_t8],     %[tmp_t5]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t9],     %[tmp_t6]  \n"
-        "shra.ph          %[tmp_t7], %[tmp_t7],     6          \n"
-        "shra.ph          %[tmp_t8], %[tmp_t8],     6          \n"
-        "shra.ph          %[tmp_t9], %[tmp_t9],     6          \n"
-        "shll_s.ph        %[tmp_t7], %[tmp_t7],     7          \n"
-        "shll_s.ph        %[tmp_t8], %[tmp_t8],     7          \n"
-        "shll_s.ph        %[tmp_t9], %[tmp_t9],     7          \n"
-        "precrqu_s.qb.ph  %[tmp_t8], %[tmp_mask],   %[tmp_t8]  \n"
-        "precrqu_s.qb.ph  %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.ph.w      %[tmp_t2], %[tmp_t8],     %[tmp_t7]  \n"
-        "ins              %[tmp_t7], %[tmp_t8],     16,     16 \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t2],     %[tmp_t7]  \n"
-        "precrq.qb.ph     %[tmp_t7], %[tmp_t2],     %[tmp_t7]  \n"
-        "ins              %[tmp_t3], %[tmp_t8],     7,      24 \n"
-        "ins              %[tmp_t3], %[tmp_t8],     10,     16 \n"
-        "ins              %[tmp_t3], %[tmp_t8],     13,     8  \n"
-        "ins              %[tmp_t4], %[tmp_t7],     7,      24 \n"
-        "ins              %[tmp_t4], %[tmp_t7],     10,     16 \n"
-        "ins              %[tmp_t4], %[tmp_t7],     13,     8  \n"
-        "precrq.ph.w      %[tmp_t8], %[tmp_t4],     %[tmp_t3]  \n"
-        "or               %[tmp_t8], %[tmp_t8],     %[tmp_mask]\n"
-        "sw               %[tmp_t8], 0(%[dst_argb1555])        \n"
-        ".set pop                                              \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [tmp_t9] "=&r"(tmp_t9)
-        : [dst_argb1555] "r"(dst_argb1555), [yg] "r"(yg), [src_u] "r"(src_u),
-          [src_v] "r"(src_v), [src_y] "r"(src_y), [tmp_ub] "r"(tmp_ub),
-          [tmp_ug] "r"(tmp_ug), [tmp_vg] "r"(tmp_vg), [tmp_vr] "r"(tmp_vr),
-          [tmp_bb] "r"(tmp_bb), [tmp_bg] "r"(tmp_bg), [tmp_br] "r"(tmp_br),
-          [tmp_yg] "r"(tmp_yg), [tmp_mask] "r"(tmp_mask));
-    src_y += 2;
-    src_u += 1;
-    src_v += 1;
-    dst_argb1555 += 4;  // Advance 2 pixels.
-  }
-}
-
-void NV12ToARGBRow_DSPR2(const uint8* src_y,
-                         const uint8* src_uv,
-                         uint8* rgb_buf,
-                         const struct YuvConstants* yuvconstants,
-                         int width) {
-  int x;
-  uint32 tmp_ub = yuvconstants->kUVToB[0];
-  uint32 tmp_ug = yuvconstants->kUVToG[0];
-  uint32 tmp_vg = yuvconstants->kUVToG[1];
-  uint32 tmp_vr = yuvconstants->kUVToR[1];
-  uint32 tmp_bb = yuvconstants->kUVBiasB[0];
-  uint32 tmp_bg = yuvconstants->kUVBiasG[0];
-  uint32 tmp_br = yuvconstants->kUVBiasR[0];
-  uint32 yg = yuvconstants->kYToRgb[0];
-  uint32 tmp_mask = 0x7fff7fff;
-  uint32 tmp_yg;
-  tmp_bb = ((uint)(tmp_bb & 0xffff) << 16) | (tmp_bb & 0xffff);
-  tmp_bg = ((uint)(tmp_bg & 0xffff) << 16) | (tmp_bg & 0xffff);
-  tmp_br = ((uint)(tmp_br & 0xffff) << 16) | (tmp_br & 0xffff);
-  tmp_yg = ((uint)(yg & 0xffff) << 16) | (yg & 0xffff);
-  tmp_ub = ~(((uint)(tmp_ub & 0xffff) << 16) | (tmp_ub & 0xffff)) + 0x00010001;
-  tmp_ug = ((uint)(tmp_ug & 0xffff) << 16) | (tmp_ug & 0xffff);
-  tmp_vg = ((uint)(tmp_vg & 0xffff) << 16) | (tmp_vg & 0xffff);
-  tmp_vr = ~(((uint)(tmp_vr & 0xffff) << 16) | (tmp_vr & 0xffff)) + 0x00010001;
-  yg = yg * 0x0101;
-
-  for (x = 0; x < width - 1; x += 2) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    uint32 tmp_t6, tmp_t7, tmp_t8, tmp_t9;
-    __asm__ __volatile__(
-        ".set push                                             \n"
-        ".set noreorder                                        \n"
-        "lbu              %[tmp_t7], 0(%[src_y])               \n"
-        "lbu              %[tmp_t1], 1(%[src_y])               \n"
-        "mul              %[tmp_t7], %[tmp_t7],     %[yg]      \n"
-        "mul              %[tmp_t1], %[tmp_t1],     %[yg]      \n"
-        "lbu              %[tmp_t2], 0(%[src_uv])              \n"
-        "lbu              %[tmp_t3], 1(%[src_uv])              \n"
-        "replv.ph         %[tmp_t2], %[tmp_t2]                 \n"
-        "replv.ph         %[tmp_t3], %[tmp_t3]                 \n"
-        "mul.ph           %[tmp_t4], %[tmp_t2],     %[tmp_ub]  \n"
-        "mul.ph           %[tmp_t5], %[tmp_t2],     %[tmp_ug]  \n"
-        "mul.ph           %[tmp_t6], %[tmp_t3],     %[tmp_vr]  \n"
-        "mul.ph           %[tmp_t3], %[tmp_t3],     %[tmp_vg]  \n"
-        "srl              %[tmp_t7], %[tmp_t7],     16         \n"
-        "ins              %[tmp_t1], %[tmp_t7],     0,      16 \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t1],     %[tmp_bb]  \n"
-        "addq_s.ph        %[tmp_t8], %[tmp_t1],     %[tmp_bg]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t1],     %[tmp_br]  \n"
-        "addq_s.ph        %[tmp_t5], %[tmp_t5],     %[tmp_t3]  \n"
-        "addq_s.ph        %[tmp_t7], %[tmp_t7],     %[tmp_t4]  \n"
-        "subq_s.ph        %[tmp_t8], %[tmp_t8],     %[tmp_t5]  \n"
-        "addq_s.ph        %[tmp_t9], %[tmp_t9],     %[tmp_t6]  \n"
-        "shra.ph          %[tmp_t7], %[tmp_t7],     6          \n"
-        "shra.ph          %[tmp_t8], %[tmp_t8],     6          \n"
-        "shra.ph          %[tmp_t9], %[tmp_t9],     6          \n"
-        "shll_s.ph        %[tmp_t7], %[tmp_t7],     7          \n"
-        "shll_s.ph        %[tmp_t8], %[tmp_t8],     7          \n"
-        "shll_s.ph        %[tmp_t9], %[tmp_t9],     7          \n"
-        "precrqu_s.qb.ph  %[tmp_t8], %[tmp_mask],   %[tmp_t8]  \n"
-        "precrqu_s.qb.ph  %[tmp_t7], %[tmp_t9],     %[tmp_t7]  \n"
-        "precrq.ph.w      %[tmp_t2], %[tmp_t8],     %[tmp_t7]  \n"
-        "ins              %[tmp_t7], %[tmp_t8],     16,     16 \n"
-        "precr.qb.ph      %[tmp_t8], %[tmp_t2],     %[tmp_t7]  \n"
-        "precrq.qb.ph     %[tmp_t7], %[tmp_t2],     %[tmp_t7]  \n"
-        "sw               %[tmp_t8], 0(%[rgb_buf])             \n"
-        "sw               %[tmp_t7], 4(%[rgb_buf])             \n"
-        ".set pop                                              \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [tmp_t9] "=&r"(tmp_t9)
-        : [src_y] "r"(src_y), [src_uv] "r"(src_uv), [yg] "r"(yg),
-          [tmp_ub] "r"(tmp_ub), [tmp_ug] "r"(tmp_ug), [tmp_vg] "r"(tmp_vg),
-          [tmp_vr] "r"(tmp_vr), [tmp_bb] "r"(tmp_bb), [tmp_bg] "r"(tmp_bg),
-          [tmp_br] "r"(tmp_br), [tmp_yg] "r"(tmp_yg), [rgb_buf] "r"(rgb_buf),
-          [tmp_mask] "r"(tmp_mask));
-
-    src_y += 2;
-    src_uv += 2;
-    rgb_buf += 8;  // Advance 2 pixels.
-  }
-}
-
-void BGRAToUVRow_DSPR2(const uint8* src_rgb0,
-                       int src_stride_rgb,
-                       uint8* dst_u,
-                       uint8* dst_v,
-                       int width) {
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;
-  int x;
-  int const1 = 0xffda0000;
-  int const2 = 0x0070ffb6;
-  int const3 = 0x00700000;
-  int const4 = 0xffeeffa2;
-  int const5 = 0x100;
-  for (x = 0; x < width - 1; x += 2) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                 \n"
-        ".set noreorder                                            \n"
-        "lw                %[tmp_t1],   0(%[src_rgb0])             \n"
-        "lw                %[tmp_t2],   4(%[src_rgb0])             \n"
-        "lw                %[tmp_t3],   0(%[src_rgb1])             \n"
-        "lw                %[tmp_t4],   4(%[src_rgb1])             \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                  \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                  \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                  \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                  \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                  \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                  \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                  \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                  \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t6]   \n"
-        "addu.ph           %[tmp_t7],   %[tmp_t7],     %[tmp_t8]   \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t2]   \n"
-        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t4]   \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t7]   \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t3]   \n"
-        "shrl.ph           %[tmp_t5],   %[tmp_t5],     2           \n"
-        "shrl.ph           %[tmp_t1],   %[tmp_t1],     2           \n"
-        "mult              $ac0,        %[const5],     %[const5]   \n"
-        "mult              $ac1,        %[const5],     %[const5]   \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t5],     %[const1]   \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t5],     %[const3]   \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t1],     %[const2]   \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t1],     %[const4]   \n"
-        "extr_r.w          %[tmp_t7],   $ac0,          9           \n"
-        "extr_r.w          %[tmp_t8],   $ac1,          9           \n"
-        "addiu             %[dst_u],    %[dst_u],    1             \n"
-        "addiu             %[dst_v],    %[dst_v],    1             \n"
-        "addiu             %[src_rgb0], %[src_rgb0], 8             \n"
-        "addiu             %[src_rgb1], %[src_rgb1], 8             \n"
-        "sb                %[tmp_t7],   -1(%[dst_u])               \n"
-        "sb                %[tmp_t8],   -1(%[dst_v])               \n"
-        ".set pop                                                  \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_rgb0] "+r"(src_rgb0), [src_rgb1] "+r"(src_rgb1),
-          [dst_u] "+r"(dst_u), [dst_v] "+r"(dst_v)
-        : [const1] "r"(const1), [const2] "r"(const2), [const3] "r"(const3),
-          [const4] "r"(const4), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi");
-  }
-}
-
-void BGRAToYRow_DSPR2(const uint8* src_argb0, uint8* dst_y, int width) {
-  int x;
-  int const1 = 0x00420000;
-  int const2 = 0x00190081;
-  int const5 = 0x40;
-  for (x = 0; x < width; x += 4) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_argb0])           \n"
-        "lw                %[tmp_t2],   4(%[src_argb0])           \n"
-        "lw                %[tmp_t3],   8(%[src_argb0])           \n"
-        "lw                %[tmp_t4],   12(%[src_argb0])          \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "mult              $ac2,        %[const5],     %[const5]  \n"
-        "mult              $ac3,        %[const5],     %[const5]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t6],     %[const1]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t7],     %[const1]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t8],     %[const1]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t2],     %[const2]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t3],     %[const2]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t4],     %[const2]  \n"
-        "extr_r.w          %[tmp_t1],   $ac0,          8          \n"
-        "extr_r.w          %[tmp_t2],   $ac1,          8          \n"
-        "extr_r.w          %[tmp_t3],   $ac2,          8          \n"
-        "extr_r.w          %[tmp_t4],   $ac3,          8          \n"
-        "addiu             %[src_argb0],%[src_argb0],  16         \n"
-        "addiu             %[dst_y],    %[dst_y],      4          \n"
-        "sb                %[tmp_t1],   -4(%[dst_y])              \n"
-        "sb                %[tmp_t2],   -3(%[dst_y])              \n"
-        "sb                %[tmp_t3],   -2(%[dst_y])              \n"
-        "sb                %[tmp_t4],   -1(%[dst_y])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_argb0] "+r"(src_argb0), [dst_y] "+r"(dst_y)
-        : [const1] "r"(const1), [const2] "r"(const2), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo",
-          "$ac3hi");
-  }
-}
-
-void ABGRToUVRow_DSPR2(const uint8* src_rgb0,
-                       int src_stride_rgb,
-                       uint8* dst_u,
-                       uint8* dst_v,
-                       int width) {
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;
-  int x;
-  int const1 = 0xffb6ffda;
-  int const2 = 0x00000070;
-  int const3 = 0xffa20070;
-  int const4 = 0x0000ffee;
-  int const5 = 0x100;
-
-  for (x = 0; x < width - 1; x += 2) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_rgb0])            \n"
-        "lw                %[tmp_t2],   4(%[src_rgb0])            \n"
-        "lw                %[tmp_t3],   0(%[src_rgb1])            \n"
-        "lw                %[tmp_t4],   4(%[src_rgb1])            \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t6]  \n"
-        "addu.ph           %[tmp_t7],   %[tmp_t7],     %[tmp_t8]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t2]  \n"
-        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t4]  \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t7]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t3]  \n"
-        "shrl.ph           %[tmp_t5],   %[tmp_t5],     2          \n"
-        "shrl.ph           %[tmp_t1],   %[tmp_t1],     2          \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t5],     %[const3]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t1],     %[const4]  \n"
-        "extr_r.w          %[tmp_t7],   $ac0,          9          \n"
-        "extr_r.w          %[tmp_t8],   $ac1,          9          \n"
-        "addiu             %[dst_u],    %[dst_u],    1            \n"
-        "addiu             %[dst_v],    %[dst_v],    1            \n"
-        "addiu             %[src_rgb0], %[src_rgb0], 8            \n"
-        "addiu             %[src_rgb1], %[src_rgb1], 8            \n"
-        "sb                %[tmp_t7],   -1(%[dst_u])              \n"
-        "sb                %[tmp_t8],   -1(%[dst_v])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_rgb0] "+r"(src_rgb0), [src_rgb1] "+r"(src_rgb1),
-          [dst_u] "+r"(dst_u), [dst_v] "+r"(dst_v)
-        : [const1] "r"(const1), [const2] "r"(const2), [const3] "r"(const3),
-          [const4] "r"(const4), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi");
-  }
-}
-
-void ARGBToYRow_DSPR2(const uint8* src_argb0, uint8* dst_y, int width) {
-  int x;
-  int const1 = 0x00810019;
-  int const2 = 0x00000042;
-  int const5 = 0x40;
-  for (x = 0; x < width; x += 4) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_argb0])           \n"
-        "lw                %[tmp_t2],   4(%[src_argb0])           \n"
-        "lw                %[tmp_t3],   8(%[src_argb0])           \n"
-        "lw                %[tmp_t4],   12(%[src_argb0])          \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "mult              $ac2,        %[const5],     %[const5]  \n"
-        "mult              $ac3,        %[const5],     %[const5]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t6],     %[const1]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t7],     %[const1]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t8],     %[const1]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t2],     %[const2]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t3],     %[const2]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t4],     %[const2]  \n"
-        "extr_r.w          %[tmp_t1],   $ac0,          8          \n"
-        "extr_r.w          %[tmp_t2],   $ac1,          8          \n"
-        "extr_r.w          %[tmp_t3],   $ac2,          8          \n"
-        "extr_r.w          %[tmp_t4],   $ac3,          8          \n"
-        "addiu             %[dst_y],    %[dst_y],      4          \n"
-        "addiu             %[src_argb0],%[src_argb0],  16         \n"
-        "sb                %[tmp_t1],   -4(%[dst_y])              \n"
-        "sb                %[tmp_t2],   -3(%[dst_y])              \n"
-        "sb                %[tmp_t3],   -2(%[dst_y])              \n"
-        "sb                %[tmp_t4],   -1(%[dst_y])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_argb0] "+r"(src_argb0), [dst_y] "+r"(dst_y)
-        : [const1] "r"(const1), [const2] "r"(const2), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo",
-          "$ac3hi");
-  }
-}
-
-void ABGRToYRow_DSPR2(const uint8* src_argb0, uint8* dst_y, int width) {
-  int x;
-  int const1 = 0x00810042;
-  int const2 = 0x00000019;
-  int const5 = 0x40;
-  for (x = 0; x < width; x += 4) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_argb0])           \n"
-        "lw                %[tmp_t2],   4(%[src_argb0])           \n"
-        "lw                %[tmp_t3],   8(%[src_argb0])           \n"
-        "lw                %[tmp_t4],   12(%[src_argb0])          \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "mult              $ac2,        %[const5],     %[const5]  \n"
-        "mult              $ac3,        %[const5],     %[const5]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t6],     %[const1]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t7],     %[const1]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t8],     %[const1]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t2],     %[const2]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t3],     %[const2]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t4],     %[const2]  \n"
-        "extr_r.w          %[tmp_t1],   $ac0,          8          \n"
-        "extr_r.w          %[tmp_t2],   $ac1,          8          \n"
-        "extr_r.w          %[tmp_t3],   $ac2,          8          \n"
-        "extr_r.w          %[tmp_t4],   $ac3,          8          \n"
-        "addiu             %[src_argb0],%[src_argb0],  16         \n"
-        "addiu             %[dst_y],    %[dst_y],      4          \n"
-        "sb                %[tmp_t1],   -4(%[dst_y])              \n"
-        "sb                %[tmp_t2],   -3(%[dst_y])              \n"
-        "sb                %[tmp_t3],   -2(%[dst_y])              \n"
-        "sb                %[tmp_t4],   -1(%[dst_y])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_argb0] "+r"(src_argb0), [dst_y] "+r"(dst_y)
-        : [const1] "r"(const1), [const2] "r"(const2), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo",
-          "$ac3hi");
-  }
-}
-
-void RGBAToUVRow_DSPR2(const uint8* src_rgb0,
-                       int src_stride_rgb,
-                       uint8* dst_u,
-                       uint8* dst_v,
-                       int width) {
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;
-  int x;
-  int const1 = 0xffb60070;
-  int const2 = 0x0000ffda;
-  int const3 = 0xffa2ffee;
-  int const4 = 0x00000070;
-  int const5 = 0x100;
-
-  for (x = 0; x < width - 1; x += 2) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "ulw               %[tmp_t1],   0+1(%[src_rgb0])          \n"
-        "ulw               %[tmp_t2],   4+1(%[src_rgb0])          \n"
-        "ulw               %[tmp_t3],   0+1(%[src_rgb1])          \n"
-        "ulw               %[tmp_t4],   4+1(%[src_rgb1])          \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t6]  \n"
-        "addu.ph           %[tmp_t7],   %[tmp_t7],     %[tmp_t8]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t2]  \n"
-        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t4]  \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t7]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t3]  \n"
-        "shrl.ph           %[tmp_t5],   %[tmp_t5],     2          \n"
-        "shrl.ph           %[tmp_t1],   %[tmp_t1],     2          \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t5],     %[const3]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t1],     %[const4]  \n"
-        "extr_r.w          %[tmp_t7],   $ac0,          9          \n"
-        "extr_r.w          %[tmp_t8],   $ac1,          9          \n"
-        "addiu             %[src_rgb0], %[src_rgb0], 8            \n"
-        "addiu             %[src_rgb1], %[src_rgb1], 8            \n"
-        "addiu             %[dst_u],    %[dst_u],    1            \n"
-        "addiu             %[dst_v],    %[dst_v],    1            \n"
-        "sb                %[tmp_t7],   -1(%[dst_u])              \n"
-        "sb                %[tmp_t8],   -1(%[dst_v])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_rgb0] "+r"(src_rgb0), [src_rgb1] "+r"(src_rgb1),
-          [dst_u] "+r"(dst_u), [dst_v] "+r"(dst_v)
-        : [const1] "r"(const1), [const2] "r"(const2), [const3] "r"(const3),
-          [const4] "r"(const4), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi");
-  }
-}
-
-void RGBAToYRow_DSPR2(const uint8* src_argb0, uint8* dst_y, int width) {
-  int x;
-  int const1 = 0x00420081;
-  int const2 = 0x00190000;
-  int const5 = 0x40;
-  for (x = 0; x < width; x += 4) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_argb0])           \n"
-        "lw                %[tmp_t2],   4(%[src_argb0])           \n"
-        "lw                %[tmp_t3],   8(%[src_argb0])           \n"
-        "lw                %[tmp_t4],   12(%[src_argb0])          \n"
-        "preceu.ph.qbl     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbr     %[tmp_t4],   %[tmp_t4]                 \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "mult              $ac2,        %[const5],     %[const5]  \n"
-        "mult              $ac3,        %[const5],     %[const5]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t6],     %[const1]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t7],     %[const1]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t8],     %[const1]  \n"
-        "dpa.w.ph          $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpa.w.ph          $ac1,        %[tmp_t2],     %[const2]  \n"
-        "dpa.w.ph          $ac2,        %[tmp_t3],     %[const2]  \n"
-        "dpa.w.ph          $ac3,        %[tmp_t4],     %[const2]  \n"
-        "extr_r.w          %[tmp_t1],   $ac0,          8          \n"
-        "extr_r.w          %[tmp_t2],   $ac1,          8          \n"
-        "extr_r.w          %[tmp_t3],   $ac2,          8          \n"
-        "extr_r.w          %[tmp_t4],   $ac3,          8          \n"
-        "addiu             %[dst_y],    %[dst_y],      4          \n"
-        "addiu             %[src_argb0],%[src_argb0],  16         \n"
-        "sb                %[tmp_t1],   -4(%[dst_y])              \n"
-        "sb                %[tmp_t2],   -3(%[dst_y])              \n"
-        "sb                %[tmp_t3],   -2(%[dst_y])              \n"
-        "sb                %[tmp_t4],   -1(%[dst_y])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_argb0] "+r"(src_argb0), [dst_y] "+r"(dst_y)
-        : [const1] "r"(const1), [const2] "r"(const2), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi", "$ac2lo", "$ac2hi", "$ac3lo",
-          "$ac3hi");
-  }
-}
-
-void ARGBToUVRow_DSPR2(const uint8* src_rgb0,
-                       int src_stride_rgb,
-                       uint8* dst_u,
-                       uint8* dst_v,
-                       int width) {
-  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;
-  int x;
-  int const1 = 0xffb60070;
-  int const2 = 0x0000ffda;
-  int const3 = 0xffa2ffee;
-  int const4 = 0x00000070;
-  int const5 = 0x100;
-
-  for (x = 0; x < width - 1; x += 2) {
-    int tmp_t1, tmp_t2, tmp_t3, tmp_t4, tmp_t5;
-    int tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t1],   0(%[src_rgb0])            \n"
-        "lw                %[tmp_t2],   4(%[src_rgb0])            \n"
-        "lw                %[tmp_t3],   0(%[src_rgb1])            \n"
-        "lw                %[tmp_t4],   4(%[src_rgb1])            \n"
-        "preceu.ph.qbr     %[tmp_t5],   %[tmp_t1]                 \n"
-        "preceu.ph.qbl     %[tmp_t1],   %[tmp_t1]                 \n"
-        "preceu.ph.qbr     %[tmp_t6],   %[tmp_t2]                 \n"
-        "preceu.ph.qbl     %[tmp_t2],   %[tmp_t2]                 \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t3]                 \n"
-        "preceu.ph.qbl     %[tmp_t3],   %[tmp_t3]                 \n"
-        "preceu.ph.qbr     %[tmp_t8],   %[tmp_t4]                 \n"
-        "preceu.ph.qbl     %[tmp_t4],   %[tmp_t4]                 \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t6]  \n"
-        "addu.ph           %[tmp_t7],   %[tmp_t7],     %[tmp_t8]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t2]  \n"
-        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t4]  \n"
-        "addu.ph           %[tmp_t5],   %[tmp_t5],     %[tmp_t7]  \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t3]  \n"
-        "shrl.ph           %[tmp_t5],   %[tmp_t5],     2          \n"
-        "shrl.ph           %[tmp_t1],   %[tmp_t1],     2          \n"
-        "mult              $ac0,        %[const5],     %[const5]  \n"
-        "mult              $ac1,        %[const5],     %[const5]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t5],     %[const1]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t5],     %[const3]  \n"
-        "dpaq_s.w.ph       $ac0,        %[tmp_t1],     %[const2]  \n"
-        "dpaq_s.w.ph       $ac1,        %[tmp_t1],     %[const4]  \n"
-        "extr_r.w          %[tmp_t7],   $ac0,          9          \n"
-        "extr_r.w          %[tmp_t8],   $ac1,          9          \n"
-        "addiu             %[src_rgb0], %[src_rgb0], 8            \n"
-        "addiu             %[src_rgb1], %[src_rgb1], 8            \n"
-        "addiu             %[dst_u],    %[dst_u],    1            \n"
-        "addiu             %[dst_v],    %[dst_v],    1            \n"
-        "sb                %[tmp_t7],   -1(%[dst_u])              \n"
-        "sb                %[tmp_t8],   -1(%[dst_v])              \n"
-        ".set pop                                                 \n"
-        : [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2),
-          [tmp_t3] "=&r"(tmp_t3), [tmp_t4] "=&r"(tmp_t4),
-          [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-          [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8),
-          [src_rgb0] "+r"(src_rgb0), [src_rgb1] "+r"(src_rgb1),
-          [dst_u] "+r"(dst_u), [dst_v] "+r"(dst_v)
-        : [const1] "r"(const1), [const2] "r"(const2), [const3] "r"(const3),
-          [const4] "r"(const4), [const5] "r"(const5)
-        : "hi", "lo", "$ac1lo", "$ac1hi");
-  }
-}
-
-#endif  // __mips_dsp_rev >= 2
-
-#endif  // defined(__mips__)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/files/source/row_gcc.cc b/files/source/row_gcc.cc
index decd3d2e..dce8c439 100644
--- a/files/source/row_gcc.cc
+++ b/files/source/row_gcc.cc
@@ -9,25 +9,26 @@
  */
 
 #include "libyuv/row.h"
-
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
 
 // Constants for ARGB
-static const vec8 kARGBToY = {13, 65, 33, 0, 13, 65, 33, 0,
-                              13, 65, 33, 0, 13, 65, 33, 0};
+static const uvec8 kARGBToY = {25u, 129u, 66u, 0u, 25u, 129u, 66u, 0u,
+                               25u, 129u, 66u, 0u, 25u, 129u, 66u, 0u};
 
 // JPeg full range.
-static const vec8 kARGBToYJ = {15, 75, 38, 0, 15, 75, 38, 0,
-                               15, 75, 38, 0, 15, 75, 38, 0};
+static const uvec8 kARGBToYJ = {29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u,
+                                29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u};
+
+static const uvec8 kRGBAToYJ = {0u, 29u, 150u, 77u, 0u, 29u, 150u, 77u,
+                                0u, 29u, 150u, 77u, 0u, 29u, 150u, 77u};
 #endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
 
 #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
@@ -45,8 +46,8 @@ static const vec8 kARGBToVJ = {-20, -107, 127, 0, -20, -107, 127, 0,
                                -20, -107, 127, 0, -20, -107, 127, 0};
 
 // Constants for BGRA
-static const vec8 kBGRAToY = {0, 33, 65, 13, 0, 33, 65, 13,
-                              0, 33, 65, 13, 0, 33, 65, 13};
+static const uvec8 kBGRAToY = {0u, 66u, 129u, 25u, 0u, 66u, 129u, 25u,
+                               0u, 66u, 129u, 25u, 0u, 66u, 129u, 25u};
 
 static const vec8 kBGRAToU = {0, -38, -74, 112, 0, -38, -74, 112,
                               0, -38, -74, 112, 0, -38, -74, 112};
@@ -55,8 +56,8 @@ static const vec8 kBGRAToV = {0, 112, -94, -18, 0, 112, -94, -18,
                               0, 112, -94, -18, 0, 112, -94, -18};
 
 // Constants for ABGR
-static const vec8 kABGRToY = {33, 65, 13, 0, 33, 65, 13, 0,
-                              33, 65, 13, 0, 33, 65, 13, 0};
+static const uvec8 kABGRToY = {66u, 129u, 25u, 0u, 66u, 129u, 25u, 0u,
+                               66u, 129u, 25u, 0u, 66u, 129u, 25u, 0u};
 
 static const vec8 kABGRToU = {-38, -74, 112, 0, -38, -74, 112, 0,
                               -38, -74, 112, 0, -38, -74, 112, 0};
@@ -65,8 +66,8 @@ static const vec8 kABGRToV = {112, -94, -18, 0, 112, -94, -18, 0,
                               112, -94, -18, 0, 112, -94, -18, 0};
 
 // Constants for RGBA.
-static const vec8 kRGBAToY = {0, 13, 65, 33, 0, 13, 65, 33,
-                              0, 13, 65, 33, 0, 13, 65, 33};
+static const uvec8 kRGBAToY = {0u, 25u, 129u, 66u, 0u, 25u, 129u, 66u,
+                               0u, 25u, 129u, 66u, 0u, 25u, 129u, 66u};
 
 static const vec8 kRGBAToU = {0, 112, -74, -38, 0, 112, -74, -38,
                               0, 112, -74, -38, 0, 112, -74, -38};
@@ -74,17 +75,15 @@ static const vec8 kRGBAToU = {0, 112, -74, -38, 0, 112, -74, -38,
 static const vec8 kRGBAToV = {0, -18, -94, 112, 0, -18, -94, 112,
                               0, -18, -94, 112, 0, -18, -94, 112};
 
-static const uvec8 kAddY16 = {16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
-                              16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u};
-
-// 7 bit fixed point 0.5.
-static const vec16 kAddYJ64 = {64, 64, 64, 64, 64, 64, 64, 64};
+static const uvec16 kAddY16 = {0x7e80u, 0x7e80u, 0x7e80u, 0x7e80u,
+                               0x7e80u, 0x7e80u, 0x7e80u, 0x7e80u};
 
 static const uvec8 kAddUV128 = {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
                                 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
 
-static const uvec16 kAddUVJ128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u,
-                                  0x8080u, 0x8080u, 0x8080u, 0x8080u};
+static const uvec16 kSub128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u,
+                               0x8080u, 0x8080u, 0x8080u, 0x8080u};
+
 #endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
 
 #ifdef HAS_RGB24TOARGBROW_SSSE3
@@ -97,6 +96,10 @@ static const uvec8 kShuffleMaskRGB24ToARGB = {
 static const uvec8 kShuffleMaskRAWToARGB = {2u, 1u, 0u, 12u, 5u,  4u,  3u, 13u,
                                             8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u};
 
+// Shuffle table for converting RAW to RGBA.
+static const uvec8 kShuffleMaskRAWToRGBA = {12u, 2u, 1u, 0u, 13u, 5u,  4u,  3u,
+                                            14u, 8u, 7u, 6u, 15u, 11u, 10u, 9u};
+
 // Shuffle table for converting RAW to RGB24.  First 8.
 static const uvec8 kShuffleMaskRAWToRGB24_0 = {
     2u,   1u,   0u,   5u,   4u,   3u,   8u,   7u,
@@ -154,24 +157,24 @@ static const lvec8 kShuffleNV21 = {
 #ifdef HAS_J400TOARGBROW_SSE2
 void J400ToARGBRow_SSE2(const uint8_t* src_y, uint8_t* dst_argb, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "pslld     $0x18,%%xmm5                    \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm0                     \n"
-      "lea       0x8(%0),%0                      \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklwd %%xmm0,%%xmm0                   \n"
-      "punpckhwd %%xmm1,%%xmm1                   \n"
-      "por       %%xmm5,%%xmm0                   \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklwd   %%xmm0,%%xmm0                 \n"
+      "punpckhwd   %%xmm1,%%xmm1                 \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -185,35 +188,35 @@ void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"  // 0xff000000
-      "pslld     $0x18,%%xmm5                    \n"
-      "movdqa    %3,%%xmm4                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm3                 \n"
-      "lea       0x30(%0),%0                     \n"
-      "movdqa    %%xmm3,%%xmm2                   \n"
-      "palignr   $0x8,%%xmm1,%%xmm2              \n"
-      "pshufb    %%xmm4,%%xmm2                   \n"
-      "por       %%xmm5,%%xmm2                   \n"
-      "palignr   $0xc,%%xmm0,%%xmm1              \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "movdqu    %%xmm2,0x20(%1)                 \n"
-      "por       %%xmm5,%%xmm0                   \n"
-      "pshufb    %%xmm4,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "palignr   $0x4,%%xmm3,%%xmm3              \n"
-      "pshufb    %%xmm4,%%xmm3                   \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "por       %%xmm5,%%xmm3                   \n"
-      "movdqu    %%xmm3,0x30(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_rgb24),              // %0
         "+r"(dst_argb),               // %1
         "+r"(width)                   // %2
@@ -223,35 +226,35 @@ void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24,
 
 void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"  // 0xff000000
-      "pslld     $0x18,%%xmm5                    \n"
-      "movdqa    %3,%%xmm4                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm3                 \n"
-      "lea       0x30(%0),%0                     \n"
-      "movdqa    %%xmm3,%%xmm2                   \n"
-      "palignr   $0x8,%%xmm1,%%xmm2              \n"
-      "pshufb    %%xmm4,%%xmm2                   \n"
-      "por       %%xmm5,%%xmm2                   \n"
-      "palignr   $0xc,%%xmm0,%%xmm1              \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "movdqu    %%xmm2,0x20(%1)                 \n"
-      "por       %%xmm5,%%xmm0                   \n"
-      "pshufb    %%xmm4,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "palignr   $0x4,%%xmm3,%%xmm3              \n"
-      "pshufb    %%xmm4,%%xmm3                   \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "por       %%xmm5,%%xmm3                   \n"
-      "movdqu    %%xmm3,0x30(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_raw),              // %0
         "+r"(dst_argb),             // %1
         "+r"(width)                 // %2
@@ -259,29 +262,68 @@ void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
+// Same code as RAWToARGB with different shuffler and A in low bits
+void RAWToRGBARow_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // 0x000000ff
+      "psrld       $0x18,%%xmm5                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm3               \n"
+      "lea         0x30(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "palignr     $0x8,%%xmm1,%%xmm2            \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "palignr     $0x4,%%xmm3,%%xmm3            \n"
+      "pshufb      %%xmm4,%%xmm3                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm3,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_raw),              // %0
+        "+r"(dst_rgba),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleMaskRAWToRGBA)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+
 void RAWToRGB24Row_SSSE3(const uint8_t* src_raw,
                          uint8_t* dst_rgb24,
                          int width) {
   asm volatile(
-      "movdqa     %3,%%xmm3                       \n"
-      "movdqa     %4,%%xmm4                       \n"
-      "movdqa     %5,%%xmm5                       \n"
+      "movdqa      %3,%%xmm3                     \n"
+      "movdqa      %4,%%xmm4                     \n"
+      "movdqa      %5,%%xmm5                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x4(%0),%%xmm1                  \n"
-      "movdqu    0x8(%0),%%xmm2                  \n"
-      "lea       0x18(%0),%0                     \n"
-      "pshufb    %%xmm3,%%xmm0                   \n"
-      "pshufb    %%xmm4,%%xmm1                   \n"
-      "pshufb    %%xmm5,%%xmm2                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq      %%xmm1,0x8(%1)                  \n"
-      "movq      %%xmm2,0x10(%1)                 \n"
-      "lea       0x18(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x4(%0),%%xmm1                \n"
+      "movdqu      0x8(%0),%%xmm2                \n"
+      "lea         0x18(%0),%0                   \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x8(%1)                \n"
+      "movq        %%xmm2,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_raw),                  // %0
         "+r"(dst_rgb24),                // %1
         "+r"(width)                     // %2
@@ -293,44 +335,44 @@ void RAWToRGB24Row_SSSE3(const uint8_t* src_raw,
 
 void RGB565ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "mov       $0x1080108,%%eax                \n"
-      "movd      %%eax,%%xmm5                    \n"
-      "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-      "mov       $0x20802080,%%eax               \n"
-      "movd      %%eax,%%xmm6                    \n"
-      "pshufd    $0x0,%%xmm6,%%xmm6              \n"
-      "pcmpeqb   %%xmm3,%%xmm3                   \n"
-      "psllw     $0xb,%%xmm3                     \n"
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "psllw     $0xa,%%xmm4                     \n"
-      "psrlw     $0x5,%%xmm4                     \n"
-      "pcmpeqb   %%xmm7,%%xmm7                   \n"
-      "psllw     $0x8,%%xmm7                     \n"
-      "sub       %0,%1                           \n"
-      "sub       %0,%1                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "pand      %%xmm3,%%xmm1                   \n"
-      "psllw     $0xb,%%xmm2                     \n"
-      "pmulhuw   %%xmm5,%%xmm1                   \n"
-      "pmulhuw   %%xmm5,%%xmm2                   \n"
-      "psllw     $0x8,%%xmm1                     \n"
-      "por       %%xmm2,%%xmm1                   \n"
-      "pand      %%xmm4,%%xmm0                   \n"
-      "pmulhuw   %%xmm6,%%xmm0                   \n"
-      "por       %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "punpcklbw %%xmm0,%%xmm1                   \n"
-      "punpckhbw %%xmm0,%%xmm2                   \n"
-      "movdqu    %%xmm1,0x00(%1,%0,2)            \n"
-      "movdqu    %%xmm2,0x10(%1,%0,2)            \n"
-      "lea       0x10(%0),%0                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "mov         $0x1080108,%%eax              \n"
+      "movd        %%eax,%%xmm5                  \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x20802080,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psllw       $0xb,%%xmm3                   \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0xa,%%xmm4                   \n"
+      "psrlw       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psllw       $0x8,%%xmm7                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "psllw       $0xb,%%xmm2                   \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "psllw       $0x8,%%xmm1                   \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "pmulhuw     %%xmm6,%%xmm0                 \n"
+      "por         %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm2,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -341,47 +383,47 @@ void RGB565ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 
 void ARGB1555ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "mov       $0x1080108,%%eax                \n"
-      "movd      %%eax,%%xmm5                    \n"
-      "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-      "mov       $0x42004200,%%eax               \n"
-      "movd      %%eax,%%xmm6                    \n"
-      "pshufd    $0x0,%%xmm6,%%xmm6              \n"
-      "pcmpeqb   %%xmm3,%%xmm3                   \n"
-      "psllw     $0xb,%%xmm3                     \n"
-      "movdqa    %%xmm3,%%xmm4                   \n"
-      "psrlw     $0x6,%%xmm4                     \n"
-      "pcmpeqb   %%xmm7,%%xmm7                   \n"
-      "psllw     $0x8,%%xmm7                     \n"
-      "sub       %0,%1                           \n"
-      "sub       %0,%1                           \n"
+      "mov         $0x1080108,%%eax              \n"
+      "movd        %%eax,%%xmm5                  \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x42004200,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psllw       $0xb,%%xmm3                   \n"
+      "movdqa      %%xmm3,%%xmm4                 \n"
+      "psrlw       $0x6,%%xmm4                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psllw       $0x8,%%xmm7                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "psllw     $0x1,%%xmm1                     \n"
-      "psllw     $0xb,%%xmm2                     \n"
-      "pand      %%xmm3,%%xmm1                   \n"
-      "pmulhuw   %%xmm5,%%xmm2                   \n"
-      "pmulhuw   %%xmm5,%%xmm1                   \n"
-      "psllw     $0x8,%%xmm1                     \n"
-      "por       %%xmm2,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "pand      %%xmm4,%%xmm0                   \n"
-      "psraw     $0x8,%%xmm2                     \n"
-      "pmulhuw   %%xmm6,%%xmm0                   \n"
-      "pand      %%xmm7,%%xmm2                   \n"
-      "por       %%xmm2,%%xmm0                   \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "punpcklbw %%xmm0,%%xmm1                   \n"
-      "punpckhbw %%xmm0,%%xmm2                   \n"
-      "movdqu    %%xmm1,0x00(%1,%0,2)            \n"
-      "movdqu    %%xmm2,0x10(%1,%0,2)            \n"
-      "lea       0x10(%0),%0                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "psllw       $0x1,%%xmm1                   \n"
+      "psllw       $0xb,%%xmm2                   \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "psllw       $0x8,%%xmm1                   \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "pmulhuw     %%xmm6,%%xmm0                 \n"
+      "pand        %%xmm7,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm2,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -392,34 +434,34 @@ void ARGB1555ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 
 void ARGB4444ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "mov       $0xf0f0f0f,%%eax                \n"
-      "movd      %%eax,%%xmm4                    \n"
-      "pshufd    $0x0,%%xmm4,%%xmm4              \n"
-      "movdqa    %%xmm4,%%xmm5                   \n"
-      "pslld     $0x4,%%xmm5                     \n"
-      "sub       %0,%1                           \n"
-      "sub       %0,%1                           \n"
+      "mov         $0xf0f0f0f,%%eax              \n"
+      "movd        %%eax,%%xmm4                  \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "pslld       $0x4,%%xmm5                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "pand      %%xmm4,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm3                   \n"
-      "psllw     $0x4,%%xmm1                     \n"
-      "psrlw     $0x4,%%xmm3                     \n"
-      "por       %%xmm1,%%xmm0                   \n"
-      "por       %%xmm3,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklbw %%xmm2,%%xmm0                   \n"
-      "punpckhbw %%xmm2,%%xmm1                   \n"
-      "movdqu    %%xmm0,0x00(%1,%0,2)            \n"
-      "movdqu    %%xmm1,0x10(%1,%0,2)            \n"
-      "lea       0x10(%0),%0                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "psllw       $0x4,%%xmm1                   \n"
+      "psrlw       $0x4,%%xmm3                   \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,0x00(%1,%0,2)          \n"
+      "movdqu      %%xmm1,0x10(%1,%0,2)          \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -430,35 +472,35 @@ void ARGB4444ToARGBRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
 
-      "movdqa    %3,%%xmm6                       \n"
+      "movdqa      %3,%%xmm6                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "pshufb    %%xmm6,%%xmm0                   \n"
-      "pshufb    %%xmm6,%%xmm1                   \n"
-      "pshufb    %%xmm6,%%xmm2                   \n"
-      "pshufb    %%xmm6,%%xmm3                   \n"
-      "movdqa    %%xmm1,%%xmm4                   \n"
-      "psrldq    $0x4,%%xmm1                     \n"
-      "pslldq    $0xc,%%xmm4                     \n"
-      "movdqa    %%xmm2,%%xmm5                   \n"
-      "por       %%xmm4,%%xmm0                   \n"
-      "pslldq    $0x8,%%xmm5                     \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "psrldq    $0x8,%%xmm2                     \n"
-      "pslldq    $0x4,%%xmm3                     \n"
-      "por       %%xmm3,%%xmm2                   \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "movdqu    %%xmm2,0x20(%1)                 \n"
-      "lea       0x30(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "lea         0x40(%0),%0                   \n"
+      "pshufb      %%xmm6,%%xmm0                 \n"
+      "pshufb      %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm2                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "psrldq      $0x4,%%xmm1                   \n"
+      "pslldq      $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pslldq      $0x8,%%xmm5                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "psrldq      $0x8,%%xmm2                   \n"
+      "pslldq      $0x4,%%xmm3                   \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src),                    // %0
         "+r"(dst),                    // %1
         "+r"(width)                   // %2
@@ -469,35 +511,35 @@ void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
 void ARGBToRAWRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
 
-      "movdqa    %3,%%xmm6                       \n"
+      "movdqa      %3,%%xmm6                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "pshufb    %%xmm6,%%xmm0                   \n"
-      "pshufb    %%xmm6,%%xmm1                   \n"
-      "pshufb    %%xmm6,%%xmm2                   \n"
-      "pshufb    %%xmm6,%%xmm3                   \n"
-      "movdqa    %%xmm1,%%xmm4                   \n"
-      "psrldq    $0x4,%%xmm1                     \n"
-      "pslldq    $0xc,%%xmm4                     \n"
-      "movdqa    %%xmm2,%%xmm5                   \n"
-      "por       %%xmm4,%%xmm0                   \n"
-      "pslldq    $0x8,%%xmm5                     \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "psrldq    $0x8,%%xmm2                     \n"
-      "pslldq    $0x4,%%xmm3                     \n"
-      "por       %%xmm3,%%xmm2                   \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "movdqu    %%xmm2,0x20(%1)                 \n"
-      "lea       0x30(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "lea         0x40(%0),%0                   \n"
+      "pshufb      %%xmm6,%%xmm0                 \n"
+      "pshufb      %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm2                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "psrldq      $0x4,%%xmm1                   \n"
+      "pslldq      $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pslldq      $0x8,%%xmm5                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "psrldq      $0x8,%%xmm2                   \n"
+      "pslldq      $0x4,%%xmm3                   \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "movdqu      %%xmm2,0x20(%1)               \n"
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src),                  // %0
         "+r"(dst),                  // %1
         "+r"(width)                 // %2
@@ -512,37 +554,37 @@ static const lvec32 kPermdRGB24_AVX = {0, 1, 2, 4, 5, 6, 3, 7};
 void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm6                  \n"
-      "vmovdqa    %4,%%ymm7                      \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "lea        0x80(%0),%0                    \n"
-      "vpshufb    %%ymm6,%%ymm0,%%ymm0           \n"  // xxx0yyy0
-      "vpshufb    %%ymm6,%%ymm1,%%ymm1           \n"
-      "vpshufb    %%ymm6,%%ymm2,%%ymm2           \n"
-      "vpshufb    %%ymm6,%%ymm3,%%ymm3           \n"
-      "vpermd     %%ymm0,%%ymm7,%%ymm0           \n"  // pack to 24 bytes
-      "vpermd     %%ymm1,%%ymm7,%%ymm1           \n"
-      "vpermd     %%ymm2,%%ymm7,%%ymm2           \n"
-      "vpermd     %%ymm3,%%ymm7,%%ymm3           \n"
-      "vpermq     $0x3f,%%ymm1,%%ymm4            \n"  // combine 24 + 8
-      "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "vpermq     $0xf9,%%ymm1,%%ymm1            \n"  // combine 16 + 16
-      "vpermq     $0x4f,%%ymm2,%%ymm4            \n"
-      "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
-      "vmovdqu    %%ymm1,0x20(%1)                \n"
-      "vpermq     $0xfe,%%ymm2,%%ymm2            \n"  // combine 8 + 24
-      "vpermq     $0x93,%%ymm3,%%ymm3            \n"
-      "vpor       %%ymm3,%%ymm2,%%ymm2           \n"
-      "vmovdqu    %%ymm2,0x40(%1)                \n"
-      "lea        0x60(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqa     %4,%%ymm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpshufb     %%ymm6,%%ymm0,%%ymm0          \n"  // xxx0yyy0
+      "vpshufb     %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm7,%%ymm0          \n"  // pack to 24 bytes
+      "vpermd      %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpermq      $0x3f,%%ymm1,%%ymm4           \n"  // combine 24 + 8
+      "vpor        %%ymm4,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vpermq      $0xf9,%%ymm1,%%ymm1           \n"  // combine 16 + 16
+      "vpermq      $0x4f,%%ymm2,%%ymm4           \n"
+      "vpor        %%ymm4,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vpermq      $0xfe,%%ymm2,%%ymm2           \n"  // combine 8 + 24
+      "vpermq      $0x93,%%ymm3,%%ymm3           \n"
+      "vpor        %%ymm3,%%ymm2,%%ymm2          \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),                     // %0
         "+r"(dst),                     // %1
@@ -571,26 +613,26 @@ static const ulvec8 kPermARGBToRGB24_2 = {
 
 void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "vmovdqa    %3,%%ymm5                      \n"
-      "vmovdqa    %4,%%ymm6                      \n"
-      "vmovdqa    %5,%%ymm7                      \n"
+      "vmovdqa     %3,%%ymm5                     \n"
+      "vmovdqa     %4,%%ymm6                     \n"
+      "vmovdqa     %5,%%ymm7                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "lea        0x80(%0),%0                    \n"
-      "vpermt2b   %%ymm1,%%ymm5,%%ymm0           \n"
-      "vpermt2b   %%ymm2,%%ymm6,%%ymm1           \n"
-      "vpermt2b   %%ymm3,%%ymm7,%%ymm2           \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "vmovdqu    %%ymm1,0x20(%1)                \n"
-      "vmovdqu    %%ymm2,0x40(%1)                \n"
-      "lea        0x60(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpermt2b    %%ymm1,%%ymm5,%%ymm0          \n"
+      "vpermt2b    %%ymm2,%%ymm6,%%ymm1          \n"
+      "vpermt2b    %%ymm3,%%ymm7,%%ymm2          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),                // %0
         "+r"(dst),                // %1
@@ -606,37 +648,37 @@ void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) {
 void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm6                  \n"
-      "vmovdqa    %4,%%ymm7                      \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "lea        0x80(%0),%0                    \n"
-      "vpshufb    %%ymm6,%%ymm0,%%ymm0           \n"  // xxx0yyy0
-      "vpshufb    %%ymm6,%%ymm1,%%ymm1           \n"
-      "vpshufb    %%ymm6,%%ymm2,%%ymm2           \n"
-      "vpshufb    %%ymm6,%%ymm3,%%ymm3           \n"
-      "vpermd     %%ymm0,%%ymm7,%%ymm0           \n"  // pack to 24 bytes
-      "vpermd     %%ymm1,%%ymm7,%%ymm1           \n"
-      "vpermd     %%ymm2,%%ymm7,%%ymm2           \n"
-      "vpermd     %%ymm3,%%ymm7,%%ymm3           \n"
-      "vpermq     $0x3f,%%ymm1,%%ymm4            \n"  // combine 24 + 8
-      "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "vpermq     $0xf9,%%ymm1,%%ymm1            \n"  // combine 16 + 16
-      "vpermq     $0x4f,%%ymm2,%%ymm4            \n"
-      "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
-      "vmovdqu    %%ymm1,0x20(%1)                \n"
-      "vpermq     $0xfe,%%ymm2,%%ymm2            \n"  // combine 8 + 24
-      "vpermq     $0x93,%%ymm3,%%ymm3            \n"
-      "vpor       %%ymm3,%%ymm2,%%ymm2           \n"
-      "vmovdqu    %%ymm2,0x40(%1)                \n"
-      "lea        0x60(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqa     %4,%%ymm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "lea         0x80(%0),%0                   \n"
+      "vpshufb     %%ymm6,%%ymm0,%%ymm0          \n"  // xxx0yyy0
+      "vpshufb     %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm7,%%ymm0          \n"  // pack to 24 bytes
+      "vpermd      %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpermq      $0x3f,%%ymm1,%%ymm4           \n"  // combine 24 + 8
+      "vpor        %%ymm4,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vpermq      $0xf9,%%ymm1,%%ymm1           \n"  // combine 16 + 16
+      "vpermq      $0x4f,%%ymm2,%%ymm4           \n"
+      "vpor        %%ymm4,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "vpermq      $0xfe,%%ymm2,%%ymm2           \n"  // combine 8 + 24
+      "vpermq      $0x93,%%ymm3,%%ymm3           \n"
+      "vpor        %%ymm3,%%ymm2,%%ymm2          \n"
+      "vmovdqu     %%ymm2,0x40(%1)               \n"
+      "lea         0x60(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),                   // %0
         "+r"(dst),                   // %1
@@ -650,34 +692,34 @@ void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 
 void ARGBToRGB565Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm3,%%xmm3                   \n"
-      "psrld     $0x1b,%%xmm3                    \n"
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "psrld     $0x1a,%%xmm4                    \n"
-      "pslld     $0x5,%%xmm4                     \n"
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "pslld     $0xb,%%xmm5                     \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psrld       $0x1b,%%xmm3                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1a,%%xmm4                  \n"
+      "pslld       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0xb,%%xmm5                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "pslld     $0x8,%%xmm0                     \n"
-      "psrld     $0x3,%%xmm1                     \n"
-      "psrld     $0x5,%%xmm2                     \n"
-      "psrad     $0x10,%%xmm0                    \n"
-      "pand      %%xmm3,%%xmm1                   \n"
-      "pand      %%xmm4,%%xmm2                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "por       %%xmm2,%%xmm1                   \n"
-      "por       %%xmm1,%%xmm0                   \n"
-      "packssdw  %%xmm0,%%xmm0                   \n"
-      "lea       0x10(%0),%0                     \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pslld       $0x8,%%xmm0                   \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x5,%%xmm2                   \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm2                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -690,40 +732,40 @@ void ARGBToRGB565DitherRow_SSE2(const uint8_t* src,
                                 const uint32_t dither4,
                                 int width) {
   asm volatile(
-      "movd       %3,%%xmm6                      \n"
-      "punpcklbw  %%xmm6,%%xmm6                  \n"
-      "movdqa     %%xmm6,%%xmm7                  \n"
-      "punpcklwd  %%xmm6,%%xmm6                  \n"
-      "punpckhwd  %%xmm7,%%xmm7                  \n"
-      "pcmpeqb    %%xmm3,%%xmm3                  \n"
-      "psrld      $0x1b,%%xmm3                   \n"
-      "pcmpeqb    %%xmm4,%%xmm4                  \n"
-      "psrld      $0x1a,%%xmm4                   \n"
-      "pslld      $0x5,%%xmm4                    \n"
-      "pcmpeqb    %%xmm5,%%xmm5                  \n"
-      "pslld      $0xb,%%xmm5                    \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu     (%0),%%xmm0                    \n"
-      "paddusb    %%xmm6,%%xmm0                  \n"
-      "movdqa     %%xmm0,%%xmm1                  \n"
-      "movdqa     %%xmm0,%%xmm2                  \n"
-      "pslld      $0x8,%%xmm0                    \n"
-      "psrld      $0x3,%%xmm1                    \n"
-      "psrld      $0x5,%%xmm2                    \n"
-      "psrad      $0x10,%%xmm0                   \n"
-      "pand       %%xmm3,%%xmm1                  \n"
-      "pand       %%xmm4,%%xmm2                  \n"
-      "pand       %%xmm5,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm1                  \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "packssdw   %%xmm0,%%xmm0                  \n"
-      "lea        0x10(%0),%0                    \n"
-      "movq       %%xmm0,(%1)                    \n"
-      "lea        0x8(%1),%1                     \n"
-      "sub        $0x4,%2                        \n"
-      "jg        1b                              \n"
+      "movd        %3,%%xmm6                     \n"
+      "punpcklbw   %%xmm6,%%xmm6                 \n"
+      "movdqa      %%xmm6,%%xmm7                 \n"
+      "punpcklwd   %%xmm6,%%xmm6                 \n"
+      "punpckhwd   %%xmm7,%%xmm7                 \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "psrld       $0x1b,%%xmm3                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1a,%%xmm4                  \n"
+      "pslld       $0x5,%%xmm4                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0xb,%%xmm5                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "paddusb     %%xmm6,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "pslld       $0x8,%%xmm0                   \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x5,%%xmm2                   \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "pand        %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm4,%%xmm2                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm1                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),    // %0
         "+r"(dst),    // %1
         "+r"(width)   // %2
@@ -739,35 +781,35 @@ void ARGBToRGB565DitherRow_AVX2(const uint8_t* src,
                                 int width) {
   asm volatile(
       "vbroadcastss %3,%%xmm6                    \n"
-      "vpunpcklbw %%xmm6,%%xmm6,%%xmm6           \n"
-      "vpermq     $0xd8,%%ymm6,%%ymm6            \n"
-      "vpunpcklwd %%ymm6,%%ymm6,%%ymm6           \n"
-      "vpcmpeqb   %%ymm3,%%ymm3,%%ymm3           \n"
-      "vpsrld     $0x1b,%%ymm3,%%ymm3            \n"
-      "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpsrld     $0x1a,%%ymm4,%%ymm4            \n"
-      "vpslld     $0x5,%%ymm4,%%ymm4             \n"
-      "vpslld     $0xb,%%ymm3,%%ymm5             \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vpaddusb   %%ymm6,%%ymm0,%%ymm0           \n"
-      "vpsrld     $0x5,%%ymm0,%%ymm2             \n"
-      "vpsrld     $0x3,%%ymm0,%%ymm1             \n"
-      "vpsrld     $0x8,%%ymm0,%%ymm0             \n"
-      "vpand      %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpand      %%ymm3,%%ymm1,%%ymm1           \n"
-      "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpor       %%ymm2,%%ymm1,%%ymm1           \n"
-      "vpor       %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpackusdw  %%ymm0,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "lea        0x20(%0),%0                    \n"
-      "vmovdqu    %%xmm0,(%1)                    \n"
-      "lea        0x10(%1),%1                    \n"
-      "sub        $0x8,%2                        \n"
-      "jg         1b                             \n"
+      "vpunpcklbw  %%xmm6,%%xmm6,%%xmm6          \n"
+      "vpermq      $0xd8,%%ymm6,%%ymm6           \n"
+      "vpunpcklwd  %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpsrld      $0x1b,%%ymm3,%%ymm3           \n"
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $0x1a,%%ymm4,%%ymm4           \n"
+      "vpslld      $0x5,%%ymm4,%%ymm4            \n"
+      "vpslld      $0xb,%%ymm3,%%ymm5            \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpaddusb    %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x5,%%ymm0,%%ymm2            \n"
+      "vpsrld      $0x3,%%ymm0,%%ymm1            \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
+      "vpand       %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpand       %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpor        %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "lea         0x20(%0),%0                   \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),    // %0
         "+r"(dst),    // %1
@@ -780,38 +822,38 @@ void ARGBToRGB565DitherRow_AVX2(const uint8_t* src,
 
 void ARGBToARGB1555Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "psrld     $0x1b,%%xmm4                    \n"
-      "movdqa    %%xmm4,%%xmm5                   \n"
-      "pslld     $0x5,%%xmm5                     \n"
-      "movdqa    %%xmm4,%%xmm6                   \n"
-      "pslld     $0xa,%%xmm6                     \n"
-      "pcmpeqb   %%xmm7,%%xmm7                   \n"
-      "pslld     $0xf,%%xmm7                     \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrld       $0x1b,%%xmm4                  \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "pslld       $0x5,%%xmm5                   \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "pslld       $0xa,%%xmm6                   \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "pslld       $0xf,%%xmm7                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm3                   \n"
-      "psrad     $0x10,%%xmm0                    \n"
-      "psrld     $0x3,%%xmm1                     \n"
-      "psrld     $0x6,%%xmm2                     \n"
-      "psrld     $0x9,%%xmm3                     \n"
-      "pand      %%xmm7,%%xmm0                   \n"
-      "pand      %%xmm4,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm2                   \n"
-      "pand      %%xmm6,%%xmm3                   \n"
-      "por       %%xmm1,%%xmm0                   \n"
-      "por       %%xmm3,%%xmm2                   \n"
-      "por       %%xmm2,%%xmm0                   \n"
-      "packssdw  %%xmm0,%%xmm0                   \n"
-      "lea       0x10(%0),%0                     \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "psrad       $0x10,%%xmm0                  \n"
+      "psrld       $0x3,%%xmm1                   \n"
+      "psrld       $0x6,%%xmm2                   \n"
+      "psrld       $0x9,%%xmm3                   \n"
+      "pand        %%xmm7,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm2                 \n"
+      "pand        %%xmm6,%%xmm3                 \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm3,%%xmm2                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -821,26 +863,26 @@ void ARGBToARGB1555Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 
 void ARGBToARGB4444Row_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "psllw     $0xc,%%xmm4                     \n"
-      "movdqa    %%xmm4,%%xmm3                   \n"
-      "psrlw     $0x8,%%xmm3                     \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0xc,%%xmm4                   \n"
+      "movdqa      %%xmm4,%%xmm3                 \n"
+      "psrlw       $0x8,%%xmm3                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pand      %%xmm3,%%xmm0                   \n"
-      "pand      %%xmm4,%%xmm1                   \n"
-      "psrlq     $0x4,%%xmm0                     \n"
-      "psrlq     $0x8,%%xmm1                     \n"
-      "por       %%xmm1,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "lea       0x10(%0),%0                     \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm3,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm1                 \n"
+      "psrlq       $0x4,%%xmm0                   \n"
+      "psrlq       $0x8,%%xmm1                   \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -884,31 +926,31 @@ static const uint32_t kMulAG10 = 64 * 65536 + 1028;
 
 void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "movdqa     %3,%%xmm2                     \n"  // shuffler for RB
-      "movd       %4,%%xmm3                     \n"  // multipler for RB
-      "movd       %5,%%xmm4                     \n"  // mask for R10 B10
-      "movd       %6,%%xmm5                     \n"  // mask for AG
-      "movd       %7,%%xmm6                     \n"  // multipler for AG
-      "pshufd     $0x0,%%xmm3,%%xmm3            \n"
-      "pshufd     $0x0,%%xmm4,%%xmm4            \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5            \n"
-      "pshufd     $0x0,%%xmm6,%%xmm6            \n"
-      "sub        %0,%1                         \n"
-
-      "1:                                       \n"
-      "movdqu     (%0),%%xmm0                   \n"  // fetch 4 ARGB pixels
-      "movdqa     %%xmm0,%%xmm1                 \n"
-      "pshufb     %%xmm2,%%xmm1                 \n"  // R0B0
-      "pand       %%xmm5,%%xmm0                 \n"  // A0G0
-      "pmulhuw    %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
-      "pmulhuw    %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
-      "pand       %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
-      "pslld      $10,%%xmm0                    \n"  // A2 x10 G10 x10
-      "por        %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
-      "movdqu     %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
-      "add        $0x10,%0                      \n"
-      "sub        $0x4,%2                       \n"
-      "jg         1b                            \n"
+      "movdqa      %3,%%xmm2                     \n"  // shuffler for RB
+      "movd        %4,%%xmm3                     \n"  // multipler for RB
+      "movd        %5,%%xmm4                     \n"  // mask for R10 B10
+      "movd        %6,%%xmm5                     \n"  // mask for AG
+      "movd        %7,%%xmm6                     \n"  // multipler for AG
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // fetch 4 ARGB pixels
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"  // R0B0
+      "pand        %%xmm5,%%xmm0                 \n"  // A0G0
+      "pmulhuw     %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
+      "pmulhuw     %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
+      "pand        %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
+      "pslld       $10,%%xmm0                    \n"  // A2 x10 G10 x10
+      "por         %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
+      "movdqu      %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
+      "add         $0x10,%0                      \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
 
       : "+r"(src),          // %0
         "+r"(dst),          // %1
@@ -923,31 +965,31 @@ void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
 
 void ABGRToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "movdqa     %3,%%xmm2                     \n"  // shuffler for RB
-      "movd       %4,%%xmm3                     \n"  // multipler for RB
-      "movd       %5,%%xmm4                     \n"  // mask for R10 B10
-      "movd       %6,%%xmm5                     \n"  // mask for AG
-      "movd       %7,%%xmm6                     \n"  // multipler for AG
-      "pshufd     $0x0,%%xmm3,%%xmm3            \n"
-      "pshufd     $0x0,%%xmm4,%%xmm4            \n"
-      "pshufd     $0x0,%%xmm5,%%xmm5            \n"
-      "pshufd     $0x0,%%xmm6,%%xmm6            \n"
-      "sub        %0,%1                         \n"
-
-      "1:                                       \n"
-      "movdqu     (%0),%%xmm0                   \n"  // fetch 4 ABGR pixels
-      "movdqa     %%xmm0,%%xmm1                 \n"
-      "pshufb     %%xmm2,%%xmm1                 \n"  // R0B0
-      "pand       %%xmm5,%%xmm0                 \n"  // A0G0
-      "pmulhuw    %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
-      "pmulhuw    %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
-      "pand       %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
-      "pslld      $10,%%xmm0                    \n"  // A2 x10 G10 x10
-      "por        %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
-      "movdqu     %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
-      "add        $0x10,%0                      \n"
-      "sub        $0x4,%2                       \n"
-      "jg         1b                            \n"
+      "movdqa      %3,%%xmm2                     \n"  // shuffler for RB
+      "movd        %4,%%xmm3                     \n"  // multipler for RB
+      "movd        %5,%%xmm4                     \n"  // mask for R10 B10
+      "movd        %6,%%xmm5                     \n"  // mask for AG
+      "movd        %7,%%xmm6                     \n"  // multipler for AG
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // fetch 4 ABGR pixels
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"  // R0B0
+      "pand        %%xmm5,%%xmm0                 \n"  // A0G0
+      "pmulhuw     %%xmm3,%%xmm1                 \n"  // X2 R16 X4  B10
+      "pmulhuw     %%xmm6,%%xmm0                 \n"  // X10 A2 X10 G10
+      "pand        %%xmm4,%%xmm1                 \n"  // X2 R10 X10 B10
+      "pslld       $10,%%xmm0                    \n"  // A2 x10 G10 x10
+      "por         %%xmm1,%%xmm0                 \n"  // A2 R10 G10 B10
+      "movdqu      %%xmm0,(%1,%0)                \n"  // store 4 AR30 pixels
+      "add         $0x10,%0                      \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
 
       : "+r"(src),          // %0
         "+r"(dst),          // %1
@@ -964,25 +1006,25 @@ void ABGRToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
 void ARGBToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm2                  \n"  // shuffler for RB
-      "vbroadcastss  %4,%%ymm3                   \n"  // multipler for RB
-      "vbroadcastss  %5,%%ymm4                   \n"  // mask for R10 B10
-      "vbroadcastss  %6,%%ymm5                   \n"  // mask for AG
-      "vbroadcastss  %7,%%ymm6                   \n"  // multipler for AG
-      "sub        %0,%1                          \n"
-
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"  // fetch 8 ARGB pixels
-      "vpshufb    %%ymm2,%%ymm0,%%ymm1           \n"  // R0B0
-      "vpand      %%ymm5,%%ymm0,%%ymm0           \n"  // A0G0
-      "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"  // X2 R16 X4  B10
-      "vpmulhuw   %%ymm6,%%ymm0,%%ymm0           \n"  // X10 A2 X10 G10
-      "vpand      %%ymm4,%%ymm1,%%ymm1           \n"  // X2 R10 X10 B10
-      "vpslld     $10,%%ymm0,%%ymm0              \n"  // A2 x10 G10 x10
-      "vpor       %%ymm1,%%ymm0,%%ymm0           \n"  // A2 R10 G10 B10
-      "vmovdqu    %%ymm0,(%1,%0)                 \n"  // store 8 AR30 pixels
-      "add        $0x20,%0                       \n"
-      "sub        $0x8,%2                        \n"
-      "jg         1b                             \n"
+      "vbroadcastss %4,%%ymm3                    \n"  // multipler for RB
+      "vbroadcastss %5,%%ymm4                    \n"  // mask for R10 B10
+      "vbroadcastss %6,%%ymm5                    \n"  // mask for AG
+      "vbroadcastss %7,%%ymm6                    \n"  // multipler for AG
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // fetch 8 ARGB pixels
+      "vpshufb     %%ymm2,%%ymm0,%%ymm1          \n"  // R0B0
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"  // A0G0
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"  // X2 R16 X4  B10
+      "vpmulhuw    %%ymm6,%%ymm0,%%ymm0          \n"  // X10 A2 X10 G10
+      "vpand       %%ymm4,%%ymm1,%%ymm1          \n"  // X2 R10 X10 B10
+      "vpslld      $10,%%ymm0,%%ymm0             \n"  // A2 x10 G10 x10
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // A2 R10 G10 B10
+      "vmovdqu     %%ymm0,(%1,%0)                \n"  // store 8 AR30 pixels
+      "add         $0x20,%0                      \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
 
       : "+r"(src),          // %0
@@ -1001,25 +1043,25 @@ void ARGBToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 void ABGRToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm2                  \n"  // shuffler for RB
-      "vbroadcastss  %4,%%ymm3                   \n"  // multipler for RB
-      "vbroadcastss  %5,%%ymm4                   \n"  // mask for R10 B10
-      "vbroadcastss  %6,%%ymm5                   \n"  // mask for AG
-      "vbroadcastss  %7,%%ymm6                   \n"  // multipler for AG
-      "sub        %0,%1                          \n"
-
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"  // fetch 8 ABGR pixels
-      "vpshufb    %%ymm2,%%ymm0,%%ymm1           \n"  // R0B0
-      "vpand      %%ymm5,%%ymm0,%%ymm0           \n"  // A0G0
-      "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"  // X2 R16 X4  B10
-      "vpmulhuw   %%ymm6,%%ymm0,%%ymm0           \n"  // X10 A2 X10 G10
-      "vpand      %%ymm4,%%ymm1,%%ymm1           \n"  // X2 R10 X10 B10
-      "vpslld     $10,%%ymm0,%%ymm0              \n"  // A2 x10 G10 x10
-      "vpor       %%ymm1,%%ymm0,%%ymm0           \n"  // A2 R10 G10 B10
-      "vmovdqu    %%ymm0,(%1,%0)                 \n"  // store 8 AR30 pixels
-      "add        $0x20,%0                       \n"
-      "sub        $0x8,%2                        \n"
-      "jg         1b                             \n"
+      "vbroadcastss %4,%%ymm3                    \n"  // multipler for RB
+      "vbroadcastss %5,%%ymm4                    \n"  // mask for R10 B10
+      "vbroadcastss %6,%%ymm5                    \n"  // mask for AG
+      "vbroadcastss %7,%%ymm6                    \n"  // multipler for AG
+      "sub         %0,%1                         \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // fetch 8 ABGR pixels
+      "vpshufb     %%ymm2,%%ymm0,%%ymm1          \n"  // R0B0
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"  // A0G0
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"  // X2 R16 X4  B10
+      "vpmulhuw    %%ymm6,%%ymm0,%%ymm0          \n"  // X10 A2 X10 G10
+      "vpand       %%ymm4,%%ymm1,%%ymm1          \n"  // X2 R10 X10 B10
+      "vpslld      $10,%%ymm0,%%ymm0             \n"  // A2 x10 G10 x10
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // A2 R10 G10 B10
+      "vmovdqu     %%ymm0,(%1,%0)                \n"  // store 8 AR30 pixels
+      "add         $0x20,%0                      \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
 
       : "+r"(src),          // %0
@@ -1034,222 +1076,490 @@ void ABGRToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 }
 #endif
 
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+static const uvec8 kShuffleARGBToAB64Lo = {2, 2, 1, 1, 0, 0, 3, 3,
+                                           6, 6, 5, 5, 4, 4, 7, 7};
+static const uvec8 kShuffleARGBToAB64Hi = {10, 10, 9,  9,  8,  8,  11, 11,
+                                           14, 14, 13, 13, 12, 12, 15, 15};
+
+void ARGBToAR64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ar64,
+                         int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+
+void ARGBToAB64Row_SSSE3(const uint8_t* src_argb,
+                         uint16_t* dst_ab64,
+                         int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm2                     \n"
+      "movdqa      %4,%%xmm3                     \n" LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm0                 \n"
+      "pshufb      %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_ab64),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleARGBToAB64Lo),  // %3
+        "m"(kShuffleARGBToAB64Hi)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+
+void AR64ToARGBRow_SSSE3(const uint16_t* src_ar64,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrlw       $8,%%xmm0                     \n"
+      "psrlw       $8,%%xmm1                     \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+
+void AB64ToARGBRow_SSSE3(const uint16_t* src_ab64,
+                         uint8_t* dst_argb,
+                         int width) {
+  asm volatile(
+
+      "movdqa      %3,%%xmm2                     \n" LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrlw       $8,%%xmm0                     \n"
+      "psrlw       $8,%%xmm1                     \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "pshufb      %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ab64),          // %0
+        "+r"(dst_argb),          // %1
+        "+r"(width)              // %2
+      : "m"(kShuffleARGBToABGR)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+
+#ifdef HAS_ARGBTOAR64ROW_AVX2
+void ARGBToAR64Row_AVX2(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+#endif
+
+#ifdef HAS_ARGBTOAB64ROW_AVX2
+void ARGBToAB64Row_AVX2(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm2                  \n"
+      "vbroadcastf128 %4,%%ymm3                  \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %%ymm3,%%ymm0,%%ymm1          \n"
+      "vpshufb     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_ab64),             // %1
+        "+r"(width)                 // %2
+      : "m"(kShuffleARGBToAB64Lo),  // %3
+        "m"(kShuffleARGBToAB64Hi)   // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_AR64TOARGBROW_AVX2
+void AR64ToARGBRow_AVX2(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpsrlw      $8,%%ymm0,%%ymm0              \n"
+      "vpsrlw      $8,%%ymm1,%%ymm1              \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x40(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
+#endif
+
+#ifdef HAS_AB64TOARGBROW_AVX2
+void AB64ToARGBRow_AVX2(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm2                  \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpsrlw      $8,%%ymm0,%%ymm0              \n"
+      "vpsrlw      $8,%%ymm1,%%ymm1              \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x40(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ab64),          // %0
+        "+r"(dst_argb),          // %1
+        "+r"(width)              // %2
+      : "m"(kShuffleARGBToABGR)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+// clang-format off
+
+// TODO(mraptis): Consider passing R, G, B multipliers as parameter.
+// round parameter is register containing value to add before shift.
+#define RGBTOY(round)                            \
+  "1:                                        \n" \
+  "movdqu    (%0),%%xmm0                     \n" \
+  "movdqu    0x10(%0),%%xmm1                 \n" \
+  "movdqu    0x20(%0),%%xmm2                 \n" \
+  "movdqu    0x30(%0),%%xmm3                 \n" \
+  "psubb     %%xmm5,%%xmm0                   \n" \
+  "psubb     %%xmm5,%%xmm1                   \n" \
+  "psubb     %%xmm5,%%xmm2                   \n" \
+  "psubb     %%xmm5,%%xmm3                   \n" \
+  "movdqu    %%xmm4,%%xmm6                   \n" \
+  "pmaddubsw %%xmm0,%%xmm6                   \n" \
+  "movdqu    %%xmm4,%%xmm0                   \n" \
+  "pmaddubsw %%xmm1,%%xmm0                   \n" \
+  "movdqu    %%xmm4,%%xmm1                   \n" \
+  "pmaddubsw %%xmm2,%%xmm1                   \n" \
+  "movdqu    %%xmm4,%%xmm2                   \n" \
+  "pmaddubsw %%xmm3,%%xmm2                   \n" \
+  "lea       0x40(%0),%0                     \n" \
+  "phaddw    %%xmm0,%%xmm6                   \n" \
+  "phaddw    %%xmm2,%%xmm1                   \n" \
+  "prefetcht0 1280(%0)                       \n" \
+  "paddw     %%" #round ",%%xmm6             \n" \
+  "paddw     %%" #round ",%%xmm1             \n" \
+  "psrlw     $0x8,%%xmm6                     \n" \
+  "psrlw     $0x8,%%xmm1                     \n" \
+  "packuswb  %%xmm1,%%xmm6                   \n" \
+  "movdqu    %%xmm6,(%1)                     \n" \
+  "lea       0x10(%1),%1                     \n" \
+  "sub       $0x10,%2                        \n" \
+  "jg        1b                              \n"
+
+#define RGBTOY_AVX2(round)                                       \
+  "1:                                        \n"                 \
+  "vmovdqu    (%0),%%ymm0                    \n"                 \
+  "vmovdqu    0x20(%0),%%ymm1                \n"                 \
+  "vmovdqu    0x40(%0),%%ymm2                \n"                 \
+  "vmovdqu    0x60(%0),%%ymm3                \n"                 \
+  "vpsubb     %%ymm5, %%ymm0, %%ymm0         \n"                 \
+  "vpsubb     %%ymm5, %%ymm1, %%ymm1         \n"                 \
+  "vpsubb     %%ymm5, %%ymm2, %%ymm2         \n"                 \
+  "vpsubb     %%ymm5, %%ymm3, %%ymm3         \n"                 \
+  "vpmaddubsw %%ymm0,%%ymm4,%%ymm0           \n"                 \
+  "vpmaddubsw %%ymm1,%%ymm4,%%ymm1           \n"                 \
+  "vpmaddubsw %%ymm2,%%ymm4,%%ymm2           \n"                 \
+  "vpmaddubsw %%ymm3,%%ymm4,%%ymm3           \n"                 \
+  "lea       0x80(%0),%0                     \n"                 \
+  "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n" /* mutates. */  \
+  "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"                 \
+  "prefetcht0 1280(%0)                       \n"                 \
+  "vpaddw     %%" #round ",%%ymm0,%%ymm0     \n" /* Add .5 for rounding. */             \
+  "vpaddw     %%" #round ",%%ymm2,%%ymm2     \n" \
+  "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"                 \
+  "vpsrlw     $0x8,%%ymm2,%%ymm2             \n"                 \
+  "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n" /* mutates. */  \
+  "vpermd     %%ymm0,%%ymm6,%%ymm0           \n" /* unmutate. */ \
+  "vmovdqu    %%ymm0,(%1)                    \n"                 \
+  "lea       0x20(%1),%1                     \n"                 \
+  "sub       $0x20,%2                        \n"                 \
+  "jg        1b                              \n"                 \
+  "vzeroupper                                \n"
+
+// clang-format on
+
 #ifdef HAS_ARGBTOYROW_SSSE3
 // Convert 16 ARGB pixels (64 bytes) to 16 Y values.
 void ARGBToYRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   asm volatile(
-      "movdqa    %3,%%xmm4                       \n"
-      "movdqa    %4,%%xmm5                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm3                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm3,%%xmm2                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
       : "+r"(src_argb),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
       : "m"(kARGBToY),   // %3
-        "m"(kAddY16)     // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYROW_SSSE3
 
 #ifdef HAS_ARGBTOYJROW_SSSE3
 // Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
-// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
+// Same as ARGBToYRow but different coefficients, no add 16.
 void ARGBToYJRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   asm volatile(
-      "movdqa    %3,%%xmm4                       \n"
-      "movdqa    %4,%%xmm5                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm3                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm3,%%xmm2                   \n"
-      "paddw     %%xmm5,%%xmm0                   \n"
-      "paddw     %%xmm5,%%xmm2                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN RGBTOY(xmm5)
       : "+r"(src_argb),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
       : "m"(kARGBToYJ),  // %3
-        "m"(kAddYJ64)    // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBTOYJROW_SSSE3
 
-#ifdef HAS_ARGBTOYROW_AVX2
+#ifdef HAS_RGBATOYJROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16.
+void RGBAToYJRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN RGBTOY(xmm5)
+      : "+r"(src_rgba),  // %0
+        "+r"(dst_y),     // %1
+        "+r"(width)      // %2
+      : "m"(kRGBAToYJ),  // %3
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif  // HAS_RGBATOYJROW_SSSE3
+
+#if defined(HAS_ARGBTOYROW_AVX2) || defined(HAS_ARGBEXTRACTALPHAROW_AVX2)
 // vpermd for vphaddw + vpackuswb vpermd.
 static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7};
+#endif
+
+#ifdef HAS_ARGBTOYROW_AVX2
 
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
 void ARGBToYRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm4                  \n"
       "vbroadcastf128 %4,%%ymm5                  \n"
-      "vmovdqu    %5,%%ymm6                      \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "lea       0x80(%0),%0                     \n"
-      "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
-      "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
-      "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
-      "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
-      "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
-      "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
-      "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"  // add 16 for Y
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
-      "vzeroupper                                \n"
+      "vbroadcastf128 %5,%%ymm7                  \n"
+      "vmovdqu     %6,%%ymm6                     \n"
+
+      LABELALIGN RGBTOY_AVX2(ymm7)
       : "+r"(src_argb),         // %0
         "+r"(dst_y),            // %1
         "+r"(width)             // %2
       : "m"(kARGBToY),          // %3
-        "m"(kAddY16),           // %4
-        "m"(kPermdARGBToY_AVX)  // %5
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+        "m"(kSub128),           // %4
+        "m"(kAddY16),           // %5
+        "m"(kPermdARGBToY_AVX)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYROW_AVX2
 
+#ifdef HAS_ABGRTOYROW_AVX2
+// Convert 32 ABGR pixels (128 bytes) to 32 Y values.
+void ABGRToYRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vbroadcastf128 %5,%%ymm7                  \n"
+      "vmovdqu     %6,%%ymm6                     \n"
+
+      LABELALIGN RGBTOY_AVX2(ymm7)
+      : "+r"(src_abgr),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kABGRToY),          // %3
+        "m"(kSub128),           // %4
+        "m"(kAddY16),           // %5
+        "m"(kPermdARGBToY_AVX)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOYROW_AVX2
+
 #ifdef HAS_ARGBTOYJROW_AVX2
 // Convert 32 ARGB pixels (128 bytes) to 32 Y values.
 void ARGBToYJRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm4                  \n"
       "vbroadcastf128 %4,%%ymm5                  \n"
-      "vmovdqu    %5,%%ymm6                      \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "lea       0x80(%0),%0                     \n"
-      "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
-      "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
-      "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"  // Add .5 for rounding.
-      "vpaddw     %%ymm5,%%ymm2,%%ymm2           \n"
-      "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
-      "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
-      "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
-      "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
-      "vzeroupper                                \n"
+      "vmovdqu     %5,%%ymm6                     \n"
+
+      LABELALIGN RGBTOY_AVX2(ymm5)
       : "+r"(src_argb),         // %0
         "+r"(dst_y),            // %1
         "+r"(width)             // %2
       : "m"(kARGBToYJ),         // %3
-        "m"(kAddYJ64),          // %4
+        "m"(kSub128),           // %4
         "m"(kPermdARGBToY_AVX)  // %5
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 #endif  // HAS_ARGBTOYJROW_AVX2
 
+#ifdef HAS_RGBATOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void RGBAToYJRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm4                  \n"
+      "vbroadcastf128 %4,%%ymm5                  \n"
+      "vmovdqu     %5,%%ymm6                     \n"
+
+      LABELALIGN RGBTOY_AVX2(
+      ymm5) "vzeroupper                                \n"
+      : "+r"(src_rgba),         // %0
+        "+r"(dst_y),            // %1
+        "+r"(width)             // %2
+      : "m"(kRGBAToYJ),         // %3
+        "m"(kSub128),           // %4
+        "m"(kPermdARGBToY_AVX)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif  // HAS_RGBATOYJROW_AVX2
+
 #ifdef HAS_ARGBTOUVROW_SSSE3
-void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
+void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width) {
   asm volatile(
-      "movdqa    %5,%%xmm3                       \n"
-      "movdqa    %6,%%xmm4                       \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x10(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm1                   \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x20(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "movdqu    0x30(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-
-      "lea       0x40(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm6,%%xmm2             \n"
-      "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm2,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm1                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm1                     \n"
-      "packsswb  %%xmm1,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movlps    %%xmm0,(%1)                     \n"
-      "movhps    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb0),                   // %0
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
@@ -1266,7 +1576,7 @@ void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
 static const lvec8 kShufARGBToUV_AVX = {
     0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
     0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
-void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
+void ARGBToUVRow_AVX2(const uint8_t* src_argb,
                       int src_stride_argb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
@@ -1275,46 +1585,46 @@ void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
       "vbroadcastf128 %5,%%ymm5                  \n"
       "vbroadcastf128 %6,%%ymm6                  \n"
       "vbroadcastf128 %7,%%ymm7                  \n"
-      "sub        %1,%2                          \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "vpavgb    0x00(%0,%4,1),%%ymm0,%%ymm0     \n"
-      "vpavgb    0x20(%0,%4,1),%%ymm1,%%ymm1     \n"
-      "vpavgb    0x40(%0,%4,1),%%ymm2,%%ymm2     \n"
-      "vpavgb    0x60(%0,%4,1),%%ymm3,%%ymm3     \n"
-      "lea        0x80(%0),%0                    \n"
-      "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
-      "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
-      "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
-      "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
-      "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
-      "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
-
-      "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
-      "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
-      "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
-      "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
-      "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
-      "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vpshufb    %8,%%ymm0,%%ymm0               \n"
-      "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+      "vpaddb      %%ymm5,%%ymm0,%%ymm0          \n"
 
       "vextractf128 $0x0,%%ymm0,(%1)             \n"
       "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
-      "lea        0x10(%1),%1                    \n"
-      "sub        $0x20,%3                       \n"
-      "jg         1b                             \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_argb0),                   // %0
+      : "+r"(src_argb),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
@@ -1328,8 +1638,71 @@ void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
 }
 #endif  // HAS_ARGBTOUVROW_AVX2
 
+#ifdef HAS_ABGRTOUVROW_AVX2
+void ABGRToUVRow_AVX2(const uint8_t* src_abgr,
+                      int src_stride_abgr,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vbroadcastf128 %6,%%ymm6                  \n"
+      "vbroadcastf128 %7,%%ymm7                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
+      "vpaddb      %%ymm5,%%ymm0,%%ymm0          \n"
+
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_abgr),                    // %0
+        "+r"(dst_u),                       // %1
+        "+r"(dst_v),                       // %2
+        "+rm"(width)                       // %3
+      : "r"((intptr_t)(src_stride_abgr)),  // %4
+        "m"(kAddUV128),                    // %5
+        "m"(kABGRToV),                     // %6
+        "m"(kABGRToU),                     // %7
+        "m"(kShufARGBToUV_AVX)             // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif  // HAS_ABGRTOUVROW_AVX2
+
 #ifdef HAS_ARGBTOUVJROW_AVX2
-void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
+void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
@@ -1338,52 +1711,52 @@ void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
       "vbroadcastf128 %5,%%ymm5                  \n"
       "vbroadcastf128 %6,%%ymm6                  \n"
       "vbroadcastf128 %7,%%ymm7                  \n"
-      "sub        %1,%2                          \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x40(%0),%%ymm2                \n"
-      "vmovdqu    0x60(%0),%%ymm3                \n"
-      "vpavgb    0x00(%0,%4,1),%%ymm0,%%ymm0     \n"
-      "vpavgb    0x20(%0,%4,1),%%ymm1,%%ymm1     \n"
-      "vpavgb    0x40(%0,%4,1),%%ymm2,%%ymm2     \n"
-      "vpavgb    0x60(%0,%4,1),%%ymm3,%%ymm3     \n"
-      "lea       0x80(%0),%0                     \n"
-      "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
-      "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
-      "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
-      "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
-      "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
-      "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
-
-      "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
-      "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
-      "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
-      "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
-      "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm5,%%ymm1,%%ymm1           \n"
-      "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
-      "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vpshufb    %8,%%ymm0,%%ymm0               \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x40(%0),%%ymm2               \n"
+      "vmovdqu     0x60(%0),%%ymm3               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "vpavgb      0x40(%0,%4,1),%%ymm2,%%ymm2   \n"
+      "vpavgb      0x60(%0,%4,1),%%ymm3,%%ymm3   \n"
+      "lea         0x80(%0),%0                   \n"
+      "vshufps     $0x88,%%ymm1,%%ymm0,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm1,%%ymm0,%%ymm0    \n"
+      "vpavgb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vshufps     $0x88,%%ymm3,%%ymm2,%%ymm4    \n"
+      "vshufps     $0xdd,%%ymm3,%%ymm2,%%ymm2    \n"
+      "vpavgb      %%ymm4,%%ymm2,%%ymm2          \n"
+
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm1          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm3          \n"
+      "vpmaddubsw  %%ymm6,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm6,%%ymm2,%%ymm2          \n"
+      "vphaddw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpsraw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsraw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpacksswb   %%ymm0,%%ymm1,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpshufb     %8,%%ymm0,%%ymm0              \n"
 
       "vextractf128 $0x0,%%ymm0,(%1)             \n"
       "vextractf128 $0x1,%%ymm0,0x0(%1,%2,1)     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_argb0),                   // %0
+      : "+r"(src_argb),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
       : "r"((intptr_t)(src_stride_argb)),  // %4
-        "m"(kAddUVJ128),                   // %5
+        "m"(kSub128),                      // %5
         "m"(kARGBToVJ),                    // %6
         "m"(kARGBToUJ),                    // %7
         "m"(kShufARGBToUV_AVX)             // %8
@@ -1393,67 +1766,67 @@ void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
 #endif  // HAS_ARGBTOUVJROW_AVX2
 
 #ifdef HAS_ARGBTOUVJROW_SSSE3
-void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
+void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
                         int src_stride_argb,
                         uint8_t* dst_u,
                         uint8_t* dst_v,
                         int width) {
   asm volatile(
-      "movdqa    %5,%%xmm3                       \n"
-      "movdqa    %6,%%xmm4                       \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x10(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm1                   \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x20(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "movdqu    0x30(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-
-      "lea       0x40(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm6,%%xmm2             \n"
-      "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm2,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm1                   \n"
-      "paddw     %%xmm5,%%xmm0                   \n"
-      "paddw     %%xmm5,%%xmm1                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm1                     \n"
-      "packsswb  %%xmm1,%%xmm0                   \n"
-      "movlps    %%xmm0,(%1)                     \n"
-      "movhps    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb0),                   // %0
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "paddw       %%xmm5,%%xmm0                 \n"
+      "paddw       %%xmm5,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
       : "r"((intptr_t)(src_stride_argb)),  // %4
         "m"(kARGBToVJ),                    // %5
         "m"(kARGBToUJ),                    // %6
-        "m"(kAddUVJ128)                    // %7
+        "m"(kSub128)                       // %7
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 #endif  // HAS_ARGBTOUVJROW_SSSE3
@@ -1464,47 +1837,47 @@ void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
                           uint8_t* dst_v,
                           int width) {
   asm volatile(
-      "movdqa    %4,%%xmm3                       \n"
-      "movdqa    %5,%%xmm4                       \n"
-      "movdqa    %6,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm6                   \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm2                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm2                     \n"
-      "packsswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "pmaddubsw %%xmm3,%%xmm0                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm2                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm2                     \n"
-      "packsswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "movdqu    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqa      %4,%%xmm3                     \n"
+      "movdqa      %5,%%xmm4                     \n"
+      "movdqa      %6,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm6                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm2                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "packsswb    %%xmm2,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm2                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm2                   \n"
+      "packsswb    %%xmm2,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "lea         0x40(%0),%0                   \n"
+      "movdqu      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1518,91 +1891,74 @@ void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
 
 void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
   asm volatile(
-      "movdqa    %4,%%xmm5                       \n"
-      "movdqa    %3,%%xmm4                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm3                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm3,%%xmm2                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
       : "+r"(src_bgra),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
       : "m"(kBGRAToY),   // %3
-        "m"(kAddY16)     // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
-void BGRAToUVRow_SSSE3(const uint8_t* src_bgra0,
+void BGRAToUVRow_SSSE3(const uint8_t* src_bgra,
                        int src_stride_bgra,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width) {
   asm volatile(
-      "movdqa    %5,%%xmm3                       \n"
-      "movdqa    %6,%%xmm4                       \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x10(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm1                   \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x20(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "movdqu    0x30(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-
-      "lea       0x40(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm6,%%xmm2             \n"
-      "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm2,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm1                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm1                     \n"
-      "packsswb  %%xmm1,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movlps    %%xmm0,(%1)                     \n"
-      "movhps    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_bgra0),                   // %0
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_bgra),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
@@ -1615,125 +1971,91 @@ void BGRAToUVRow_SSSE3(const uint8_t* src_bgra0,
 
 void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
   asm volatile(
-      "movdqa    %4,%%xmm5                       \n"
-      "movdqa    %3,%%xmm4                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm3                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm3,%%xmm2                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
       : "+r"(src_abgr),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
       : "m"(kABGRToY),   // %3
-        "m"(kAddY16)     // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
 void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
   asm volatile(
-      "movdqa    %4,%%xmm5                       \n"
-      "movdqa    %3,%%xmm4                       \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm4,%%xmm3                   \n"
-      "lea       0x40(%0),%0                     \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "phaddw    %%xmm3,%%xmm2                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN RGBTOY(xmm7)
       : "+r"(src_rgba),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
       : "m"(kRGBAToY),   // %3
-        "m"(kAddY16)     // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128),    // %4
+        "m"(kAddY16)     // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
 }
 
-void ABGRToUVRow_SSSE3(const uint8_t* src_abgr0,
+void ABGRToUVRow_SSSE3(const uint8_t* src_abgr,
                        int src_stride_abgr,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width) {
   asm volatile(
-      "movdqa    %5,%%xmm3                       \n"
-      "movdqa    %6,%%xmm4                       \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x10(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm1                   \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x20(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "movdqu    0x30(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-
-      "lea       0x40(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm6,%%xmm2             \n"
-      "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm2,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm1                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm1                     \n"
-      "packsswb  %%xmm1,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movlps    %%xmm0,(%1)                     \n"
-      "movhps    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_abgr0),                   // %0
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_abgr),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
@@ -1744,59 +2066,59 @@ void ABGRToUVRow_SSSE3(const uint8_t* src_abgr0,
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm6", "xmm7");
 }
 
-void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
+void RGBAToUVRow_SSSE3(const uint8_t* src_rgba,
                        int src_stride_rgba,
                        uint8_t* dst_u,
                        uint8_t* dst_v,
                        int width) {
   asm volatile(
-      "movdqa    %5,%%xmm3                       \n"
-      "movdqa    %6,%%xmm4                       \n"
-      "movdqa    %7,%%xmm5                       \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x10(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm1                   \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x20(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqu    0x30(%0),%%xmm6                 \n"
-      "movdqu    0x30(%0,%4,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-
-      "lea       0x40(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm7                   \n"
-      "shufps    $0x88,%%xmm6,%%xmm2             \n"
-      "shufps    $0xdd,%%xmm6,%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm2                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "phaddw    %%xmm2,%%xmm0                   \n"
-      "phaddw    %%xmm6,%%xmm1                   \n"
-      "psraw     $0x8,%%xmm0                     \n"
-      "psraw     $0x8,%%xmm1                     \n"
-      "packsswb  %%xmm1,%%xmm0                   \n"
-      "paddb     %%xmm5,%%xmm0                   \n"
-      "movlps    %%xmm0,(%1)                     \n"
-      "movhps    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_rgba0),                   // %0
+      "movdqa      %5,%%xmm3                     \n"
+      "movdqa      %6,%%xmm4                     \n"
+      "movdqa      %7,%%xmm5                     \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x10(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm1                 \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x20(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqu      0x30(%0),%%xmm6               \n"
+      "movdqu      0x30(%0,%4,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+
+      "lea         0x40(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "shufps      $0x88,%%xmm6,%%xmm2           \n"
+      "shufps      $0xdd,%%xmm6,%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "phaddw      %%xmm2,%%xmm0                 \n"
+      "phaddw      %%xmm6,%%xmm1                 \n"
+      "psraw       $0x8,%%xmm0                   \n"
+      "psraw       $0x8,%%xmm1                   \n"
+      "packsswb    %%xmm1,%%xmm0                 \n"
+      "paddb       %%xmm5,%%xmm0                 \n"
+      "movlps      %%xmm0,(%1)                   \n"
+      "movhps      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgba),                    // %0
         "+r"(dst_u),                       // %1
         "+r"(dst_v),                       // %2
         "+rm"(width)                       // %3
@@ -1811,21 +2133,21 @@ void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
 
 // Read 8 UV from 444
 #define READYUV444                                                \
-  "movq       (%[u_buf]),%%xmm0                               \n" \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
   "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
   "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
-  "punpcklbw  %%xmm1,%%xmm0                                   \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
   "movq       (%[y_buf]),%%xmm4                               \n" \
   "punpcklbw  %%xmm4,%%xmm4                                   \n" \
   "lea        0x8(%[y_buf]),%[y_buf]                          \n"
 
 // Read 4 UV from 422, upsample to 8 UV
 #define READYUV422                                                \
-  "movd       (%[u_buf]),%%xmm0                               \n" \
+  "movd       (%[u_buf]),%%xmm3                               \n" \
   "movd       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
   "lea        0x4(%[u_buf]),%[u_buf]                          \n" \
-  "punpcklbw  %%xmm1,%%xmm0                                   \n" \
-  "punpcklwd  %%xmm0,%%xmm0                                   \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
   "movq       (%[y_buf]),%%xmm4                               \n" \
   "punpcklbw  %%xmm4,%%xmm4                                   \n" \
   "lea        0x8(%[y_buf]),%[y_buf]                          \n"
@@ -1835,24 +2157,99 @@ void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
 // TODO(fbarchard): Consider pmulhuw to replace psraw
 // TODO(fbarchard): Consider pmullw to replace psllw and allow different bits.
 #define READYUV210                                                \
-  "movq       (%[u_buf]),%%xmm0                               \n" \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
   "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
   "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
-  "punpcklwd  %%xmm1,%%xmm0                                   \n" \
-  "psraw      $0x2,%%xmm0                                     \n" \
-  "packuswb   %%xmm0,%%xmm0                                   \n" \
-  "punpcklwd  %%xmm0,%%xmm0                                   \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+#define READYUVA210                                               \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n" \
+  "movdqu     (%[a_buf]),%%xmm5                               \n" \
+  "psraw      $2,%%xmm5                                       \n" \
+  "packuswb   %%xmm5,%%xmm5                                   \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                         \n"
+
+// Read 8 UV from 444 10 bit
+#define READYUV410                                                \
+  "movdqu     (%[u_buf]),%%xmm3                               \n" \
+  "movdqu     0x00(%[u_buf],%[v_buf],1),%%xmm2                \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                         \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "psraw      $2,%%xmm2                                       \n" \
+  "movdqa     %%xmm3,%%xmm1                                   \n" \
+  "punpcklwd  %%xmm2,%%xmm3                                   \n" \
+  "punpckhwd  %%xmm2,%%xmm1                                   \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "psllw      $6,%%xmm4                                       \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+// Read 8 UV from 444 10 bit.  With 8 Alpha.
+#define READYUVA410                                               \
+  "movdqu     (%[u_buf]),%%xmm3                               \n" \
+  "movdqu     0x00(%[u_buf],%[v_buf],1),%%xmm2                \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                         \n" \
+  "psraw      $2,%%xmm3                                       \n" \
+  "psraw      $2,%%xmm2                                       \n" \
+  "movdqa     %%xmm3,%%xmm1                                   \n" \
+  "punpcklwd  %%xmm2,%%xmm3                                   \n" \
+  "punpckhwd  %%xmm2,%%xmm1                                   \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
   "movdqu     (%[y_buf]),%%xmm4                               \n" \
   "psllw      $0x6,%%xmm4                                     \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n" \
+  "movdqu     (%[a_buf]),%%xmm5                               \n" \
+  "psraw      $2,%%xmm5                                       \n" \
+  "packuswb   %%xmm5,%%xmm5                                   \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                         \n"
+
+// Read 4 UV from 422 12 bit, upsample to 8 UV
+#define READYUV212                                                \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklwd  %%xmm1,%%xmm3                                   \n" \
+  "psraw      $0x4,%%xmm3                                     \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "psllw      $0x4,%%xmm4                                     \n" \
   "lea        0x10(%[y_buf]),%[y_buf]                         \n"
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
 #define READYUVA422                                               \
-  "movd       (%[u_buf]),%%xmm0                               \n" \
+  "movd       (%[u_buf]),%%xmm3                               \n" \
   "movd       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
   "lea        0x4(%[u_buf]),%[u_buf]                          \n" \
-  "punpcklbw  %%xmm1,%%xmm0                                   \n" \
-  "punpcklwd  %%xmm0,%%xmm0                                   \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movq       (%[y_buf]),%%xmm4                               \n" \
+  "punpcklbw  %%xmm4,%%xmm4                                   \n" \
+  "lea        0x8(%[y_buf]),%[y_buf]                          \n" \
+  "movq       (%[a_buf]),%%xmm5                               \n" \
+  "lea        0x8(%[a_buf]),%[a_buf]                          \n"
+
+// Read 8 UV from 444.  With 8 Alpha.
+#define READYUVA444                                               \
+  "movq       (%[u_buf]),%%xmm3                               \n" \
+  "movq       0x00(%[u_buf],%[v_buf],1),%%xmm1                \n" \
+  "lea        0x8(%[u_buf]),%[u_buf]                          \n" \
+  "punpcklbw  %%xmm1,%%xmm3                                   \n" \
   "movq       (%[y_buf]),%%xmm4                               \n" \
   "punpcklbw  %%xmm4,%%xmm4                                   \n" \
   "lea        0x8(%[y_buf]),%[y_buf]                          \n" \
@@ -1861,18 +2258,18 @@ void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
 
 // Read 4 UV from NV12, upsample to 8 UV
 #define READNV12                                                  \
-  "movq       (%[uv_buf]),%%xmm0                              \n" \
+  "movq       (%[uv_buf]),%%xmm3                              \n" \
   "lea        0x8(%[uv_buf]),%[uv_buf]                        \n" \
-  "punpcklwd  %%xmm0,%%xmm0                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
   "movq       (%[y_buf]),%%xmm4                               \n" \
   "punpcklbw  %%xmm4,%%xmm4                                   \n" \
   "lea        0x8(%[y_buf]),%[y_buf]                          \n"
 
 // Read 4 VU from NV21, upsample to 8 UV
 #define READNV21                                                  \
-  "movq       (%[vu_buf]),%%xmm0                              \n" \
+  "movq       (%[vu_buf]),%%xmm3                              \n" \
   "lea        0x8(%[vu_buf]),%[vu_buf]                        \n" \
-  "pshufb     %[kShuffleNV21], %%xmm0                         \n" \
+  "pshufb     %[kShuffleNV21], %%xmm3                         \n" \
   "movq       (%[y_buf]),%%xmm4                               \n" \
   "punpcklbw  %%xmm4,%%xmm4                                   \n" \
   "lea        0x8(%[y_buf]),%[y_buf]                          \n"
@@ -1881,68 +2278,92 @@ void RGBAToUVRow_SSSE3(const uint8_t* src_rgba0,
 #define READYUY2                                                  \
   "movdqu     (%[yuy2_buf]),%%xmm4                            \n" \
   "pshufb     %[kShuffleYUY2Y], %%xmm4                        \n" \
-  "movdqu     (%[yuy2_buf]),%%xmm0                            \n" \
-  "pshufb     %[kShuffleYUY2UV], %%xmm0                       \n" \
+  "movdqu     (%[yuy2_buf]),%%xmm3                            \n" \
+  "pshufb     %[kShuffleYUY2UV], %%xmm3                       \n" \
   "lea        0x10(%[yuy2_buf]),%[yuy2_buf]                   \n"
 
 // Read 4 UYVY with 8 Y and update 4 UV to 8 UV.
 #define READUYVY                                                  \
   "movdqu     (%[uyvy_buf]),%%xmm4                            \n" \
   "pshufb     %[kShuffleUYVYY], %%xmm4                        \n" \
-  "movdqu     (%[uyvy_buf]),%%xmm0                            \n" \
-  "pshufb     %[kShuffleUYVYUV], %%xmm0                       \n" \
+  "movdqu     (%[uyvy_buf]),%%xmm3                            \n" \
+  "pshufb     %[kShuffleUYVYUV], %%xmm3                       \n" \
   "lea        0x10(%[uyvy_buf]),%[uyvy_buf]                   \n"
 
+// Read 4 UV from P210, upsample to 8 UV
+#define READP210                                                  \
+  "movdqu     (%[uv_buf]),%%xmm3                              \n" \
+  "lea        0x10(%[uv_buf]),%[uv_buf]                       \n" \
+  "psrlw      $0x8,%%xmm3                                     \n" \
+  "packuswb   %%xmm3,%%xmm3                                   \n" \
+  "punpcklwd  %%xmm3,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
+// Read 8 UV from P410
+#define READP410                                                  \
+  "movdqu     (%[uv_buf]),%%xmm3                              \n" \
+  "movdqu     0x10(%[uv_buf]),%%xmm1                          \n" \
+  "lea        0x20(%[uv_buf]),%[uv_buf]                       \n" \
+  "psrlw      $0x8,%%xmm3                                     \n" \
+  "psrlw      $0x8,%%xmm1                                     \n" \
+  "packuswb   %%xmm1,%%xmm3                                   \n" \
+  "movdqu     (%[y_buf]),%%xmm4                               \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                         \n"
+
 #if defined(__x86_64__)
 #define YUVTORGB_SETUP(yuvconstants)                              \
+  "pcmpeqb    %%xmm13,%%xmm13                                 \n" \
   "movdqa     (%[yuvconstants]),%%xmm8                        \n" \
+  "pxor       %%xmm12,%%xmm12                                 \n" \
   "movdqa     32(%[yuvconstants]),%%xmm9                      \n" \
+  "psllw      $7,%%xmm13                                      \n" \
   "movdqa     64(%[yuvconstants]),%%xmm10                     \n" \
+  "pshufb     %%xmm12,%%xmm13                                 \n" \
   "movdqa     96(%[yuvconstants]),%%xmm11                     \n" \
-  "movdqa     128(%[yuvconstants]),%%xmm12                    \n" \
-  "movdqa     160(%[yuvconstants]),%%xmm13                    \n" \
-  "movdqa     192(%[yuvconstants]),%%xmm14                    \n"
+  "movdqa     128(%[yuvconstants]),%%xmm12                    \n"
+
 // Convert 8 pixels: 8 UV and 8 Y
 #define YUVTORGB16(yuvconstants)                                  \
-  "movdqa     %%xmm0,%%xmm1                                   \n" \
-  "movdqa     %%xmm0,%%xmm2                                   \n" \
-  "movdqa     %%xmm0,%%xmm3                                   \n" \
-  "movdqa     %%xmm11,%%xmm0                                  \n" \
-  "pmaddubsw  %%xmm8,%%xmm1                                   \n" \
-  "psubw      %%xmm1,%%xmm0                                   \n" \
-  "movdqa     %%xmm12,%%xmm1                                  \n" \
-  "pmaddubsw  %%xmm9,%%xmm2                                   \n" \
-  "psubw      %%xmm2,%%xmm1                                   \n" \
-  "movdqa     %%xmm13,%%xmm2                                  \n" \
-  "pmaddubsw  %%xmm10,%%xmm3                                  \n" \
-  "psubw      %%xmm3,%%xmm2                                   \n" \
-  "pmulhuw    %%xmm14,%%xmm4                                  \n" \
+  "psubb      %%xmm13,%%xmm3                                  \n" \
+  "pmulhuw    %%xmm11,%%xmm4                                  \n" \
+  "movdqa     %%xmm8,%%xmm0                                   \n" \
+  "movdqa     %%xmm9,%%xmm1                                   \n" \
+  "movdqa     %%xmm10,%%xmm2                                  \n" \
+  "paddw      %%xmm12,%%xmm4                                  \n" \
+  "pmaddubsw  %%xmm3,%%xmm0                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm1                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm2                                   \n" \
   "paddsw     %%xmm4,%%xmm0                                   \n" \
-  "paddsw     %%xmm4,%%xmm1                                   \n" \
-  "paddsw     %%xmm4,%%xmm2                                   \n"
-#define YUVTORGB_REGS \
-  "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+  "paddsw     %%xmm4,%%xmm2                                   \n" \
+  "psubsw     %%xmm1,%%xmm4                                   \n" \
+  "movdqa     %%xmm4,%%xmm1                                   \n"
+
+#define YUVTORGB_REGS "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",
 
 #else
 #define YUVTORGB_SETUP(yuvconstants)
 // Convert 8 pixels: 8 UV and 8 Y
 #define YUVTORGB16(yuvconstants)                                  \
-  "movdqa     %%xmm0,%%xmm1                                   \n" \
-  "movdqa     %%xmm0,%%xmm2                                   \n" \
-  "movdqa     %%xmm0,%%xmm3                                   \n" \
-  "movdqa     96(%[yuvconstants]),%%xmm0                      \n" \
-  "pmaddubsw  (%[yuvconstants]),%%xmm1                        \n" \
-  "psubw      %%xmm1,%%xmm0                                   \n" \
-  "movdqa     128(%[yuvconstants]),%%xmm1                     \n" \
-  "pmaddubsw  32(%[yuvconstants]),%%xmm2                      \n" \
-  "psubw      %%xmm2,%%xmm1                                   \n" \
-  "movdqa     160(%[yuvconstants]),%%xmm2                     \n" \
-  "pmaddubsw  64(%[yuvconstants]),%%xmm3                      \n" \
-  "psubw      %%xmm3,%%xmm2                                   \n" \
-  "pmulhuw    192(%[yuvconstants]),%%xmm4                     \n" \
+  "pcmpeqb    %%xmm0,%%xmm0                                   \n" \
+  "pxor       %%xmm1,%%xmm1                                   \n" \
+  "psllw      $7,%%xmm0                                       \n" \
+  "pshufb     %%xmm1,%%xmm0                                   \n" \
+  "psubb      %%xmm0,%%xmm3                                   \n" \
+  "pmulhuw    96(%[yuvconstants]),%%xmm4                      \n" \
+  "movdqa     (%[yuvconstants]),%%xmm0                        \n" \
+  "movdqa     32(%[yuvconstants]),%%xmm1                      \n" \
+  "movdqa     64(%[yuvconstants]),%%xmm2                      \n" \
+  "pmaddubsw  %%xmm3,%%xmm0                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm1                                   \n" \
+  "pmaddubsw  %%xmm3,%%xmm2                                   \n" \
+  "movdqa     128(%[yuvconstants]),%%xmm3                     \n" \
+  "paddw      %%xmm3,%%xmm4                                   \n" \
   "paddsw     %%xmm4,%%xmm0                                   \n" \
-  "paddsw     %%xmm4,%%xmm1                                   \n" \
-  "paddsw     %%xmm4,%%xmm2                                   \n"
+  "paddsw     %%xmm4,%%xmm2                                   \n" \
+  "psubsw     %%xmm1,%%xmm4                                   \n" \
+  "movdqa     %%xmm4,%%xmm1                                   \n"
+
 #define YUVTORGB_REGS
 #endif
 
@@ -2012,16 +2433,16 @@ void OMITFP I444ToARGBRow_SSSE3(const uint8_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV444
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2033,6 +2454,44 @@ void OMITFP I444ToARGBRow_SSSE3(const uint8_t* y_buf,
   );
 }
 
+#ifdef HAS_I444ALPHATOARGBROW_SSSE3
+void OMITFP I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                                     const uint8_t* u_buf,
+                                     const uint8_t* v_buf,
+                                     const uint8_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  // clang-format off
+  asm volatile (
+  YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+  LABELALIGN
+      "1:                                        \n"
+  READYUVA444
+  YUVTORGB(yuvconstants)
+  STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [a_buf]"+r"(a_buf),    // %[a_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_I444ALPHATOARGBROW_SSSE3
+
 void OMITFP I422ToRGB24Row_SSSE3(const uint8_t* y_buf,
                                  const uint8_t* u_buf,
                                  const uint8_t* v_buf,
@@ -2041,27 +2500,27 @@ void OMITFP I422ToRGB24Row_SSSE3(const uint8_t* y_buf,
                                  int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
-    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
-    "sub       %[u_buf],%[v_buf]               \n"
+      "movdqa      %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+      "movdqa      %[kShuffleMaskARGBToRGB24],%%xmm6 \n"
+      "sub         %[u_buf],%[v_buf]             \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
-    "punpcklbw %%xmm1,%%xmm0                   \n"
-    "punpcklbw %%xmm2,%%xmm2                   \n"
-    "movdqa    %%xmm0,%%xmm1                   \n"
-    "punpcklwd %%xmm2,%%xmm0                   \n"
-    "punpckhwd %%xmm2,%%xmm1                   \n"
-    "pshufb    %%xmm5,%%xmm0                   \n"
-    "pshufb    %%xmm6,%%xmm1                   \n"
-    "palignr   $0xc,%%xmm0,%%xmm1              \n"
-    "movq      %%xmm0,(%[dst_rgb24])           \n"
-    "movdqu    %%xmm1,0x8(%[dst_rgb24])        \n"
-    "lea       0x18(%[dst_rgb24]),%[dst_rgb24] \n"
-    "subl      $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpcklbw   %%xmm2,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklwd   %%xmm2,%%xmm0                 \n"
+      "punpckhwd   %%xmm2,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm6,%%xmm1                 \n"
+      "palignr     $0xc,%%xmm0,%%xmm1            \n"
+      "movq        %%xmm0,(%[dst_rgb24])         \n"
+      "movdqu      %%xmm1,0x8(%[dst_rgb24])      \n"
+      "lea         0x18(%[dst_rgb24]),%[dst_rgb24] \n"
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2087,16 +2546,16 @@ void OMITFP I422ToARGBRow_SSSE3(const uint8_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2116,21 +2575,21 @@ void OMITFP I422ToAR30Row_SSSE3(const uint8_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // AR30 constants
-    "psrlw     $14,%%xmm5                      \n"
-    "psllw     $4,%%xmm5                       \n"  // 2 alpha bits
-    "pxor      %%xmm6,%%xmm6                   \n"
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"  // 0 for min
-    "psrlw     $6,%%xmm7                       \n"  // 1023 for max
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"  // AR30 constants
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB16(yuvconstants)
     STOREAR30
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2151,16 +2610,46 @@ void OMITFP I210ToARGBRow_SSSE3(const uint16_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV210
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+// 12 bit YUV to ARGB
+void OMITFP I212ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2181,21 +2670,21 @@ void OMITFP I210ToAR30Row_SSSE3(const uint16_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
-    "psrlw     $14,%%xmm5                      \n"
-    "psllw     $4,%%xmm5                       \n"  // 2 alpha bits
-    "pxor      %%xmm6,%%xmm6                   \n"
-    "pcmpeqb   %%xmm7,%%xmm7                   \n"  // 0 for min
-    "psrlw     $6,%%xmm7                       \n"  // 1023 for max
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV210
     YUVTORGB16(yuvconstants)
     STOREAR30
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2207,6 +2696,176 @@ void OMITFP I210ToAR30Row_SSSE3(const uint16_t* y_buf,
   );
 }
 
+// 12 bit YUV to AR30
+void OMITFP I212ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// 10 bit YUV to ARGB
+void OMITFP I410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+#ifdef HAS_I210ALPHATOARGBROW_SSSE3
+// 10 bit YUVA to ARGB
+void OMITFP I210AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                                     const uint16_t* u_buf,
+                                     const uint16_t* v_buf,
+                                     const uint16_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA210
+          YUVTORGB(yuvconstants) STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),  // %[y_buf]
+        [u_buf] "+r"(u_buf),  // %[u_buf]
+        [v_buf] "+r"(v_buf),  // %[v_buf]
+        [a_buf] "+r"(a_buf),
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+#endif
+
+#ifdef HAS_I410ALPHATOARGBROW_SSSE3
+// 10 bit YUVA to ARGB
+void OMITFP I410AlphaToARGBRow_SSSE3(const uint16_t* y_buf,
+                                     const uint16_t* u_buf,
+                                     const uint16_t* v_buf,
+                                     const uint16_t* a_buf,
+                                     uint8_t* dst_argb,
+                                     const struct YuvConstants* yuvconstants,
+                                     int width) {
+  // clang-format off
+  asm volatile(
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUVA410
+    YUVTORGB(yuvconstants)
+    STOREARGB
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+    : [y_buf] "+r"(y_buf),  // %[y_buf]
+      [u_buf] "+r"(u_buf),  // %[u_buf]
+      [v_buf] "+r"(v_buf),  // %[v_buf]
+      [a_buf] "+r"(a_buf),
+      [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+      [width] "+m"(width)  // %[width]
+#else
+      [width] "+rm"(width)  // %[width]
+#endif
+    : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+    : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+      "xmm5");
+  // clang-format on
+}
+#endif
+
+// 10 bit YUV to AR30
+void OMITFP I410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* u_buf,
+                                const uint16_t* v_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)   // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
 #ifdef HAS_I422ALPHATOARGBROW_SSSE3
 void OMITFP I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
                                      const uint8_t* u_buf,
@@ -2218,15 +2877,15 @@ void OMITFP I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
+      "sub         %[u_buf],%[v_buf]             \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUVA422
     YUVTORGB(yuvconstants)
     STOREARGB
-    "subl      $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "subl        $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2253,15 +2912,15 @@ void OMITFP NV12ToARGBRow_SSSE3(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READNV12
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [uv_buf]"+r"(uv_buf),    // %[uv_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
@@ -2281,15 +2940,15 @@ void OMITFP NV21ToARGBRow_SSSE3(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READNV21
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [vu_buf]"+r"(vu_buf),    // %[vu_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
@@ -2309,15 +2968,15 @@ void OMITFP YUY2ToARGBRow_SSSE3(const uint8_t* yuy2_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUY2
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [yuy2_buf]"+r"(yuy2_buf),    // %[yuy2_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
@@ -2337,15 +2996,15 @@ void OMITFP UYVYToARGBRow_SSSE3(const uint8_t* uyvy_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READUYVY
     YUVTORGB(yuvconstants)
     STOREARGB
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [uyvy_buf]"+r"(uyvy_buf),    // %[uyvy_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
@@ -2358,6 +3017,112 @@ void OMITFP UYVYToARGBRow_SSSE3(const uint8_t* uyvy_buf,
   // clang-format on
 }
 
+void OMITFP P210ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN "1:                                        \n" READP210
+          YUVTORGB(yuvconstants) STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),              // %[y_buf]
+        [uv_buf] "+r"(uv_buf),            // %[u_buf]
+        [dst_argb] "+r"(dst_argb),        // %[dst_argb]
+        [width] "+rm"(width)              // %[width]
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+
+void OMITFP P410ToARGBRow_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_argb,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile(
+      YUVTORGB_SETUP(
+      yuvconstants) "pcmpeqb     %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN "1:                                        \n" READP410
+          YUVTORGB(yuvconstants) STOREARGB
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+      : [y_buf] "+r"(y_buf),              // %[y_buf]
+        [uv_buf] "+r"(uv_buf),            // %[u_buf]
+        [dst_argb] "+r"(dst_argb),        // %[dst_argb]
+        [width] "+rm"(width)              // %[width]
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4",
+        "xmm5");
+}
+
+void OMITFP P210ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READP210
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),              // %[y_buf]
+    [uv_buf]"+r"(uv_buf),            // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),        // %[dst_ar30]
+    [width]"+rm"(width)              // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void OMITFP P410ToAR30Row_SSSE3(const uint16_t* y_buf,
+                                const uint16_t* uv_buf,
+                                uint8_t* dst_ar30,
+                                const struct YuvConstants* yuvconstants,
+                                int width) {
+  asm volatile (
+    YUVTORGB_SETUP(yuvconstants)
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $14,%%xmm5                    \n"
+      "psllw       $4,%%xmm5                     \n"  // 2 alpha bits
+      "pxor        %%xmm6,%%xmm6                 \n"  // 0 for min
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $6,%%xmm7                     \n"  // 1023 for max
+
+    LABELALIGN
+      "1:                                        \n"
+    READP410
+    YUVTORGB16(yuvconstants)
+    STOREAR30
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
+  : [y_buf]"+r"(y_buf),              // %[y_buf]
+    [uv_buf]"+r"(uv_buf),            // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),        // %[dst_ar30]
+    [width]"+rm"(width)              // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
 void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
                                 const uint8_t* u_buf,
                                 const uint8_t* v_buf,
@@ -2366,16 +3131,16 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
                                 int width) {
   asm volatile (
     YUVTORGB_SETUP(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422
     YUVTORGB(yuvconstants)
     STORERGBA
-    "sub       $0x8,%[width]                   \n"
-    "jg        1b                              \n"
+      "sub         $0x8,%[width]                 \n"
+      "jg          1b                            \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2391,12 +3156,12 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
 
 // Read 16 UV from 444
 #define READYUV444_AVX2                                               \
-  "vmovdqu    (%[u_buf]),%%xmm0                                   \n" \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
   "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
   "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
   "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n" \
-  "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
   "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
   "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
   "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
@@ -2404,42 +3169,139 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
 
 // Read 8 UV from 422, upsample to 16 UV.
 #define READYUV422_AVX2                                               \
-  "vmovq      (%[u_buf]),%%xmm0                                   \n" \
+  "vmovq      (%[u_buf]),%%xmm3                                   \n" \
   "vmovq      0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
   "lea        0x8(%[u_buf]),%[u_buf]                              \n" \
-  "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
-  "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
   "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
   "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
   "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
   "lea        0x10(%[y_buf]),%[y_buf]                             \n"
 
-// Read 8 UV from 210 10 bit, upsample to 16 UV
+#define READYUV422_AVX512BW                                           \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "vpermq     %%zmm3,%%zmm16,%%zmm3                               \n" \
+  "vpermq     %%zmm1,%%zmm16,%%zmm1                               \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
+  "vpunpcklbw %%zmm1,%%zmm3,%%zmm3                                \n" \
+  "vpermq     $0xd8,%%zmm3,%%zmm3                                 \n" \
+  "vpunpcklwd %%zmm3,%%zmm3,%%zmm3                                \n" \
+  "vmovdqu8   (%[y_buf]),%%ymm4                                   \n" \
+  "vpermq     %%zmm4,%%zmm17,%%zmm4                               \n" \
+  "vpermq     $0xd8,%%zmm4,%%zmm4                                 \n" \
+  "vpunpcklbw %%zmm4,%%zmm4,%%zmm4                                \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
+
+// Read 8 UV from 210, upsample to 16 UV
 // TODO(fbarchard): Consider vshufb to replace pack/unpack
 // TODO(fbarchard): Consider vunpcklpd to combine the 2 registers into 1.
 #define READYUV210_AVX2                                            \
-  "vmovdqu    (%[u_buf]),%%xmm0                                \n" \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
   "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
   "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                              \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
   "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
-  "vpunpcklwd %%ymm1,%%ymm0,%%ymm0                             \n" \
-  "vpsraw     $0x2,%%ymm0,%%ymm0                               \n" \
-  "vpackuswb  %%ymm0,%%ymm0,%%ymm0                             \n" \
-  "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                             \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
   "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
-  "vpsllw     $0x6,%%ymm4,%%ymm4                               \n" \
+  "vpsllw     $6,%%ymm4,%%ymm4                                 \n" \
   "lea        0x20(%[y_buf]),%[y_buf]                          \n"
 
+// Read 8 UV from 210, upsample to 16 UV. With 16 Alpha.
+#define READYUVA210_AVX2                                           \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm4                                 \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n" \
+  "vmovdqu    (%[a_buf]),%%ymm5                                \n" \
+  "vpsraw     $2,%%ymm5,%%ymm5                                 \n" \
+  "vpackuswb  %%ymm5,%%ymm5,%%ymm5                             \n" \
+  "lea        0x20(%[a_buf]),%[a_buf]                          \n"
+
+// Read 16 UV from 410
+#define READYUV410_AVX2                                            \
+  "vmovdqu    (%[u_buf]),%%ymm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%ymm2                 \n" \
+  "lea        0x20(%[u_buf]),%[u_buf]                          \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpsraw     $2,%%ymm2,%%ymm2                                 \n" \
+  "vpunpckhwd %%ymm2,%%ymm3,%%ymm1                             \n" \
+  "vpunpcklwd %%ymm2,%%ymm3,%%ymm3                             \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm4                                 \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n"
+
+// Read 8 UV from 212 12 bit, upsample to 16 UV
+#define READYUV212_AVX2                                            \
+  "vmovdqu    (%[u_buf]),%%xmm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                 \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                          \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                              \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                              \n" \
+  "vpunpcklwd %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vpsraw     $0x4,%%ymm3,%%ymm3                               \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $0x4,%%ymm4,%%ymm4                               \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n"
+
+// Read 16 UV from 410. With 16 Alpha.
+#define READYUVA410_AVX2                                           \
+  "vmovdqu    (%[u_buf]),%%ymm3                                \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%ymm2                 \n" \
+  "lea        0x20(%[u_buf]),%[u_buf]                          \n" \
+  "vpsraw     $2,%%ymm3,%%ymm3                                 \n" \
+  "vpsraw     $2,%%ymm2,%%ymm2                                 \n" \
+  "vpunpckhwd %%ymm2,%%ymm3,%%ymm1                             \n" \
+  "vpunpcklwd %%ymm2,%%ymm3,%%ymm3                             \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                             \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                \n" \
+  "vpsllw     $6,%%ymm4,%%ymm4                                 \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                          \n" \
+  "vmovdqu    (%[a_buf]),%%ymm5                                \n" \
+  "vpsraw     $2,%%ymm5,%%ymm5                                 \n" \
+  "vpackuswb  %%ymm5,%%ymm5,%%ymm5                             \n" \
+  "lea        0x20(%[a_buf]),%[a_buf]                          \n"
+
+// Read 16 UV from 444.  With 16 Alpha.
+#define READYUVA444_AVX2                                              \
+  "vmovdqu    (%[u_buf]),%%xmm3                                   \n" \
+  "vmovdqu    0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
+  "lea        0x10(%[u_buf]),%[u_buf]                             \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpermq     $0xd8,%%ymm1,%%ymm1                                 \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
+  "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
+  "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
+  "lea        0x10(%[y_buf]),%[y_buf]                             \n" \
+  "vmovdqu    (%[a_buf]),%%xmm5                                   \n" \
+  "vpermq     $0xd8,%%ymm5,%%ymm5                                 \n" \
+  "lea        0x10(%[a_buf]),%[a_buf]                             \n"
+
 // Read 8 UV from 422, upsample to 16 UV.  With 16 Alpha.
 #define READYUVA422_AVX2                                              \
-  "vmovq      (%[u_buf]),%%xmm0                                   \n" \
+  "vmovq      (%[u_buf]),%%xmm3                                   \n" \
   "vmovq      0x00(%[u_buf],%[v_buf],1),%%xmm1                    \n" \
   "lea        0x8(%[u_buf]),%[u_buf]                              \n" \
-  "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
-  "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n" \
+  "vpunpcklbw %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
   "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
   "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
   "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
@@ -2450,10 +3312,10 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
 
 // Read 8 UV from NV12, upsample to 16 UV.
 #define READNV12_AVX2                                                 \
-  "vmovdqu    (%[uv_buf]),%%xmm0                                  \n" \
+  "vmovdqu    (%[uv_buf]),%%xmm3                                  \n" \
   "lea        0x10(%[uv_buf]),%[uv_buf]                           \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
-  "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
   "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
   "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
   "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
@@ -2461,73 +3323,130 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
 
 // Read 8 VU from NV21, upsample to 16 UV.
 #define READNV21_AVX2                                                 \
-  "vmovdqu    (%[vu_buf]),%%xmm0                                  \n" \
+  "vmovdqu    (%[vu_buf]),%%xmm3                                  \n" \
   "lea        0x10(%[vu_buf]),%[vu_buf]                           \n" \
-  "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n" \
-  "vpshufb     %[kShuffleNV21], %%ymm0, %%ymm0                    \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vpshufb     %[kShuffleNV21], %%ymm3, %%ymm3                    \n" \
   "vmovdqu    (%[y_buf]),%%xmm4                                   \n" \
   "vpermq     $0xd8,%%ymm4,%%ymm4                                 \n" \
   "vpunpcklbw %%ymm4,%%ymm4,%%ymm4                                \n" \
   "lea        0x10(%[y_buf]),%[y_buf]                             \n"
 
+// Read 4 UV from P210, upsample to 8 UV
+#define READP210_AVX2                                                 \
+  "vmovdqu    (%[uv_buf]),%%ymm3                                  \n" \
+  "lea        0x20(%[uv_buf]),%[uv_buf]                           \n" \
+  "vpsrlw     $0x8,%%ymm3,%%ymm3                                  \n" \
+  "vpackuswb  %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vpunpcklwd %%ymm3,%%ymm3,%%ymm3                                \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                   \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
+
+// Read 8 UV from P410
+#define READP410_AVX2                                                 \
+  "vmovdqu    (%[uv_buf]),%%ymm3                                  \n" \
+  "vmovdqu    0x20(%[uv_buf]),%%ymm1                              \n" \
+  "lea        0x40(%[uv_buf]),%[uv_buf]                           \n" \
+  "vpsrlw     $0x8,%%ymm3,%%ymm3                                  \n" \
+  "vpsrlw     $0x8,%%ymm1,%%ymm1                                  \n" \
+  "vpackuswb  %%ymm1,%%ymm3,%%ymm3                                \n" \
+  "vpermq     $0xd8,%%ymm3,%%ymm3                                 \n" \
+  "vmovdqu    (%[y_buf]),%%ymm4                                   \n" \
+  "lea        0x20(%[y_buf]),%[y_buf]                             \n"
+
 // Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV.
 #define READYUY2_AVX2                                                 \
   "vmovdqu    (%[yuy2_buf]),%%ymm4                                \n" \
   "vpshufb    %[kShuffleYUY2Y], %%ymm4, %%ymm4                    \n" \
-  "vmovdqu    (%[yuy2_buf]),%%ymm0                                \n" \
-  "vpshufb    %[kShuffleYUY2UV], %%ymm0, %%ymm0                   \n" \
+  "vmovdqu    (%[yuy2_buf]),%%ymm3                                \n" \
+  "vpshufb    %[kShuffleYUY2UV], %%ymm3, %%ymm3                   \n" \
   "lea        0x20(%[yuy2_buf]),%[yuy2_buf]                       \n"
 
 // Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV.
 #define READUYVY_AVX2                                                 \
   "vmovdqu    (%[uyvy_buf]),%%ymm4                                \n" \
   "vpshufb    %[kShuffleUYVYY], %%ymm4, %%ymm4                    \n" \
-  "vmovdqu    (%[uyvy_buf]),%%ymm0                                \n" \
-  "vpshufb    %[kShuffleUYVYUV], %%ymm0, %%ymm0                   \n" \
+  "vmovdqu    (%[uyvy_buf]),%%ymm3                                \n" \
+  "vpshufb    %[kShuffleUYVYUV], %%ymm3, %%ymm3                   \n" \
   "lea        0x20(%[uyvy_buf]),%[uyvy_buf]                       \n"
 
+// TODO(fbarchard): Remove broadcastb
 #if defined(__x86_64__)
-#define YUVTORGB_SETUP_AVX2(yuvconstants)                            \
-  "vmovdqa     (%[yuvconstants]),%%ymm8                          \n" \
-  "vmovdqa     32(%[yuvconstants]),%%ymm9                        \n" \
-  "vmovdqa     64(%[yuvconstants]),%%ymm10                       \n" \
-  "vmovdqa     96(%[yuvconstants]),%%ymm11                       \n" \
-  "vmovdqa     128(%[yuvconstants]),%%ymm12                      \n" \
-  "vmovdqa     160(%[yuvconstants]),%%ymm13                      \n" \
-  "vmovdqa     192(%[yuvconstants]),%%ymm14                      \n"
+#define YUVTORGB_SETUP_AVX2(yuvconstants)                             \
+  "vpcmpeqb    %%xmm13,%%xmm13,%%xmm13                            \n" \
+  "vmovdqa     (%[yuvconstants]),%%ymm8                           \n" \
+  "vpsllw      $7,%%xmm13,%%xmm13                                 \n" \
+  "vmovdqa     32(%[yuvconstants]),%%ymm9                         \n" \
+  "vpbroadcastb %%xmm13,%%ymm13                                   \n" \
+  "vmovdqa     64(%[yuvconstants]),%%ymm10                        \n" \
+  "vmovdqa     96(%[yuvconstants]),%%ymm11                        \n" \
+  "vmovdqa     128(%[yuvconstants]),%%ymm12                       \n"
+
+#define YUVTORGB_SETUP_AVX512BW(yuvconstants)                         \
+  "vpcmpeqb   %%xmm13,%%xmm13,%%xmm13                             \n" \
+  "movdqa     (%[yuvconstants]),%%xmm8                            \n" \
+  "vpbroadcastq %%xmm8, %%zmm8                                    \n" \
+  "vpsllw     $7,%%xmm13,%%xmm13                                  \n" \
+  "vpbroadcastb %%xmm13,%%zmm13                                   \n" \
+  "movq     32(%[yuvconstants]),%%xmm9                            \n" \
+  "vpbroadcastq %%xmm9,%%zmm9                                     \n" \
+  "movq     64(%[yuvconstants]),%%xmm10                           \n" \
+  "vpbroadcastq %%xmm10,%%zmm10                                   \n" \
+  "movq     96(%[yuvconstants]),%%xmm11                           \n" \
+  "vpbroadcastq %%xmm11,%%zmm11                                   \n" \
+  "movq     128(%[yuvconstants]),%%xmm12                          \n" \
+  "vpbroadcastq %%xmm12,%%zmm12                                   \n" \
+  "vmovdqu8 (%[quadsplitperm]),%%zmm16                            \n" \
+  "vmovdqu8 (%[dquadsplitperm]),%%zmm17                           \n" \
+  "vmovdqu8 (%[unperm]),%%zmm18                                   \n"
 
 #define YUVTORGB16_AVX2(yuvconstants)                                 \
-  "vpmaddubsw  %%ymm10,%%ymm0,%%ymm2                              \n" \
-  "vpmaddubsw  %%ymm9,%%ymm0,%%ymm1                               \n" \
-  "vpmaddubsw  %%ymm8,%%ymm0,%%ymm0                               \n" \
-  "vpsubw      %%ymm2,%%ymm13,%%ymm2                              \n" \
-  "vpsubw      %%ymm1,%%ymm12,%%ymm1                              \n" \
-  "vpsubw      %%ymm0,%%ymm11,%%ymm0                              \n" \
-  "vpmulhuw    %%ymm14,%%ymm4,%%ymm4                              \n" \
+  "vpsubb      %%ymm13,%%ymm3,%%ymm3                              \n" \
+  "vpmulhuw    %%ymm11,%%ymm4,%%ymm4                              \n" \
+  "vpmaddubsw  %%ymm3,%%ymm8,%%ymm0                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm9,%%ymm1                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm10,%%ymm2                              \n" \
+  "vpaddw      %%ymm4,%%ymm12,%%ymm4                              \n" \
   "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n" \
-  "vpaddsw     %%ymm4,%%ymm1,%%ymm1                               \n" \
+  "vpsubsw     %%ymm1,%%ymm4,%%ymm1                               \n" \
   "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"
 
-#define YUVTORGB_REGS_AVX2 \
-  "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14",
+#define YUVTORGB16_AVX512BW(yuvconstants)                             \
+  "vpsubb      %%zmm13,%%zmm3,%%zmm3                              \n" \
+  "vpmulhuw    %%zmm11,%%zmm4,%%zmm4                              \n" \
+  "vpmaddubsw  %%zmm3,%%zmm8,%%zmm0                               \n" \
+  "vpmaddubsw  %%zmm3,%%zmm9,%%zmm1                               \n" \
+  "vpmaddubsw  %%zmm3,%%zmm10,%%zmm2                              \n" \
+  "vpaddw      %%zmm4,%%zmm12,%%zmm4                              \n" \
+  "vpaddsw     %%zmm4,%%zmm0,%%zmm0                               \n" \
+  "vpsubsw     %%zmm1,%%zmm4,%%zmm1                               \n" \
+  "vpaddsw     %%zmm4,%%zmm2,%%zmm2                               \n"
+
+#define YUVTORGB_REGS_AVX2 "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",
+#define YUVTORGB_REGS_AVX512BW \
+  "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm16", "xmm17", "xmm18",
 
 #else  // Convert 16 pixels: 16 UV and 16 Y.
 
 #define YUVTORGB_SETUP_AVX2(yuvconstants)
 #define YUVTORGB16_AVX2(yuvconstants)                                 \
-  "vpmaddubsw  64(%[yuvconstants]),%%ymm0,%%ymm2                  \n" \
-  "vpmaddubsw  32(%[yuvconstants]),%%ymm0,%%ymm1                  \n" \
-  "vpmaddubsw  (%[yuvconstants]),%%ymm0,%%ymm0                    \n" \
-  "vmovdqu     160(%[yuvconstants]),%%ymm3                        \n" \
-  "vpsubw      %%ymm2,%%ymm3,%%ymm2                               \n" \
-  "vmovdqu     128(%[yuvconstants]),%%ymm3                        \n" \
-  "vpsubw      %%ymm1,%%ymm3,%%ymm1                               \n" \
-  "vmovdqu     96(%[yuvconstants]),%%ymm3                         \n" \
-  "vpsubw      %%ymm0,%%ymm3,%%ymm0                               \n" \
-  "vpmulhuw    192(%[yuvconstants]),%%ymm4,%%ymm4                 \n" \
+  "vpcmpeqb    %%xmm0,%%xmm0,%%xmm0                               \n" \
+  "vpsllw      $7,%%xmm0,%%xmm0                                   \n" \
+  "vpbroadcastb %%xmm0,%%ymm0                                     \n" \
+  "vpsubb      %%ymm0,%%ymm3,%%ymm3                               \n" \
+  "vpmulhuw    96(%[yuvconstants]),%%ymm4,%%ymm4                  \n" \
+  "vmovdqa     (%[yuvconstants]),%%ymm0                           \n" \
+  "vmovdqa     32(%[yuvconstants]),%%ymm1                         \n" \
+  "vmovdqa     64(%[yuvconstants]),%%ymm2                         \n" \
+  "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm1,%%ymm1                               \n" \
+  "vpmaddubsw  %%ymm3,%%ymm2,%%ymm2                               \n" \
+  "vmovdqa     128(%[yuvconstants]),%%ymm3                        \n" \
+  "vpaddw      %%ymm4,%%ymm3,%%ymm4                               \n" \
   "vpaddsw     %%ymm4,%%ymm0,%%ymm0                               \n" \
-  "vpaddsw     %%ymm4,%%ymm1,%%ymm1                               \n" \
+  "vpsubsw     %%ymm1,%%ymm4,%%ymm1                               \n" \
   "vpaddsw     %%ymm4,%%ymm2,%%ymm2                               \n"
+
 #define YUVTORGB_REGS_AVX2
 #endif
 
@@ -2540,6 +3459,15 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
   "vpackuswb   %%ymm1,%%ymm1,%%ymm1                               \n" \
   "vpackuswb   %%ymm2,%%ymm2,%%ymm2                               \n"
 
+#define YUVTORGB_AVX512BW(yuvconstants)                               \
+  YUVTORGB16_AVX512BW(yuvconstants)                                   \
+  "vpsraw     $0x6,%%zmm0,%%zmm0                                  \n" \
+  "vpsraw     $0x6,%%zmm1,%%zmm1                                  \n" \
+  "vpsraw     $0x6,%%zmm2,%%zmm2                                  \n" \
+  "vpackuswb  %%zmm0,%%zmm0,%%zmm0                                \n" \
+  "vpackuswb  %%zmm1,%%zmm1,%%zmm1                                \n" \
+  "vpackuswb  %%zmm2,%%zmm2,%%zmm2                                \n"
+
 // Store 16 ARGB values.
 #define STOREARGB_AVX2                                                \
   "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n" \
@@ -2550,7 +3478,19 @@ void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf,
   "vpunpckhwd %%ymm2,%%ymm0,%%ymm0                                \n" \
   "vmovdqu    %%ymm1,(%[dst_argb])                                \n" \
   "vmovdqu    %%ymm0,0x20(%[dst_argb])                            \n" \
-  "lea       0x40(%[dst_argb]), %[dst_argb]                       \n"
+  "lea        0x40(%[dst_argb]), %[dst_argb]                      \n"
+
+// Store 32 ARGB values.
+#define STOREARGB_AVX512BW                                            \
+  "vpunpcklbw %%zmm1,%%zmm0,%%zmm0                                \n" \
+  "vpermq     %%zmm0,%%zmm18,%%zmm0                               \n" \
+  "vpunpcklbw %%zmm5,%%zmm2,%%zmm2                                \n" \
+  "vpermq     %%zmm2,%%zmm18,%%zmm2                               \n" \
+  "vpunpcklwd %%zmm2,%%zmm0,%%zmm1                                \n" \
+  "vpunpckhwd %%zmm2,%%zmm0,%%zmm0                                \n" \
+  "vmovdqu8   %%zmm1,(%[dst_argb])                                \n" \
+  "vmovdqu8   %%zmm0,0x40(%[dst_argb])                            \n" \
+  "lea        0x80(%[dst_argb]), %[dst_argb]                      \n"
 
 // Store 16 AR30 values.
 #define STOREAR30_AVX2                                                \
@@ -2590,17 +3530,17 @@ void OMITFP I444ToARGBRow_AVX2(const uint8_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV444_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2624,18 +3564,18 @@ void OMITFP I422ToARGBRow_AVX2(const uint8_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
 
-    "vzeroupper                                \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2648,6 +3588,50 @@ void OMITFP I422ToARGBRow_AVX2(const uint8_t* y_buf,
 }
 #endif  // HAS_I422TOARGBROW_AVX2
 
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+static const uint64_t kSplitQuadWords[8] = {0, 2, 2, 2, 1, 2, 2, 2};
+static const uint64_t kSplitDoubleQuadWords[8] = {0, 1, 4, 4, 2, 3, 4, 4};
+static const uint64_t kUnpermuteAVX512[8] = {0, 4, 1, 5, 2, 6, 3, 7};
+
+// 32 pixels
+// 16 UV values upsampled to 32 UV, mixed with 32 Y producing 32 ARGB (128
+// bytes).
+void OMITFP I422ToARGBRow_AVX512BW(const uint8_t* y_buf,
+                                   const uint8_t* u_buf,
+                                   const uint8_t* v_buf,
+                                   uint8_t* dst_argb,
+                                   const struct YuvConstants* yuvconstants,
+                                   int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX512BW(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%xmm5,%%xmm5,%%xmm5          \n"
+      "vpbroadcastq %%xmm5,%%zmm5                \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV422_AVX512BW
+    YUVTORGB_AVX512BW(yuvconstants)
+    STOREARGB_AVX512BW
+      "sub         $0x20,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),                         // %[y_buf]
+    [u_buf]"+r"(u_buf),                         // %[u_buf]
+    [v_buf]"+r"(v_buf),                         // %[v_buf]
+    [dst_argb]"+r"(dst_argb),                   // %[dst_argb]
+    [width]"+rm"(width)                         // %[width]
+  : [yuvconstants]"r"(yuvconstants),            // %[yuvconstants]
+    [quadsplitperm]"r"(kSplitQuadWords),        // %[quadsplitperm]
+    [dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm]
+    [unperm]"r"(kUnpermuteAVX512)               // %[unperm]
+  : "memory", "cc", YUVTORGB_REGS_AVX512BW
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I422TOARGBROW_AVX512BW
+
 #if defined(HAS_I422TOAR30ROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
@@ -2659,23 +3643,23 @@ void OMITFP I422ToAR30Row_AVX2(const uint8_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"  // AR30 constants
-    "vpsrlw    $14,%%ymm5,%%ymm5               \n"
-    "vpsllw    $4,%%ymm5,%%ymm5                \n"  // 2 alpha bits
-    "vpxor     %%ymm6,%%ymm6,%%ymm6            \n"  // 0 for min
-    "vpcmpeqb  %%ymm7,%%ymm7,%%ymm7            \n"  // 1023 for max
-    "vpsrlw    $6,%%ymm7,%%ymm7                \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422_AVX2
     YUVTORGB16_AVX2(yuvconstants)
     STOREAR30_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
 
-    "vzeroupper                                \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2699,18 +3683,18 @@ void OMITFP I210ToARGBRow_AVX2(const uint16_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV210_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
 
-    "vzeroupper                                \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2723,6 +3707,41 @@ void OMITFP I210ToARGBRow_AVX2(const uint16_t* y_buf,
 }
 #endif  // HAS_I210TOARGBROW_AVX2
 
+#if defined(HAS_I212TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I212ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I212TOARGBROW_AVX2
+
 #if defined(HAS_I210TOAR30ROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
@@ -2734,23 +3753,23 @@ void OMITFP I210ToAR30Row_AVX2(const uint16_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"  // AR30 constants
-    "vpsrlw    $14,%%ymm5,%%ymm5               \n"
-    "vpsllw    $4,%%ymm5,%%ymm5                \n"  // 2 alpha bits
-    "vpxor     %%ymm6,%%ymm6,%%ymm6            \n"  // 0 for min
-    "vpcmpeqb  %%ymm7,%%ymm7,%%ymm7            \n"  // 1023 for max
-    "vpsrlw    $6,%%ymm7,%%ymm7                \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV210_AVX2
     YUVTORGB16_AVX2(yuvconstants)
     STOREAR30_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
 
-    "vzeroupper                                \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2758,11 +3777,239 @@ void OMITFP I210ToAR30Row_AVX2(const uint16_t* y_buf,
     [width]"+rm"(width)    // %[width]
   : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
   : "memory", "cc", YUVTORGB_REGS_AVX2
-    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
   );
 }
 #endif  // HAS_I210TOAR30ROW_AVX2
 
+#if defined(HAS_I212TOAR30ROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I212ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV212_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I212TOAR30ROW_AVX2
+
+#if defined(HAS_I410TOARGBROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I410ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_I410TOARGBROW_AVX2
+
+#if defined(HAS_I210ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 8 UV, 16 Y and 16 A producing 16 ARGB (64 bytes).
+void OMITFP I210AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                                    const uint16_t* u_buf,
+                                    const uint16_t* v_buf,
+                                    const uint16_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  asm volatile(
+      YUVTORGB_SETUP_AVX2(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA210_AVX2
+          YUVTORGB_AVX2(yuvconstants) STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : [y_buf] "+r"(y_buf),        // %[y_buf]
+        [u_buf] "+r"(u_buf),        // %[u_buf]
+        [v_buf] "+r"(v_buf),        // %[v_buf]
+        [a_buf] "+r"(a_buf),        // %[a_buf]
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", "xmm3",
+        "xmm4", "xmm5");
+}
+#endif  // HAS_I210TOARGBROW_AVX2
+
+#if defined(HAS_I410ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 16 UV, 16 Y and 16 A producing 16 ARGB (64 bytes).
+void OMITFP I410AlphaToARGBRow_AVX2(const uint16_t* y_buf,
+                                    const uint16_t* u_buf,
+                                    const uint16_t* v_buf,
+                                    const uint16_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  asm volatile(
+      YUVTORGB_SETUP_AVX2(
+      yuvconstants) "sub         %[u_buf],%[v_buf]             \n"
+
+      LABELALIGN "1:                                        \n" READYUVA410_AVX2
+          YUVTORGB_AVX2(yuvconstants) STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+
+      : [y_buf] "+r"(y_buf),        // %[y_buf]
+        [u_buf] "+r"(u_buf),        // %[u_buf]
+        [v_buf] "+r"(v_buf),        // %[v_buf]
+        [a_buf] "+r"(a_buf),        // %[a_buf]
+        [dst_argb] "+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+        [width] "+m"(width)  // %[width]
+#else
+        [width] "+rm"(width)  // %[width]
+#endif
+      : [yuvconstants] "r"(yuvconstants)  // %[yuvconstants]
+      : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", "xmm3",
+        "xmm4", "xmm5");
+}
+#endif  // HAS_I410TOARGBROW_AVX2
+
+#if defined(HAS_I410TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP I410ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* u_buf,
+                               const uint16_t* v_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READYUV410_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_I410TOAR30ROW_AVX2
+
+#if defined(HAS_I444ALPHATOARGBROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y and 16 A producing 16 ARGB.
+void OMITFP I444AlphaToARGBRow_AVX2(const uint8_t* y_buf,
+                                    const uint8_t* u_buf,
+                                    const uint8_t* v_buf,
+                                    const uint8_t* a_buf,
+                                    uint8_t* dst_argb,
+                                    const struct YuvConstants* yuvconstants,
+                                    int width) {
+  // clang-format off
+  asm volatile (
+  YUVTORGB_SETUP_AVX2(yuvconstants)
+      "sub         %[u_buf],%[v_buf]             \n"
+
+  LABELALIGN
+      "1:                                        \n"
+  READYUVA444_AVX2
+  YUVTORGB_AVX2(yuvconstants)
+  STOREARGB_AVX2
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [a_buf]"+r"(a_buf),    // %[a_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+#if defined(__i386__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_I444ALPHATOARGBROW_AVX2
+
 #if defined(HAS_I422ALPHATOARGBROW_AVX2)
 // 16 pixels
 // 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB.
@@ -2776,16 +4023,16 @@ void OMITFP I422AlphaToARGBRow_AVX2(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
+      "sub         %[u_buf],%[v_buf]             \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUVA422_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "subl      $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "subl        $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [u_buf]"+r"(u_buf),    // %[u_buf]
     [v_buf]"+r"(v_buf),    // %[v_buf]
@@ -2815,11 +4062,11 @@ void OMITFP I422ToRGBARow_AVX2(const uint8_t* y_buf,
                                int width) {
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "sub       %[u_buf],%[v_buf]               \n"
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "sub         %[u_buf],%[v_buf]             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUV422_AVX2
     YUVTORGB_AVX2(yuvconstants)
 
@@ -2859,16 +4106,16 @@ void OMITFP NV12ToARGBRow_AVX2(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READNV12_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [uv_buf]"+r"(uv_buf),    // %[uv_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
@@ -2892,16 +4139,16 @@ void OMITFP NV21ToARGBRow_AVX2(const uint8_t* y_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READNV21_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [y_buf]"+r"(y_buf),    // %[y_buf]
     [vu_buf]"+r"(vu_buf),    // %[vu_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
@@ -2925,16 +4172,16 @@ void OMITFP YUY2ToARGBRow_AVX2(const uint8_t* yuy2_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READYUY2_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [yuy2_buf]"+r"(yuy2_buf),    // %[yuy2_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
@@ -2958,16 +4205,16 @@ void OMITFP UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf,
   // clang-format off
   asm volatile (
     YUVTORGB_SETUP_AVX2(yuvconstants)
-    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
 
     LABELALIGN
-    "1:                                        \n"
+      "1:                                        \n"
     READUYVY_AVX2
     YUVTORGB_AVX2(yuvconstants)
     STOREARGB_AVX2
-    "sub       $0x10,%[width]                  \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : [uyvy_buf]"+r"(uyvy_buf),    // %[uyvy_buf]
     [dst_argb]"+r"(dst_argb),  // %[dst_argb]
     [width]"+rm"(width)    // %[width]
@@ -2981,18 +4228,156 @@ void OMITFP UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf,
 }
 #endif  // HAS_UYVYTOARGBROW_AVX2
 
+#if defined(HAS_P210TOARGBROW_AVX2)
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP P210ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  // clang-format off
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP210_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_P210TOARGBROW_AVX2
+
+#if defined(HAS_P410TOARGBROW_AVX2)
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP P410ToARGBRow_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_argb,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  // clang-format off
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP410_AVX2
+    YUVTORGB_AVX2(yuvconstants)
+    STOREARGB_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+  // clang-format on
+}
+#endif  // HAS_P410TOARGBROW_AVX2
+
+#if defined(HAS_P210TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP P210ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP210_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_P210TOAR30ROW_AVX2
+
+#if defined(HAS_P410TOAR30ROW_AVX2)
+// 16 pixels
+// 16 UV values with 16 Y producing 16 AR30 (64 bytes).
+void OMITFP P410ToAR30Row_AVX2(const uint16_t* y_buf,
+                               const uint16_t* uv_buf,
+                               uint8_t* dst_ar30,
+                               const struct YuvConstants* yuvconstants,
+                               int width) {
+  asm volatile (
+    YUVTORGB_SETUP_AVX2(yuvconstants)
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpxor       %%ymm6,%%ymm6,%%ymm6          \n"  // 0 for min
+      "vpcmpeqb    %%ymm7,%%ymm7,%%ymm7          \n"  // 1023 for max
+      "vpsrlw      $6,%%ymm7,%%ymm7              \n"
+
+    LABELALIGN
+      "1:                                        \n"
+    READP410_AVX2
+    YUVTORGB16_AVX2(yuvconstants)
+    STOREAR30_AVX2
+      "sub         $0x10,%[width]                \n"
+      "jg          1b                            \n"
+
+      "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_ar30]"+r"(dst_ar30),  // %[dst_ar30]
+    [width]"+rm"(width)    // %[width]
+  : [yuvconstants]"r"(yuvconstants)  // %[yuvconstants]
+  : "memory", "cc", YUVTORGB_REGS_AVX2
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_P410TOAR30ROW_AVX2
+
 #ifdef HAS_I400TOARGBROW_SSE2
-void I400ToARGBRow_SSE2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
+void I400ToARGBRow_SSE2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
   asm volatile(
-      "mov       $0x4a354a35,%%eax               \n"  // 4a35 = 18997 = 1.164
-      "movd      %%eax,%%xmm2                    \n"
-      "pshufd    $0x0,%%xmm2,%%xmm2              \n"
-      "mov       $0x04880488,%%eax               \n"  // 0488 = 1160 = 1.164 *
-                                                      // 16
-      "movd      %%eax,%%xmm3                    \n"
-      "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "pslld     $0x18,%%xmm4                    \n"
+      "movdqa      96(%3),%%xmm2                 \n"  // yg = 18997 = 1.164
+      "movdqa      128(%3),%%xmm3                \n"  // ygb = 1160 = 1.164 * 16
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"  // 0xff000000
+      "pslld       $0x18,%%xmm4                  \n"
 
       LABELALIGN
       "1:                                        \n"
@@ -3001,8 +4386,8 @@ void I400ToARGBRow_SSE2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
       "lea       0x8(%0),%0                      \n"
       "punpcklbw %%xmm0,%%xmm0                   \n"
       "pmulhuw   %%xmm2,%%xmm0                   \n"
-      "psubusw   %%xmm3,%%xmm0                   \n"
-      "psrlw     $6, %%xmm0                      \n"
+      "paddsw    %%xmm3,%%xmm0                   \n"
+      "psraw     $6, %%xmm0                      \n"
       "packuswb  %%xmm0,%%xmm0                   \n"
 
       // Step 2: Weave into ARGB
@@ -3018,28 +4403,26 @@ void I400ToARGBRow_SSE2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
 
       "sub       $0x8,%2                         \n"
       "jg        1b                              \n"
-      : "+r"(y_buf),     // %0
-        "+r"(dst_argb),  // %1
-        "+rm"(width)     // %2
-      :
-      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+      : "+r"(y_buf),       // %0
+        "+r"(dst_argb),    // %1
+        "+rm"(width)       // %2
+      : "r"(yuvconstants)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
 }
 #endif  // HAS_I400TOARGBROW_SSE2
 
 #ifdef HAS_I400TOARGBROW_AVX2
 // 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
 // note: vpunpcklbw mutates and vpackuswb unmutates.
-void I400ToARGBRow_AVX2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
+void I400ToARGBRow_AVX2(const uint8_t* y_buf,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
   asm volatile(
-      "mov        $0x4a354a35,%%eax              \n"  // 0488 = 1160 = 1.164 *
-                                                      // 16
-      "vmovd      %%eax,%%xmm2                   \n"
-      "vbroadcastss %%xmm2,%%ymm2                \n"
-      "mov        $0x4880488,%%eax               \n"  // 4a35 = 18997 = 1.164
-      "vmovd      %%eax,%%xmm3                   \n"
-      "vbroadcastss %%xmm3,%%ymm3                \n"
-      "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpslld     $0x18,%%ymm4,%%ymm4            \n"
+      "vmovdqa     96(%3),%%ymm2                 \n"  // yg = 18997 = 1.164
+      "vmovdqa     128(%3),%%ymm3                \n"  // ygb = -1160 = 1.164*16
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"  // 0xff000000
+      "vpslld      $0x18,%%ymm4,%%ymm4           \n"
 
       LABELALIGN
       "1:                                        \n"
@@ -3049,8 +4432,8 @@ void I400ToARGBRow_AVX2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
       "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
       "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
       "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpsubusw   %%ymm3,%%ymm0,%%ymm0           \n"
-      "vpsrlw     $0x6,%%ymm0,%%ymm0             \n"
+      "vpaddsw    %%ymm3,%%ymm0,%%ymm0           \n"
+      "vpsraw     $0x6,%%ymm0,%%ymm0             \n"
       "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
       "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
       "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
@@ -3060,15 +4443,15 @@ void I400ToARGBRow_AVX2(const uint8_t* y_buf, uint8_t* dst_argb, int width) {
       "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
       "vmovdqu    %%ymm0,(%1)                    \n"
       "vmovdqu    %%ymm1,0x20(%1)                \n"
-      "lea       0x40(%1),%1                     \n"
+      "lea        0x40(%1),%1                     \n"
       "sub        $0x10,%2                       \n"
       "jg        1b                              \n"
       "vzeroupper                                \n"
-      : "+r"(y_buf),     // %0
-        "+r"(dst_argb),  // %1
-        "+rm"(width)     // %2
-      :
-      : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+      : "+r"(y_buf),       // %0
+        "+r"(dst_argb),    // %1
+        "+rm"(width)       // %2
+      : "r"(yuvconstants)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
 }
 #endif  // HAS_I400TOARGBROW_AVX2
 
@@ -3081,16 +4464,16 @@ void MirrorRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile(
 
-      "movdqa    %3,%%xmm5                       \n"
+      "movdqa      %3,%%xmm5                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    -0x10(%0,%2,1),%%xmm0           \n"
-      "pshufb    %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      -0x10(%0,%2,1),%%xmm0         \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src),           // %0
         "+r"(dst),           // %1
         "+r"(temp_width)     // %2
@@ -3108,13 +4491,13 @@ void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    -0x20(%0,%2,1),%%ymm0          \n"
-      "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpermq     $0x4e,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     -0x20(%0,%2,1),%%ymm0         \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpermq      $0x4e,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),           // %0
         "+r"(dst),           // %1
@@ -3125,37 +4508,136 @@ void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
 #endif  // HAS_MIRRORROW_AVX2
 
 #ifdef HAS_MIRRORUVROW_SSSE3
+// Shuffle table for reversing the UV.
+static const uvec8 kShuffleMirrorUV = {14u, 15u, 12u, 13u, 10u, 11u, 8u, 9u,
+                                       6u,  7u,  4u,  5u,  2u,  3u,  0u, 1u};
+
+void MirrorUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile(
+
+      "movdqa      %3,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      -0x10(%0,%2,2),%%xmm0         \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),          // %0
+        "+r"(dst_uv),          // %1
+        "+r"(temp_width)       // %2
+      : "m"(kShuffleMirrorUV)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
+}
+#endif  // HAS_MIRRORUVROW_SSSE3
+
+#ifdef HAS_MIRRORUVROW_AVX2
+void MirrorUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     -0x20(%0,%2,2),%%ymm0         \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpermq      $0x4e,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uv),          // %0
+        "+r"(dst_uv),          // %1
+        "+r"(temp_width)       // %2
+      : "m"(kShuffleMirrorUV)  // %3
+      : "memory", "cc", "xmm0", "xmm5");
+}
+#endif  // HAS_MIRRORUVROW_AVX2
+
+#ifdef HAS_MIRRORSPLITUVROW_SSSE3
 // Shuffle table for reversing the bytes of UV channels.
-static const uvec8 kShuffleMirrorUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u,
-                                       15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u};
-void MirrorUVRow_SSSE3(const uint8_t* src,
-                       uint8_t* dst_u,
-                       uint8_t* dst_v,
-                       int width) {
+static const uvec8 kShuffleMirrorSplitUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u,
+                                            15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u};
+void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile(
-      "movdqa    %4,%%xmm1                       \n"
-      "lea       -0x10(%0,%3,2),%0               \n"
-      "sub       %1,%2                           \n"
+      "movdqa      %4,%%xmm1                     \n"
+      "lea         -0x10(%0,%3,2),%0             \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "lea       -0x10(%0),%0                    \n"
-      "pshufb    %%xmm1,%%xmm0                   \n"
-      "movlpd    %%xmm0,(%1)                     \n"
-      "movhpd    %%xmm0,0x00(%1,%2,1)            \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $8,%3                           \n"
-      "jg        1b                              \n"
-      : "+r"(src),             // %0
-        "+r"(dst_u),           // %1
-        "+r"(dst_v),           // %2
-        "+r"(temp_width)       // %3
-      : "m"(kShuffleMirrorUV)  // %4
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         -0x10(%0),%0                  \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"
+      "movlpd      %%xmm0,(%1)                   \n"
+      "movhpd      %%xmm0,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $8,%3                         \n"
+      "jg          1b                            \n"
+      : "+r"(src),                  // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(temp_width)            // %3
+      : "m"(kShuffleMirrorSplitUV)  // %4
       : "memory", "cc", "xmm0", "xmm1");
 }
-#endif  // HAS_MIRRORUVROW_SSSE3
+#endif  // HAS_MIRRORSPLITUVROW_SSSE3
+
+#ifdef HAS_RGB24MIRRORROW_SSSE3
+
+// Shuffle first 5 pixels to last 5 mirrored.  first byte zero
+static const uvec8 kShuffleMirrorRGB0 = {128u, 12u, 13u, 14u, 9u, 10u, 11u, 6u,
+                                         7u,   8u,  3u,  4u,  5u, 0u,  1u,  2u};
+
+// Shuffle last 5 pixels to first 5 mirrored.  last byte zero
+static const uvec8 kShuffleMirrorRGB1 = {
+    13u, 14u, 15u, 10u, 11u, 12u, 7u, 8u, 9u, 4u, 5u, 6u, 1u, 2u, 3u, 128u};
+
+// Shuffle 5 pixels at a time (15 bytes)
+void RGB24MirrorRow_SSSE3(const uint8_t* src_rgb24,
+                          uint8_t* dst_rgb24,
+                          int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  src_rgb24 += width * 3 - 48;
+  asm volatile(
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // first 5
+      "movdqu      15(%0),%%xmm1                 \n"  // next 5
+      "movdqu      30(%0),%%xmm2                 \n"  // next 5
+      "movdqu      32(%0),%%xmm3                 \n"  // last 1 special
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm2                 \n"
+      "pshufb      %%xmm5,%%xmm3                 \n"
+      "lea         -0x30(%0),%0                  \n"
+      "movdqu      %%xmm0,32(%1)                 \n"  // last 5
+      "movdqu      %%xmm1,17(%1)                 \n"  // next 5
+      "movdqu      %%xmm2,2(%1)                  \n"  // next 5
+      "movlpd      %%xmm3,0(%1)                  \n"  // first 1
+      "lea         0x30(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_rgb24),          // %0
+        "+r"(dst_rgb24),          // %1
+        "+r"(temp_width)          // %2
+      : "m"(kShuffleMirrorRGB0),  // %3
+        "m"(kShuffleMirrorRGB1)   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_RGB24MIRRORROW_SSSE3
 
 #ifdef HAS_ARGBMIRRORROW_SSE2
 
@@ -3163,17 +4645,17 @@ void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile(
 
-      "lea       -0x10(%0,%2,4),%0               \n"
+      "lea         -0x10(%0,%2,4),%0             \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "pshufd    $0x1b,%%xmm0,%%xmm0             \n"
-      "lea       -0x10(%0),%0                    \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "pshufd      $0x1b,%%xmm0,%%xmm0           \n"
+      "lea         -0x10(%0),%0                  \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),        // %0
         "+r"(dst),        // %1
         "+r"(temp_width)  // %2
@@ -3189,15 +4671,15 @@ void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   intptr_t temp_width = (intptr_t)(width);
   asm volatile(
 
-      "vmovdqu    %3,%%ymm5                      \n"
+      "vmovdqu     %3,%%ymm5                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vpermd    -0x20(%0,%2,4),%%ymm5,%%ymm0    \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x8,%2                        \n"
-      "jg         1b                             \n"
+      "vpermd      -0x20(%0,%2,4),%%ymm5,%%ymm0  \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),                    // %0
         "+r"(dst),                    // %1
@@ -3213,28 +4695,28 @@ void SplitUVRow_AVX2(const uint8_t* src_uv,
                      uint8_t* dst_v,
                      int width) {
   asm volatile(
-      "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-      "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
-      "sub        %1,%2                          \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm2             \n"
-      "vpsrlw     $0x8,%%ymm1,%%ymm3             \n"
-      "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpand      %%ymm5,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpackuswb  %%ymm3,%%ymm2,%%ymm2           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "vmovdqu    %%ymm2,0x00(%1,%2,1)            \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x20,%3                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm2            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm3            \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpackuswb   %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm2,%%ymm2           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x00(%1,%2,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_uv),  // %0
         "+r"(dst_u),   // %1
@@ -3251,28 +4733,28 @@ void SplitUVRow_SSE2(const uint8_t* src_uv,
                      uint8_t* dst_v,
                      int width) {
   asm volatile(
-      "pcmpeqb    %%xmm5,%%xmm5                  \n"
-      "psrlw      $0x8,%%xmm5                    \n"
-      "sub        %1,%2                          \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     0x10(%0),%%xmm1                \n"
-      "lea        0x20(%0),%0                    \n"
-      "movdqa     %%xmm0,%%xmm2                  \n"
-      "movdqa     %%xmm1,%%xmm3                  \n"
-      "pand       %%xmm5,%%xmm0                  \n"
-      "pand       %%xmm5,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                  \n"
-      "psrlw      $0x8,%%xmm2                    \n"
-      "psrlw      $0x8,%%xmm3                    \n"
-      "packuswb   %%xmm3,%%xmm2                  \n"
-      "movdqu     %%xmm0,(%1)                    \n"
-      "movdqu    %%xmm2,0x00(%1,%2,1)            \n"
-      "lea        0x10(%1),%1                    \n"
-      "sub        $0x10,%3                       \n"
-      "jg         1b                             \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "psrlw       $0x8,%%xmm3                   \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm2,0x00(%1,%2,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_uv),  // %0
         "+r"(dst_u),   // %1
         "+r"(dst_v),   // %2
@@ -3282,6 +4764,63 @@ void SplitUVRow_SSE2(const uint8_t* src_uv,
 }
 #endif  // HAS_SPLITUVROW_SSE2
 
+#ifdef HAS_DETILEROW_SSE2
+void DetileRow_SSE2(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "sub         $0x10,%2                      \n"
+      "lea         (%0,%3),%0                    \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "jg          1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "xmm0");
+}
+#endif  // HAS_DETILEROW_SSE2
+
+#ifdef HAS_DETILESPLITUVROW_SSSE3
+// TODO(greenjustin): Look into generating these constants instead of loading
+// them since this can cause branch mispredicts for fPIC code on 32-bit
+// machines.
+static const uvec8 kDeinterlaceUV = {0, 2, 4, 6, 8, 10, 12, 14,
+                                     1, 3, 5, 7, 9, 11, 13, 15};
+
+// TODO(greenjustin): Research alternatives to pshufb, since pshufb can be very
+// slow on older SSE2 processors.
+void DetileSplitUVRow_SSSE3(const uint8_t* src_uv,
+                            ptrdiff_t src_tile_stride,
+                            uint8_t* dst_u,
+                            uint8_t* dst_v,
+                            int width) {
+  asm volatile(
+      "movdqu      %4,%%xmm1                     \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         (%0, %5),%0                   \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "movhps      %%xmm0,(%2)                   \n"
+      "lea         0x8(%2),%2                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),         // %0
+        "+r"(dst_u),          // %1
+        "+r"(dst_v),          // %2
+        "+r"(width)           // %3
+      : "m"(kDeinterlaceUV),  // %4
+        "r"(src_tile_stride)  // %5
+      : "cc", "memory", "xmm0", "xmm1");
+}
+#endif  // HAS_DETILESPLITUVROW_SSSE3
+
 #ifdef HAS_MERGEUVROW_AVX2
 void MergeUVRow_AVX2(const uint8_t* src_u,
                      const uint8_t* src_v,
@@ -3289,22 +4828,22 @@ void MergeUVRow_AVX2(const uint8_t* src_u,
                      int width) {
   asm volatile(
 
-      "sub       %0,%1                           \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu    0x00(%0,%1,1),%%ymm1           \n"
-      "lea       0x20(%0),%0                     \n"
-      "vpunpcklbw %%ymm1,%%ymm0,%%ymm2           \n"
-      "vpunpckhbw %%ymm1,%%ymm0,%%ymm0           \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x00(%0,%1,1),%%ymm1          \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpckhbw  %%ymm1,%%ymm0,%%ymm0          \n"
       "vextractf128 $0x0,%%ymm2,(%2)             \n"
       "vextractf128 $0x0,%%ymm0,0x10(%2)         \n"
       "vextractf128 $0x1,%%ymm2,0x20(%2)         \n"
       "vextractf128 $0x1,%%ymm0,0x30(%2)         \n"
-      "lea       0x40(%2),%2                     \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_u),   // %0
         "+r"(src_v),   // %1
@@ -3322,21 +4861,21 @@ void MergeUVRow_SSE2(const uint8_t* src_u,
                      int width) {
   asm volatile(
 
-      "sub       %0,%1                           \n"
+      "sub         %0,%1                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%1,1),%%xmm1            \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "punpcklbw %%xmm1,%%xmm0                   \n"
-      "punpckhbw %%xmm1,%%xmm2                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "movdqu    %%xmm2,0x10(%2)                 \n"
-      "lea       0x20(%2),%2                     \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "movdqu      %%xmm2,0x10(%2)               \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_u),   // %0
         "+r"(src_v),   // %1
         "+r"(dst_uv),  // %2
@@ -3346,53 +4885,94 @@ void MergeUVRow_SSE2(const uint8_t* src_u,
 }
 #endif  // HAS_MERGEUVROW_SSE2
 
-// Use scale to convert lsb formats to msb, depending how many bits there are:
-// 128 = 9 bits
-// 64 = 10 bits
-// 16 = 12 bits
-// 1 = 16 bits
 #ifdef HAS_MERGEUVROW_16_AVX2
 void MergeUVRow_16_AVX2(const uint16_t* src_u,
                         const uint16_t* src_v,
                         uint16_t* dst_uv,
-                        int scale,
+                        int depth,
                         int width) {
+  depth = 16 - depth;
   // clang-format off
   asm volatile (
-    "vmovd      %4,%%xmm3                      \n"
-    "vpunpcklwd %%xmm3,%%xmm3,%%xmm3           \n"
-    "vbroadcastss %%xmm3,%%ymm3                \n"
-    "sub       %0,%1                           \n"
+      "vmovd       %4,%%xmm3                     \n"
+      "sub         %0,%1                         \n"
 
     // 16 pixels per loop.
     LABELALIGN
-    "1:                                        \n"
-    "vmovdqu   (%0),%%ymm0                     \n"
-    "vmovdqu   (%0,%1,1),%%ymm1                \n"
-    "add        $0x20,%0                       \n"
-
-    "vpmullw   %%ymm3,%%ymm0,%%ymm0            \n"
-    "vpmullw   %%ymm3,%%ymm1,%%ymm1            \n"
-    "vpunpcklwd %%ymm1,%%ymm0,%%ymm2           \n"  // mutates
-    "vpunpckhwd %%ymm1,%%ymm0,%%ymm0           \n"
-    "vextractf128 $0x0,%%ymm2,(%2)             \n"
-    "vextractf128 $0x0,%%ymm0,0x10(%2)         \n"
-    "vextractf128 $0x1,%%ymm2,0x20(%2)         \n"
-    "vextractf128 $0x1,%%ymm0,0x30(%2)         \n"
-    "add       $0x40,%2                        \n"
-    "sub       $0x10,%3                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     (%0,%1,1),%%ymm1              \n"
+      "add         $0x20,%0                      \n"
+
+      "vpsllw      %%xmm3,%%ymm0,%%ymm0          \n"
+      "vpsllw      %%xmm3,%%ymm1,%%ymm1          \n"
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm2          \n"  // mutates
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vextractf128 $0x0,%%ymm2,(%2)             \n"
+      "vextractf128 $0x0,%%ymm0,0x10(%2)         \n"
+      "vextractf128 $0x1,%%ymm2,0x20(%2)         \n"
+      "vextractf128 $0x1,%%ymm0,0x30(%2)         \n"
+      "add         $0x40,%2                      \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : "+r"(src_u),   // %0
     "+r"(src_v),   // %1
     "+r"(dst_uv),  // %2
     "+r"(width)    // %3
-  : "r"(scale)     // %4
+  : "r"(depth)     // %4
   : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
   // clang-format on
 }
 #endif  // HAS_MERGEUVROW_AVX2
 
+#ifdef HAS_SPLITUVROW_16_AVX2
+const uvec8 kSplitUVShuffle16 = {0, 1, 4, 5, 8,  9,  12, 13,
+                                 2, 3, 6, 7, 10, 11, 14, 15};
+void SplitUVRow_16_AVX2(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  depth = 16 - depth;
+  // clang-format off
+  asm volatile (
+      "vmovd       %4,%%xmm3                     \n"
+      "vbroadcastf128 %5,%%ymm4                  \n"
+      "sub         %1,%2                         \n"
+
+    // 16 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "add         $0x40,%0                      \n"
+
+      "vpsrlw      %%xmm3,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm3,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vextractf128 $0x0,%%ymm0,(%1)             \n"
+      "vextractf128 $0x0,%%ymm1,0x10(%1)         \n"
+      "vextractf128 $0x1,%%ymm0,(%1,%2)          \n"
+      "vextractf128 $0x1,%%ymm1,0x10(%1,%2)      \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_uv),   // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(width)     // %3
+  : "r"(depth),     // %4
+    "m"(kSplitUVShuffle16) // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+  // clang-format on
+}
+#endif  // HAS_SPLITUVROW_16_AVX2
+
 // Use scale to convert lsb formats to msb, depending how many bits there are:
 // 128 = 9 bits
 // 64 = 10 bits
@@ -3405,24 +4985,24 @@ void MultiplyRow_16_AVX2(const uint16_t* src_y,
                          int width) {
   // clang-format off
   asm volatile (
-    "vmovd      %3,%%xmm3                      \n"
-    "vpunpcklwd %%xmm3,%%xmm3,%%xmm3           \n"
-    "vbroadcastss %%xmm3,%%ymm3                \n"
-    "sub       %0,%1                           \n"
+      "vmovd       %3,%%xmm3                     \n"
+      "vpunpcklwd  %%xmm3,%%xmm3,%%xmm3          \n"
+      "vbroadcastss %%xmm3,%%ymm3                \n"
+      "sub         %0,%1                         \n"
 
-    // 16 pixels per loop.
+    // 32 pixels per loop.
     LABELALIGN
-    "1:                                        \n"
-    "vmovdqu   (%0),%%ymm0                     \n"
-    "vmovdqu   0x20(%0),%%ymm1                 \n"
-    "vpmullw   %%ymm3,%%ymm0,%%ymm0            \n"
-    "vpmullw   %%ymm3,%%ymm1,%%ymm1            \n"
-    "vmovdqu   %%ymm0,(%0,%1)                  \n"
-    "vmovdqu   %%ymm1,0x20(%0,%1)              \n"
-    "add        $0x40,%0                       \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpmullw     %%ymm3,%%ymm0,%%ymm0          \n"
+      "vpmullw     %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%0,%1)                \n"
+      "vmovdqu     %%ymm1,0x20(%0,%1)            \n"
+      "add         $0x40,%0                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : "+r"(src_y),   // %0
     "+r"(dst_y),   // %1
     "+r"(width)    // %2
@@ -3432,6 +5012,46 @@ void MultiplyRow_16_AVX2(const uint16_t* src_y,
 }
 #endif  // HAS_MULTIPLYROW_16_AVX2
 
+// Use scale to convert msb formats to lsb, depending how many bits there are:
+// 512 = 9 bits
+// 1024 = 10 bits
+// 4096 = 12 bits
+// 65536 = 16 bits
+#ifdef HAS_DIVIDEROW_16_AVX2
+void DivideRow_16_AVX2(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  // clang-format off
+  asm volatile (
+      "vmovd       %3,%%xmm3                     \n"
+      "vpunpcklwd  %%xmm3,%%xmm3,%%xmm3          \n"
+      "vbroadcastss %%xmm3,%%ymm3                \n"
+      "sub         %0,%1                         \n"
+
+    // 32 pixels per loop.
+    LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpmulhuw    %%ymm3,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%0,%1)                \n"
+      "vmovdqu     %%ymm1,0x20(%0,%1)            \n"
+      "add         $0x40,%0                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+  : "+r"(src_y),   // %0
+    "+r"(dst_y),   // %1
+    "+r"(width),    // %2
+    "+r"(scale)     // %3
+  :
+  : "memory", "cc", "xmm0", "xmm1", "xmm3");
+  // clang-format on
+}
+#endif  // HAS_MULTIPLYROW_16_AVX2
+
 // Use scale to convert lsb formats to msb, depending how many bits there are:
 // 32768 = 9 bits
 // 16384 = 10 bits
@@ -3443,23 +5063,23 @@ void Convert16To8Row_SSSE3(const uint16_t* src_y,
                            int width) {
   // clang-format off
   asm volatile (
-    "movd      %3,%%xmm2                      \n"
-    "punpcklwd %%xmm2,%%xmm2                  \n"
-    "pshufd    $0x0,%%xmm2,%%xmm2             \n"
+      "movd        %3,%%xmm2                     \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
 
     // 32 pixels per loop.
     LABELALIGN
-    "1:                                       \n"
-    "movdqu    (%0),%%xmm0                    \n"
-    "movdqu    0x10(%0),%%xmm1                \n"
-    "add       $0x20,%0                       \n"
-    "pmulhuw   %%xmm2,%%xmm0                  \n"
-    "pmulhuw   %%xmm2,%%xmm1                  \n"
-    "packuswb  %%xmm1,%%xmm0                  \n"
-    "movdqu    %%xmm0,(%1)                    \n"
-    "add       $0x10,%1                       \n"
-    "sub       $0x10,%2                       \n"
-    "jg        1b                             \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "add         $0x20,%0                      \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "add         $0x10,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
   : "+r"(src_y),   // %0
     "+r"(dst_y),   // %1
     "+r"(width)    // %2
@@ -3475,25 +5095,25 @@ void Convert16To8Row_AVX2(const uint16_t* src_y,
                           int width) {
   // clang-format off
   asm volatile (
-    "vmovd      %3,%%xmm2                      \n"
-    "vpunpcklwd %%xmm2,%%xmm2,%%xmm2           \n"
-    "vbroadcastss %%xmm2,%%ymm2                \n"
+      "vmovd       %3,%%xmm2                     \n"
+      "vpunpcklwd  %%xmm2,%%xmm2,%%xmm2          \n"
+      "vbroadcastss %%xmm2,%%ymm2                \n"
 
     // 32 pixels per loop.
     LABELALIGN
-    "1:                                        \n"
-    "vmovdqu   (%0),%%ymm0                     \n"
-    "vmovdqu   0x20(%0),%%ymm1                 \n"
-    "add       $0x40,%0                        \n"
-    "vpmulhuw  %%ymm2,%%ymm0,%%ymm0            \n"
-    "vpmulhuw  %%ymm2,%%ymm1,%%ymm1            \n"
-    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"  // mutates
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "vmovdqu   %%ymm0,(%1)                     \n"
-    "add       $0x20,%1                        \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "add         $0x40,%0                      \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"  // mutates
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : "+r"(src_y),   // %0
     "+r"(dst_y),   // %1
     "+r"(width)    // %2
@@ -3514,25 +5134,25 @@ void Convert8To16Row_SSE2(const uint8_t* src_y,
                           int width) {
   // clang-format off
   asm volatile (
-    "movd      %3,%%xmm2                      \n"
-    "punpcklwd %%xmm2,%%xmm2                  \n"
-    "pshufd    $0x0,%%xmm2,%%xmm2             \n"
+      "movd        %3,%%xmm2                     \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
 
     // 32 pixels per loop.
     LABELALIGN
-    "1:                                       \n"
-    "movdqu    (%0),%%xmm0                    \n"
-    "movdqa    %%xmm0,%%xmm1                  \n"
-    "punpcklbw %%xmm0,%%xmm0                  \n"
-    "punpckhbw %%xmm1,%%xmm1                  \n"
-    "add       $0x10,%0                       \n"
-    "pmulhuw   %%xmm2,%%xmm0                  \n"
-    "pmulhuw   %%xmm2,%%xmm1                  \n"
-    "movdqu    %%xmm0,(%1)                    \n"
-    "movdqu    %%xmm1,0x10(%1)                \n"
-    "add       $0x20,%1                       \n"
-    "sub       $0x10,%2                       \n"
-    "jg        1b                             \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "add         $0x10,%0                      \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "add         $0x20,%1                      \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
   : "+r"(src_y),   // %0
     "+r"(dst_y),   // %1
     "+r"(width)    // %2
@@ -3548,26 +5168,26 @@ void Convert8To16Row_AVX2(const uint8_t* src_y,
                           int width) {
   // clang-format off
   asm volatile (
-    "vmovd      %3,%%xmm2                      \n"
-    "vpunpcklwd %%xmm2,%%xmm2,%%xmm2           \n"
-    "vbroadcastss %%xmm2,%%ymm2                \n"
+      "vmovd       %3,%%xmm2                     \n"
+      "vpunpcklwd  %%xmm2,%%xmm2,%%xmm2          \n"
+      "vbroadcastss %%xmm2,%%ymm2                \n"
 
     // 32 pixels per loop.
     LABELALIGN
-    "1:                                        \n"
-    "vmovdqu   (%0),%%ymm0                     \n"
-    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-    "add       $0x20,%0                        \n"
-    "vpunpckhbw %%ymm0,%%ymm0,%%ymm1           \n"
-    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
-    "vpmulhuw  %%ymm2,%%ymm0,%%ymm0            \n"
-    "vpmulhuw  %%ymm2,%%ymm1,%%ymm1            \n"
-    "vmovdqu   %%ymm0,(%1)                     \n"
-    "vmovdqu   %%ymm1,0x20(%1)                 \n"
-    "add       $0x40,%1                        \n"
-    "sub       $0x20,%2                        \n"
-    "jg        1b                              \n"
-    "vzeroupper                                \n"
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "add         $0x20,%0                      \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm2,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "add         $0x40,%1                      \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
   : "+r"(src_y),   // %0
     "+r"(dst_y),   // %1
     "+r"(width)    // %2
@@ -3578,37 +5198,26 @@ void Convert8To16Row_AVX2(const uint8_t* src_y,
 #endif  // HAS_CONVERT8TO16ROW_AVX2
 
 #ifdef HAS_SPLITRGBROW_SSSE3
-
 // Shuffle table for converting RGB to Planar.
-static const uvec8 kShuffleMaskRGBToR0 = {0u,   3u,   6u,   9u,   12u,  15u,
-                                          128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToR1 = {128u, 128u, 128u, 128u, 128u, 128u,
-                                          2u,   5u,   8u,   11u,  14u,  128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToR2 = {128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u, 128u, 1u,
-                                          4u,   7u,   10u,  13u};
-
-static const uvec8 kShuffleMaskRGBToG0 = {1u,   4u,   7u,   10u,  13u,  128u,
-                                          128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToG1 = {128u, 128u, 128u, 128u, 128u, 0u,
-                                          3u,   6u,   9u,   12u,  15u,  128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToG2 = {128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u, 128u, 2u,
-                                          5u,   8u,   11u,  14u};
-
-static const uvec8 kShuffleMaskRGBToB0 = {2u,   5u,   8u,   11u,  14u,  128u,
-                                          128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToB1 = {128u, 128u, 128u, 128u, 128u, 1u,
-                                          4u,   7u,   10u,  13u,  128u, 128u,
-                                          128u, 128u, 128u, 128u};
-static const uvec8 kShuffleMaskRGBToB2 = {128u, 128u, 128u, 128u, 128u, 128u,
-                                          128u, 128u, 128u, 128u, 0u,   3u,
-                                          6u,   9u,   12u,  15u};
+static const uvec8 kSplitRGBShuffle[9] = {
+    {0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 1u, 4u,
+     7u, 10u, 13u},
+    {1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u,
+     8u, 11u, 14u},
+    {2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u,
+     128u, 128u},
+    {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u,
+     12u, 15u}};
 
 void SplitRGBRow_SSSE3(const uint8_t* src_rgb,
                        uint8_t* dst_r,
@@ -3619,91 +5228,72 @@ void SplitRGBRow_SSSE3(const uint8_t* src_rgb,
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     0x10(%0),%%xmm1                \n"
-      "movdqu     0x20(%0),%%xmm2                \n"
-      "pshufb     %5, %%xmm0                     \n"
-      "pshufb     %6, %%xmm1                     \n"
-      "pshufb     %7, %%xmm2                     \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,(%1)                    \n"
-      "lea        0x10(%1),%1                    \n"
-
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     0x10(%0),%%xmm1                \n"
-      "movdqu     0x20(%0),%%xmm2                \n"
-      "pshufb     %8, %%xmm0                     \n"
-      "pshufb     %9, %%xmm1                     \n"
-      "pshufb     %10, %%xmm2                    \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,(%2)                    \n"
-      "lea        0x10(%2),%2                    \n"
-
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     0x10(%0),%%xmm1                \n"
-      "movdqu     0x20(%0),%%xmm2                \n"
-      "pshufb     %11, %%xmm0                    \n"
-      "pshufb     %12, %%xmm1                    \n"
-      "pshufb     %13, %%xmm2                    \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,(%3)                    \n"
-      "lea        0x10(%3),%3                    \n"
-      "lea        0x30(%0),%0                    \n"
-      "sub        $0x10,%4                       \n"
-      "jg         1b                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      0(%5), %%xmm0                 \n"
+      "pshufb      16(%5), %%xmm1                \n"
+      "pshufb      32(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      48(%5),%%xmm0                 \n"
+      "pshufb      64(%5),%%xmm1                 \n"
+      "pshufb      80(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "pshufb      96(%5), %%xmm0                \n"
+      "pshufb      112(%5), %%xmm1               \n"
+      "pshufb      128(%5), %%xmm2               \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+      "lea         0x10(%3),%3                   \n"
+      "lea         0x30(%0),%0                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
       : "+r"(src_rgb),             // %0
         "+r"(dst_r),               // %1
         "+r"(dst_g),               // %2
         "+r"(dst_b),               // %3
         "+r"(width)                // %4
-      : "m"(kShuffleMaskRGBToR0),  // %5
-        "m"(kShuffleMaskRGBToR1),  // %6
-        "m"(kShuffleMaskRGBToR2),  // %7
-        "m"(kShuffleMaskRGBToG0),  // %8
-        "m"(kShuffleMaskRGBToG1),  // %9
-        "m"(kShuffleMaskRGBToG2),  // %10
-        "m"(kShuffleMaskRGBToB0),  // %11
-        "m"(kShuffleMaskRGBToB1),  // %12
-        "m"(kShuffleMaskRGBToB2)   // %13
+      : "r"(&kSplitRGBShuffle[0])  // %5
       : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_SPLITRGBROW_SSSE3
 
 #ifdef HAS_MERGERGBROW_SSSE3
-
-// Shuffle table for converting RGB to Planar.
-static const uvec8 kShuffleMaskRToRGB0 = {0u, 128u, 128u, 1u, 128u, 128u,
-                                          2u, 128u, 128u, 3u, 128u, 128u,
-                                          4u, 128u, 128u, 5u};
-static const uvec8 kShuffleMaskGToRGB0 = {128u, 0u, 128u, 128u, 1u, 128u,
-                                          128u, 2u, 128u, 128u, 3u, 128u,
-                                          128u, 4u, 128u, 128u};
-static const uvec8 kShuffleMaskBToRGB0 = {128u, 128u, 0u, 128u, 128u, 1u,
-                                          128u, 128u, 2u, 128u, 128u, 3u,
-                                          128u, 128u, 4u, 128u};
-
-static const uvec8 kShuffleMaskGToRGB1 = {5u, 128u, 128u, 6u, 128u, 128u,
-                                          7u, 128u, 128u, 8u, 128u, 128u,
-                                          9u, 128u, 128u, 10u};
-static const uvec8 kShuffleMaskBToRGB1 = {128u, 5u, 128u, 128u, 6u, 128u,
-                                          128u, 7u, 128u, 128u, 8u, 128u,
-                                          128u, 9u, 128u, 128u};
-static const uvec8 kShuffleMaskRToRGB1 = {128u, 128u, 6u,  128u, 128u, 7u,
-                                          128u, 128u, 8u,  128u, 128u, 9u,
-                                          128u, 128u, 10u, 128u};
-
-static const uvec8 kShuffleMaskBToRGB2 = {10u, 128u, 128u, 11u, 128u, 128u,
-                                          12u, 128u, 128u, 13u, 128u, 128u,
-                                          14u, 128u, 128u, 15u};
-static const uvec8 kShuffleMaskRToRGB2 = {128u, 11u, 128u, 128u, 12u, 128u,
-                                          128u, 13u, 128u, 128u, 14u, 128u,
-                                          128u, 15u, 128u, 128u};
-static const uvec8 kShuffleMaskGToRGB2 = {128u, 128u, 11u, 128u, 128u, 12u,
-                                          128u, 128u, 13u, 128u, 128u, 14u,
-                                          128u, 128u, 15u, 128u};
+// Shuffle table for converting Planar to RGB.
+static const uvec8 kMergeRGBShuffle[9] = {
+    {0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u, 128u,
+     128u, 5u},
+    {128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u,
+     128u, 128u},
+    {128u, 128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u,
+     4u, 128u},
+    {128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u, 128u,
+     10u, 128u},
+    {5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u,
+     128u, 10u},
+    {128u, 5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u,
+     128u, 128u},
+    {128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u, 128u,
+     15u, 128u, 128u},
+    {128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u,
+     128u, 15u, 128u},
+    {10u, 128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u,
+     128u, 128u, 15u}};
 
 void MergeRGBRow_SSSE3(const uint8_t* src_r,
                        const uint8_t* src_g,
@@ -3714,92 +5304,858 @@ void MergeRGBRow_SSSE3(const uint8_t* src_r,
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     (%1),%%xmm1                    \n"
-      "movdqu     (%2),%%xmm2                    \n"
-      "pshufb     %5, %%xmm0                     \n"
-      "pshufb     %6, %%xmm1                     \n"
-      "pshufb     %7, %%xmm2                     \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,(%3)                    \n"
-
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     (%1),%%xmm1                    \n"
-      "movdqu     (%2),%%xmm2                    \n"
-      "pshufb     %8, %%xmm0                     \n"
-      "pshufb     %9, %%xmm1                     \n"
-      "pshufb     %10, %%xmm2                    \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,16(%3)                  \n"
-
-      "movdqu     (%0),%%xmm0                    \n"
-      "movdqu     (%1),%%xmm1                    \n"
-      "movdqu     (%2),%%xmm2                    \n"
-      "pshufb     %11, %%xmm0                    \n"
-      "pshufb     %12, %%xmm1                    \n"
-      "pshufb     %13, %%xmm2                    \n"
-      "por        %%xmm1,%%xmm0                  \n"
-      "por        %%xmm2,%%xmm0                  \n"
-      "movdqu     %%xmm0,32(%3)                  \n"
-
-      "lea        0x10(%0),%0                    \n"
-      "lea        0x10(%1),%1                    \n"
-      "lea        0x10(%2),%2                    \n"
-      "lea        0x30(%3),%3                    \n"
-      "sub        $0x10,%4                       \n"
-      "jg         1b                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      (%5), %%xmm0                  \n"
+      "pshufb      16(%5), %%xmm1                \n"
+      "pshufb      32(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      48(%5), %%xmm0                \n"
+      "pshufb      64(%5), %%xmm1                \n"
+      "pshufb      80(%5), %%xmm2                \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,16(%3)                 \n"
+
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "pshufb      96(%5), %%xmm0                \n"
+      "pshufb      112(%5), %%xmm1               \n"
+      "pshufb      128(%5), %%xmm2               \n"
+      "por         %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,32(%3)                 \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "lea         0x30(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
       : "+r"(src_r),               // %0
         "+r"(src_g),               // %1
         "+r"(src_b),               // %2
         "+r"(dst_rgb),             // %3
         "+r"(width)                // %4
-      : "m"(kShuffleMaskRToRGB0),  // %5
-        "m"(kShuffleMaskGToRGB0),  // %6
-        "m"(kShuffleMaskBToRGB0),  // %7
-        "m"(kShuffleMaskRToRGB1),  // %8
-        "m"(kShuffleMaskGToRGB1),  // %9
-        "m"(kShuffleMaskBToRGB1),  // %10
-        "m"(kShuffleMaskRToRGB2),  // %11
-        "m"(kShuffleMaskGToRGB2),  // %12
-        "m"(kShuffleMaskBToRGB2)   // %13
+      : "r"(&kMergeRGBShuffle[0])  // %5
       : "memory", "cc", "xmm0", "xmm1", "xmm2");
 }
 #endif  // HAS_MERGERGBROW_SSSE3
 
+#ifdef HAS_MERGEARGBROW_SSE2
+void MergeARGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movq        (%0,%2),%%xmm0                \n"  // B
+      "movq        (%0),%%xmm1                   \n"  // R
+      "movq        (%0,%1),%%xmm2                \n"  // G
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // BR
+      "movq        (%0,%3),%%xmm1                \n"  // A
+      "punpcklbw   %%xmm1,%%xmm2                 \n"  // GA
+      "movdqa      %%xmm0,%%xmm1                 \n"  // BR
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // BGRA (hi)
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // BGRA (lo)
+      "movdqu      %%xmm0,(%4)                   \n"
+      "movdqu      %%xmm1,16(%4)                 \n"
+
+      "lea         8(%0),%0                      \n"
+      "lea         32(%4),%4                     \n"
+      "sub         $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_MERGEXRGBROW_SSE2
+void MergeXRGBRow_SSE2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movq        (%2),%%xmm0                   \n"  // B
+      "movq        (%0),%%xmm1                   \n"  // R
+      "movq        (%1),%%xmm2                   \n"  // G
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // BR
+      "pcmpeqd     %%xmm1,%%xmm1                 \n"  // A(255)
+      "punpcklbw   %%xmm1,%%xmm2                 \n"  // GA
+      "movdqa      %%xmm0,%%xmm1                 \n"  // BR
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // BGRA (hi)
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // BGRA (lo)
+      "movdqu      %%xmm0,(%3)                   \n"
+      "movdqu      %%xmm1,16(%3)                 \n"
+
+      "lea         8(%0),%0                      \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         32(%3),%3                     \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGEARGBROW_SSE2
+
+#ifdef HAS_MERGEARGBROW_AVX2
+void MergeARGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0,%2),%%xmm0                \n"  // B
+      "vmovdqu     (%0,%1),%%xmm1                \n"  // R
+      "vinserti128 $1,(%0),%%ymm0,%%ymm0         \n"  // G
+      "vinserti128 $1,(%0,%3),%%ymm1,%%ymm1      \n"  // A
+      "vpunpckhbw  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vmovdqu     %%ymm0,(%4)                   \n"  // First 8
+      "vmovdqu     %%ymm1,32(%4)                 \n"  // Next 8
+
+      "lea         16(%0),%0                     \n"
+      "lea         64(%4),%4                     \n"
+      "sub         $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_MERGEXRGBROW_AVX2
+void MergeXRGBRow_AVX2(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%2),%%xmm0                   \n"  // B
+      "vpcmpeqd    %%ymm1,%%ymm1,%%ymm1          \n"  // A(255)
+      "vinserti128 $0,(%1),%%ymm1,%%ymm1         \n"  // R
+      "vinserti128 $1,(%0),%%ymm0,%%ymm0         \n"  // G
+      "vpunpckhbw  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm0,%%ymm1    \n"
+      "vperm2i128  $0x20,%%ymm2,%%ymm0,%%ymm0    \n"
+      "vmovdqu     %%ymm0,(%3)                   \n"  // First 8
+      "vmovdqu     %%ymm1,32(%3)                 \n"  // Next 8
+
+      "lea         16(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "lea         16(%2),%2                     \n"
+      "lea         64(%3),%3                     \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif  // HAS_MERGEARGBROW_AVX2
+
+#ifdef HAS_SPLITARGBROW_SSE2
+void SplitARGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 00-07 10-17
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 08-0F 18-1F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 08192A3B4C5D6E7F (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 08192A3B4C5D6E7F (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 08192A3B08192A3B
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 4C5D6E7F4C5D6E7F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%1,%3)                \n"  // B
+      "movhps      %%xmm0,(%1,%2)                \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+      "movhps      %%xmm2,(%1,%4)                \n"  // A
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "sub         $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+        "+rm"(width)     // %5
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_SSE2
+void SplitXRGBRow_SSE2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 00-07 10-17
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 08-0F 18-1F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 08192A3B4C5D6E7F (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 08192A3B4C5D6E7F (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"  // 08192A3B08192A3B
+      "punpckhqdq  %%xmm1,%%xmm2                 \n"  // 4C5D6E7F4C5D6E7F
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "punpckhbw   %%xmm2,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%3)                   \n"  // B
+      "movhps      %%xmm0,(%2)                   \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         8(%3),%3                      \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+rm"(width)     // %4
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2");
+}
+#endif
+
+static const uvec8 kShuffleMaskARGBSplit = {0, 4, 8,  12, 1, 5, 9,  13,
+                                            2, 6, 10, 14, 3, 7, 11, 15};
+#ifdef HAS_SPLITARGBROW_SSSE3
+void SplitARGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        uint8_t* dst_a,
+                        int width) {
+  asm volatile(
+
+      "movdqa      %6,%%xmm3                     \n"
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "pshufb      %%xmm3,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "pshufb      %%xmm3,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%1,%3)                \n"  // B
+      "movhps      %%xmm0,(%1,%2)                \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+      "movhps      %%xmm2,(%1,%4)                \n"  // A
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "subl        $0x8,%5                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(kShuffleMaskARGBSplit)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_SSSE3
+void SplitXRGBRow_SSSE3(const uint8_t* src_argb,
+                        uint8_t* dst_r,
+                        uint8_t* dst_g,
+                        uint8_t* dst_b,
+                        int width) {
+  asm volatile(
+
+      "movdqa      %5,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "movdqu      (%0),%%xmm0                   \n"  // 00-0F
+      "movdqu      16(%0),%%xmm1                 \n"  // 10-1F
+      "pshufb      %%xmm3,%%xmm0                 \n"  // 048C159D26AE37BF (lo)
+      "pshufb      %%xmm3,%%xmm1                 \n"  // 048C159D26AE37BF (hi)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 048C048C159D159D (BG)
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 26AE26AE37BF37BF (RA)
+      "movlps      %%xmm0,(%3)                   \n"  // B
+      "movhps      %%xmm0,(%2)                   \n"  // G
+      "movlps      %%xmm2,(%1)                   \n"  // R
+
+      "lea         32(%0),%0                     \n"
+      "lea         8(%1),%1                      \n"
+      "lea         8(%2),%2                      \n"
+      "lea         8(%3),%3                      \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),             // %0
+        "+r"(dst_r),                // %1
+        "+r"(dst_g),                // %2
+        "+r"(dst_b),                // %3
+        "+r"(width)                 // %4
+      : "m"(kShuffleMaskARGBSplit)  // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+}
+#endif
+
+#ifdef HAS_SPLITARGBROW_AVX2
+static const ulvec32 kShuffleMaskARGBPermute = {0, 4, 1, 5, 2, 6, 3, 7};
+void SplitARGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+
+      "sub         %1,%2                         \n"
+      "sub         %1,%3                         \n"
+      "sub         %1,%4                         \n"
+      "vmovdqa     %7,%%ymm3                     \n"
+      "vbroadcastf128 %6,%%ymm4                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00-0F
+      "vmovdqu     16(%0),%%xmm1                 \n"  // 10-1F
+      "vinserti128 $1,32(%0),%%ymm0,%%ymm0       \n"  // 00-0F 20-2F
+      "vinserti128 $1,48(%0),%%ymm1,%%ymm1       \n"  // 10-1F 30-3F
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermd      %%ymm0,%%ymm3,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm3,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"  // GA
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"  // BR
+      "vmovdqu     %%xmm0,(%1,%3)                \n"  // B
+      "vextracti128 $1,%%ymm0,(%1)               \n"  // R
+      "vmovdqu     %%xmm2,(%1,%2)                \n"  // G
+      "vextracti128 $1,%%ymm2,(%1,%4)            \n"  // A
+      "lea         64(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_r),     // %1
+        "+r"(dst_g),     // %2
+        "+r"(dst_b),     // %3
+        "+r"(dst_a),     // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(kShuffleMaskARGBSplit),   // %6
+        "m"(kShuffleMaskARGBPermute)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SPLITXRGBROW_AVX2
+void SplitXRGBRow_AVX2(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+
+      "vmovdqa     %6,%%ymm3                     \n"
+      "vbroadcastf128 %5,%%ymm4                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00-0F
+      "vmovdqu     16(%0),%%xmm1                 \n"  // 10-1F
+      "vinserti128 $1,32(%0),%%ymm0,%%ymm0       \n"  // 00-0F 20-2F
+      "vinserti128 $1,48(%0),%%ymm1,%%ymm1       \n"  // 10-1F 30-3F
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpermd      %%ymm0,%%ymm3,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm3,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"  // GA
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"  // BR
+      "vmovdqu     %%xmm0,(%3)                   \n"  // B
+      "vextracti128 $1,%%ymm0,(%1)               \n"  // R
+      "vmovdqu     %%xmm2,(%2)                   \n"  // G
+
+      "lea         64(%0),%0                     \n"
+      "lea         16(%1),%1                     \n"
+      "lea         16(%2),%2                     \n"
+      "lea         16(%3),%3                     \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_argb),               // %0
+        "+r"(dst_r),                  // %1
+        "+r"(dst_g),                  // %2
+        "+r"(dst_b),                  // %3
+        "+r"(width)                   // %4
+      : "m"(kShuffleMaskARGBSplit),   // %5
+        "m"(kShuffleMaskARGBPermute)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_MERGEXR30ROW_AVX2
+void MergeXR30Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = depth - 10;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"  // AR30 constants
+      "vpsrlw      $14,%%ymm5,%%ymm5             \n"
+      "vpsllw      $4,%%ymm5,%%ymm5              \n"  // 2 alpha bits
+      "vpcmpeqb    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $6,%%ymm6,%%ymm6              \n"
+      "vmovd       %5,%%xmm4                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     (%0,%1),%%ymm1                \n"
+      "vmovdqu     (%0,%2),%%ymm2                \n"
+      "vpsrlw      %%xmm4,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm4,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm4,%%ymm2,%%ymm2          \n"
+      "vpminuw     %%ymm0,%%ymm6,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm6,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm6,%%ymm2          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm2,%%ymm2           \n"
+      "vpsllw      $0x4,%%ymm0,%%ymm0            \n"  // Shift R to target bit
+      "vpunpckhwd  %%ymm0,%%ymm2,%%ymm3          \n"  // RB
+      "vpunpcklwd  %%ymm0,%%ymm2,%%ymm0          \n"
+      "vpunpckhwd  %%ymm5,%%ymm1,%%ymm2          \n"  // AG
+      "vpunpcklwd  %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpslld      $0xa,%%ymm1,%%ymm1            \n"  // Shift AG to target bit
+      "vpslld      $0xa,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm0,%%ymm0          \n"  // Combine
+      "vpor        %%ymm2,%%ymm3,%%ymm3          \n"
+      "vmovdqu     %%ymm0,(%3)                   \n"
+      "vmovdqu     %%ymm3,0x20(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+#if defined(__i386__)
+      : "m"(shift)  // %5
+#else
+      : "rm"(shift)           // %5
+#endif
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_MERGEAR64ROW_AVX2
+static const lvec32 MergeAR64Permute = {0, 4, 2, 6, 1, 5, 3, 7};
+void MergeAR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  mask = (mask << 16) + mask;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "vmovdqa     %8,%%ymm5                     \n"
+      "vmovd       %6,%%xmm6                     \n"
+      "vbroadcastss %7,%%ymm7                    \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vmovdqu     (%0,%3),%%ymm3                \n"  // A
+      "vpminuw     %%ymm0,%%ymm7,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpminuw     %%ymm3,%%ymm7,%%ymm3          \n"
+      "vpsllw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsllw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsllw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpsllw      %%xmm6,%%ymm3,%%ymm3          \n"
+      "vpermd      %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm5,%%ymm2          \n"
+      "vpermd      %%ymm3,%%ymm5,%%ymm3          \n"
+      "vpunpcklwd  %%ymm1,%%ymm2,%%ymm4          \n"  // BG(low)
+      "vpunpckhwd  %%ymm1,%%ymm2,%%ymm1          \n"  // BG(hi)
+      "vpunpcklwd  %%ymm3,%%ymm0,%%ymm2          \n"  // RA(low)
+      "vpunpckhwd  %%ymm3,%%ymm0,%%ymm0          \n"  // RA(hi)
+      "vpunpckldq  %%ymm2,%%ymm4,%%ymm3          \n"  // BGRA(1)
+      "vpunpckhdq  %%ymm2,%%ymm4,%%ymm4          \n"  // BGRA(3)
+      "vpunpckldq  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA(2)
+      "vpunpckhdq  %%ymm0,%%ymm1,%%ymm1          \n"  // BGRA(4)
+      "vmovdqu     %%ymm3,(%4)                   \n"
+      "vmovdqu     %%ymm2,0x20(%4)               \n"
+      "vmovdqu     %%ymm4,0x40(%4)               \n"
+      "vmovdqu     %%ymm1,0x60(%4)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x80(%4),%4                   \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(shift),            // %6
+        "m"(mask),             // %7
+        "m"(MergeAR64Permute)  // %8
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_MERGEXR64ROW_AVX2
+void MergeXR64Row_AVX2(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  mask = (mask << 16) + mask;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vmovdqa     %7,%%ymm5                     \n"
+      "vmovd       %5,%%xmm6                     \n"
+      "vbroadcastss %6,%%ymm7                    \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vpminuw     %%ymm0,%%ymm7,%%ymm0          \n"
+      "vpminuw     %%ymm1,%%ymm7,%%ymm1          \n"
+      "vpminuw     %%ymm2,%%ymm7,%%ymm2          \n"
+      "vpsllw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsllw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsllw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpermd      %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpermd      %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpermd      %%ymm2,%%ymm5,%%ymm2          \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"  // A (0xffff)
+      "vpunpcklwd  %%ymm1,%%ymm2,%%ymm4          \n"  // BG(low)
+      "vpunpckhwd  %%ymm1,%%ymm2,%%ymm1          \n"  // BG(hi)
+      "vpunpcklwd  %%ymm3,%%ymm0,%%ymm2          \n"  // RA(low)
+      "vpunpckhwd  %%ymm3,%%ymm0,%%ymm0          \n"  // RA(hi)
+      "vpunpckldq  %%ymm2,%%ymm4,%%ymm3          \n"  // BGRA(1)
+      "vpunpckhdq  %%ymm2,%%ymm4,%%ymm4          \n"  // BGRA(3)
+      "vpunpckldq  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA(2)
+      "vpunpckhdq  %%ymm0,%%ymm1,%%ymm1          \n"  // BGRA(4)
+      "vmovdqu     %%ymm3,(%3)                   \n"
+      "vmovdqu     %%ymm2,0x20(%3)               \n"
+      "vmovdqu     %%ymm4,0x40(%3)               \n"
+      "vmovdqu     %%ymm1,0x60(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x80(%3),%3                   \n"
+      "subl        $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),           // %0
+        "+r"(src_g),           // %1
+        "+r"(src_b),           // %2
+        "+r"(dst_ar64),        // %3
+        "+r"(width)            // %4
+      : "m"(shift),            // %5
+        "m"(mask),             // %6
+        "m"(MergeAR64Permute)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_MERGEARGB16TO8ROW_AVX2
+static const uvec8 MergeARGB16To8Shuffle = {0, 8,  1, 9,  2, 10, 3, 11,
+                                            4, 12, 5, 13, 6, 14, 7, 15};
+void MergeARGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = depth - 8;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "vbroadcastf128 %7,%%ymm5                  \n"
+      "vmovd       %6,%%xmm6                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vmovdqu     (%0,%3),%%ymm3                \n"  // A
+      "vpsrlw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpsrlw      %%xmm6,%%ymm3,%%ymm3          \n"
+      "vpackuswb   %%ymm1,%%ymm2,%%ymm1          \n"  // BG (planar)
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"  // RA (planar)
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"  // BG (interleave)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // RA (interleave)
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpcklwd  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA (low)
+      "vpunpckhwd  %%ymm0,%%ymm1,%%ymm0          \n"  // BGRA (hi)
+      "vmovdqu     %%ymm2,(%4)                   \n"
+      "vmovdqu     %%ymm0,0x20(%4)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%4),%4                   \n"
+      "subl        $0x10,%5                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+#if defined(__i386__)
+        "+m"(width)  // %5
+#else
+        "+rm"(width)          // %5
+#endif
+      : "m"(shift),                 // %6
+        "m"(MergeARGB16To8Shuffle)  // %7
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_MERGEXRGB16TO8ROW_AVX2
+void MergeXRGB16To8Row_AVX2(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = depth - 8;
+  asm volatile(
+
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "vbroadcastf128 %6,%%ymm5                  \n"
+      "vmovd       %5,%%xmm6                     \n"
+      "vpcmpeqb    %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpsrlw      $8,%%ymm3,%%ymm3              \n"  // A (0xff)
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // R
+      "vmovdqu     (%0,%1),%%ymm1                \n"  // G
+      "vmovdqu     (%0,%2),%%ymm2                \n"  // B
+      "vpsrlw      %%xmm6,%%ymm0,%%ymm0          \n"
+      "vpsrlw      %%xmm6,%%ymm1,%%ymm1          \n"
+      "vpsrlw      %%xmm6,%%ymm2,%%ymm2          \n"
+      "vpackuswb   %%ymm1,%%ymm2,%%ymm1          \n"  // BG (planar)
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"  // RA (planar)
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"  // BG (interleave)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // RA (interleave)
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpunpcklwd  %%ymm0,%%ymm1,%%ymm2          \n"  // BGRA (low)
+      "vpunpckhwd  %%ymm0,%%ymm1,%%ymm0          \n"  // BGRA (hi)
+      "vmovdqu     %%ymm2,(%3)                   \n"
+      "vmovdqu     %%ymm0,0x20(%3)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%3),%3                   \n"
+      "subl        $0x10,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_r),                // %0
+        "+r"(src_g),                // %1
+        "+r"(src_b),                // %2
+        "+r"(dst_argb),             // %3
+        "+r"(width)                 // %4
+      : "m"(shift),                 // %5
+        "m"(MergeARGB16To8Shuffle)  // %6
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
 #ifdef HAS_COPYROW_SSE2
 void CopyRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "test       $0xf,%0                        \n"
-      "jne        2f                             \n"
-      "test       $0xf,%1                        \n"
-      "jne        2f                             \n"
+      "test        $0xf,%0                       \n"
+      "jne         2f                            \n"
+      "test        $0xf,%1                       \n"
+      "jne         2f                            \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqa    (%0),%%xmm0                     \n"
-      "movdqa    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "movdqa    %%xmm0,(%1)                     \n"
-      "movdqa    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
-      "jmp       9f                              \n"
+      "movdqa      (%0),%%xmm0                   \n"
+      "movdqa      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,(%1)                   \n"
+      "movdqa      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         9f                            \n"
 
       LABELALIGN
       "2:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        2b                              \n"
-
-      LABELALIGN "9:                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          2b                            \n"
+
+      LABELALIGN "9:                                        \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -3814,14 +6170,14 @@ void CopyRow_AVX(const uint8_t* src, uint8_t* dst, int width) {
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vmovdqu   %%ymm0,(%1)                     \n"
-      "vmovdqu   %%ymm1,0x20(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x40,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x40,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -3836,7 +6192,7 @@ void CopyRow_ERMS(const uint8_t* src, uint8_t* dst, int width) {
   size_t width_tmp = (size_t)(width);
   asm volatile(
 
-      "rep movsb                      \n"
+      "rep         movsb                         \n"
       : "+S"(src),       // %0
         "+D"(dst),       // %1
         "+c"(width_tmp)  // %2
@@ -3849,29 +6205,29 @@ void CopyRow_ERMS(const uint8_t* src, uint8_t* dst, int width) {
 // width in pixels
 void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm0,%%xmm0                   \n"
-      "pslld     $0x18,%%xmm0                    \n"
-      "pcmpeqb   %%xmm1,%%xmm1                   \n"
-      "psrld     $0x8,%%xmm1                     \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "movdqu    (%0),%%xmm2                     \n"
-      "movdqu    0x10(%0),%%xmm3                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "movdqu    (%1),%%xmm4                     \n"
-      "movdqu    0x10(%1),%%xmm5                 \n"
-      "pand      %%xmm0,%%xmm2                   \n"
-      "pand      %%xmm0,%%xmm3                   \n"
-      "pand      %%xmm1,%%xmm4                   \n"
-      "pand      %%xmm1,%%xmm5                   \n"
-      "por       %%xmm4,%%xmm2                   \n"
-      "por       %%xmm5,%%xmm3                   \n"
-      "movdqu    %%xmm2,(%1)                     \n"
-      "movdqu    %%xmm3,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "pcmpeqb     %%xmm0,%%xmm0                 \n"
+      "pslld       $0x18,%%xmm0                  \n"
+      "pcmpeqb     %%xmm1,%%xmm1                 \n"
+      "psrld       $0x8,%%xmm1                   \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      0x10(%0),%%xmm3               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqu      (%1),%%xmm4                   \n"
+      "movdqu      0x10(%1),%%xmm5               \n"
+      "pand        %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm0,%%xmm3                 \n"
+      "pand        %%xmm1,%%xmm4                 \n"
+      "pand        %%xmm1,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -3884,21 +6240,21 @@ void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 // width in pixels
 void ARGBCopyAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+      "vpcmpeqb    %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm1                     \n"
-      "vmovdqu   0x20(%0),%%ymm2                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpblendvb %%ymm0,(%1),%%ymm1,%%ymm1       \n"
-      "vpblendvb %%ymm0,0x20(%1),%%ymm2,%%ymm2   \n"
-      "vmovdqu   %%ymm1,(%1)                     \n"
-      "vmovdqu   %%ymm2,0x20(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm1                   \n"
+      "vmovdqu     0x20(%0),%%ymm2               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpblendvb   %%ymm0,(%1),%%ymm1,%%ymm1     \n"
+      "vpblendvb   %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n"
+      "vmovdqu     %%ymm1,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
@@ -3917,17 +6273,17 @@ void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb,
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0), %%xmm0                    \n"
-      "movdqu    0x10(%0), %%xmm1                \n"
-      "lea       0x20(%0), %0                    \n"
-      "psrld     $0x18, %%xmm0                   \n"
-      "psrld     $0x18, %%xmm1                   \n"
-      "packssdw  %%xmm1, %%xmm0                  \n"
-      "packuswb  %%xmm0, %%xmm0                  \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1), %1                     \n"
-      "sub       $0x8, %2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0), %%xmm0                  \n"
+      "movdqu      0x10(%0), %%xmm1              \n"
+      "lea         0x20(%0), %0                  \n"
+      "psrld       $0x18, %%xmm0                 \n"
+      "psrld       $0x18, %%xmm1                 \n"
+      "packssdw    %%xmm1, %%xmm0                \n"
+      "packuswb    %%xmm0, %%xmm0                \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1), %1                   \n"
+      "sub         $0x8, %2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_a),     // %1
         "+rm"(width)     // %2
@@ -3945,28 +6301,28 @@ void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb,
                               uint8_t* dst_a,
                               int width) {
   asm volatile(
-      "vmovdqa    %3,%%ymm4                      \n"
+      "vmovdqa     %3,%%ymm4                     \n"
       "vbroadcastf128 %4,%%ymm5                  \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0), %%ymm0                    \n"
-      "vmovdqu   0x20(%0), %%ymm1                \n"
-      "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"  // vpsrld $0x18, %%ymm0
-      "vpshufb    %%ymm5,%%ymm1,%%ymm1           \n"
-      "vmovdqu   0x40(%0), %%ymm2                \n"
-      "vmovdqu   0x60(%0), %%ymm3                \n"
-      "lea       0x80(%0), %0                    \n"
-      "vpackssdw  %%ymm1, %%ymm0, %%ymm0         \n"  // mutates
-      "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
-      "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
-      "vpackssdw  %%ymm3, %%ymm2, %%ymm2         \n"  // mutates
-      "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
-      "vpermd     %%ymm0,%%ymm4,%%ymm0           \n"  // unmutate.
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub        $0x20, %2                      \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0), %%ymm0                  \n"
+      "vmovdqu     0x20(%0), %%ymm1              \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"  // vpsrld $0x18, %%ymm0
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x40(%0), %%ymm2              \n"
+      "vmovdqu     0x60(%0), %%ymm3              \n"
+      "lea         0x80(%0), %0                  \n"
+      "vpackssdw   %%ymm1, %%ymm0, %%ymm0        \n"  // mutates
+      "vpshufb     %%ymm5,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpackssdw   %%ymm3, %%ymm2, %%ymm2        \n"  // mutates
+      "vpackuswb   %%ymm2,%%ymm0,%%ymm0          \n"  // mutates.
+      "vpermd      %%ymm0,%%ymm4,%%ymm0          \n"  // unmutate.
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20, %2                     \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_argb),               // %0
         "+r"(dst_a),                  // %1
@@ -3981,31 +6337,31 @@ void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb,
 // width in pixels
 void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm0,%%xmm0                   \n"
-      "pslld     $0x18,%%xmm0                    \n"
-      "pcmpeqb   %%xmm1,%%xmm1                   \n"
-      "psrld     $0x8,%%xmm1                     \n"
+      "pcmpeqb     %%xmm0,%%xmm0                 \n"
+      "pslld       $0x18,%%xmm0                  \n"
+      "pcmpeqb     %%xmm1,%%xmm1                 \n"
+      "psrld       $0x8,%%xmm1                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm2                     \n"
-      "lea       0x8(%0),%0                      \n"
-      "punpcklbw %%xmm2,%%xmm2                   \n"
-      "punpckhwd %%xmm2,%%xmm3                   \n"
-      "punpcklwd %%xmm2,%%xmm2                   \n"
-      "movdqu    (%1),%%xmm4                     \n"
-      "movdqu    0x10(%1),%%xmm5                 \n"
-      "pand      %%xmm0,%%xmm2                   \n"
-      "pand      %%xmm0,%%xmm3                   \n"
-      "pand      %%xmm1,%%xmm4                   \n"
-      "pand      %%xmm1,%%xmm5                   \n"
-      "por       %%xmm4,%%xmm2                   \n"
-      "por       %%xmm5,%%xmm3                   \n"
-      "movdqu    %%xmm2,(%1)                     \n"
-      "movdqu    %%xmm3,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm2                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm2,%%xmm2                 \n"
+      "punpckhwd   %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm2,%%xmm2                 \n"
+      "movdqu      (%1),%%xmm4                   \n"
+      "movdqu      0x10(%1),%%xmm5               \n"
+      "pand        %%xmm0,%%xmm2                 \n"
+      "pand        %%xmm0,%%xmm3                 \n"
+      "pand        %%xmm1,%%xmm4                 \n"
+      "pand        %%xmm1,%%xmm5                 \n"
+      "por         %%xmm4,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -4018,23 +6374,23 @@ void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) {
 // width in pixels
 void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+      "vpcmpeqb    %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpsrld      $0x8,%%ymm0,%%ymm0            \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vpmovzxbd (%0),%%ymm1                     \n"
-      "vpmovzxbd 0x8(%0),%%ymm2                  \n"
-      "lea       0x10(%0),%0                     \n"
-      "vpslld    $0x18,%%ymm1,%%ymm1             \n"
-      "vpslld    $0x18,%%ymm2,%%ymm2             \n"
-      "vpblendvb %%ymm0,(%1),%%ymm1,%%ymm1       \n"
-      "vpblendvb %%ymm0,0x20(%1),%%ymm2,%%ymm2   \n"
-      "vmovdqu   %%ymm1,(%1)                     \n"
-      "vmovdqu   %%ymm2,0x20(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "vpmovzxbd   (%0),%%ymm1                   \n"
+      "vpmovzxbd   0x8(%0),%%ymm2                \n"
+      "lea         0x10(%0),%0                   \n"
+      "vpslld      $0x18,%%ymm1,%%ymm1           \n"
+      "vpslld      $0x18,%%ymm2,%%ymm2           \n"
+      "vpblendvb   %%ymm0,(%1),%%ymm1,%%ymm1     \n"
+      "vpblendvb   %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n"
+      "vmovdqu     %%ymm1,(%1)                   \n"
+      "vmovdqu     %%ymm2,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
@@ -4050,7 +6406,7 @@ void SetRow_X86(uint8_t* dst, uint8_t v8, int width) {
   const uint32_t v32 = v8 * 0x01010101u;  // Duplicate byte to all bytes.
   asm volatile(
 
-      "rep stosl                      \n"
+      "rep         stosl                         \n"
       : "+D"(dst),       // %0
         "+c"(width_tmp)  // %1
       : "a"(v32)         // %2
@@ -4061,7 +6417,7 @@ void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width) {
   size_t width_tmp = (size_t)(width);
   asm volatile(
 
-      "rep stosb                      \n"
+      "rep         stosb                         \n"
       : "+D"(dst),       // %0
         "+c"(width_tmp)  // %1
       : "a"(v8)          // %2
@@ -4072,7 +6428,7 @@ void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width) {
   size_t width_tmp = (size_t)(width);
   asm volatile(
 
-      "rep stosl                      \n"
+      "rep         stosl                         \n"
       : "+D"(dst_argb),  // %0
         "+c"(width_tmp)  // %1
       : "a"(v32)         // %2
@@ -4083,21 +6439,21 @@ void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width) {
 #ifdef HAS_YUY2TOYROW_SSE2
 void YUY2ToYRow_SSE2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrlw     $0x8,%%xmm5                     \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -4111,32 +6467,32 @@ void YUY2ToUVRow_SSE2(const uint8_t* src_yuy2,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrlw     $0x8,%%xmm5                     \n"
-      "sub       %1,%2                           \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x00(%0,%4,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%4,1),%%xmm3            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "pavgb     %%xmm3,%%xmm1                   \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq    %%xmm1,0x00(%1,%2,1)              \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_yuy2),               // %0
         "+r"(dst_u),                  // %1
         "+r"(dst_v),                  // %2
@@ -4150,28 +6506,28 @@ void YUY2ToUV422Row_SSE2(const uint8_t* src_yuy2,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrlw     $0x8,%%xmm5                     \n"
-      "sub       %1,%2                           \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq    %%xmm1,0x00(%1,%2,1)              \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -4185,16 +6541,16 @@ void UYVYToYRow_SSE2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -4208,32 +6564,32 @@ void UYVYToUVRow_SSE2(const uint8_t* src_uyvy,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrlw     $0x8,%%xmm5                     \n"
-      "sub       %1,%2                           \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x00(%0,%4,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%4,1),%%xmm3            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "pavgb     %%xmm3,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq    %%xmm1,0x00(%1,%2,1)              \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_uyvy),               // %0
         "+r"(dst_u),                  // %1
         "+r"(dst_v),                  // %2
@@ -4247,28 +6603,28 @@ void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrlw     $0x8,%%xmm5                     \n"
-      "sub       %1,%2                           \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrlw       $0x8,%%xmm5                   \n"
+      "sub         %1,%2                         \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq    %%xmm1,0x00(%1,%2,1)              \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x00(%1,%2,1)          \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -4281,22 +6637,22 @@ void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy,
 #ifdef HAS_YUY2TOYROW_AVX2
 void YUY2ToYRow_AVX2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-      "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-      "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vmovdqu   %%ymm0,(%1)                     \n"
-      "lea      0x20(%1),%1                      \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_y),     // %1
@@ -4311,32 +6667,32 @@ void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-      "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "vpavgb    0x00(%0,%4,1),%%ymm0,%%ymm0     \n"
-      "vpavgb    0x20(%0,%4,1),%%ymm1,%%ymm1     \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
       "vextractf128 $0x0,%%ymm1,(%1)             \n"
       "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
-      "lea      0x10(%1),%1                      \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_yuy2),               // %0
         "+r"(dst_u),                  // %1
@@ -4351,30 +6707,30 @@ void YUY2ToUV422Row_AVX2(const uint8_t* src_yuy2,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-      "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
       "vextractf128 $0x0,%%ymm1,(%1)             \n"
       "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
-      "lea      0x10(%1),%1                      \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_u),     // %1
@@ -4389,17 +6745,17 @@ void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vmovdqu   %%ymm0,(%1)                     \n"
-      "lea      0x20(%1),%1                      \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_y),     // %1
@@ -4413,32 +6769,32 @@ void UYVYToUVRow_AVX2(const uint8_t* src_uyvy,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
-      "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "vpavgb    0x00(%0,%4,1),%%ymm0,%%ymm0     \n"
-      "vpavgb    0x20(%0,%4,1),%%ymm1,%%ymm1     \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-      "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vpavgb      0x00(%0,%4,1),%%ymm0,%%ymm0   \n"
+      "vpavgb      0x20(%0,%4,1),%%ymm1,%%ymm1   \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
       "vextractf128 $0x0,%%ymm1,(%1)             \n"
       "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
-      "lea      0x10(%1),%1                      \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_uyvy),               // %0
         "+r"(dst_u),                  // %1
@@ -4453,30 +6809,30 @@ void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-      "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
-      "sub       %1,%2                           \n"
-
-      LABELALIGN
-      "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
-      "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
-      "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
-      "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
-      "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
-      "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
-      "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
-      "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrlw      $0x8,%%ymm5,%%ymm5            \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm1          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm1,%%ymm1           \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
       "vextractf128 $0x0,%%ymm1,(%1)             \n"
       "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1)    \n"
-      "lea      0x10(%1),%1                      \n"
-      "sub       $0x20,%3                        \n"
-      "jg        1b                              \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x20,%3                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_u),     // %1
@@ -4493,78 +6849,78 @@ static const uvec8 kShuffleAlpha = {3u,  0x80, 3u,  0x80, 7u,  0x80, 7u,  0x80,
                                     11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80};
 
 // Blend 8 pixels at a time
-void ARGBBlendRow_SSSE3(const uint8_t* src_argb0,
+void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
                         const uint8_t* src_argb1,
                         uint8_t* dst_argb,
                         int width) {
   asm volatile(
-      "pcmpeqb   %%xmm7,%%xmm7                   \n"
-      "psrlw     $0xf,%%xmm7                     \n"
-      "pcmpeqb   %%xmm6,%%xmm6                   \n"
-      "psrlw     $0x8,%%xmm6                     \n"
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psllw     $0x8,%%xmm5                     \n"
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "pslld     $0x18,%%xmm4                    \n"
-      "sub       $0x4,%3                         \n"
-      "jl        49f                             \n"
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $0xf,%%xmm7                   \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x8,%%xmm6                   \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psllw       $0x8,%%xmm5                   \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "pslld       $0x18,%%xmm4                  \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
 
       // 4 pixel loop.
       LABELALIGN
       "40:                                       \n"
-      "movdqu    (%0),%%xmm3                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqa    %%xmm3,%%xmm0                   \n"
-      "pxor      %%xmm4,%%xmm3                   \n"
-      "movdqu    (%1),%%xmm2                     \n"
-      "pshufb    %4,%%xmm3                       \n"
-      "pand      %%xmm6,%%xmm2                   \n"
-      "paddw     %%xmm7,%%xmm3                   \n"
-      "pmullw    %%xmm3,%%xmm2                   \n"
-      "movdqu    (%1),%%xmm1                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "por       %%xmm4,%%xmm0                   \n"
-      "pmullw    %%xmm3,%%xmm1                   \n"
-      "psrlw     $0x8,%%xmm2                     \n"
-      "paddusb   %%xmm2,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jge       40b                             \n"
+      "movdqu      (%0),%%xmm3                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm3,%%xmm0                 \n"
+      "pxor        %%xmm4,%%xmm3                 \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "pshufb      %4,%%xmm3                     \n"
+      "pand        %%xmm6,%%xmm2                 \n"
+      "paddw       %%xmm7,%%xmm3                 \n"
+      "pmullw      %%xmm3,%%xmm2                 \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "paddusb     %%xmm2,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
 
       "49:                                       \n"
-      "add       $0x3,%3                         \n"
-      "jl        99f                             \n"
+      "add         $0x3,%3                       \n"
+      "jl          99f                           \n"
 
       // 1 pixel loop.
       "91:                                       \n"
-      "movd      (%0),%%xmm3                     \n"
-      "lea       0x4(%0),%0                      \n"
-      "movdqa    %%xmm3,%%xmm0                   \n"
-      "pxor      %%xmm4,%%xmm3                   \n"
-      "movd      (%1),%%xmm2                     \n"
-      "pshufb    %4,%%xmm3                       \n"
-      "pand      %%xmm6,%%xmm2                   \n"
-      "paddw     %%xmm7,%%xmm3                   \n"
-      "pmullw    %%xmm3,%%xmm2                   \n"
-      "movd      (%1),%%xmm1                     \n"
-      "lea       0x4(%1),%1                      \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "por       %%xmm4,%%xmm0                   \n"
-      "pmullw    %%xmm3,%%xmm1                   \n"
-      "psrlw     $0x8,%%xmm2                     \n"
-      "paddusb   %%xmm2,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movd      %%xmm0,(%2)                     \n"
-      "lea       0x4(%2),%2                      \n"
-      "sub       $0x1,%3                         \n"
-      "jge       91b                             \n"
+      "movd        (%0),%%xmm3                   \n"
+      "lea         0x4(%0),%0                    \n"
+      "movdqa      %%xmm3,%%xmm0                 \n"
+      "pxor        %%xmm4,%%xmm3                 \n"
+      "movd        (%1),%%xmm2                   \n"
+      "pshufb      %4,%%xmm3                     \n"
+      "pand        %%xmm6,%%xmm2                 \n"
+      "paddw       %%xmm7,%%xmm3                 \n"
+      "pmullw      %%xmm3,%%xmm2                 \n"
+      "movd        (%1),%%xmm1                   \n"
+      "lea         0x4(%1),%1                    \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "por         %%xmm4,%%xmm0                 \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "paddusb     %%xmm2,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x4(%2),%2                    \n"
+      "sub         $0x1,%3                       \n"
+      "jge         91b                           \n"
       "99:                                       \n"
-      : "+r"(src_argb0),    // %0
+      : "+r"(src_argb),     // %0
         "+r"(src_argb1),    // %1
         "+r"(dst_argb),     // %2
         "+r"(width)         // %3
@@ -4586,36 +6942,36 @@ void BlendPlaneRow_SSSE3(const uint8_t* src0,
                          uint8_t* dst,
                          int width) {
   asm volatile(
-      "pcmpeqb    %%xmm5,%%xmm5                  \n"
-      "psllw      $0x8,%%xmm5                    \n"
-      "mov        $0x80808080,%%eax              \n"
-      "movd       %%eax,%%xmm6                   \n"
-      "pshufd     $0x0,%%xmm6,%%xmm6             \n"
-      "mov        $0x807f807f,%%eax              \n"
-      "movd       %%eax,%%xmm7                   \n"
-      "pshufd     $0x0,%%xmm7,%%xmm7             \n"
-      "sub        %2,%0                          \n"
-      "sub        %2,%1                          \n"
-      "sub        %2,%3                          \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psllw       $0x8,%%xmm5                   \n"
+      "mov         $0x80808080,%%eax             \n"
+      "movd        %%eax,%%xmm6                  \n"
+      "pshufd      $0x0,%%xmm6,%%xmm6            \n"
+      "mov         $0x807f807f,%%eax             \n"
+      "movd        %%eax,%%xmm7                  \n"
+      "pshufd      $0x0,%%xmm7,%%xmm7            \n"
+      "sub         %2,%0                         \n"
+      "sub         %2,%1                         \n"
+      "sub         %2,%3                         \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movq       (%2),%%xmm0                    \n"
-      "punpcklbw  %%xmm0,%%xmm0                  \n"
-      "pxor       %%xmm5,%%xmm0                  \n"
-      "movq       (%0,%2,1),%%xmm1               \n"
-      "movq       (%1,%2,1),%%xmm2               \n"
-      "punpcklbw  %%xmm2,%%xmm1                  \n"
-      "psubb      %%xmm6,%%xmm1                  \n"
-      "pmaddubsw  %%xmm1,%%xmm0                  \n"
-      "paddw      %%xmm7,%%xmm0                  \n"
-      "psrlw      $0x8,%%xmm0                    \n"
-      "packuswb   %%xmm0,%%xmm0                  \n"
-      "movq       %%xmm0,(%3,%2,1)               \n"
-      "lea        0x8(%2),%2                     \n"
-      "sub        $0x8,%4                        \n"
-      "jg        1b                              \n"
+      "movq        (%2),%%xmm0                   \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm5,%%xmm0                 \n"
+      "movq        (%0,%2,1),%%xmm1              \n"
+      "movq        (%1,%2,1),%%xmm2              \n"
+      "punpcklbw   %%xmm2,%%xmm1                 \n"
+      "psubb       %%xmm6,%%xmm1                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm7,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%3,%2,1)              \n"
+      "lea         0x8(%2),%2                    \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
       : "+r"(src0),   // %0
         "+r"(src1),   // %1
         "+r"(alpha),  // %2
@@ -4638,43 +6994,43 @@ void BlendPlaneRow_AVX2(const uint8_t* src0,
                         uint8_t* dst,
                         int width) {
   asm volatile(
-      "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-      "vpsllw     $0x8,%%ymm5,%%ymm5             \n"
-      "mov        $0x80808080,%%eax              \n"
-      "vmovd      %%eax,%%xmm6                   \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsllw      $0x8,%%ymm5,%%ymm5            \n"
+      "mov         $0x80808080,%%eax             \n"
+      "vmovd       %%eax,%%xmm6                  \n"
       "vbroadcastss %%xmm6,%%ymm6                \n"
-      "mov        $0x807f807f,%%eax              \n"
-      "vmovd      %%eax,%%xmm7                   \n"
+      "mov         $0x807f807f,%%eax             \n"
+      "vmovd       %%eax,%%xmm7                  \n"
       "vbroadcastss %%xmm7,%%ymm7                \n"
-      "sub        %2,%0                          \n"
-      "sub        %2,%1                          \n"
-      "sub        %2,%3                          \n"
+      "sub         %2,%0                         \n"
+      "sub         %2,%1                         \n"
+      "sub         %2,%3                         \n"
 
       // 32 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%2),%%ymm0                    \n"
-      "vpunpckhbw %%ymm0,%%ymm0,%%ymm3           \n"
-      "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
-      "vpxor      %%ymm5,%%ymm3,%%ymm3           \n"
-      "vpxor      %%ymm5,%%ymm0,%%ymm0           \n"
-      "vmovdqu    (%0,%2,1),%%ymm1               \n"
-      "vmovdqu    (%1,%2,1),%%ymm2               \n"
-      "vpunpckhbw %%ymm2,%%ymm1,%%ymm4           \n"
-      "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
-      "vpsubb     %%ymm6,%%ymm4,%%ymm4           \n"
-      "vpsubb     %%ymm6,%%ymm1,%%ymm1           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "vpmaddubsw %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm7,%%ymm3,%%ymm3           \n"
-      "vpaddw     %%ymm7,%%ymm0,%%ymm0           \n"
-      "vpsrlw     $0x8,%%ymm3,%%ymm3             \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpackuswb  %%ymm3,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,(%3,%2,1)               \n"
-      "lea        0x20(%2),%2                    \n"
-      "sub        $0x20,%4                       \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%2),%%ymm0                   \n"
+      "vpunpckhbw  %%ymm0,%%ymm0,%%ymm3          \n"
+      "vpunpcklbw  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpxor       %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpxor       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vmovdqu     (%0,%2,1),%%ymm1              \n"
+      "vmovdqu     (%1,%2,1),%%ymm2              \n"
+      "vpunpckhbw  %%ymm2,%%ymm1,%%ymm4          \n"
+      "vpunpcklbw  %%ymm2,%%ymm1,%%ymm1          \n"
+      "vpsubb      %%ymm6,%%ymm4,%%ymm4          \n"
+      "vpsubb      %%ymm6,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpmaddubsw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm7,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm7,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x8,%%ymm3,%%ymm3            \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm3,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%3,%2,1)              \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src0),   // %0
         "+r"(src1),   // %1
@@ -4688,7 +7044,7 @@ void BlendPlaneRow_AVX2(const uint8_t* src0,
 #endif  // HAS_BLENDPLANEROW_AVX2
 
 #ifdef HAS_ARGBATTENUATEROW_SSSE3
-// Shuffle table duplicating alpha
+// Shuffle table duplicating alpha.
 static const uvec8 kShuffleAlpha0 = {3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u,
                                      7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u};
 static const uvec8 kShuffleAlpha1 = {11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
@@ -4698,35 +7054,35 @@ void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb,
                             uint8_t* dst_argb,
                             int width) {
   asm volatile(
-      "pcmpeqb   %%xmm3,%%xmm3                   \n"
-      "pslld     $0x18,%%xmm3                    \n"
-      "movdqa    %3,%%xmm4                       \n"
-      "movdqa    %4,%%xmm5                       \n"
+      "pcmpeqb     %%xmm3,%%xmm3                 \n"
+      "pslld       $0x18,%%xmm3                  \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "punpcklbw %%xmm1,%%xmm1                   \n"
-      "pmulhuw   %%xmm1,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "pshufb    %%xmm5,%%xmm1                   \n"
-      "movdqu    (%0),%%xmm2                     \n"
-      "punpckhbw %%xmm2,%%xmm2                   \n"
-      "pmulhuw   %%xmm2,%%xmm1                   \n"
-      "movdqu    (%0),%%xmm2                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "pand      %%xmm3,%%xmm2                   \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "por       %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "punpcklbw   %%xmm1,%%xmm1                 \n"
+      "pmulhuw     %%xmm1,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "punpckhbw   %%xmm2,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "pand        %%xmm3,%%xmm2                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "por         %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),       // %0
         "+r"(dst_argb),       // %1
         "+r"(width)           // %2
@@ -4747,29 +7103,29 @@ void ARGBAttenuateRow_AVX2(const uint8_t* src_argb,
                            int width) {
   asm volatile(
       "vbroadcastf128 %3,%%ymm4                  \n"
-      "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-      "vpslld     $0x18,%%ymm5,%%ymm5            \n"
-      "sub        %0,%1                          \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpslld      $0x18,%%ymm5,%%ymm5           \n"
+      "sub         %0,%1                         \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm6                    \n"
-      "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
-      "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
-      "vpshufb    %%ymm4,%%ymm0,%%ymm2           \n"
-      "vpshufb    %%ymm4,%%ymm1,%%ymm3           \n"
-      "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-      "vpand      %%ymm5,%%ymm6,%%ymm6           \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpor       %%ymm6,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,0x00(%0,%1,1)           \n"
-      "lea       0x20(%0),%0                     \n"
-      "sub        $0x8,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm6                   \n"
+      "vpunpcklbw  %%ymm6,%%ymm6,%%ymm0          \n"
+      "vpunpckhbw  %%ymm6,%%ymm6,%%ymm1          \n"
+      "vpshufb     %%ymm4,%%ymm0,%%ymm2          \n"
+      "vpshufb     %%ymm4,%%ymm1,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpand       %%ymm5,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpor        %%ymm6,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%0,%1,1)          \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_argb),          // %0
         "+r"(dst_argb),          // %1
@@ -4789,32 +7145,32 @@ void ARGBUnattenuateRow_SSE2(const uint8_t* src_argb,
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movzb     0x03(%0),%3                     \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "movd      0x00(%4,%3,4),%%xmm2            \n"
-      "movzb     0x07(%0),%3                     \n"
-      "movd      0x00(%4,%3,4),%%xmm3            \n"
-      "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-      "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-      "movlhps   %%xmm3,%%xmm2                   \n"
-      "pmulhuw   %%xmm2,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "movzb     0x0b(%0),%3                     \n"
-      "punpckhbw %%xmm1,%%xmm1                   \n"
-      "movd      0x00(%4,%3,4),%%xmm2            \n"
-      "movzb     0x0f(%0),%3                     \n"
-      "movd      0x00(%4,%3,4),%%xmm3            \n"
-      "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-      "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-      "movlhps   %%xmm3,%%xmm2                   \n"
-      "pmulhuw   %%xmm2,%%xmm1                   \n"
-      "lea       0x10(%0),%0                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movzb       0x03(%0),%3                   \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "movd        0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x07(%0),%3                   \n"
+      "movd        0x00(%4,%3,4),%%xmm3          \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "movlhps     %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "movzb       0x0b(%0),%3                   \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movd        0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x0f(%0),%3                   \n"
+      "movd        0x00(%4,%3,4),%%xmm3          \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "movlhps     %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "lea         0x10(%0),%0                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),     // %0
         "+r"(dst_argb),     // %1
         "+r"(width),        // %2
@@ -4834,52 +7190,52 @@ void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
                              int width) {
   uintptr_t alpha;
   asm volatile(
-      "sub        %0,%1                          \n"
+      "sub         %0,%1                         \n"
       "vbroadcastf128 %5,%%ymm5                  \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
       // replace VPGATHER
-      "movzb     0x03(%0),%3                     \n"
-      "vmovd     0x00(%4,%3,4),%%xmm0            \n"
-      "movzb     0x07(%0),%3                     \n"
-      "vmovd     0x00(%4,%3,4),%%xmm1            \n"
-      "movzb     0x0b(%0),%3                     \n"
-      "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
-      "vmovd     0x00(%4,%3,4),%%xmm2            \n"
-      "movzb     0x0f(%0),%3                     \n"
-      "vmovd     0x00(%4,%3,4),%%xmm3            \n"
-      "movzb     0x13(%0),%3                     \n"
-      "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
-      "vmovd     0x00(%4,%3,4),%%xmm0            \n"
-      "movzb     0x17(%0),%3                     \n"
-      "vmovd     0x00(%4,%3,4),%%xmm1            \n"
-      "movzb     0x1b(%0),%3                     \n"
-      "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
-      "vmovd     0x00(%4,%3,4),%%xmm2            \n"
-      "movzb     0x1f(%0),%3                     \n"
-      "vmovd     0x00(%4,%3,4),%%xmm3            \n"
-      "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
+      "movzb       0x03(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm0          \n"
+      "movzb       0x07(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm1          \n"
+      "movzb       0x0b(%0),%3                   \n"
+      "vpunpckldq  %%xmm1,%%xmm0,%%xmm6          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x0f(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm3          \n"
+      "movzb       0x13(%0),%3                   \n"
+      "vpunpckldq  %%xmm3,%%xmm2,%%xmm7          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm0          \n"
+      "movzb       0x17(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm1          \n"
+      "movzb       0x1b(%0),%3                   \n"
+      "vpunpckldq  %%xmm1,%%xmm0,%%xmm0          \n"
+      "vmovd       0x00(%4,%3,4),%%xmm2          \n"
+      "movzb       0x1f(%0),%3                   \n"
+      "vmovd       0x00(%4,%3,4),%%xmm3          \n"
+      "vpunpckldq  %%xmm3,%%xmm2,%%xmm2          \n"
       "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
       "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
       "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
       // end of VPGATHER
 
-      "vmovdqu    (%0),%%ymm6                    \n"
-      "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
-      "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
-      "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
-      "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
-      "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
-      "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
-      "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,0x00(%0,%1,1)           \n"
-      "lea       0x20(%0),%0                     \n"
-      "sub        $0x8,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm6                   \n"
+      "vpunpcklbw  %%ymm6,%%ymm6,%%ymm0          \n"
+      "vpunpckhbw  %%ymm6,%%ymm6,%%ymm1          \n"
+      "vpunpcklwd  %%ymm3,%%ymm3,%%ymm2          \n"
+      "vpunpckhwd  %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpshufb     %%ymm5,%%ymm2,%%ymm2          \n"
+      "vpshufb     %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%0,%1,1)          \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_argb),                 // %0
         "+r"(dst_argb),                 // %1
@@ -4896,44 +7252,48 @@ void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 void ARGBGrayRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   asm volatile(
-      "movdqa    %3,%%xmm4                       \n"
-      "movdqa    %4,%%xmm5                       \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "pmaddubsw %%xmm4,%%xmm0                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "phaddw    %%xmm1,%%xmm0                   \n"
-      "paddw     %%xmm5,%%xmm0                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm2                     \n"
-      "movdqu    0x10(%0),%%xmm3                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "psrld     $0x18,%%xmm2                    \n"
-      "psrld     $0x18,%%xmm3                    \n"
-      "packuswb  %%xmm3,%%xmm2                   \n"
-      "packuswb  %%xmm2,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm3                   \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "punpcklbw %%xmm2,%%xmm3                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklwd %%xmm3,%%xmm0                   \n"
-      "punpckhwd %%xmm3,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psubb       %%xmm5,%%xmm0                 \n"
+      "psubb       %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "movdqu      %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "phaddw      %%xmm0,%%xmm6                 \n"
+      "paddw       %%xmm5,%%xmm6                 \n"
+      "psrlw       $0x8,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "movdqu      0x10(%0),%%xmm3               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrld       $0x18,%%xmm2                  \n"
+      "psrld       $0x18,%%xmm3                  \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm2                 \n"
+      "movdqa      %%xmm6,%%xmm3                 \n"
+      "punpcklbw   %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm6,%%xmm1                 \n"
+      "punpcklwd   %%xmm3,%%xmm6                 \n"
+      "punpckhwd   %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm6,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
       : "m"(kARGBToYJ),  // %3
-        "m"(kAddYJ64)    // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+        "m"(kSub128)     // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
 }
 #endif  // HAS_ARGBGRAYROW_SSSE3
 
@@ -4954,50 +7314,50 @@ static const vec8 kARGBToSepiaR = {24, 98, 50, 0, 24, 98, 50, 0,
 // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
 void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width) {
   asm volatile(
-      "movdqa    %2,%%xmm2                       \n"
-      "movdqa    %3,%%xmm3                       \n"
-      "movdqa    %4,%%xmm4                       \n"
+      "movdqa      %2,%%xmm2                     \n"
+      "movdqa      %3,%%xmm3                     \n"
+      "movdqa      %4,%%xmm4                     \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm6                 \n"
-      "pmaddubsw %%xmm2,%%xmm0                   \n"
-      "pmaddubsw %%xmm2,%%xmm6                   \n"
-      "phaddw    %%xmm6,%%xmm0                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm5                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "pmaddubsw %%xmm3,%%xmm5                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "phaddw    %%xmm1,%%xmm5                   \n"
-      "psrlw     $0x7,%%xmm5                     \n"
-      "packuswb  %%xmm5,%%xmm5                   \n"
-      "punpcklbw %%xmm5,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm5                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "pmaddubsw %%xmm4,%%xmm5                   \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "phaddw    %%xmm1,%%xmm5                   \n"
-      "psrlw     $0x7,%%xmm5                     \n"
-      "packuswb  %%xmm5,%%xmm5                   \n"
-      "movdqu    (%0),%%xmm6                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "psrld     $0x18,%%xmm6                    \n"
-      "psrld     $0x18,%%xmm1                    \n"
-      "packuswb  %%xmm1,%%xmm6                   \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "punpcklbw %%xmm6,%%xmm5                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklwd %%xmm5,%%xmm0                   \n"
-      "punpckhwd %%xmm5,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%0)                     \n"
-      "movdqu    %%xmm1,0x10(%0)                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "sub       $0x8,%1                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "pmaddubsw   %%xmm2,%%xmm6                 \n"
+      "phaddw      %%xmm6,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm5                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm3,%%xmm5                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm5                 \n"
+      "psrlw       $0x7,%%xmm5                   \n"
+      "packuswb    %%xmm5,%%xmm5                 \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm5                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm4,%%xmm5                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm5                 \n"
+      "psrlw       $0x7,%%xmm5                   \n"
+      "packuswb    %%xmm5,%%xmm5                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "psrld       $0x18,%%xmm6                  \n"
+      "psrld       $0x18,%%xmm1                  \n"
+      "packuswb    %%xmm1,%%xmm6                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm5                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklwd   %%xmm5,%%xmm0                 \n"
+      "punpckhwd   %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%1                       \n"
+      "jg          1b                            \n"
       : "+r"(dst_argb),      // %0
         "+r"(width)          // %1
       : "m"(kARGBToSepiaB),  // %2
@@ -5015,54 +7375,54 @@ void ARGBColorMatrixRow_SSSE3(const uint8_t* src_argb,
                               const int8_t* matrix_argb,
                               int width) {
   asm volatile(
-      "movdqu    (%3),%%xmm5                     \n"
-      "pshufd    $0x00,%%xmm5,%%xmm2             \n"
-      "pshufd    $0x55,%%xmm5,%%xmm3             \n"
-      "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
-      "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+      "movdqu      (%3),%%xmm5                   \n"
+      "pshufd      $0x00,%%xmm5,%%xmm2           \n"
+      "pshufd      $0x55,%%xmm5,%%xmm3           \n"
+      "pshufd      $0xaa,%%xmm5,%%xmm4           \n"
+      "pshufd      $0xff,%%xmm5,%%xmm5           \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm7                 \n"
-      "pmaddubsw %%xmm2,%%xmm0                   \n"
-      "pmaddubsw %%xmm2,%%xmm7                   \n"
-      "movdqu    (%0),%%xmm6                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "pmaddubsw %%xmm3,%%xmm6                   \n"
-      "pmaddubsw %%xmm3,%%xmm1                   \n"
-      "phaddsw   %%xmm7,%%xmm0                   \n"
-      "phaddsw   %%xmm1,%%xmm6                   \n"
-      "psraw     $0x6,%%xmm0                     \n"
-      "psraw     $0x6,%%xmm6                     \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "punpcklbw %%xmm6,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "movdqu    0x10(%0),%%xmm7                 \n"
-      "pmaddubsw %%xmm4,%%xmm1                   \n"
-      "pmaddubsw %%xmm4,%%xmm7                   \n"
-      "phaddsw   %%xmm7,%%xmm1                   \n"
-      "movdqu    (%0),%%xmm6                     \n"
-      "movdqu    0x10(%0),%%xmm7                 \n"
-      "pmaddubsw %%xmm5,%%xmm6                   \n"
-      "pmaddubsw %%xmm5,%%xmm7                   \n"
-      "phaddsw   %%xmm7,%%xmm6                   \n"
-      "psraw     $0x6,%%xmm1                     \n"
-      "psraw     $0x6,%%xmm6                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "punpcklbw %%xmm6,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm6                   \n"
-      "punpcklwd %%xmm1,%%xmm0                   \n"
-      "punpckhwd %%xmm1,%%xmm6                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm6,0x10(%1)                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "pmaddubsw   %%xmm2,%%xmm7                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "pmaddubsw   %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm3,%%xmm1                 \n"
+      "phaddsw     %%xmm7,%%xmm0                 \n"
+      "phaddsw     %%xmm1,%%xmm6                 \n"
+      "psraw       $0x6,%%xmm0                   \n"
+      "psraw       $0x6,%%xmm6                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm7                 \n"
+      "phaddsw     %%xmm7,%%xmm1                 \n"
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x10(%0),%%xmm7               \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm7                 \n"
+      "phaddsw     %%xmm7,%%xmm6                 \n"
+      "psraw       $0x6,%%xmm1                   \n"
+      "psraw       $0x6,%%xmm6                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "punpcklbw   %%xmm6,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "punpcklwd   %%xmm1,%%xmm0                 \n"
+      "punpckhwd   %%xmm1,%%xmm6                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm6,0x10(%1)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),   // %0
         "+r"(dst_argb),   // %1
         "+r"(width)       // %2
@@ -5080,40 +7440,40 @@ void ARGBQuantizeRow_SSE2(uint8_t* dst_argb,
                           int interval_offset,
                           int width) {
   asm volatile(
-      "movd      %2,%%xmm2                       \n"
-      "movd      %3,%%xmm3                       \n"
-      "movd      %4,%%xmm4                       \n"
-      "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
-      "pshufd    $0x44,%%xmm2,%%xmm2             \n"
-      "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
-      "pshufd    $0x44,%%xmm3,%%xmm3             \n"
-      "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
-      "pshufd    $0x44,%%xmm4,%%xmm4             \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
-      "pcmpeqb   %%xmm6,%%xmm6                   \n"
-      "pslld     $0x18,%%xmm6                    \n"
+      "movd        %2,%%xmm2                     \n"
+      "movd        %3,%%xmm3                     \n"
+      "movd        %4,%%xmm4                     \n"
+      "pshuflw     $0x40,%%xmm2,%%xmm2           \n"
+      "pshufd      $0x44,%%xmm2,%%xmm2           \n"
+      "pshuflw     $0x40,%%xmm3,%%xmm3           \n"
+      "pshufd      $0x44,%%xmm3,%%xmm3           \n"
+      "pshuflw     $0x40,%%xmm4,%%xmm4           \n"
+      "pshufd      $0x44,%%xmm4,%%xmm4           \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "pslld       $0x18,%%xmm6                  \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "punpcklbw %%xmm5,%%xmm0                   \n"
-      "pmulhuw   %%xmm2,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm1                     \n"
-      "punpckhbw %%xmm5,%%xmm1                   \n"
-      "pmulhuw   %%xmm2,%%xmm1                   \n"
-      "pmullw    %%xmm3,%%xmm0                   \n"
-      "movdqu    (%0),%%xmm7                     \n"
-      "pmullw    %%xmm3,%%xmm1                   \n"
-      "pand      %%xmm6,%%xmm7                   \n"
-      "paddw     %%xmm4,%%xmm0                   \n"
-      "paddw     %%xmm4,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "por       %%xmm7,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%0)                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "sub       $0x4,%1                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm1                   \n"
+      "punpckhbw   %%xmm5,%%xmm1                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "pmullw      %%xmm3,%%xmm0                 \n"
+      "movdqu      (%0),%%xmm7                   \n"
+      "pmullw      %%xmm3,%%xmm1                 \n"
+      "pand        %%xmm6,%%xmm7                 \n"
+      "paddw       %%xmm4,%%xmm0                 \n"
+      "paddw       %%xmm4,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "por         %%xmm7,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "sub         $0x4,%1                       \n"
+      "jg          1b                            \n"
       : "+r"(dst_argb),       // %0
         "+r"(width)           // %1
       : "r"(scale),           // %2
@@ -5131,27 +7491,27 @@ void ARGBShadeRow_SSE2(const uint8_t* src_argb,
                        int width,
                        uint32_t value) {
   asm volatile(
-      "movd      %3,%%xmm2                       \n"
-      "punpcklbw %%xmm2,%%xmm2                   \n"
-      "punpcklqdq %%xmm2,%%xmm2                  \n"
+      "movd        %3,%%xmm2                     \n"
+      "punpcklbw   %%xmm2,%%xmm2                 \n"
+      "punpcklqdq  %%xmm2,%%xmm2                 \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "punpckhbw %%xmm1,%%xmm1                   \n"
-      "pmulhuw   %%xmm2,%%xmm0                   \n"
-      "pmulhuw   %%xmm2,%%xmm1                   \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm2,%%xmm1                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -5162,35 +7522,35 @@ void ARGBShadeRow_SSE2(const uint8_t* src_argb,
 
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
 // Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
+void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
 
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqu    (%1),%%xmm2                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "movdqu    %%xmm0,%%xmm1                   \n"
-      "movdqu    %%xmm2,%%xmm3                   \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "punpckhbw %%xmm1,%%xmm1                   \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "punpckhbw %%xmm5,%%xmm3                   \n"
-      "pmulhuw   %%xmm2,%%xmm0                   \n"
-      "pmulhuw   %%xmm3,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb0),  // %0
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm2                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqu      %%xmm0,%%xmm1                 \n"
+      "movdqu      %%xmm2,%%xmm3                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpckhbw   %%xmm5,%%xmm3                 \n"
+      "pmulhuw     %%xmm2,%%xmm0                 \n"
+      "pmulhuw     %%xmm3,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -5201,50 +7561,45 @@ void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBMULTIPLYROW_AVX2
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_AVX2(const uint8_t* src_argb0,
+void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
 
-      "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm1                    \n"
-      "lea        0x20(%0),%0                    \n"
-      "vmovdqu    (%1),%%ymm3                    \n"
-      "lea        0x20(%1),%1                    \n"
-      "vpunpcklbw %%ymm1,%%ymm1,%%ymm0           \n"
-      "vpunpckhbw %%ymm1,%%ymm1,%%ymm1           \n"
-      "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
-      "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
-      "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,(%2)                    \n"
-      "lea       0x20(%2),%2                     \n"
-      "sub        $0x8,%3                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm1                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vmovdqu     (%1),%%ymm3                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "vpunpcklbw  %%ymm1,%%ymm1,%%ymm0          \n"
+      "vpunpckhbw  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpcklbw  %%ymm5,%%ymm3,%%ymm2          \n"
+      "vpunpckhbw  %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpmulhuw    %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpmulhuw    %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_argb0),  // %0
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
       :
-      : "memory", "cc"
-#if defined(__AVX2__)
-        ,
-        "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
-#endif
-  );
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_ARGBMULTIPLYROW_AVX2
 
 #ifdef HAS_ARGBADDROW_SSE2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBAddRow_SSE2(const uint8_t* src_argb0,
+void ARGBAddRow_SSE2(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width) {
@@ -5252,16 +7607,16 @@ void ARGBAddRow_SSE2(const uint8_t* src_argb0,
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqu    (%1),%%xmm1                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb0),  // %0
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -5272,7 +7627,7 @@ void ARGBAddRow_SSE2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBADDROW_AVX2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
-void ARGBAddRow_AVX2(const uint8_t* src_argb0,
+void ARGBAddRow_AVX2(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width) {
@@ -5280,16 +7635,16 @@ void ARGBAddRow_AVX2(const uint8_t* src_argb0,
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "lea        0x20(%0),%0                    \n"
-      "vpaddusb   (%1),%%ymm0,%%ymm0             \n"
-      "lea        0x20(%1),%1                    \n"
-      "vmovdqu    %%ymm0,(%2)                    \n"
-      "lea        0x20(%2),%2                    \n"
-      "sub        $0x8,%3                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpaddusb    (%1),%%ymm0,%%ymm0            \n"
+      "lea         0x20(%1),%1                   \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_argb0),  // %0
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -5300,7 +7655,7 @@ void ARGBAddRow_AVX2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBSUBTRACTROW_SSE2
 // Subtract 2 rows of ARGB pixels, 4 pixels at a time.
-void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
+void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
@@ -5308,16 +7663,16 @@ void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqu    (%1),%%xmm1                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "psubusb   %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb0),  // %0
+      "movdqu      (%0),%%xmm0                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqu      (%1),%%xmm1                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "psubusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -5328,7 +7683,7 @@ void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBSUBTRACTROW_AVX2
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_AVX2(const uint8_t* src_argb0,
+void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
@@ -5336,16 +7691,16 @@ void ARGBSubtractRow_AVX2(const uint8_t* src_argb0,
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "lea        0x20(%0),%0                    \n"
-      "vpsubusb   (%1),%%ymm0,%%ymm0             \n"
-      "lea        0x20(%1),%1                    \n"
-      "vmovdqu    %%ymm0,(%2)                    \n"
-      "lea        0x20(%2),%2                    \n"
-      "sub        $0x8,%3                        \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpsubusb    (%1),%%ymm0,%%ymm0            \n"
+      "lea         0x20(%1),%1                   \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_argb0),  // %0
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -5365,40 +7720,40 @@ void SobelXRow_SSE2(const uint8_t* src_y0,
                     uint8_t* dst_sobelx,
                     int width) {
   asm volatile(
-      "sub       %0,%1                           \n"
-      "sub       %0,%2                           \n"
-      "sub       %0,%3                           \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "sub         %0,%3                         \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm0                     \n"
-      "movq      0x2(%0),%%xmm1                  \n"
-      "punpcklbw %%xmm5,%%xmm0                   \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "psubw     %%xmm1,%%xmm0                   \n"
-      "movq      0x00(%0,%1,1),%%xmm1            \n"
-      "movq      0x02(%0,%1,1),%%xmm2            \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "psubw     %%xmm2,%%xmm1                   \n"
-      "movq      0x00(%0,%2,1),%%xmm2            \n"
-      "movq      0x02(%0,%2,1),%%xmm3            \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "punpcklbw %%xmm5,%%xmm3                   \n"
-      "psubw     %%xmm3,%%xmm2                   \n"
-      "paddw     %%xmm2,%%xmm0                   \n"
-      "paddw     %%xmm1,%%xmm0                   \n"
-      "paddw     %%xmm1,%%xmm0                   \n"
-      "pxor      %%xmm1,%%xmm1                   \n"
-      "psubw     %%xmm0,%%xmm1                   \n"
-      "pmaxsw    %%xmm1,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movq      %%xmm0,0x00(%0,%3,1)            \n"
-      "lea       0x8(%0),%0                      \n"
-      "sub       $0x8,%4                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        0x2(%0),%%xmm1                \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "psubw       %%xmm1,%%xmm0                 \n"
+      "movq        0x00(%0,%1,1),%%xmm1          \n"
+      "movq        0x02(%0,%1,1),%%xmm2          \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "psubw       %%xmm2,%%xmm1                 \n"
+      "movq        0x00(%0,%2,1),%%xmm2          \n"
+      "movq        0x02(%0,%2,1),%%xmm3          \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "psubw       %%xmm3,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "psubw       %%xmm0,%%xmm1                 \n"
+      "pmaxsw      %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,0x00(%0,%3,1)          \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x8,%4                       \n"
+      "jg          1b                            \n"
       : "+r"(src_y0),      // %0
         "+r"(src_y1),      // %1
         "+r"(src_y2),      // %2
@@ -5419,39 +7774,39 @@ void SobelYRow_SSE2(const uint8_t* src_y0,
                     uint8_t* dst_sobely,
                     int width) {
   asm volatile(
-      "sub       %0,%1                           \n"
-      "sub       %0,%2                           \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "sub         %0,%1                         \n"
+      "sub         %0,%2                         \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm0                     \n"
-      "movq      0x00(%0,%1,1),%%xmm1            \n"
-      "punpcklbw %%xmm5,%%xmm0                   \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "psubw     %%xmm1,%%xmm0                   \n"
-      "movq      0x1(%0),%%xmm1                  \n"
-      "movq      0x01(%0,%1,1),%%xmm2            \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "psubw     %%xmm2,%%xmm1                   \n"
-      "movq      0x2(%0),%%xmm2                  \n"
-      "movq      0x02(%0,%1,1),%%xmm3            \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "punpcklbw %%xmm5,%%xmm3                   \n"
-      "psubw     %%xmm3,%%xmm2                   \n"
-      "paddw     %%xmm2,%%xmm0                   \n"
-      "paddw     %%xmm1,%%xmm0                   \n"
-      "paddw     %%xmm1,%%xmm0                   \n"
-      "pxor      %%xmm1,%%xmm1                   \n"
-      "psubw     %%xmm0,%%xmm1                   \n"
-      "pmaxsw    %%xmm1,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movq      %%xmm0,0x00(%0,%2,1)            \n"
-      "lea       0x8(%0),%0                      \n"
-      "sub       $0x8,%3                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movq        0x00(%0,%1,1),%%xmm1          \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "psubw       %%xmm1,%%xmm0                 \n"
+      "movq        0x1(%0),%%xmm1                \n"
+      "movq        0x01(%0,%1,1),%%xmm2          \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "psubw       %%xmm2,%%xmm1                 \n"
+      "movq        0x2(%0),%%xmm2                \n"
+      "movq        0x02(%0,%1,1),%%xmm3          \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "psubw       %%xmm3,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm1,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "psubw       %%xmm0,%%xmm1                 \n"
+      "pmaxsw      %%xmm1,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,0x00(%0,%2,1)          \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x8,%3                       \n"
+      "jg          1b                            \n"
       : "+r"(src_y0),      // %0
         "+r"(src_y1),      // %1
         "+r"(dst_sobely),  // %2
@@ -5472,37 +7827,37 @@ void SobelRow_SSE2(const uint8_t* src_sobelx,
                    uint8_t* dst_argb,
                    int width) {
   asm volatile(
-      "sub       %0,%1                           \n"
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "pslld     $0x18,%%xmm5                    \n"
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%1,1),%%xmm1            \n"
-      "lea       0x10(%0),%0                     \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "punpcklbw %%xmm0,%%xmm2                   \n"
-      "punpckhbw %%xmm0,%%xmm0                   \n"
-      "movdqa    %%xmm2,%%xmm1                   \n"
-      "punpcklwd %%xmm2,%%xmm1                   \n"
-      "punpckhwd %%xmm2,%%xmm2                   \n"
-      "por       %%xmm5,%%xmm1                   \n"
-      "por       %%xmm5,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm3                   \n"
-      "punpcklwd %%xmm0,%%xmm3                   \n"
-      "punpckhwd %%xmm0,%%xmm0                   \n"
-      "por       %%xmm5,%%xmm3                   \n"
-      "por       %%xmm5,%%xmm0                   \n"
-      "movdqu    %%xmm1,(%2)                     \n"
-      "movdqu    %%xmm2,0x10(%2)                 \n"
-      "movdqu    %%xmm3,0x20(%2)                 \n"
-      "movdqu    %%xmm0,0x30(%2)                 \n"
-      "lea       0x40(%2),%2                     \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpcklbw   %%xmm0,%%xmm2                 \n"
+      "punpckhbw   %%xmm0,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "punpcklwd   %%xmm2,%%xmm1                 \n"
+      "punpckhwd   %%xmm2,%%xmm2                 \n"
+      "por         %%xmm5,%%xmm1                 \n"
+      "por         %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "punpcklwd   %%xmm0,%%xmm3                 \n"
+      "punpckhwd   %%xmm0,%%xmm0                 \n"
+      "por         %%xmm5,%%xmm3                 \n"
+      "por         %%xmm5,%%xmm0                 \n"
+      "movdqu      %%xmm1,(%2)                   \n"
+      "movdqu      %%xmm2,0x10(%2)               \n"
+      "movdqu      %%xmm3,0x20(%2)               \n"
+      "movdqu      %%xmm0,0x30(%2)               \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -5519,21 +7874,21 @@ void SobelToPlaneRow_SSE2(const uint8_t* src_sobelx,
                           uint8_t* dst_y,
                           int width) {
   asm volatile(
-      "sub       %0,%1                           \n"
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "pslld     $0x18,%%xmm5                    \n"
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "pslld       $0x18,%%xmm5                  \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%1,1),%%xmm1            \n"
-      "lea       0x10(%0),%0                     \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_y),       // %2
@@ -5554,36 +7909,36 @@ void SobelXYRow_SSE2(const uint8_t* src_sobelx,
                      uint8_t* dst_argb,
                      int width) {
   asm volatile(
-      "sub       %0,%1                           \n"
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
+      "sub         %0,%1                         \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
 
       // 8 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%1,1),%%xmm1            \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "paddusb   %%xmm1,%%xmm2                   \n"
-      "movdqa    %%xmm0,%%xmm3                   \n"
-      "punpcklbw %%xmm5,%%xmm3                   \n"
-      "punpckhbw %%xmm5,%%xmm0                   \n"
-      "movdqa    %%xmm1,%%xmm4                   \n"
-      "punpcklbw %%xmm2,%%xmm4                   \n"
-      "punpckhbw %%xmm2,%%xmm1                   \n"
-      "movdqa    %%xmm4,%%xmm6                   \n"
-      "punpcklwd %%xmm3,%%xmm6                   \n"
-      "punpckhwd %%xmm3,%%xmm4                   \n"
-      "movdqa    %%xmm1,%%xmm7                   \n"
-      "punpcklwd %%xmm0,%%xmm7                   \n"
-      "punpckhwd %%xmm0,%%xmm1                   \n"
-      "movdqu    %%xmm6,(%2)                     \n"
-      "movdqu    %%xmm4,0x10(%2)                 \n"
-      "movdqu    %%xmm7,0x20(%2)                 \n"
-      "movdqu    %%xmm1,0x30(%2)                 \n"
-      "lea       0x40(%2),%2                     \n"
-      "sub       $0x10,%3                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%1,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "paddusb     %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm0,%%xmm3                 \n"
+      "punpcklbw   %%xmm5,%%xmm3                 \n"
+      "punpckhbw   %%xmm5,%%xmm0                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm2,%%xmm4                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm4,%%xmm6                 \n"
+      "punpcklwd   %%xmm3,%%xmm6                 \n"
+      "punpckhwd   %%xmm3,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "punpcklwd   %%xmm0,%%xmm7                 \n"
+      "punpckhwd   %%xmm0,%%xmm1                 \n"
+      "movdqu      %%xmm6,(%2)                   \n"
+      "movdqu      %%xmm4,0x10(%2)               \n"
+      "movdqu      %%xmm7,0x20(%2)               \n"
+      "movdqu      %%xmm1,0x30(%2)               \n"
+      "lea         0x40(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"
+      "jg          1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -5602,67 +7957,67 @@ void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
                                   const int32_t* previous_cumsum,
                                   int width) {
   asm volatile(
-      "pxor      %%xmm0,%%xmm0                   \n"
-      "pxor      %%xmm1,%%xmm1                   \n"
-      "sub       $0x4,%3                         \n"
-      "jl        49f                             \n"
-      "test      $0xf,%1                         \n"
-      "jne       49f                             \n"
+      "pxor        %%xmm0,%%xmm0                 \n"
+      "pxor        %%xmm1,%%xmm1                 \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
+      "test        $0xf,%1                       \n"
+      "jne         49f                           \n"
 
       // 4 pixel loop.
       LABELALIGN
       "40:                                       \n"
-      "movdqu    (%0),%%xmm2                     \n"
-      "lea       0x10(%0),%0                     \n"
-      "movdqa    %%xmm2,%%xmm4                   \n"
-      "punpcklbw %%xmm1,%%xmm2                   \n"
-      "movdqa    %%xmm2,%%xmm3                   \n"
-      "punpcklwd %%xmm1,%%xmm2                   \n"
-      "punpckhwd %%xmm1,%%xmm3                   \n"
-      "punpckhbw %%xmm1,%%xmm4                   \n"
-      "movdqa    %%xmm4,%%xmm5                   \n"
-      "punpcklwd %%xmm1,%%xmm4                   \n"
-      "punpckhwd %%xmm1,%%xmm5                   \n"
-      "paddd     %%xmm2,%%xmm0                   \n"
-      "movdqu    (%2),%%xmm2                     \n"
-      "paddd     %%xmm0,%%xmm2                   \n"
-      "paddd     %%xmm3,%%xmm0                   \n"
-      "movdqu    0x10(%2),%%xmm3                 \n"
-      "paddd     %%xmm0,%%xmm3                   \n"
-      "paddd     %%xmm4,%%xmm0                   \n"
-      "movdqu    0x20(%2),%%xmm4                 \n"
-      "paddd     %%xmm0,%%xmm4                   \n"
-      "paddd     %%xmm5,%%xmm0                   \n"
-      "movdqu    0x30(%2),%%xmm5                 \n"
-      "lea       0x40(%2),%2                     \n"
-      "paddd     %%xmm0,%%xmm5                   \n"
-      "movdqu    %%xmm2,(%1)                     \n"
-      "movdqu    %%xmm3,0x10(%1)                 \n"
-      "movdqu    %%xmm4,0x20(%1)                 \n"
-      "movdqu    %%xmm5,0x30(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x4,%3                         \n"
-      "jge       40b                             \n"
+      "movdqu      (%0),%%xmm2                   \n"
+      "lea         0x10(%0),%0                   \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "punpcklwd   %%xmm1,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm3                 \n"
+      "punpckhbw   %%xmm1,%%xmm4                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "punpcklwd   %%xmm1,%%xmm4                 \n"
+      "punpckhwd   %%xmm1,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "movdqu      0x10(%2),%%xmm3               \n"
+      "paddd       %%xmm0,%%xmm3                 \n"
+      "paddd       %%xmm4,%%xmm0                 \n"
+      "movdqu      0x20(%2),%%xmm4               \n"
+      "paddd       %%xmm0,%%xmm4                 \n"
+      "paddd       %%xmm5,%%xmm0                 \n"
+      "movdqu      0x30(%2),%%xmm5               \n"
+      "lea         0x40(%2),%2                   \n"
+      "paddd       %%xmm0,%%xmm5                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "movdqu      %%xmm3,0x10(%1)               \n"
+      "movdqu      %%xmm4,0x20(%1)               \n"
+      "movdqu      %%xmm5,0x30(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
 
       "49:                                       \n"
-      "add       $0x3,%3                         \n"
-      "jl        19f                             \n"
+      "add         $0x3,%3                       \n"
+      "jl          19f                           \n"
 
       // 1 pixel loop.
       LABELALIGN
       "10:                                       \n"
-      "movd      (%0),%%xmm2                     \n"
-      "lea       0x4(%0),%0                      \n"
-      "punpcklbw %%xmm1,%%xmm2                   \n"
-      "punpcklwd %%xmm1,%%xmm2                   \n"
-      "paddd     %%xmm2,%%xmm0                   \n"
-      "movdqu    (%2),%%xmm2                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "paddd     %%xmm0,%%xmm2                   \n"
-      "movdqu    %%xmm2,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x1,%3                         \n"
-      "jge       10b                             \n"
+      "movd        (%0),%%xmm2                   \n"
+      "lea         0x4(%0),%0                    \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "punpcklwd   %%xmm1,%%xmm2                 \n"
+      "paddd       %%xmm2,%%xmm0                 \n"
+      "movdqu      (%2),%%xmm2                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm2,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x1,%3                       \n"
+      "jge         10b                           \n"
 
       "19:                                       \n"
       : "+r"(row),              // %0
@@ -5682,119 +8037,119 @@ void CumulativeSumToAverageRow_SSE2(const int32_t* topleft,
                                     uint8_t* dst,
                                     int count) {
   asm volatile(
-      "movd      %5,%%xmm5                       \n"
-      "cvtdq2ps  %%xmm5,%%xmm5                   \n"
-      "rcpss     %%xmm5,%%xmm4                   \n"
-      "pshufd    $0x0,%%xmm4,%%xmm4              \n"
-      "sub       $0x4,%3                         \n"
-      "jl        49f                             \n"
-      "cmpl      $0x80,%5                        \n"
-      "ja        40f                             \n"
-
-      "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-      "pcmpeqb   %%xmm6,%%xmm6                   \n"
-      "psrld     $0x10,%%xmm6                    \n"
-      "cvtdq2ps  %%xmm6,%%xmm6                   \n"
-      "addps     %%xmm6,%%xmm5                   \n"
-      "mulps     %%xmm4,%%xmm5                   \n"
-      "cvtps2dq  %%xmm5,%%xmm5                   \n"
-      "packssdw  %%xmm5,%%xmm5                   \n"
+      "movd        %5,%%xmm5                     \n"
+      "cvtdq2ps    %%xmm5,%%xmm5                 \n"
+      "rcpss       %%xmm5,%%xmm4                 \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
+      "sub         $0x4,%3                       \n"
+      "jl          49f                           \n"
+      "cmpl        $0x80,%5                      \n"
+      "ja          40f                           \n"
+
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrld       $0x10,%%xmm6                  \n"
+      "cvtdq2ps    %%xmm6,%%xmm6                 \n"
+      "addps       %%xmm6,%%xmm5                 \n"
+      "mulps       %%xmm4,%%xmm5                 \n"
+      "cvtps2dq    %%xmm5,%%xmm5                 \n"
+      "packssdw    %%xmm5,%%xmm5                 \n"
 
       // 4 pixel small loop.
       LABELALIGN
       "4:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "psubd     0x00(%0,%4,4),%%xmm0            \n"
-      "psubd     0x10(%0,%4,4),%%xmm1            \n"
-      "psubd     0x20(%0,%4,4),%%xmm2            \n"
-      "psubd     0x30(%0,%4,4),%%xmm3            \n"
-      "lea       0x40(%0),%0                     \n"
-      "psubd     (%1),%%xmm0                     \n"
-      "psubd     0x10(%1),%%xmm1                 \n"
-      "psubd     0x20(%1),%%xmm2                 \n"
-      "psubd     0x30(%1),%%xmm3                 \n"
-      "paddd     0x00(%1,%4,4),%%xmm0            \n"
-      "paddd     0x10(%1,%4,4),%%xmm1            \n"
-      "paddd     0x20(%1,%4,4),%%xmm2            \n"
-      "paddd     0x30(%1,%4,4),%%xmm3            \n"
-      "lea       0x40(%1),%1                     \n"
-      "packssdw  %%xmm1,%%xmm0                   \n"
-      "packssdw  %%xmm3,%%xmm2                   \n"
-      "pmulhuw   %%xmm5,%%xmm0                   \n"
-      "pmulhuw   %%xmm5,%%xmm2                   \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jge       4b                              \n"
-      "jmp       49f                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "psubd       0x10(%0,%4,4),%%xmm1          \n"
+      "psubd       0x20(%0,%4,4),%%xmm2          \n"
+      "psubd       0x30(%0,%4,4),%%xmm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "psubd       0x10(%1),%%xmm1               \n"
+      "psubd       0x20(%1),%%xmm2               \n"
+      "psubd       0x30(%1),%%xmm3               \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "paddd       0x10(%1,%4,4),%%xmm1          \n"
+      "paddd       0x20(%1,%4,4),%%xmm2          \n"
+      "paddd       0x30(%1,%4,4),%%xmm3          \n"
+      "lea         0x40(%1),%1                   \n"
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm3,%%xmm2                 \n"
+      "pmulhuw     %%xmm5,%%xmm0                 \n"
+      "pmulhuw     %%xmm5,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         4b                            \n"
+      "jmp         49f                           \n"
 
       // 4 pixel loop
       LABELALIGN
       "40:                                       \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x20(%0),%%xmm2                 \n"
-      "movdqu    0x30(%0),%%xmm3                 \n"
-      "psubd     0x00(%0,%4,4),%%xmm0            \n"
-      "psubd     0x10(%0,%4,4),%%xmm1            \n"
-      "psubd     0x20(%0,%4,4),%%xmm2            \n"
-      "psubd     0x30(%0,%4,4),%%xmm3            \n"
-      "lea       0x40(%0),%0                     \n"
-      "psubd     (%1),%%xmm0                     \n"
-      "psubd     0x10(%1),%%xmm1                 \n"
-      "psubd     0x20(%1),%%xmm2                 \n"
-      "psubd     0x30(%1),%%xmm3                 \n"
-      "paddd     0x00(%1,%4,4),%%xmm0            \n"
-      "paddd     0x10(%1,%4,4),%%xmm1            \n"
-      "paddd     0x20(%1,%4,4),%%xmm2            \n"
-      "paddd     0x30(%1,%4,4),%%xmm3            \n"
-      "lea       0x40(%1),%1                     \n"
-      "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-      "cvtdq2ps  %%xmm1,%%xmm1                   \n"
-      "mulps     %%xmm4,%%xmm0                   \n"
-      "mulps     %%xmm4,%%xmm1                   \n"
-      "cvtdq2ps  %%xmm2,%%xmm2                   \n"
-      "cvtdq2ps  %%xmm3,%%xmm3                   \n"
-      "mulps     %%xmm4,%%xmm2                   \n"
-      "mulps     %%xmm4,%%xmm3                   \n"
-      "cvtps2dq  %%xmm0,%%xmm0                   \n"
-      "cvtps2dq  %%xmm1,%%xmm1                   \n"
-      "cvtps2dq  %%xmm2,%%xmm2                   \n"
-      "cvtps2dq  %%xmm3,%%xmm3                   \n"
-      "packssdw  %%xmm1,%%xmm0                   \n"
-      "packssdw  %%xmm3,%%xmm2                   \n"
-      "packuswb  %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jge       40b                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x20(%0),%%xmm2               \n"
+      "movdqu      0x30(%0),%%xmm3               \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "psubd       0x10(%0,%4,4),%%xmm1          \n"
+      "psubd       0x20(%0,%4,4),%%xmm2          \n"
+      "psubd       0x30(%0,%4,4),%%xmm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "psubd       0x10(%1),%%xmm1               \n"
+      "psubd       0x20(%1),%%xmm2               \n"
+      "psubd       0x30(%1),%%xmm3               \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "paddd       0x10(%1,%4,4),%%xmm1          \n"
+      "paddd       0x20(%1,%4,4),%%xmm2          \n"
+      "paddd       0x30(%1,%4,4),%%xmm3          \n"
+      "lea         0x40(%1),%1                   \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "cvtdq2ps    %%xmm1,%%xmm1                 \n"
+      "mulps       %%xmm4,%%xmm0                 \n"
+      "mulps       %%xmm4,%%xmm1                 \n"
+      "cvtdq2ps    %%xmm2,%%xmm2                 \n"
+      "cvtdq2ps    %%xmm3,%%xmm3                 \n"
+      "mulps       %%xmm4,%%xmm2                 \n"
+      "mulps       %%xmm4,%%xmm3                 \n"
+      "cvtps2dq    %%xmm0,%%xmm0                 \n"
+      "cvtps2dq    %%xmm1,%%xmm1                 \n"
+      "cvtps2dq    %%xmm2,%%xmm2                 \n"
+      "cvtps2dq    %%xmm3,%%xmm3                 \n"
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "packssdw    %%xmm3,%%xmm2                 \n"
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jge         40b                           \n"
 
       "49:                                       \n"
-      "add       $0x3,%3                         \n"
-      "jl        19f                             \n"
+      "add         $0x3,%3                       \n"
+      "jl          19f                           \n"
 
       // 1 pixel loop
       LABELALIGN
       "10:                                       \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "psubd     0x00(%0,%4,4),%%xmm0            \n"
-      "lea       0x10(%0),%0                     \n"
-      "psubd     (%1),%%xmm0                     \n"
-      "paddd     0x00(%1,%4,4),%%xmm0            \n"
-      "lea       0x10(%1),%1                     \n"
-      "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-      "mulps     %%xmm4,%%xmm0                   \n"
-      "cvtps2dq  %%xmm0,%%xmm0                   \n"
-      "packssdw  %%xmm0,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movd      %%xmm0,(%2)                     \n"
-      "lea       0x4(%2),%2                      \n"
-      "sub       $0x1,%3                         \n"
-      "jge       10b                             \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "psubd       0x00(%0,%4,4),%%xmm0          \n"
+      "lea         0x10(%0),%0                   \n"
+      "psubd       (%1),%%xmm0                   \n"
+      "paddd       0x00(%1,%4,4),%%xmm0          \n"
+      "lea         0x10(%1),%1                   \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "mulps       %%xmm4,%%xmm0                 \n"
+      "cvtps2dq    %%xmm0,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x4(%2),%2                    \n"
+      "sub         $0x1,%3                       \n"
+      "jge         10b                           \n"
       "19:                                       \n"
       : "+r"(topleft),           // %0
         "+r"(botleft),           // %1
@@ -5817,70 +8172,70 @@ void ARGBAffineRow_SSE2(const uint8_t* src_argb,
   intptr_t src_argb_stride_temp = src_argb_stride;
   intptr_t temp;
   asm volatile(
-      "movq      (%3),%%xmm2                     \n"
-      "movq      0x08(%3),%%xmm7                 \n"
-      "shl       $0x10,%1                        \n"
-      "add       $0x4,%1                         \n"
-      "movd      %1,%%xmm5                       \n"
-      "sub       $0x4,%4                         \n"
-      "jl        49f                             \n"
-
-      "pshufd    $0x44,%%xmm7,%%xmm7             \n"
-      "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-      "movdqa    %%xmm2,%%xmm0                   \n"
-      "addps     %%xmm7,%%xmm0                   \n"
-      "movlhps   %%xmm0,%%xmm2                   \n"
-      "movdqa    %%xmm7,%%xmm4                   \n"
-      "addps     %%xmm4,%%xmm4                   \n"
-      "movdqa    %%xmm2,%%xmm3                   \n"
-      "addps     %%xmm4,%%xmm3                   \n"
-      "addps     %%xmm4,%%xmm4                   \n"
+      "movq        (%3),%%xmm2                   \n"
+      "movq        0x08(%3),%%xmm7               \n"
+      "shl         $0x10,%1                      \n"
+      "add         $0x4,%1                       \n"
+      "movd        %1,%%xmm5                     \n"
+      "sub         $0x4,%4                       \n"
+      "jl          49f                           \n"
+
+      "pshufd      $0x44,%%xmm7,%%xmm7           \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "addps       %%xmm7,%%xmm0                 \n"
+      "movlhps     %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm7,%%xmm4                 \n"
+      "addps       %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "addps       %%xmm4,%%xmm3                 \n"
+      "addps       %%xmm4,%%xmm4                 \n"
 
       // 4 pixel loop
       LABELALIGN
       "40:                                       \n"
-      "cvttps2dq %%xmm2,%%xmm0                   \n"  // x,y float->int first 2
-      "cvttps2dq %%xmm3,%%xmm1                   \n"  // x,y float->int next 2
-      "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
-      "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x*4 + y*stride
-      "movd      %%xmm0,%k1                      \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-      "movd      %%xmm0,%k5                      \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-      "movd      0x00(%0,%1,1),%%xmm1            \n"
-      "movd      0x00(%0,%5,1),%%xmm6            \n"
-      "punpckldq %%xmm6,%%xmm1                   \n"
-      "addps     %%xmm4,%%xmm2                   \n"
-      "movq      %%xmm1,(%2)                     \n"
-      "movd      %%xmm0,%k1                      \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-      "movd      %%xmm0,%k5                      \n"
-      "movd      0x00(%0,%1,1),%%xmm0            \n"
-      "movd      0x00(%0,%5,1),%%xmm6            \n"
-      "punpckldq %%xmm6,%%xmm0                   \n"
-      "addps     %%xmm4,%%xmm3                   \n"
-      "movq      %%xmm0,0x08(%2)                 \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%4                         \n"
-      "jge       40b                             \n"
+      "cvttps2dq   %%xmm2,%%xmm0                 \n"  // x,y float->int first 2
+      "cvttps2dq   %%xmm3,%%xmm1                 \n"  // x,y float->int next 2
+      "packssdw    %%xmm1,%%xmm0                 \n"  // x, y as 8 shorts
+      "pmaddwd     %%xmm5,%%xmm0                 \n"  // off = x*4 + y*stride
+      "movd        %%xmm0,%k1                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        %%xmm0,%k5                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        0x00(%0,%1,1),%%xmm1          \n"
+      "movd        0x00(%0,%5,1),%%xmm6          \n"
+      "punpckldq   %%xmm6,%%xmm1                 \n"
+      "addps       %%xmm4,%%xmm2                 \n"
+      "movq        %%xmm1,(%2)                   \n"
+      "movd        %%xmm0,%k1                    \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+      "movd        %%xmm0,%k5                    \n"
+      "movd        0x00(%0,%1,1),%%xmm0          \n"
+      "movd        0x00(%0,%5,1),%%xmm6          \n"
+      "punpckldq   %%xmm6,%%xmm0                 \n"
+      "addps       %%xmm4,%%xmm3                 \n"
+      "movq        %%xmm0,0x08(%2)               \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%4                       \n"
+      "jge         40b                           \n"
 
       "49:                                       \n"
-      "add       $0x3,%4                         \n"
-      "jl        19f                             \n"
+      "add         $0x3,%4                       \n"
+      "jl          19f                           \n"
 
       // 1 pixel loop
       LABELALIGN
       "10:                                       \n"
-      "cvttps2dq %%xmm2,%%xmm0                   \n"
-      "packssdw  %%xmm0,%%xmm0                   \n"
-      "pmaddwd   %%xmm5,%%xmm0                   \n"
-      "addps     %%xmm7,%%xmm2                   \n"
-      "movd      %%xmm0,%k1                      \n"
-      "movd      0x00(%0,%1,1),%%xmm0            \n"
-      "movd      %%xmm0,(%2)                     \n"
-      "lea       0x04(%2),%2                     \n"
-      "sub       $0x1,%4                         \n"
-      "jge       10b                             \n"
+      "cvttps2dq   %%xmm2,%%xmm0                 \n"
+      "packssdw    %%xmm0,%%xmm0                 \n"
+      "pmaddwd     %%xmm5,%%xmm0                 \n"
+      "addps       %%xmm7,%%xmm2                 \n"
+      "movd        %%xmm0,%k1                    \n"
+      "movd        0x00(%0,%1,1),%%xmm0          \n"
+      "movd        %%xmm0,(%2)                   \n"
+      "lea         0x04(%2),%2                   \n"
+      "sub         $0x1,%4                       \n"
+      "jge         10b                           \n"
       "19:                                       \n"
       : "+r"(src_argb),              // %0
         "+r"(src_argb_stride_temp),  // %1
@@ -5899,76 +8254,76 @@ void ARGBAffineRow_SSE2(const uint8_t* src_argb,
 void InterpolateRow_SSSE3(uint8_t* dst_ptr,
                           const uint8_t* src_ptr,
                           ptrdiff_t src_stride,
-                          int dst_width,
+                          int width,
                           int source_y_fraction) {
   asm volatile(
-      "sub       %1,%0                           \n"
-      "cmp       $0x0,%3                         \n"
-      "je        100f                            \n"
-      "cmp       $0x80,%3                        \n"
-      "je        50f                             \n"
-
-      "movd      %3,%%xmm0                       \n"
-      "neg       %3                              \n"
-      "add       $0x100,%3                       \n"
-      "movd      %3,%%xmm5                       \n"
-      "punpcklbw %%xmm0,%%xmm5                   \n"
-      "punpcklwd %%xmm5,%%xmm5                   \n"
-      "pshufd    $0x0,%%xmm5,%%xmm5              \n"
-      "mov       $0x80808080,%%eax               \n"
-      "movd      %%eax,%%xmm4                    \n"
-      "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+      "sub         %1,%0                         \n"
+      "cmp         $0x0,%3                       \n"
+      "je          100f                          \n"
+      "cmp         $0x80,%3                      \n"
+      "je          50f                           \n"
+
+      "movd        %3,%%xmm0                     \n"
+      "neg         %3                            \n"
+      "add         $0x100,%3                     \n"
+      "movd        %3,%%xmm5                     \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"
+      "punpcklwd   %%xmm5,%%xmm5                 \n"
+      "pshufd      $0x0,%%xmm5,%%xmm5            \n"
+      "mov         $0x80808080,%%eax             \n"
+      "movd        %%eax,%%xmm4                  \n"
+      "pshufd      $0x0,%%xmm4,%%xmm4            \n"
 
       // General purpose row blend.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%1),%%xmm0                     \n"
-      "movdqu    0x00(%1,%4,1),%%xmm2            \n"
-      "movdqa     %%xmm0,%%xmm1                  \n"
-      "punpcklbw  %%xmm2,%%xmm0                  \n"
-      "punpckhbw  %%xmm2,%%xmm1                  \n"
-      "psubb      %%xmm4,%%xmm0                  \n"
-      "psubb      %%xmm4,%%xmm1                  \n"
-      "movdqa     %%xmm5,%%xmm2                  \n"
-      "movdqa     %%xmm5,%%xmm3                  \n"
-      "pmaddubsw  %%xmm0,%%xmm2                  \n"
-      "pmaddubsw  %%xmm1,%%xmm3                  \n"
-      "paddw      %%xmm4,%%xmm2                  \n"
-      "paddw      %%xmm4,%%xmm3                  \n"
-      "psrlw      $0x8,%%xmm2                    \n"
-      "psrlw      $0x8,%%xmm3                    \n"
-      "packuswb   %%xmm3,%%xmm2                  \n"
-      "movdqu    %%xmm2,0x00(%1,%0,1)            \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
-      "jmp       99f                             \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x00(%1,%4,1),%%xmm2          \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "psubb       %%xmm4,%%xmm0                 \n"
+      "psubb       %%xmm4,%%xmm1                 \n"
+      "movdqa      %%xmm5,%%xmm2                 \n"
+      "movdqa      %%xmm5,%%xmm3                 \n"
+      "pmaddubsw   %%xmm0,%%xmm2                 \n"
+      "pmaddubsw   %%xmm1,%%xmm3                 \n"
+      "paddw       %%xmm4,%%xmm2                 \n"
+      "paddw       %%xmm4,%%xmm3                 \n"
+      "psrlw       $0x8,%%xmm2                   \n"
+      "psrlw       $0x8,%%xmm3                   \n"
+      "packuswb    %%xmm3,%%xmm2                 \n"
+      "movdqu      %%xmm2,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         99f                           \n"
 
       // Blend 50 / 50.
       LABELALIGN
       "50:                                       \n"
-      "movdqu    (%1),%%xmm0                     \n"
-      "movdqu    0x00(%1,%4,1),%%xmm1            \n"
-      "pavgb     %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,0x00(%1,%0,1)            \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        50b                             \n"
-      "jmp       99f                             \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x00(%1,%4,1),%%xmm1          \n"
+      "pavgb       %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          50b                           \n"
+      "jmp         99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       LABELALIGN
       "100:                                      \n"
-      "movdqu    (%1),%%xmm0                     \n"
-      "movdqu    %%xmm0,0x00(%1,%0,1)            \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        100b                            \n"
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      %%xmm0,0x00(%1,%0,1)          \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          100b                          \n"
 
       "99:                                       \n"
       : "+r"(dst_ptr),               // %0
         "+r"(src_ptr),               // %1
-        "+rm"(dst_width),            // %2
+        "+rm"(width),                // %2
         "+r"(source_y_fraction)      // %3
       : "r"((intptr_t)(src_stride))  // %4
       : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
@@ -5980,71 +8335,73 @@ void InterpolateRow_SSSE3(uint8_t* dst_ptr,
 void InterpolateRow_AVX2(uint8_t* dst_ptr,
                          const uint8_t* src_ptr,
                          ptrdiff_t src_stride,
-                         int dst_width,
+                         int width,
                          int source_y_fraction) {
   asm volatile(
-      "cmp       $0x0,%3                         \n"
-      "je        100f                            \n"
-      "sub       %1,%0                           \n"
-      "cmp       $0x80,%3                        \n"
-      "je        50f                             \n"
-
-      "vmovd      %3,%%xmm0                      \n"
-      "neg        %3                             \n"
-      "add        $0x100,%3                      \n"
-      "vmovd      %3,%%xmm5                      \n"
-      "vpunpcklbw %%xmm0,%%xmm5,%%xmm5           \n"
-      "vpunpcklwd %%xmm5,%%xmm5,%%xmm5           \n"
+      "sub         %1,%0                         \n"
+      "cmp         $0x0,%3                       \n"
+      "je          100f                          \n"
+      "cmp         $0x80,%3                      \n"
+      "je          50f                           \n"
+
+      "vmovd       %3,%%xmm0                     \n"
+      "neg         %3                            \n"
+      "add         $0x100,%3                     \n"
+      "vmovd       %3,%%xmm5                     \n"
+      "vpunpcklbw  %%xmm0,%%xmm5,%%xmm5          \n"
+      "vpunpcklwd  %%xmm5,%%xmm5,%%xmm5          \n"
       "vbroadcastss %%xmm5,%%ymm5                \n"
-      "mov        $0x80808080,%%eax              \n"
-      "vmovd      %%eax,%%xmm4                   \n"
+      "mov         $0x80808080,%%eax             \n"
+      "vmovd       %%eax,%%xmm4                  \n"
       "vbroadcastss %%xmm4,%%ymm4                \n"
 
       // General purpose row blend.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%1),%%ymm0                    \n"
-      "vmovdqu    0x00(%1,%4,1),%%ymm2           \n"
-      "vpunpckhbw %%ymm2,%%ymm0,%%ymm1           \n"
-      "vpunpcklbw %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpsubb     %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpsubb     %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm1,%%ymm5,%%ymm1           \n"
-      "vpmaddubsw %%ymm0,%%ymm5,%%ymm0           \n"
-      "vpaddw     %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpaddw     %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vmovdqu    %%ymm0,0x00(%1,%0,1)           \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
-      "jmp        99f                            \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vmovdqu     0x00(%1,%4,1),%%ymm2          \n"
+      "vpunpckhbw  %%ymm2,%%ymm0,%%ymm1          \n"
+      "vpunpcklbw  %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpsubb      %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpsubb      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm1,%%ymm5,%%ymm1          \n"
+      "vpmaddubsw  %%ymm0,%%ymm5,%%ymm0          \n"
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "jmp         99f                           \n"
 
       // Blend 50 / 50.
       LABELALIGN
       "50:                                       \n"
-      "vmovdqu   (%1),%%ymm0                     \n"
-      "vpavgb    0x00(%1,%4,1),%%ymm0,%%ymm0     \n"
-      "vmovdqu   %%ymm0,0x00(%1,%0,1)            \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        50b                             \n"
-      "jmp       99f                             \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vpavgb      0x00(%1,%4,1),%%ymm0,%%ymm0   \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          50b                           \n"
+      "jmp         99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       LABELALIGN
       "100:                                      \n"
-      "rep movsb                                 \n"
-      "jmp       999f                            \n"
+      "vmovdqu     (%1),%%ymm0                   \n"
+      "vmovdqu     %%ymm0,0x00(%1,%0,1)          \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          100b                          \n"
 
       "99:                                       \n"
       "vzeroupper                                \n"
-      "999:                                      \n"
-      : "+D"(dst_ptr),               // %0
-        "+S"(src_ptr),               // %1
-        "+cm"(dst_width),            // %2
+      : "+r"(dst_ptr),               // %0
+        "+r"(src_ptr),               // %1
+        "+r"(width),                 // %2
         "+r"(source_y_fraction)      // %3
       : "r"((intptr_t)(src_stride))  // %4
       : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm4", "xmm5");
@@ -6059,20 +8416,20 @@ void ARGBShuffleRow_SSSE3(const uint8_t* src_argb,
                           int width) {
   asm volatile(
 
-      "movdqu    (%3),%%xmm5                     \n"
+      "movdqu      (%3),%%xmm5                   \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pshufb    %%xmm5,%%xmm0                   \n"
-      "pshufb    %%xmm5,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -6093,16 +8450,16 @@ void ARGBShuffleRow_AVX2(const uint8_t* src_argb,
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu   (%0),%%ymm0                     \n"
-      "vmovdqu   0x20(%0),%%ymm1                 \n"
-      "lea       0x40(%0),%0                     \n"
-      "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
-      "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
-      "vmovdqu   %%ymm0,(%1)                     \n"
-      "vmovdqu   %%ymm1,0x20(%1)                 \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
@@ -6120,24 +8477,24 @@ void I422ToYUY2Row_SSE2(const uint8_t* src_y,
                         int width) {
   asm volatile(
 
-      "sub       %1,%2                             \n"
-
-      LABELALIGN
-      "1:                                          \n"
-      "movq      (%1),%%xmm2                       \n"
-      "movq      0x00(%1,%2,1),%%xmm1              \n"
-      "add       $0x8,%1                           \n"
-      "punpcklbw %%xmm1,%%xmm2                     \n"
-      "movdqu    (%0),%%xmm0                       \n"
-      "add       $0x10,%0                          \n"
-      "movdqa    %%xmm0,%%xmm1                     \n"
-      "punpcklbw %%xmm2,%%xmm0                     \n"
-      "punpckhbw %%xmm2,%%xmm1                     \n"
-      "movdqu    %%xmm0,(%3)                       \n"
-      "movdqu    %%xmm1,0x10(%3)                   \n"
-      "lea       0x20(%3),%3                       \n"
-      "sub       $0x10,%4                          \n"
-      "jg         1b                               \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%1),%%xmm2                   \n"
+      "movq        0x00(%1,%2,1),%%xmm1          \n"
+      "add         $0x8,%1                       \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "add         $0x10,%0                      \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm2,%%xmm0                 \n"
+      "punpckhbw   %%xmm2,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%3)                   \n"
+      "movdqu      %%xmm1,0x10(%3)               \n"
+      "lea         0x20(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -6156,24 +8513,24 @@ void I422ToUYVYRow_SSE2(const uint8_t* src_y,
                         int width) {
   asm volatile(
 
-      "sub        %1,%2                            \n"
-
-      LABELALIGN
-      "1:                                          \n"
-      "movq      (%1),%%xmm2                       \n"
-      "movq      0x00(%1,%2,1),%%xmm1              \n"
-      "add       $0x8,%1                           \n"
-      "punpcklbw %%xmm1,%%xmm2                     \n"
-      "movdqu    (%0),%%xmm0                       \n"
-      "movdqa    %%xmm2,%%xmm1                     \n"
-      "add       $0x10,%0                          \n"
-      "punpcklbw %%xmm0,%%xmm1                     \n"
-      "punpckhbw %%xmm0,%%xmm2                     \n"
-      "movdqu    %%xmm1,(%3)                       \n"
-      "movdqu    %%xmm2,0x10(%3)                   \n"
-      "lea       0x20(%3),%3                       \n"
-      "sub       $0x10,%4                          \n"
-      "jg         1b                               \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%1),%%xmm2                   \n"
+      "movq        0x00(%1,%2,1),%%xmm1          \n"
+      "add         $0x8,%1                       \n"
+      "punpcklbw   %%xmm1,%%xmm2                 \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "add         $0x10,%0                      \n"
+      "punpcklbw   %%xmm0,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm2                 \n"
+      "movdqu      %%xmm1,(%3)                   \n"
+      "movdqu      %%xmm2,0x10(%3)               \n"
+      "lea         0x20(%3),%3                   \n"
+      "sub         $0x10,%4                      \n"
+      "jg          1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -6192,27 +8549,27 @@ void I422ToYUY2Row_AVX2(const uint8_t* src_y,
                         int width) {
   asm volatile(
 
-      "sub       %1,%2                             \n"
-
-      LABELALIGN
-      "1:                                          \n"
-      "vpmovzxbw  (%1),%%ymm1                      \n"
-      "vpmovzxbw  0x00(%1,%2,1),%%ymm2             \n"
-      "add        $0x10,%1                         \n"
-      "vpsllw     $0x8,%%ymm2,%%ymm2               \n"
-      "vpor       %%ymm1,%%ymm2,%%ymm2             \n"
-      "vmovdqu    (%0),%%ymm0                      \n"
-      "add        $0x20,%0                         \n"
-      "vpunpcklbw %%ymm2,%%ymm0,%%ymm1             \n"
-      "vpunpckhbw %%ymm2,%%ymm0,%%ymm2             \n"
-      "vextractf128 $0x0,%%ymm1,(%3)               \n"
-      "vextractf128 $0x0,%%ymm2,0x10(%3)           \n"
-      "vextractf128 $0x1,%%ymm1,0x20(%3)           \n"
-      "vextractf128 $0x1,%%ymm2,0x30(%3)           \n"
-      "lea        0x40(%3),%3                      \n"
-      "sub        $0x20,%4                         \n"
-      "jg         1b                               \n"
-      "vzeroupper                                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%1),%%ymm1                   \n"
+      "vpmovzxbw   0x00(%1,%2,1),%%ymm2          \n"
+      "add         $0x10,%1                      \n"
+      "vpsllw      $0x8,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm2,%%ymm2          \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "add         $0x20,%0                      \n"
+      "vpunpcklbw  %%ymm2,%%ymm0,%%ymm1          \n"
+      "vpunpckhbw  %%ymm2,%%ymm0,%%ymm2          \n"
+      "vextractf128 $0x0,%%ymm1,(%3)             \n"
+      "vextractf128 $0x0,%%ymm2,0x10(%3)         \n"
+      "vextractf128 $0x1,%%ymm1,0x20(%3)         \n"
+      "vextractf128 $0x1,%%ymm2,0x30(%3)         \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -6231,27 +8588,27 @@ void I422ToUYVYRow_AVX2(const uint8_t* src_y,
                         int width) {
   asm volatile(
 
-      "sub        %1,%2                            \n"
-
-      LABELALIGN
-      "1:                                          \n"
-      "vpmovzxbw  (%1),%%ymm1                      \n"
-      "vpmovzxbw  0x00(%1,%2,1),%%ymm2             \n"
-      "add        $0x10,%1                         \n"
-      "vpsllw     $0x8,%%ymm2,%%ymm2               \n"
-      "vpor       %%ymm1,%%ymm2,%%ymm2             \n"
-      "vmovdqu    (%0),%%ymm0                      \n"
-      "add        $0x20,%0                         \n"
-      "vpunpcklbw %%ymm0,%%ymm2,%%ymm1             \n"
-      "vpunpckhbw %%ymm0,%%ymm2,%%ymm2             \n"
-      "vextractf128 $0x0,%%ymm1,(%3)               \n"
-      "vextractf128 $0x0,%%ymm2,0x10(%3)           \n"
-      "vextractf128 $0x1,%%ymm1,0x20(%3)           \n"
-      "vextractf128 $0x1,%%ymm2,0x30(%3)           \n"
-      "lea        0x40(%3),%3                      \n"
-      "sub        $0x20,%4                         \n"
-      "jg         1b                               \n"
-      "vzeroupper                                  \n"
+      "sub         %1,%2                         \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vpmovzxbw   (%1),%%ymm1                   \n"
+      "vpmovzxbw   0x00(%1,%2,1),%%ymm2          \n"
+      "add         $0x10,%1                      \n"
+      "vpsllw      $0x8,%%ymm2,%%ymm2            \n"
+      "vpor        %%ymm1,%%ymm2,%%ymm2          \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "add         $0x20,%0                      \n"
+      "vpunpcklbw  %%ymm0,%%ymm2,%%ymm1          \n"
+      "vpunpckhbw  %%ymm0,%%ymm2,%%ymm2          \n"
+      "vextractf128 $0x0,%%ymm1,(%3)             \n"
+      "vextractf128 $0x0,%%ymm2,0x10(%3)         \n"
+      "vextractf128 $0x1,%%ymm1,0x20(%3)         \n"
+      "vextractf128 $0x1,%%ymm2,0x30(%3)         \n"
+      "lea         0x40(%3),%3                   \n"
+      "sub         $0x20,%4                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -6269,47 +8626,47 @@ void ARGBPolynomialRow_SSE2(const uint8_t* src_argb,
                             int width) {
   asm volatile(
 
-      "pxor      %%xmm3,%%xmm3                   \n"
+      "pxor        %%xmm3,%%xmm3                 \n"
 
       // 2 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm0                     \n"
-      "lea       0x8(%0),%0                      \n"
-      "punpcklbw %%xmm3,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm4                   \n"
-      "punpcklwd %%xmm3,%%xmm0                   \n"
-      "punpckhwd %%xmm3,%%xmm4                   \n"
-      "cvtdq2ps  %%xmm0,%%xmm0                   \n"
-      "cvtdq2ps  %%xmm4,%%xmm4                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "movdqa    %%xmm4,%%xmm5                   \n"
-      "mulps     0x10(%3),%%xmm0                 \n"
-      "mulps     0x10(%3),%%xmm4                 \n"
-      "addps     (%3),%%xmm0                     \n"
-      "addps     (%3),%%xmm4                     \n"
-      "movdqa    %%xmm1,%%xmm2                   \n"
-      "movdqa    %%xmm5,%%xmm6                   \n"
-      "mulps     %%xmm1,%%xmm2                   \n"
-      "mulps     %%xmm5,%%xmm6                   \n"
-      "mulps     %%xmm2,%%xmm1                   \n"
-      "mulps     %%xmm6,%%xmm5                   \n"
-      "mulps     0x20(%3),%%xmm2                 \n"
-      "mulps     0x20(%3),%%xmm6                 \n"
-      "mulps     0x30(%3),%%xmm1                 \n"
-      "mulps     0x30(%3),%%xmm5                 \n"
-      "addps     %%xmm2,%%xmm0                   \n"
-      "addps     %%xmm6,%%xmm4                   \n"
-      "addps     %%xmm1,%%xmm0                   \n"
-      "addps     %%xmm5,%%xmm4                   \n"
-      "cvttps2dq %%xmm0,%%xmm0                   \n"
-      "cvttps2dq %%xmm4,%%xmm4                   \n"
-      "packuswb  %%xmm4,%%xmm0                   \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x2,%2                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "punpcklbw   %%xmm3,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "punpcklwd   %%xmm3,%%xmm0                 \n"
+      "punpckhwd   %%xmm3,%%xmm4                 \n"
+      "cvtdq2ps    %%xmm0,%%xmm0                 \n"
+      "cvtdq2ps    %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "mulps       0x10(%3),%%xmm0               \n"
+      "mulps       0x10(%3),%%xmm4               \n"
+      "addps       (%3),%%xmm0                   \n"
+      "addps       (%3),%%xmm4                   \n"
+      "movdqa      %%xmm1,%%xmm2                 \n"
+      "movdqa      %%xmm5,%%xmm6                 \n"
+      "mulps       %%xmm1,%%xmm2                 \n"
+      "mulps       %%xmm5,%%xmm6                 \n"
+      "mulps       %%xmm2,%%xmm1                 \n"
+      "mulps       %%xmm6,%%xmm5                 \n"
+      "mulps       0x20(%3),%%xmm2               \n"
+      "mulps       0x20(%3),%%xmm6               \n"
+      "mulps       0x30(%3),%%xmm1               \n"
+      "mulps       0x30(%3),%%xmm5               \n"
+      "addps       %%xmm2,%%xmm0                 \n"
+      "addps       %%xmm6,%%xmm4                 \n"
+      "addps       %%xmm1,%%xmm0                 \n"
+      "addps       %%xmm5,%%xmm4                 \n"
+      "cvttps2dq   %%xmm0,%%xmm0                 \n"
+      "cvttps2dq   %%xmm4,%%xmm4                 \n"
+      "packuswb    %%xmm4,%%xmm0                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x2,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -6405,27 +8762,27 @@ void HalfFloatRow_AVX2(const uint16_t* src,
                        int width) {
   scale *= kScaleBias;
   asm volatile(
-      "vbroadcastss  %3, %%ymm4                  \n"
-      "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
-      "sub        %0,%1                          \n"
+      "vbroadcastss %3, %%ymm4                   \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+      "sub         %0,%1                         \n"
 
       // 16 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm2                    \n"  // 16 shorts
-      "add        $0x20,%0                       \n"
-      "vpunpckhwd %%ymm5,%%ymm2,%%ymm3           \n"  // mutates
-      "vpunpcklwd %%ymm5,%%ymm2,%%ymm2           \n"
-      "vcvtdq2ps  %%ymm3,%%ymm3                  \n"
-      "vcvtdq2ps  %%ymm2,%%ymm2                  \n"
-      "vmulps     %%ymm3,%%ymm4,%%ymm3           \n"
-      "vmulps     %%ymm2,%%ymm4,%%ymm2           \n"
-      "vpsrld     $0xd,%%ymm3,%%ymm3             \n"
-      "vpsrld     $0xd,%%ymm2,%%ymm2             \n"
-      "vpackssdw  %%ymm3, %%ymm2, %%ymm2         \n"  // unmutates
-      "vmovdqu    %%ymm2,-0x20(%0,%1,1)          \n"
-      "sub        $0x10,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // 16 shorts
+      "add         $0x20,%0                      \n"
+      "vpunpckhwd  %%ymm5,%%ymm2,%%ymm3          \n"  // mutates
+      "vpunpcklwd  %%ymm5,%%ymm2,%%ymm2          \n"
+      "vcvtdq2ps   %%ymm3,%%ymm3                 \n"
+      "vcvtdq2ps   %%ymm2,%%ymm2                 \n"
+      "vmulps      %%ymm3,%%ymm4,%%ymm3          \n"
+      "vmulps      %%ymm2,%%ymm4,%%ymm2          \n"
+      "vpsrld      $0xd,%%ymm3,%%ymm3            \n"
+      "vpsrld      $0xd,%%ymm2,%%ymm2            \n"
+      "vpackssdw   %%ymm3, %%ymm2, %%ymm2        \n"  // unmutates
+      "vmovdqu     %%ymm2,-0x20(%0,%1,1)         \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
 
       "vzeroupper                                \n"
       : "+r"(src),   // %0
@@ -6434,7 +8791,7 @@ void HalfFloatRow_AVX2(const uint16_t* src,
 #if defined(__x86_64__)
       : "x"(scale)  // %3
 #else
-      : "m"(scale)  // %3
+      : "m"(scale)            // %3
 #endif
       : "memory", "cc", "xmm2", "xmm3", "xmm4", "xmm5");
 }
@@ -6446,8 +8803,8 @@ void HalfFloatRow_F16C(const uint16_t* src,
                        float scale,
                        int width) {
   asm volatile(
-      "vbroadcastss  %3, %%ymm4                  \n"
-      "sub        %0,%1                          \n"
+      "vbroadcastss %3, %%ymm4                   \n"
+      "sub         %0,%1                         \n"
 
       // 16 pixel loop.
       LABELALIGN
@@ -6472,7 +8829,7 @@ void HalfFloatRow_F16C(const uint16_t* src,
 #if defined(__x86_64__)
       : "x"(scale)  // %3
 #else
-      : "m"(scale)  // %3
+      : "m"(scale)            // %3
 #endif
       : "memory", "cc", "xmm2", "xmm3", "xmm4");
 }
@@ -6481,7 +8838,7 @@ void HalfFloatRow_F16C(const uint16_t* src,
 #ifdef HAS_HALFFLOATROW_F16C
 void HalfFloat1Row_F16C(const uint16_t* src, uint16_t* dst, float, int width) {
   asm volatile(
-      "sub        %0,%1                          \n"
+      "sub         %0,%1                         \n"
       // 16 pixel loop.
       LABELALIGN
       "1:                                        \n"
@@ -6515,21 +8872,21 @@ void ARGBColorTableRow_X86(uint8_t* dst_argb,
       // 1 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movzb     (%0),%1                         \n"
-      "lea       0x4(%0),%0                      \n"
-      "movzb     0x00(%3,%1,4),%1                \n"
-      "mov       %b1,-0x4(%0)                    \n"
-      "movzb     -0x3(%0),%1                     \n"
-      "movzb     0x01(%3,%1,4),%1                \n"
-      "mov       %b1,-0x3(%0)                    \n"
-      "movzb     -0x2(%0),%1                     \n"
-      "movzb     0x02(%3,%1,4),%1                \n"
-      "mov       %b1,-0x2(%0)                    \n"
-      "movzb     -0x1(%0),%1                     \n"
-      "movzb     0x03(%3,%1,4),%1                \n"
-      "mov       %b1,-0x1(%0)                    \n"
-      "dec       %2                              \n"
-      "jg        1b                              \n"
+      "movzb       (%0),%1                       \n"
+      "lea         0x4(%0),%0                    \n"
+      "movzb       0x00(%3,%1,4),%1              \n"
+      "mov         %b1,-0x4(%0)                  \n"
+      "movzb       -0x3(%0),%1                   \n"
+      "movzb       0x01(%3,%1,4),%1              \n"
+      "mov         %b1,-0x3(%0)                  \n"
+      "movzb       -0x2(%0),%1                   \n"
+      "movzb       0x02(%3,%1,4),%1              \n"
+      "mov         %b1,-0x2(%0)                  \n"
+      "movzb       -0x1(%0),%1                   \n"
+      "movzb       0x03(%3,%1,4),%1              \n"
+      "mov         %b1,-0x1(%0)                  \n"
+      "dec         %2                            \n"
+      "jg          1b                            \n"
       : "+r"(dst_argb),     // %0
         "=&d"(pixel_temp),  // %1
         "+r"(width)         // %2
@@ -6548,18 +8905,18 @@ void RGBColorTableRow_X86(uint8_t* dst_argb,
       // 1 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movzb     (%0),%1                         \n"
-      "lea       0x4(%0),%0                      \n"
-      "movzb     0x00(%3,%1,4),%1                \n"
-      "mov       %b1,-0x4(%0)                    \n"
-      "movzb     -0x3(%0),%1                     \n"
-      "movzb     0x01(%3,%1,4),%1                \n"
-      "mov       %b1,-0x3(%0)                    \n"
-      "movzb     -0x2(%0),%1                     \n"
-      "movzb     0x02(%3,%1,4),%1                \n"
-      "mov       %b1,-0x2(%0)                    \n"
-      "dec       %2                              \n"
-      "jg        1b                              \n"
+      "movzb       (%0),%1                       \n"
+      "lea         0x4(%0),%0                    \n"
+      "movzb       0x00(%3,%1,4),%1              \n"
+      "mov         %b1,-0x4(%0)                  \n"
+      "movzb       -0x3(%0),%1                   \n"
+      "movzb       0x01(%3,%1,4),%1              \n"
+      "mov         %b1,-0x3(%0)                  \n"
+      "movzb       -0x2(%0),%1                   \n"
+      "movzb       0x02(%3,%1,4),%1              \n"
+      "mov         %b1,-0x2(%0)                  \n"
+      "dec         %2                            \n"
+      "jg          1b                            \n"
       : "+r"(dst_argb),     // %0
         "=&d"(pixel_temp),  // %1
         "+r"(width)         // %2
@@ -6578,86 +8935,86 @@ void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
   uintptr_t pixel_temp;
   uintptr_t table_temp;
   asm volatile(
-      "movd      %6,%%xmm3                       \n"
-      "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-      "pcmpeqb   %%xmm4,%%xmm4                   \n"
-      "psllw     $0x8,%%xmm4                     \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "movd        %6,%%xmm3                     \n"
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psllw       $0x8,%%xmm4                   \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       // 4 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%2),%%xmm0                     \n"
-      "pmaddubsw %%xmm3,%%xmm0                   \n"
-      "phaddw    %%xmm0,%%xmm0                   \n"
-      "pand      %%xmm4,%%xmm0                   \n"
-      "punpcklwd %%xmm5,%%xmm0                   \n"
-      "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-      "add       %5,%1                           \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-      "movzb     (%2),%0                         \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,(%3)                        \n"
-      "movzb     0x1(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x1(%3)                     \n"
-      "movzb     0x2(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x2(%3)                     \n"
-      "movzb     0x3(%2),%0                      \n"
-      "mov       %b0,0x3(%3)                     \n"
-
-      "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-      "add       %5,%1                           \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-      "movzb     0x4(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x4(%3)                     \n"
-      "movzb     0x5(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x5(%3)                     \n"
-      "movzb     0x6(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x6(%3)                     \n"
-      "movzb     0x7(%2),%0                      \n"
-      "mov       %b0,0x7(%3)                     \n"
-
-      "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-      "add       %5,%1                           \n"
-      "pshufd    $0x39,%%xmm0,%%xmm0             \n"
-
-      "movzb     0x8(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x8(%3)                     \n"
-      "movzb     0x9(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0x9(%3)                     \n"
-      "movzb     0xa(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0xa(%3)                     \n"
-      "movzb     0xb(%2),%0                      \n"
-      "mov       %b0,0xb(%3)                     \n"
-
-      "movd      %%xmm0,%k1                      \n"  // 32 bit offset
-      "add       %5,%1                           \n"
-
-      "movzb     0xc(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0xc(%3)                     \n"
-      "movzb     0xd(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0xd(%3)                     \n"
-      "movzb     0xe(%2),%0                      \n"
-      "movzb     0x00(%1,%0,1),%0                \n"
-      "mov       %b0,0xe(%3)                     \n"
-      "movzb     0xf(%2),%0                      \n"
-      "mov       %b0,0xf(%3)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "lea       0x10(%3),%3                     \n"
-      "sub       $0x4,%4                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%2),%%xmm0                   \n"
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"
+      "phaddw      %%xmm0,%%xmm0                 \n"
+      "pand        %%xmm4,%%xmm0                 \n"
+      "punpcklwd   %%xmm5,%%xmm0                 \n"
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       (%2),%0                       \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,(%3)                      \n"
+      "movzb       0x1(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x1(%3)                   \n"
+      "movzb       0x2(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x2(%3)                   \n"
+      "movzb       0x3(%2),%0                    \n"
+      "mov         %b0,0x3(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       0x4(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x4(%3)                   \n"
+      "movzb       0x5(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x5(%3)                   \n"
+      "movzb       0x6(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x6(%3)                   \n"
+      "movzb       0x7(%2),%0                    \n"
+      "mov         %b0,0x7(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+      "pshufd      $0x39,%%xmm0,%%xmm0           \n"
+
+      "movzb       0x8(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x8(%3)                   \n"
+      "movzb       0x9(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0x9(%3)                   \n"
+      "movzb       0xa(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xa(%3)                   \n"
+      "movzb       0xb(%2),%0                    \n"
+      "mov         %b0,0xb(%3)                   \n"
+
+      "movd        %%xmm0,%k1                    \n"  // 32 bit offset
+      "add         %5,%1                         \n"
+
+      "movzb       0xc(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xc(%3)                   \n"
+      "movzb       0xd(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xd(%3)                   \n"
+      "movzb       0xe(%2),%0                    \n"
+      "movzb       0x00(%1,%0,1),%0              \n"
+      "mov         %b0,0xe(%3)                   \n"
+      "movzb       0xf(%2),%0                    \n"
+      "mov         %b0,0xf(%3)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "lea         0x10(%3),%3                   \n"
+      "sub         $0x4,%4                       \n"
+      "jg          1b                            \n"
       : "=&d"(pixel_temp),  // %0
         "=&a"(table_temp),  // %1
         "+r"(src_argb),     // %2
@@ -6669,126 +9026,306 @@ void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb,
 }
 #endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
 
-#ifdef HAS_NV21TOYUV24ROW_AVX2
-
-// begin NV21ToYUV24Row_C avx2 constants
-static const ulvec8 kBLEND0 = {0x80, 0x00, 0x80, 0x80, 0x00, 0x80, 0x80, 0x00,
-                               0x80, 0x80, 0x00, 0x80, 0x80, 0x00, 0x80, 0x80,
-                               0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80,
-                               0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00};
-
-static const ulvec8 kBLEND1 = {0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00,
-                               0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00,
-                               0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00,
-                               0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80};
-
-static const ulvec8 kBLEND2 = {0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00,
-                               0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80,
-                               0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00, 0x00,
-                               0x80, 0x00, 0x00, 0x80, 0x00, 0x00, 0x80, 0x00};
-
-static const ulvec8 kSHUF0 = {0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02, 0x0d,
-                              0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80, 0x05,
-                              0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02, 0x0d,
-                              0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80, 0x05};
-
-static const ulvec8 kSHUF1 = {0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02,
-                              0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80,
-                              0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80, 0x02,
-                              0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f, 0x80};
-
-static const ulvec8 kSHUF2 = {0x0a, 0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80,
-                              0x02, 0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f,
-                              0x0a, 0x80, 0x00, 0x0b, 0x80, 0x01, 0x0c, 0x80,
-                              0x02, 0x0d, 0x80, 0x03, 0x0e, 0x80, 0x04, 0x0f};
-
-static const ulvec8 kSHUF3 = {0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80, 0x80,
-                              0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a, 0x80,
-                              0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80, 0x80,
-                              0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a, 0x80};
-
-static const ulvec8 kSHUF4 = {0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80,
-                              0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a,
-                              0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07, 0x80,
-                              0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80, 0x0a};
-
-static const ulvec8 kSHUF5 = {0x80, 0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07,
-                              0x80, 0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80,
-                              0x80, 0x05, 0x80, 0x80, 0x06, 0x80, 0x80, 0x07,
-                              0x80, 0x80, 0x08, 0x80, 0x80, 0x09, 0x80, 0x80};
-
-// NV21ToYUV24Row_AVX2
+static const uvec8 kYUV24Shuffle[3] = {
+    {8, 9, 0, 8, 9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12},
+    {9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15},
+    {2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15, 6, 14, 15, 7}};
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
+void NV21ToYUV24Row_SSSE3(const uint8_t* src_y,
+                          const uint8_t* src_vu,
+                          uint8_t* dst_yuv24,
+                          int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "movdqa      (%4),%%xmm4                   \n"  // 3 shuffler constants
+      "movdqa      16(%4),%%xmm5                 \n"
+      "movdqa      32(%4),%%xmm6                 \n"
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm2                   \n"  // load 16 Y values
+      "movdqu      (%0,%1),%%xmm3                \n"  // load 8 VU values
+      "lea         16(%0),%0                     \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "shufps      $0x44,%%xmm3,%%xmm0           \n"  // Y 0..7,  UV 0..3
+      "shufps      $0x99,%%xmm3,%%xmm1           \n"  // Y 4..11, UV 2..5
+      "shufps      $0xee,%%xmm3,%%xmm2           \n"  // Y 8..15, UV 4..7
+      "pshufb      %%xmm4, %%xmm0                \n"  // weave into YUV24
+      "pshufb      %%xmm5, %%xmm1                \n"
+      "pshufb      %%xmm6, %%xmm2                \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "movdqu      %%xmm1,16(%2)                 \n"
+      "movdqu      %%xmm2,32(%2)                 \n"
+      "lea         48(%2),%2                     \n"
+      "sub         $16,%3                        \n"  // 16 pixels per loop
+      "jg          1b                            \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
 void NV21ToYUV24Row_AVX2(const uint8_t* src_y,
                          const uint8_t* src_vu,
                          uint8_t* dst_yuv24,
                          int width) {
-  uint8_t* src_y_ptr;
-  uint64_t src_offset = 0;
-  uint64_t width64;
-
-  width64 = width;
-  src_y_ptr = (uint8_t*)src_y;
-
-  asm volatile(
-      "vmovdqu     %5, %%ymm0 \n"  // init blend value
-      "vmovdqu     %6, %%ymm1 \n"  // init blend value
-      "vmovdqu     %7, %%ymm2 \n"  // init blend value
-      //      "sub         $0x20, %3  \n"  //sub 32 from width for final loop
-
-      LABELALIGN
-      "1:                                             \n"  // label 1
-      "vmovdqu     (%0,%4), %%ymm3                    \n"  // src_y
-      "vmovdqu     1(%1,%4), %%ymm4                   \n"  // src_uv+1
-      "vmovdqu     (%1), %%ymm5                       \n"  // src_uv
-      "vpshufb     %8, %%ymm3, %%ymm13                \n"  // y, kSHUF0 for shuf
-      "vpshufb     %9, %%ymm4, %%ymm14                \n"  // uv+1, kSHUF1 for
-                                                           // shuf
-      "vpshufb     %10, %%ymm5, %%ymm15               \n"  // uv, kSHUF2 for
-                                                           // shuf
-      "vpshufb     %11, %%ymm3, %%ymm3                \n"  // y kSHUF3 for shuf
-      "vpshufb     %12, %%ymm4, %%ymm4                \n"  // uv+1 kSHUF4 for
-                                                           // shuf
-      "vpblendvb   %%ymm0, %%ymm14, %%ymm13, %%ymm12  \n"  // blend 0
-      "vpblendvb   %%ymm0, %%ymm13, %%ymm14, %%ymm14  \n"  // blend 0
-      "vpblendvb   %%ymm2, %%ymm15, %%ymm12, %%ymm12  \n"  // blend 2
-      "vpblendvb   %%ymm1, %%ymm15, %%ymm14, %%ymm13  \n"  // blend 1
-      "vpshufb     %13, %%ymm5, %%ymm15               \n"  // shuffle const
-      "vpor        %%ymm4, %%ymm3, %%ymm5             \n"  // get results
-      "vmovdqu     %%ymm12, 0x20(%2)                  \n"  // store dst_yuv+20h
-      "vpor        %%ymm15, %%ymm5, %%ymm3            \n"  // get results
-      "add         $0x20, %4                          \n"  // add to src buffer
-                                                           // ptr
-      "vinserti128 $0x1, %%xmm3, %%ymm13, %%ymm4      \n"  // insert
-      "vperm2i128  $0x31, %%ymm13, %%ymm3, %%ymm5     \n"  // insert
-      "vmovdqu     %%ymm4, (%2)                       \n"  // store dst_yuv
-      "vmovdqu     %%ymm5, 0x40(%2)                   \n"  // store dst_yuv+40h
-      "add         $0x60,%2                           \n"  // add to dst buffer
-                                                           // ptr
-      //      "cmp         %3, %4                             \n" //(width64 -
-      //      32 bytes) and src_offset
-      "sub         $0x20,%3                           \n"  // 32 pixels per loop
-      "jg          1b                                 \n"
-      "vzeroupper                                     \n"  // sse-avx2
-                                                           // transistions
-
-      : "+r"(src_y),      //%0
-        "+r"(src_vu),     //%1
-        "+r"(dst_yuv24),  //%2
-        "+r"(width64),    //%3
-        "+r"(src_offset)  //%4
-      : "m"(kBLEND0),     //%5
-        "m"(kBLEND1),     //%6
-        "m"(kBLEND2),     //%7
-        "m"(kSHUF0),      //%8
-        "m"(kSHUF1),      //%9
-        "m"(kSHUF2),      //%10
-        "m"(kSHUF3),      //%11
-        "m"(kSHUF4),      //%12
-        "m"(kSHUF5)       //%13
-      : "memory", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm12",
-        "xmm13", "xmm14", "xmm15");
-}
-#endif  // HAS_NV21TOYUV24ROW_AVX2
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "vbroadcastf128 (%4),%%ymm4                \n"  // 3 shuffler constants
+      "vbroadcastf128 16(%4),%%ymm5              \n"
+      "vbroadcastf128 32(%4),%%ymm6              \n"
+
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // load 32 Y values
+      "vmovdqu     (%0,%1),%%ymm3                \n"  // load 16 VU values
+      "lea         32(%0),%0                     \n"
+      "vshufps     $0x44,%%ymm3,%%ymm2,%%ymm0    \n"  // Y 0..7,  UV 0..3
+      "vshufps     $0x99,%%ymm3,%%ymm2,%%ymm1    \n"  // Y 4..11, UV 2..5
+      "vshufps     $0xee,%%ymm3,%%ymm2,%%ymm2    \n"  // Y 8..15, UV 4..7
+      "vpshufb     %%ymm4,%%ymm0,%%ymm0          \n"  // weave into YUV24
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpshufb     %%ymm6,%%ymm2,%%ymm2          \n"
+      "vperm2i128  $0x20,%%ymm1,%%ymm0,%%ymm3    \n"
+      "vperm2i128  $0x30,%%ymm0,%%ymm2,%%ymm0    \n"
+      "vperm2i128  $0x31,%%ymm2,%%ymm1,%%ymm1    \n"
+      "vmovdqu     %%ymm3,(%2)                   \n"
+      "vmovdqu     %%ymm0,32(%2)                 \n"
+      "vmovdqu     %%ymm1,64(%2)                 \n"
+      "lea         96(%2),%2                     \n"
+      "sub         $32,%3                        \n"  // 32 pixels per loop
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+#ifdef HAS_NV21ToYUV24ROW_AVX512
+// The following VMBI VEX256 code tests okay with the intelsde emulator.
+static const lvec8 kYUV24Perm[3] = {
+    {32, 33, 0,  32, 33, 1,  34, 35, 2,  34, 35, 3,  36, 37, 4,  36,
+     37, 5,  38, 39, 6,  38, 39, 7,  40, 41, 8,  40, 41, 9,  42, 43},
+    {10, 42, 43, 11, 44, 45, 12, 44, 45, 13, 46, 47, 14, 46, 47, 15,
+     48, 49, 16, 48, 49, 17, 50, 51, 18, 50, 51, 19, 52, 53, 20, 52},
+    {53, 21, 54, 55, 22, 54, 55, 23, 56, 57, 24, 56, 57, 25, 58, 59,
+     26, 58, 59, 27, 60, 61, 28, 60, 61, 29, 62, 63, 30, 62, 63, 31}};
+
+void NV21ToYUV24Row_AVX512(const uint8_t* src_y,
+                           const uint8_t* src_vu,
+                           uint8_t* dst_yuv24,
+                           int width) {
+  asm volatile(
+      "sub         %0,%1                         \n"
+      "vmovdqa     (%4),%%ymm4                   \n"  // 3 shuffler constants
+      "vmovdqa     32(%4),%%ymm5                 \n"
+      "vmovdqa     64(%4),%%ymm6                 \n" LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm2                   \n"  // load 32 Y values
+      "vmovdqu     (%0,%1),%%ymm3                \n"  // load 16 VU values
+      "lea         32(%0),%0                     \n"
+      "vmovdqa     %%ymm2, %%ymm0                \n"
+      "vmovdqa     %%ymm2, %%ymm1                \n"
+      "vpermt2b    %%ymm3,%%ymm4,%%ymm0          \n"
+      "vpermt2b    %%ymm3,%%ymm5,%%ymm1          \n"
+      "vpermt2b    %%ymm3,%%ymm6,%%ymm2          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"
+      "vmovdqu     %%ymm1,32(%2)                 \n"
+      "vmovdqu     %%ymm2,64(%2)                 \n"
+      "lea         96(%2),%2                     \n"
+      "sub         $32,%3                        \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_y),         // %0
+        "+r"(src_vu),        // %1
+        "+r"(dst_yuv24),     // %2
+        "+r"(width)          // %3
+      : "r"(&kYUV24Perm[0])  // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+
+#endif  // HAS_NV21ToYUV24ROW_AVX512
+
+#ifdef HAS_SWAPUVROW_SSSE3
+
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleUVToVU = {1u, 0u, 3u,  2u,  5u,  4u,  7u,  6u,
+                                     9u, 8u, 11u, 10u, 13u, 12u, 15u, 14u};
+
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+
+      "movdqu      %3,%%xmm5                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_uv),        // %0
+        "+r"(dst_vu),        // %1
+        "+r"(width)          // %2
+      : "m"(kShuffleUVToVU)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
+}
+#endif  // HAS_SWAPUVROW_SSSE3
+
+#ifdef HAS_SWAPUVROW_AVX2
+void SwapUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+
+      "vbroadcastf128 %3,%%ymm5                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpshufb     %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm5,%%ymm1,%%ymm1          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_uv),        // %0
+        "+r"(dst_vu),        // %1
+        "+r"(width)          // %2
+      : "m"(kShuffleUVToVU)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm5");
+}
+#endif  // HAS_SWAPUVROW_AVX2
+
+void HalfMergeUVRow_SSSE3(const uint8_t* src_u,
+                          int src_stride_u,
+                          const uint8_t* src_v,
+                          int src_stride_v,
+                          uint8_t* dst_uv,
+                          int width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // load 16 U values
+      "movdqu      (%1),%%xmm1                   \n"  // load 16 V values
+      "movdqu      0(%0,%4,1),%%xmm2             \n"  // 16 from next row
+      "movdqu      0(%1,%5,1),%%xmm3             \n"
+      "lea         0x10(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"  // half size
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "lea         0x10(%1),%1                   \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x1,%%xmm0                   \n"
+      "psrlw       $0x1,%%xmm1                   \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "punpcklbw   %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"  // store 8 UV pixels
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x10,%3                      \n"  // 16 src pixels per loop
+      "jg          1b                            \n"
+      : "+r"(src_u),                    // %0
+        "+r"(src_v),                    // %1
+        "+r"(dst_uv),                   // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride_u)),  // %4
+        "r"((intptr_t)(src_stride_v))   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+
+void HalfMergeUVRow_AVX2(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // load 32 U values
+      "vmovdqu     (%1),%%ymm1                   \n"  // load 32 V values
+      "vmovdqu     0(%0,%4,1),%%ymm2             \n"  // 32 from next row
+      "vmovdqu     0(%1,%5,1),%%ymm3             \n"
+      "lea         0x20(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"  // half size
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "lea         0x20(%1),%1                   \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x1,%%ymm1,%%ymm1            \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpackuswb   %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%2)                   \n"  // store 16 UV pixels
+      "lea         0x20(%2),%2                   \n"
+      "sub         $0x20,%3                      \n"  // 32 src pixels per loop
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_u),                    // %0
+        "+r"(src_v),                    // %1
+        "+r"(dst_uv),                   // %2
+        "+r"(width)                     // %3
+      : "r"((intptr_t)(src_stride_u)),  // %4
+        "r"((intptr_t)(src_stride_v))   // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+
+void ClampFloatToZero_SSE2(const float* src_x, float* dst_y, int width) {
+  asm volatile(
+      "pxor        %%xmm1,%%xmm1                 \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movd        (%0),%%xmm0                   \n"  // load float
+      "maxss       %%xmm1, %%xmm0                \n"  // clamp to zero
+      "add         4, %0                         \n"
+      "movd        %%xmm0, (%1)                  \n"  // store float
+      "add         4, %1                         \n"
+      "sub         $0x4,%2                       \n"  // 1 float per loop
+      "jg          1b                            \n"
+      : "+r"(src_x),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1");
+}
 
 #endif  // defined(__x86_64__) || defined(__i386__)
 
diff --git a/files/source/row_lasx.cc b/files/source/row_lasx.cc
new file mode 100644
index 00000000..7dd18f40
--- /dev/null
+++ b/files/source/row_lasx.cc
@@ -0,0 +1,2230 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ALPHA_VAL (-1)
+
+// Fill YUV -> RGB conversion constants into vectors
+#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
+  {                                                      \
+    ub = __lasx_xvreplgr2vr_h(yuvconst->kUVToB[0]);      \
+    vr = __lasx_xvreplgr2vr_h(yuvconst->kUVToR[1]);      \
+    ug = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[0]);      \
+    vg = __lasx_xvreplgr2vr_h(yuvconst->kUVToG[1]);      \
+    yg = __lasx_xvreplgr2vr_h(yuvconst->kYToRgb[0]);     \
+    yb = __lasx_xvreplgr2vr_w(yuvconst->kYBiasToRgb[0]); \
+  }
+
+// Load 32 YUV422 pixel data
+#define READYUV422_D(psrc_y, psrc_u, psrc_v, out_y, uv_l, uv_h) \
+  {                                                             \
+    __m256i temp0, temp1;                                       \
+                                                                \
+    DUP2_ARG2(__lasx_xvld, psrc_y, 0, psrc_u, 0, out_y, temp0); \
+    temp1 = __lasx_xvld(psrc_v, 0);                             \
+    temp0 = __lasx_xvsub_b(temp0, const_0x80);                  \
+    temp1 = __lasx_xvsub_b(temp1, const_0x80);                  \
+    temp0 = __lasx_vext2xv_h_b(temp0);                          \
+    temp1 = __lasx_vext2xv_h_b(temp1);                          \
+    uv_l = __lasx_xvilvl_h(temp0, temp1);                       \
+    uv_h = __lasx_xvilvh_h(temp0, temp1);                       \
+  }
+
+// Load 16 YUV422 pixel data
+#define READYUV422(psrc_y, psrc_u, psrc_v, out_y, uv) \
+  {                                                   \
+    __m256i temp0, temp1;                             \
+                                                      \
+    out_y = __lasx_xvld(psrc_y, 0);                   \
+    temp0 = __lasx_xvldrepl_d(psrc_u, 0);             \
+    temp1 = __lasx_xvldrepl_d(psrc_v, 0);             \
+    uv = __lasx_xvilvl_b(temp0, temp1);               \
+    uv = __lasx_xvsub_b(uv, const_0x80);              \
+    uv = __lasx_vext2xv_h_b(uv);                      \
+  }
+
+// Convert 16 pixels of YUV420 to RGB.
+#define YUVTORGB_D(in_y, in_uvl, in_uvh, ubvr, ugvg, yg, yb, b_l, b_h, g_l,   \
+                   g_h, r_l, r_h)                                             \
+  {                                                                           \
+    __m256i u_l, u_h, v_l, v_h;                                               \
+    __m256i yl_ev, yl_od, yh_ev, yh_od;                                       \
+    __m256i temp0, temp1, temp2, temp3;                                       \
+                                                                              \
+    temp0 = __lasx_xvilvl_b(in_y, in_y);                                      \
+    temp1 = __lasx_xvilvh_b(in_y, in_y);                                      \
+    yl_ev = __lasx_xvmulwev_w_hu_h(temp0, yg);                                \
+    yl_od = __lasx_xvmulwod_w_hu_h(temp0, yg);                                \
+    yh_ev = __lasx_xvmulwev_w_hu_h(temp1, yg);                                \
+    yh_od = __lasx_xvmulwod_w_hu_h(temp1, yg);                                \
+    DUP4_ARG2(__lasx_xvsrai_w, yl_ev, 16, yl_od, 16, yh_ev, 16, yh_od, 16,    \
+              yl_ev, yl_od, yh_ev, yh_od);                                    \
+    yl_ev = __lasx_xvadd_w(yl_ev, yb);                                        \
+    yl_od = __lasx_xvadd_w(yl_od, yb);                                        \
+    yh_ev = __lasx_xvadd_w(yh_ev, yb);                                        \
+    yh_od = __lasx_xvadd_w(yh_od, yb);                                        \
+    v_l = __lasx_xvmulwev_w_h(in_uvl, ubvr);                                  \
+    u_l = __lasx_xvmulwod_w_h(in_uvl, ubvr);                                  \
+    v_h = __lasx_xvmulwev_w_h(in_uvh, ubvr);                                  \
+    u_h = __lasx_xvmulwod_w_h(in_uvh, ubvr);                                  \
+    temp0 = __lasx_xvadd_w(yl_ev, u_l);                                       \
+    temp1 = __lasx_xvadd_w(yl_od, u_l);                                       \
+    temp2 = __lasx_xvadd_w(yh_ev, u_h);                                       \
+    temp3 = __lasx_xvadd_w(yh_od, u_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    b_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    b_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+    temp0 = __lasx_xvadd_w(yl_ev, v_l);                                       \
+    temp1 = __lasx_xvadd_w(yl_od, v_l);                                       \
+    temp2 = __lasx_xvadd_w(yh_ev, v_h);                                       \
+    temp3 = __lasx_xvadd_w(yh_od, v_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    r_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    r_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+    DUP2_ARG2(__lasx_xvdp2_w_h, in_uvl, ugvg, in_uvh, ugvg, u_l, u_h);        \
+    temp0 = __lasx_xvsub_w(yl_ev, u_l);                                       \
+    temp1 = __lasx_xvsub_w(yl_od, u_l);                                       \
+    temp2 = __lasx_xvsub_w(yh_ev, u_h);                                       \
+    temp3 = __lasx_xvsub_w(yh_od, u_h);                                       \
+    DUP4_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp2, 6, temp3, 6, temp0, \
+              temp1, temp2, temp3);                                           \
+    DUP4_ARG1(__lasx_xvclip255_w, temp0, temp1, temp2, temp3, temp0, temp1,   \
+              temp2, temp3);                                                  \
+    g_l = __lasx_xvpackev_h(temp1, temp0);                                    \
+    g_h = __lasx_xvpackev_h(temp3, temp2);                                    \
+  }
+
+// Convert 8 pixels of YUV420 to RGB.
+#define YUVTORGB(in_y, in_uv, ubvr, ugvg, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    __m256i u_l, v_l, yl_ev, yl_od;                                    \
+    __m256i temp0, temp1;                                              \
+                                                                       \
+    in_y = __lasx_xvpermi_d(in_y, 0xD8);                               \
+    temp0 = __lasx_xvilvl_b(in_y, in_y);                               \
+    yl_ev = __lasx_xvmulwev_w_hu_h(temp0, yg);                         \
+    yl_od = __lasx_xvmulwod_w_hu_h(temp0, yg);                         \
+    DUP2_ARG2(__lasx_xvsrai_w, yl_ev, 16, yl_od, 16, yl_ev, yl_od);    \
+    yl_ev = __lasx_xvadd_w(yl_ev, yb);                                 \
+    yl_od = __lasx_xvadd_w(yl_od, yb);                                 \
+    v_l = __lasx_xvmulwev_w_h(in_uv, ubvr);                            \
+    u_l = __lasx_xvmulwod_w_h(in_uv, ubvr);                            \
+    temp0 = __lasx_xvadd_w(yl_ev, u_l);                                \
+    temp1 = __lasx_xvadd_w(yl_od, u_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_b = __lasx_xvpackev_h(temp1, temp0);                           \
+    temp0 = __lasx_xvadd_w(yl_ev, v_l);                                \
+    temp1 = __lasx_xvadd_w(yl_od, v_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_r = __lasx_xvpackev_h(temp1, temp0);                           \
+    u_l = __lasx_xvdp2_w_h(in_uv, ugvg);                               \
+    temp0 = __lasx_xvsub_w(yl_ev, u_l);                                \
+    temp1 = __lasx_xvsub_w(yl_od, u_l);                                \
+    DUP2_ARG2(__lasx_xvsrai_w, temp0, 6, temp1, 6, temp0, temp1);      \
+    DUP2_ARG1(__lasx_xvclip255_w, temp0, temp1, temp0, temp1);         \
+    out_g = __lasx_xvpackev_h(temp1, temp0);                           \
+  }
+
+// Pack and Store 16 ARGB values.
+#define STOREARGB_D(a_l, a_h, r_l, r_h, g_l, g_h, b_l, b_h, pdst_argb) \
+  {                                                                    \
+    __m256i temp0, temp1, temp2, temp3;                                \
+                                                                       \
+    temp0 = __lasx_xvpackev_b(g_l, b_l);                               \
+    temp1 = __lasx_xvpackev_b(a_l, r_l);                               \
+    temp2 = __lasx_xvpackev_b(g_h, b_h);                               \
+    temp3 = __lasx_xvpackev_b(a_h, r_h);                               \
+    r_l = __lasx_xvilvl_h(temp1, temp0);                               \
+    r_h = __lasx_xvilvh_h(temp1, temp0);                               \
+    g_l = __lasx_xvilvl_h(temp3, temp2);                               \
+    g_h = __lasx_xvilvh_h(temp3, temp2);                               \
+    temp0 = __lasx_xvpermi_q(r_h, r_l, 0x20);                          \
+    temp1 = __lasx_xvpermi_q(g_h, g_l, 0x20);                          \
+    temp2 = __lasx_xvpermi_q(r_h, r_l, 0x31);                          \
+    temp3 = __lasx_xvpermi_q(g_h, g_l, 0x31);                          \
+    __lasx_xvst(temp0, pdst_argb, 0);                                  \
+    __lasx_xvst(temp1, pdst_argb, 32);                                 \
+    __lasx_xvst(temp2, pdst_argb, 64);                                 \
+    __lasx_xvst(temp3, pdst_argb, 96);                                 \
+    pdst_argb += 128;                                                  \
+  }
+
+// Pack and Store 8 ARGB values.
+#define STOREARGB(in_a, in_r, in_g, in_b, pdst_argb) \
+  {                                                  \
+    __m256i temp0, temp1, temp2, temp3;              \
+                                                     \
+    temp0 = __lasx_xvpackev_b(in_g, in_b);           \
+    temp1 = __lasx_xvpackev_b(in_a, in_r);           \
+    temp2 = __lasx_xvilvl_h(temp1, temp0);           \
+    temp3 = __lasx_xvilvh_h(temp1, temp0);           \
+    temp0 = __lasx_xvpermi_q(temp3, temp2, 0x20);    \
+    temp1 = __lasx_xvpermi_q(temp3, temp2, 0x31);    \
+    __lasx_xvst(temp0, pdst_argb, 0);                \
+    __lasx_xvst(temp1, pdst_argb, 32);               \
+    pdst_argb += 64;                                 \
+  }
+
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _reg0, _reg1) \
+  {                                                                     \
+    __m256i _tmp0, _tmp1, _tmp2, _tmp3;                                 \
+    _tmp0 = __lasx_xvaddwev_h_bu(_tmpb, _nexb);                         \
+    _tmp1 = __lasx_xvaddwod_h_bu(_tmpb, _nexb);                         \
+    _tmp2 = __lasx_xvaddwev_h_bu(_tmpg, _nexg);                         \
+    _tmp3 = __lasx_xvaddwod_h_bu(_tmpg, _nexg);                         \
+    _reg0 = __lasx_xvaddwev_h_bu(_tmpr, _nexr);                         \
+    _reg1 = __lasx_xvaddwod_h_bu(_tmpr, _nexr);                         \
+    _tmpb = __lasx_xvavgr_hu(_tmp0, _tmp1);                             \
+    _tmpg = __lasx_xvavgr_hu(_tmp2, _tmp3);                             \
+    _tmpr = __lasx_xvavgr_hu(_reg0, _reg1);                             \
+    _reg0 = __lasx_xvmadd_h(const_8080, const_112, _tmpb);              \
+    _reg1 = __lasx_xvmadd_h(const_8080, const_112, _tmpr);              \
+    _reg0 = __lasx_xvmsub_h(_reg0, const_74, _tmpg);                    \
+    _reg1 = __lasx_xvmsub_h(_reg1, const_94, _tmpg);                    \
+    _reg0 = __lasx_xvmsub_h(_reg0, const_38, _tmpr);                    \
+    _reg1 = __lasx_xvmsub_h(_reg1, const_18, _tmpb);                    \
+  }
+
+void MirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  int len = width / 64;
+  __m256i src0, src1;
+  __m256i shuffler = {0x08090A0B0C0D0E0F, 0x0001020304050607,
+                      0x08090A0B0C0D0E0F, 0x0001020304050607};
+  src += width - 64;
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src, 0, src, 32, src0, src1);
+    DUP2_ARG3(__lasx_xvshuf_b, src0, src0, shuffler, src1, src1, shuffler, src0,
+              src1);
+    src0 = __lasx_xvpermi_q(src0, src0, 0x01);
+    src1 = __lasx_xvpermi_q(src1, src1, 0x01);
+    __lasx_xvst(src1, dst, 0);
+    __lasx_xvst(src0, dst, 32);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void MirrorUVRow_LASX(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src, dst;
+  __m256i shuffler = {0x0004000500060007, 0x0000000100020003,
+                      0x0004000500060007, 0x0000000100020003};
+
+  src_uv += (width - 16) << 1;
+  for (x = 0; x < len; x++) {
+    src = __lasx_xvld(src_uv, 0);
+    dst = __lasx_xvshuf_h(shuffler, src, src);
+    dst = __lasx_xvpermi_q(dst, dst, 0x01);
+    __lasx_xvst(dst, dst_uv, 0);
+    src_uv -= 32;
+    dst_uv += 32;
+  }
+}
+
+void ARGBMirrorRow_LASX(const uint8_t* src, uint8_t* dst, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1;
+  __m256i dst0, dst1;
+  __m256i shuffler = {0x0B0A09080F0E0D0C, 0x0302010007060504,
+                      0x0B0A09080F0E0D0C, 0x0302010007060504};
+  src += (width * 4) - 64;
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src, 0, src, 32, src0, src1);
+    DUP2_ARG3(__lasx_xvshuf_b, src0, src0, shuffler, src1, src1, shuffler, src0,
+              src1);
+    dst1 = __lasx_xvpermi_q(src0, src0, 0x01);
+    dst0 = __lasx_xvpermi_q(src1, src1, 0x01);
+    __lasx_xvst(dst0, dst, 0);
+    __lasx_xvst(dst1, dst, 32);
+    dst += 64;
+    src -= 64;
+  }
+}
+
+void I422ToYUY2Row_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_yuy2,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i src_u0, src_v0, src_y0, vec_uv0;
+  __m256i vec_yuy2_0, vec_yuy2_1;
+  __m256i dst_yuy2_0, dst_yuy2_1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_u, 0, src_v, 0, src_u0, src_v0);
+    src_y0 = __lasx_xvld(src_y, 0);
+    src_u0 = __lasx_xvpermi_d(src_u0, 0xD8);
+    src_v0 = __lasx_xvpermi_d(src_v0, 0xD8);
+    vec_uv0 = __lasx_xvilvl_b(src_v0, src_u0);
+    vec_yuy2_0 = __lasx_xvilvl_b(vec_uv0, src_y0);
+    vec_yuy2_1 = __lasx_xvilvh_b(vec_uv0, src_y0);
+    dst_yuy2_0 = __lasx_xvpermi_q(vec_yuy2_1, vec_yuy2_0, 0x20);
+    dst_yuy2_1 = __lasx_xvpermi_q(vec_yuy2_1, vec_yuy2_0, 0x31);
+    __lasx_xvst(dst_yuy2_0, dst_yuy2, 0);
+    __lasx_xvst(dst_yuy2_1, dst_yuy2, 32);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_yuy2 += 64;
+  }
+}
+
+void I422ToUYVYRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_uyvy,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i src_u0, src_v0, src_y0, vec_uv0;
+  __m256i vec_uyvy0, vec_uyvy1;
+  __m256i dst_uyvy0, dst_uyvy1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_u, 0, src_v, 0, src_u0, src_v0);
+    src_y0 = __lasx_xvld(src_y, 0);
+    src_u0 = __lasx_xvpermi_d(src_u0, 0xD8);
+    src_v0 = __lasx_xvpermi_d(src_v0, 0xD8);
+    vec_uv0 = __lasx_xvilvl_b(src_v0, src_u0);
+    vec_uyvy0 = __lasx_xvilvl_b(src_y0, vec_uv0);
+    vec_uyvy1 = __lasx_xvilvh_b(src_y0, vec_uv0);
+    dst_uyvy0 = __lasx_xvpermi_q(vec_uyvy1, vec_uyvy0, 0x20);
+    dst_uyvy1 = __lasx_xvpermi_q(vec_uyvy1, vec_uyvy0, 0x31);
+    __lasx_xvst(dst_uyvy0, dst_uyvy, 0);
+    __lasx_xvst(dst_uyvy1, dst_uyvy, 32);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_uyvy += 64;
+  }
+}
+
+void I422ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_ug, vec_vr, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(alpha, alpha, r_l, r_h, g_l, g_h, b_l, b_h, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422ToRGBARow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(r_l, r_h, g_l, g_h, b_l, b_h, alpha, alpha, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422AlphaToARGBRow_LASX(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  int x;
+  int len = width / 32;
+  int res = width & 31;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h, a_l, a_h;
+
+    y = __lasx_xvld(src_a, 0);
+    a_l = __lasx_xvilvl_b(zero, y);
+    a_h = __lasx_xvilvh_b(zero, y);
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    STOREARGB_D(a_l, a_h, r_l, r_h, g_l, g_h, b_l, b_h, dst_argb);
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+    src_a += 32;
+  }
+  if (res) {
+    __m256i y, uv, r, g, b, a;
+    a = __lasx_xvld(src_a, 0);
+    a = __lasx_vext2xv_hu_bu(a);
+    READYUV422(src_y, src_u, src_v, y, uv);
+    YUVTORGB(y, uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b, g, r);
+    STOREARGB(a, r, g, b, dst_argb);
+  }
+}
+
+void I422ToRGB24Row_LASX(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int32_t width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i shuffler0 = {0x0504120302100100, 0x0A18090816070614,
+                       0x0504120302100100, 0x0A18090816070614};
+  __m256i shuffler1 = {0x1E0F0E1C0D0C1A0B, 0x1E0F0E1C0D0C1A0B,
+                       0x1E0F0E1C0D0C1A0B, 0x1E0F0E1C0D0C1A0B};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i temp0, temp1, temp2, temp3;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    temp0 = __lasx_xvpackev_b(g_l, b_l);
+    temp1 = __lasx_xvpackev_b(g_h, b_h);
+    DUP4_ARG3(__lasx_xvshuf_b, r_l, temp0, shuffler1, r_h, temp1, shuffler1,
+              r_l, temp0, shuffler0, r_h, temp1, shuffler0, temp2, temp3, temp0,
+              temp1);
+
+    b_l = __lasx_xvilvl_d(temp1, temp2);
+    b_h = __lasx_xvilvh_d(temp3, temp1);
+    temp1 = __lasx_xvpermi_q(b_l, temp0, 0x20);
+    temp2 = __lasx_xvpermi_q(temp0, b_h, 0x30);
+    temp3 = __lasx_xvpermi_q(b_h, b_l, 0x31);
+    __lasx_xvst(temp1, dst_argb, 0);
+    __lasx_xvst(temp2, dst_argb, 32);
+    __lasx_xvst(temp3, dst_argb, 64);
+    dst_argb += 96;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 5 upper bits of R.
+void I422ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_u,
+                          const uint8_t* src_v,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 3);
+    b_h = __lasx_xvsrli_h(b_h, 3);
+    g_l = __lasx_xvsrli_h(g_l, 2);
+    g_h = __lasx_xvsrli_h(g_h, 2);
+    r_l = __lasx_xvsrli_h(r_l, 3);
+    r_h = __lasx_xvsrli_h(r_h, 3);
+    r_l = __lasx_xvslli_h(r_l, 11);
+    r_h = __lasx_xvslli_h(r_h, 11);
+    g_l = __lasx_xvslli_h(g_l, 5);
+    g_h = __lasx_xvslli_h(g_h, 5);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_rgb565, 0);
+    __lasx_xvst(dst_h, dst_rgb565, 32);
+    dst_rgb565 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+// TODO(fbarchard): Consider AND instead of shift to isolate 4 upper bits of G.
+void I422ToARGB4444Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb4444,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = {0xF000F000F000F000, 0xF000F000F000F000, 0xF000F000F000F000,
+                   0xF000F000F000F000};
+  __m256i mask = {0x00F000F000F000F0, 0x00F000F000F000F0, 0x00F000F000F000F0,
+                  0x00F000F000F000F0};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 4);
+    b_h = __lasx_xvsrli_h(b_h, 4);
+    r_l = __lasx_xvsrli_h(r_l, 4);
+    r_h = __lasx_xvsrli_h(r_h, 4);
+    g_l = __lasx_xvand_v(g_l, mask);
+    g_h = __lasx_xvand_v(g_h, mask);
+    r_l = __lasx_xvslli_h(r_l, 8);
+    r_h = __lasx_xvslli_h(r_h, 8);
+    r_l = __lasx_xvor_v(r_l, alpha);
+    r_h = __lasx_xvor_v(r_h, alpha);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_argb4444, 0);
+    __lasx_xvst(dst_h, dst_argb4444, 32);
+    dst_argb4444 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I422ToARGB1555Row_LASX(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            uint8_t* dst_argb1555,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i vec_yb, vec_yg, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = {0x8000800080008000, 0x8000800080008000, 0x8000800080008000,
+                   0x8000800080008000};
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    __m256i y, uv_l, uv_h, b_l, b_h, g_l, g_h, r_l, r_h;
+    __m256i dst_l, dst_h;
+
+    READYUV422_D(src_y, src_u, src_v, y, uv_l, uv_h);
+    YUVTORGB_D(y, uv_l, uv_h, vec_ubvr, vec_ugvg, vec_yg, vec_yb, b_l, b_h, g_l,
+               g_h, r_l, r_h);
+    b_l = __lasx_xvsrli_h(b_l, 3);
+    b_h = __lasx_xvsrli_h(b_h, 3);
+    g_l = __lasx_xvsrli_h(g_l, 3);
+    g_h = __lasx_xvsrli_h(g_h, 3);
+    g_l = __lasx_xvslli_h(g_l, 5);
+    g_h = __lasx_xvslli_h(g_h, 5);
+    r_l = __lasx_xvsrli_h(r_l, 3);
+    r_h = __lasx_xvsrli_h(r_h, 3);
+    r_l = __lasx_xvslli_h(r_l, 10);
+    r_h = __lasx_xvslli_h(r_h, 10);
+    r_l = __lasx_xvor_v(r_l, alpha);
+    r_h = __lasx_xvor_v(r_h, alpha);
+    r_l = __lasx_xvor_v(r_l, g_l);
+    r_h = __lasx_xvor_v(r_h, g_h);
+    r_l = __lasx_xvor_v(r_l, b_l);
+    r_h = __lasx_xvor_v(r_h, b_h);
+    dst_l = __lasx_xvpermi_q(r_h, r_l, 0x20);
+    dst_h = __lasx_xvpermi_q(r_h, r_l, 0x31);
+    __lasx_xvst(dst_l, dst_argb1555, 0);
+    __lasx_xvst(dst_h, dst_argb1555, 32);
+    dst_argb1555 += 64;
+    src_y += 32;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void YUY2ToYRow_LASX(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src0, src1);
+    dst0 = __lasx_xvpickev_b(src1, src0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_yuy2 += 64;
+    dst_y += 32;
+  }
+}
+
+void YUY2ToUVRow_LASX(const uint8_t* src_yuy2,
+                      int src_stride_yuy2,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_yuy2_next = src_yuy2 + src_stride_yuy2;
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src_yuy2_next, 0,
+              src_yuy2_next, 32, src0, src1, src2, src3);
+    src0 = __lasx_xvpickod_b(src1, src0);
+    src1 = __lasx_xvpickod_b(src3, src2);
+    tmp0 = __lasx_xvavgr_bu(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_yuy2 += 64;
+    src_yuy2_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void YUY2ToUV422Row_LASX(const uint8_t* src_yuy2,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_yuy2, 0, src_yuy2, 32, src0, src1);
+    tmp0 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_yuy2 += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToYRow_LASX(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src0, src1);
+    dst0 = __lasx_xvpickod_b(src1, src0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_uyvy += 64;
+    dst_y += 32;
+  }
+}
+
+void UYVYToUVRow_LASX(const uint8_t* src_uyvy,
+                      int src_stride_uyvy,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_uyvy_next = src_uyvy + src_stride_uyvy;
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src_uyvy_next, 0,
+              src_uyvy_next, 32, src0, src1, src2, src3);
+    src0 = __lasx_xvpickev_b(src1, src0);
+    src1 = __lasx_xvpickev_b(src3, src2);
+    tmp0 = __lasx_xvavgr_bu(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_uyvy += 64;
+    src_uyvy_next += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void UYVYToUV422Row_LASX(const uint8_t* src_uyvy,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, tmp0, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_uyvy, 0, src_uyvy, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp0 = __lasx_xvpermi_d(tmp0, 0xD8);
+    dst0 = __lasx_xvpickev_b(tmp0, tmp0);
+    dst1 = __lasx_xvpickod_b(tmp0, tmp0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_uyvy += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToYRow_LASX(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3, vec0, vec1, vec2, vec3;
+  __m256i tmp0, tmp1, dst0;
+  __m256i const_19 = __lasx_xvldi(0x19);
+  __m256i const_42 = __lasx_xvldi(0x42);
+  __m256i const_81 = __lasx_xvldi(0x81);
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+  __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                     0x0000000700000003};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb0, 0, src_argb0, 32, src_argb0, 64,
+              src_argb0, 96, src0, src1, src2, src3);
+    vec0 = __lasx_xvpickev_b(src1, src0);
+    vec1 = __lasx_xvpickev_b(src3, src2);
+    vec2 = __lasx_xvpickod_b(src1, src0);
+    vec3 = __lasx_xvpickod_b(src3, src2);
+    tmp0 = __lasx_xvmaddwev_h_bu(const_1080, vec0, const_19);
+    tmp1 = __lasx_xvmaddwev_h_bu(const_1080, vec1, const_19);
+    tmp0 = __lasx_xvmaddwev_h_bu(tmp0, vec2, const_81);
+    tmp1 = __lasx_xvmaddwev_h_bu(tmp1, vec3, const_81);
+    tmp0 = __lasx_xvmaddwod_h_bu(tmp0, vec0, const_42);
+    tmp1 = __lasx_xvmaddwod_h_bu(tmp1, vec1, const_42);
+    dst0 = __lasx_xvssrani_b_h(tmp1, tmp0, 8);
+    dst0 = __lasx_xvperm_w(dst0, control);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_argb0 += 128;
+    dst_y += 32;
+  }
+}
+
+void ARGBToUVRow_LASX(const uint8_t* src_argb0,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* src_argb1 = src_argb0 + src_stride_argb;
+
+  __m256i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m256i vec0, vec1, vec2, vec3;
+  __m256i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, dst0, dst1;
+  __m256i const_0x70 = {0x0038003800380038, 0x0038003800380038,
+                        0x0038003800380038, 0x0038003800380038};
+  __m256i const_0x4A = {0x0025002500250025, 0x0025002500250025,
+                        0x0025002500250025, 0x0025002500250025};
+  __m256i const_0x26 = {0x0013001300130013, 0x0013001300130013,
+                        0x0013001300130013, 0x0013001300130013};
+  __m256i const_0x5E = {0x002f002f002f002f, 0x002f002f002f002f,
+                        0x002f002f002f002f, 0x002f002f002f002f};
+  __m256i const_0x12 = {0x0009000900090009, 0x0009000900090009,
+                        0x0009000900090009, 0x0009000900090009};
+  __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                     0x0000000700000003};
+  __m256i const_0x8080 = {0x8080808080808080, 0x8080808080808080,
+                          0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb0, 0, src_argb0, 32, src_argb0, 64,
+              src_argb0, 96, src0, src1, src2, src3);
+    DUP4_ARG2(__lasx_xvld, src_argb1, 0, src_argb1, 32, src_argb1, 64,
+              src_argb1, 96, src4, src5, src6, src7);
+    vec0 = __lasx_xvaddwev_h_bu(src0, src4);
+    vec1 = __lasx_xvaddwev_h_bu(src1, src5);
+    vec2 = __lasx_xvaddwev_h_bu(src2, src6);
+    vec3 = __lasx_xvaddwev_h_bu(src3, src7);
+    tmp0 = __lasx_xvpickev_h(vec1, vec0);
+    tmp1 = __lasx_xvpickev_h(vec3, vec2);
+    tmp2 = __lasx_xvpickod_h(vec1, vec0);
+    tmp3 = __lasx_xvpickod_h(vec3, vec2);
+    vec0 = __lasx_xvaddwod_h_bu(src0, src4);
+    vec1 = __lasx_xvaddwod_h_bu(src1, src5);
+    vec2 = __lasx_xvaddwod_h_bu(src2, src6);
+    vec3 = __lasx_xvaddwod_h_bu(src3, src7);
+    tmp4 = __lasx_xvpickev_h(vec1, vec0);
+    tmp5 = __lasx_xvpickev_h(vec3, vec2);
+    vec0 = __lasx_xvpickev_h(tmp1, tmp0);
+    vec1 = __lasx_xvpickod_h(tmp1, tmp0);
+    src0 = __lasx_xvavgr_h(vec0, vec1);
+    vec0 = __lasx_xvpickev_h(tmp3, tmp2);
+    vec1 = __lasx_xvpickod_h(tmp3, tmp2);
+    src1 = __lasx_xvavgr_h(vec0, vec1);
+    vec0 = __lasx_xvpickev_h(tmp5, tmp4);
+    vec1 = __lasx_xvpickod_h(tmp5, tmp4);
+    src2 = __lasx_xvavgr_h(vec0, vec1);
+    dst0 = __lasx_xvmadd_h(const_0x8080, src0, const_0x70);
+    dst0 = __lasx_xvmsub_h(dst0, src2, const_0x4A);
+    dst0 = __lasx_xvmsub_h(dst0, src1, const_0x26);
+    dst1 = __lasx_xvmadd_h(const_0x8080, src1, const_0x70);
+    dst1 = __lasx_xvmsub_h(dst1, src2, const_0x5E);
+    dst1 = __lasx_xvmsub_h(dst1, src0, const_0x12);
+    dst0 = __lasx_xvperm_w(dst0, control);
+    dst1 = __lasx_xvperm_w(dst1, control);
+    dst0 = __lasx_xvssrani_b_h(dst0, dst0, 8);
+    dst1 = __lasx_xvssrani_b_h(dst1, dst1, 8);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst1, dst_v, 0, 0);
+    __lasx_xvstelm_d(dst1, dst_v, 8, 2);
+    src_argb0 += 128;
+    src_argb1 += 128;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void ARGBToRGB24Row_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  int len = (width / 32) - 1;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i shuf = {0x0908060504020100, 0x000000000E0D0C0A, 0x0908060504020100,
+                  0x000000000E0D0C0A};
+  __m256i control = {0x0000000100000000, 0x0000000400000002, 0x0000000600000005,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+    tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+    tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+    tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+    tmp0 = __lasx_xvperm_w(tmp0, control);
+    tmp1 = __lasx_xvperm_w(tmp1, control);
+    tmp2 = __lasx_xvperm_w(tmp2, control);
+    tmp3 = __lasx_xvperm_w(tmp3, control);
+    __lasx_xvst(tmp0, dst_rgb, 0);
+    __lasx_xvst(tmp1, dst_rgb, 24);
+    __lasx_xvst(tmp2, dst_rgb, 48);
+    __lasx_xvst(tmp3, dst_rgb, 72);
+    dst_rgb += 96;
+    src_argb += 128;
+  }
+  DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb, 96,
+            src0, src1, src2, src3);
+  tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+  tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+  tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+  tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+  tmp0 = __lasx_xvperm_w(tmp0, control);
+  tmp1 = __lasx_xvperm_w(tmp1, control);
+  tmp2 = __lasx_xvperm_w(tmp2, control);
+  tmp3 = __lasx_xvperm_w(tmp3, control);
+  __lasx_xvst(tmp0, dst_rgb, 0);
+  __lasx_xvst(tmp1, dst_rgb, 24);
+  __lasx_xvst(tmp2, dst_rgb, 48);
+  dst_rgb += 72;
+  __lasx_xvstelm_d(tmp3, dst_rgb, 0, 0);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 8, 1);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 16, 2);
+}
+
+void ARGBToRAWRow_LASX(const uint8_t* src_argb, uint8_t* dst_rgb, int width) {
+  int x;
+  int len = (width / 32) - 1;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i shuf = {0x090A040506000102, 0x000000000C0D0E08, 0x090A040506000102,
+                  0x000000000C0D0E08};
+  __m256i control = {0x0000000100000000, 0x0000000400000002, 0x0000000600000005,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+    tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+    tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+    tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+    tmp0 = __lasx_xvperm_w(tmp0, control);
+    tmp1 = __lasx_xvperm_w(tmp1, control);
+    tmp2 = __lasx_xvperm_w(tmp2, control);
+    tmp3 = __lasx_xvperm_w(tmp3, control);
+    __lasx_xvst(tmp0, dst_rgb, 0);
+    __lasx_xvst(tmp1, dst_rgb, 24);
+    __lasx_xvst(tmp2, dst_rgb, 48);
+    __lasx_xvst(tmp3, dst_rgb, 72);
+    dst_rgb += 96;
+    src_argb += 128;
+  }
+  DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb, 96,
+            src0, src1, src2, src3);
+  tmp0 = __lasx_xvshuf_b(src0, src0, shuf);
+  tmp1 = __lasx_xvshuf_b(src1, src1, shuf);
+  tmp2 = __lasx_xvshuf_b(src2, src2, shuf);
+  tmp3 = __lasx_xvshuf_b(src3, src3, shuf);
+  tmp0 = __lasx_xvperm_w(tmp0, control);
+  tmp1 = __lasx_xvperm_w(tmp1, control);
+  tmp2 = __lasx_xvperm_w(tmp2, control);
+  tmp3 = __lasx_xvperm_w(tmp3, control);
+  __lasx_xvst(tmp0, dst_rgb, 0);
+  __lasx_xvst(tmp1, dst_rgb, 24);
+  __lasx_xvst(tmp2, dst_rgb, 48);
+  dst_rgb += 72;
+  __lasx_xvstelm_d(tmp3, dst_rgb, 0, 0);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 8, 1);
+  __lasx_xvstelm_d(tmp3, dst_rgb, 16, 2);
+}
+
+void ARGBToRGB565Row_LASX(const uint8_t* src_argb,
+                          uint8_t* dst_rgb,
+                          int width) {
+  int x;
+  int len = width / 16;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, tmp0, tmp1, dst0;
+  __m256i shift = {0x0300030003000300, 0x0300030003000300, 0x0300030003000300,
+                   0x0300030003000300};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 3);
+    tmp1 = __lasx_xvpackev_b(zero, tmp1);
+    tmp1 = __lasx_xvsrli_h(tmp1, 2);
+    tmp0 = __lasx_xvsll_b(tmp0, shift);
+    tmp1 = __lasx_xvslli_h(tmp1, 5);
+    dst0 = __lasx_xvor_v(tmp0, tmp1);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToARGB1555Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width) {
+  int x;
+  int len = width / 16;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, tmp0, tmp1, tmp2, tmp3, dst0;
+  __m256i shift1 = {0x0703070307030703, 0x0703070307030703, 0x0703070307030703,
+                    0x0703070307030703};
+  __m256i shift2 = {0x0200020002000200, 0x0200020002000200, 0x0200020002000200,
+                    0x0200020002000200};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 3);
+    tmp1 = __lasx_xvsrl_b(tmp1, shift1);
+    tmp0 = __lasx_xvsll_b(tmp0, shift2);
+    tmp2 = __lasx_xvpackev_b(zero, tmp1);
+    tmp3 = __lasx_xvpackod_b(zero, tmp1);
+    tmp2 = __lasx_xvslli_h(tmp2, 5);
+    tmp3 = __lasx_xvslli_h(tmp3, 15);
+    dst0 = __lasx_xvor_v(tmp0, tmp2);
+    dst0 = __lasx_xvor_v(dst0, tmp3);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToARGB4444Row_LASX(const uint8_t* src_argb,
+                            uint8_t* dst_rgb,
+                            int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp1 = __lasx_xvandi_b(tmp1, 0xF0);
+    tmp0 = __lasx_xvsrli_b(tmp0, 4);
+    dst0 = __lasx_xvor_v(tmp1, tmp0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    dst_rgb += 32;
+    src_argb += 64;
+  }
+}
+
+void ARGBToUV444Row_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int32_t width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, reg3, dst0, dst1;
+  __m256i const_112 = __lasx_xvldi(112);
+  __m256i const_74 = __lasx_xvldi(74);
+  __m256i const_38 = __lasx_xvldi(38);
+  __m256i const_94 = __lasx_xvldi(94);
+  __m256i const_18 = __lasx_xvldi(18);
+  __m256i const_0x8080 = {0x8080808080808080, 0x8080808080808080,
+                          0x8080808080808080, 0x8080808080808080};
+  __m256i control = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                     0x0000000700000003};
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvpickev_h(src1, src0);
+    tmp1 = __lasx_xvpickod_h(src1, src0);
+    tmp2 = __lasx_xvpickev_h(src3, src2);
+    tmp3 = __lasx_xvpickod_h(src3, src2);
+    reg0 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp0, const_112);
+    reg1 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp2, const_112);
+    reg2 = __lasx_xvmulwod_h_bu(tmp0, const_74);
+    reg3 = __lasx_xvmulwod_h_bu(tmp2, const_74);
+    reg2 = __lasx_xvmaddwev_h_bu(reg2, tmp1, const_38);
+    reg3 = __lasx_xvmaddwev_h_bu(reg3, tmp3, const_38);
+    reg0 = __lasx_xvsub_h(reg0, reg2);
+    reg1 = __lasx_xvsub_h(reg1, reg3);
+    dst0 = __lasx_xvssrani_b_h(reg1, reg0, 8);
+    dst0 = __lasx_xvperm_w(dst0, control);
+    reg0 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp1, const_112);
+    reg1 = __lasx_xvmaddwev_h_bu(const_0x8080, tmp3, const_112);
+    reg2 = __lasx_xvmulwev_h_bu(tmp0, const_18);
+    reg3 = __lasx_xvmulwev_h_bu(tmp2, const_18);
+    reg2 = __lasx_xvmaddwod_h_bu(reg2, tmp0, const_94);
+    reg3 = __lasx_xvmaddwod_h_bu(reg3, tmp2, const_94);
+    reg0 = __lasx_xvsub_h(reg0, reg2);
+    reg1 = __lasx_xvsub_h(reg1, reg3);
+    dst1 = __lasx_xvssrani_b_h(reg1, reg0, 8);
+    dst1 = __lasx_xvperm_w(dst1, control);
+    __lasx_xvst(dst0, dst_u, 0);
+    __lasx_xvst(dst1, dst_v, 0);
+    dst_u += 32;
+    dst_v += 32;
+    src_argb += 128;
+  }
+}
+
+void ARGBMultiplyRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 8;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i src0, src1, dst0, dst1;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    tmp0 = __lasx_xvilvl_b(src0, src0);
+    tmp1 = __lasx_xvilvh_b(src0, src0);
+    tmp2 = __lasx_xvilvl_b(zero, src1);
+    tmp3 = __lasx_xvilvh_b(zero, src1);
+    dst0 = __lasx_xvmuh_hu(tmp0, tmp2);
+    dst1 = __lasx_xvmuh_hu(tmp1, tmp3);
+    dst0 = __lasx_xvpickev_b(dst1, dst0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBAddRow_LASX(const uint8_t* src_argb0,
+                     const uint8_t* src_argb1,
+                     uint8_t* dst_argb,
+                     int width) {
+  int x;
+  int len = width / 8;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    dst0 = __lasx_xvsadd_bu(src0, src1);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBSubtractRow_LASX(const uint8_t* src_argb0,
+                          const uint8_t* src_argb1,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 8;
+  __m256i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb0, 0, src_argb1, 0, src0, src1);
+    dst0 = __lasx_xvssub_bu(src0, src1);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb0 += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBAttenuateRow_LASX(const uint8_t* src_argb,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, reg2, reg3, reg4, reg5;
+  __m256i b, g, r, a, dst0, dst1;
+  __m256i control = {0x0005000100040000, 0x0007000300060002, 0x0005000100040000,
+                     0x0007000300060002};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    b = __lasx_xvpackev_b(tmp0, tmp0);
+    r = __lasx_xvpackod_b(tmp0, tmp0);
+    g = __lasx_xvpackev_b(tmp1, tmp1);
+    a = __lasx_xvpackod_b(tmp1, tmp1);
+    reg0 = __lasx_xvmulwev_w_hu(b, a);
+    reg1 = __lasx_xvmulwod_w_hu(b, a);
+    reg2 = __lasx_xvmulwev_w_hu(r, a);
+    reg3 = __lasx_xvmulwod_w_hu(r, a);
+    reg4 = __lasx_xvmulwev_w_hu(g, a);
+    reg5 = __lasx_xvmulwod_w_hu(g, a);
+    reg0 = __lasx_xvssrani_h_w(reg1, reg0, 24);
+    reg2 = __lasx_xvssrani_h_w(reg3, reg2, 24);
+    reg4 = __lasx_xvssrani_h_w(reg5, reg4, 24);
+    reg0 = __lasx_xvshuf_h(control, reg0, reg0);
+    reg2 = __lasx_xvshuf_h(control, reg2, reg2);
+    reg4 = __lasx_xvshuf_h(control, reg4, reg4);
+    tmp0 = __lasx_xvpackev_b(reg4, reg0);
+    tmp1 = __lasx_xvpackev_b(a, reg2);
+    dst0 = __lasx_xvilvl_h(tmp1, tmp0);
+    dst1 = __lasx_xvilvh_h(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    dst_argb += 64;
+    src_argb += 64;
+  }
+}
+
+void ARGBToRGB565DitherRow_LASX(const uint8_t* src_argb,
+                                uint8_t* dst_rgb,
+                                const uint32_t dither4,
+                                int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1, dst0;
+  __m256i b, g, r;
+  __m256i zero = __lasx_xvldi(0);
+  __m256i vec_dither = __lasx_xvldrepl_w(&dither4, 0);
+
+  vec_dither = __lasx_xvilvl_b(zero, vec_dither);
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    b = __lasx_xvpackev_b(zero, tmp0);
+    r = __lasx_xvpackod_b(zero, tmp0);
+    g = __lasx_xvpackev_b(zero, tmp1);
+    b = __lasx_xvadd_h(b, vec_dither);
+    g = __lasx_xvadd_h(g, vec_dither);
+    r = __lasx_xvadd_h(r, vec_dither);
+    DUP2_ARG1(__lasx_xvclip255_h, b, g, b, g);
+    r = __lasx_xvclip255_h(r);
+    b = __lasx_xvsrai_h(b, 3);
+    g = __lasx_xvsrai_h(g, 2);
+    r = __lasx_xvsrai_h(r, 3);
+    g = __lasx_xvslli_h(g, 5);
+    r = __lasx_xvslli_h(r, 11);
+    dst0 = __lasx_xvor_v(b, g);
+    dst0 = __lasx_xvor_v(dst0, r);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_rgb, 0);
+    src_argb += 64;
+    dst_rgb += 32;
+  }
+}
+
+void ARGBShuffleRow_LASX(const uint8_t* src_argb,
+                         uint8_t* dst_argb,
+                         const uint8_t* shuffler,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, dst0, dst1;
+  __m256i shuf = {0x0404040400000000, 0x0C0C0C0C08080808, 0x0404040400000000,
+                  0x0C0C0C0C08080808};
+  __m256i temp = __lasx_xvldrepl_w(shuffler, 0);
+
+  shuf = __lasx_xvadd_b(shuf, temp);
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    dst0 = __lasx_xvshuf_b(src0, src0, shuf);
+    dst1 = __lasx_xvshuf_b(src1, src1, shuf);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    src_argb += 64;
+    dst_argb += 64;
+  }
+}
+
+void ARGBShadeRow_LASX(const uint8_t* src_argb,
+                       uint8_t* dst_argb,
+                       int width,
+                       uint32_t value) {
+  int x;
+  int len = width / 8;
+  __m256i src0, dst0, tmp0, tmp1;
+  __m256i vec_value = __lasx_xvreplgr2vr_w(value);
+
+  vec_value = __lasx_xvilvl_b(vec_value, vec_value);
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb, 0);
+    tmp0 = __lasx_xvilvl_b(src0, src0);
+    tmp1 = __lasx_xvilvh_b(src0, src0);
+    tmp0 = __lasx_xvmuh_hu(tmp0, vec_value);
+    tmp1 = __lasx_xvmuh_hu(tmp1, vec_value);
+    dst0 = __lasx_xvpickod_b(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBGrayRow_LASX(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, reg2, dst0, dst1;
+  __m256i const_128 = __lasx_xvldi(0x480);
+  __m256i const_150 = __lasx_xvldi(0x96);
+  __m256i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D,
+                      0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    reg0 = __lasx_xvdp2_h_bu(tmp0, const_br);
+    reg1 = __lasx_xvmaddwev_h_bu(const_128, tmp1, const_150);
+    reg2 = __lasx_xvadd_h(reg0, reg1);
+    tmp0 = __lasx_xvpackod_b(reg2, reg2);
+    tmp1 = __lasx_xvpackod_b(tmp1, reg2);
+    dst0 = __lasx_xvilvl_h(tmp1, tmp0);
+    dst1 = __lasx_xvilvh_h(tmp1, tmp0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    src_argb += 64;
+    dst_argb += 64;
+  }
+}
+
+void ARGBSepiaRow_LASX(uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m256i src0, src1, tmp0, tmp1;
+  __m256i reg0, reg1, spb, spg, spr;
+  __m256i dst0, dst1;
+  __m256i spb_g = __lasx_xvldi(68);
+  __m256i spg_g = __lasx_xvldi(88);
+  __m256i spr_g = __lasx_xvldi(98);
+  __m256i spb_br = {0x2311231123112311, 0x2311231123112311, 0x2311231123112311,
+                    0x2311231123112311};
+  __m256i spg_br = {0x2D162D162D162D16, 0x2D162D162D162D16, 0x2D162D162D162D16,
+                    0x2D162D162D162D16};
+  __m256i spr_br = {0x3218321832183218, 0x3218321832183218, 0x3218321832183218,
+                    0x3218321832183218};
+  __m256i shuff = {0x1706150413021100, 0x1F0E1D0C1B0A1908, 0x1706150413021100,
+                   0x1F0E1D0C1B0A1908};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lasx_xvld, dst_argb, 0, dst_argb, 32, src0, src1);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    DUP2_ARG2(__lasx_xvdp2_h_bu, tmp0, spb_br, tmp0, spg_br, spb, spg);
+    spr = __lasx_xvdp2_h_bu(tmp0, spr_br);
+    spb = __lasx_xvmaddwev_h_bu(spb, tmp1, spb_g);
+    spg = __lasx_xvmaddwev_h_bu(spg, tmp1, spg_g);
+    spr = __lasx_xvmaddwev_h_bu(spr, tmp1, spr_g);
+    spb = __lasx_xvsrli_h(spb, 7);
+    spg = __lasx_xvsrli_h(spg, 7);
+    spr = __lasx_xvsrli_h(spr, 7);
+    spg = __lasx_xvsat_hu(spg, 7);
+    spr = __lasx_xvsat_hu(spr, 7);
+    reg0 = __lasx_xvpackev_b(spg, spb);
+    reg1 = __lasx_xvshuf_b(tmp1, spr, shuff);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    dst_argb += 64;
+  }
+}
+
+void ARGB4444ToARGBRow_LASX(const uint8_t* src_argb4444,
+                            uint8_t* dst_argb,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb4444, 0);
+    src1 = __lasx_xvld(src_argb4444, 32);
+    DUP4_ARG2(__lasx_xvandi_b, src0, 0x0F, src0, 0xF0, src1, 0x0F, src1, 0xF0,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lasx_xvslli_b, tmp0, 4, tmp2, 4, reg0, reg2);
+    DUP2_ARG2(__lasx_xvsrli_b, tmp1, 4, tmp3, 4, reg1, reg3);
+    DUP4_ARG2(__lasx_xvor_v, tmp0, reg0, tmp1, reg1, tmp2, reg2, tmp3, reg3,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lasx_xvilvl_b, tmp1, tmp0, tmp3, tmp2, reg0, reg2);
+    DUP2_ARG2(__lasx_xvilvh_b, tmp1, tmp0, tmp3, tmp2, reg1, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg1, reg0, 0x31, reg3, reg2,
+              0x20, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_argb4444 += 64;
+    dst_argb += 128;
+  }
+}
+
+void ARGB1555ToARGBRow_LASX(const uint8_t* src_argb1555,
+                            uint8_t* dst_argb,
+                            int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr, tmpa;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb1555, 0);
+    src1 = __lasx_xvld(src_argb1555, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    tmpa = __lasx_xvsrli_b(tmp1, 7);
+    tmpa = __lasx_xvneg_b(tmpa);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    DUP2_ARG2(__lasx_xvilvl_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    DUP2_ARG2(__lasx_xvilvh_b, tmpg, tmpb, tmpa, tmpr, reg2, reg3);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    dst2 = __lasx_xvilvl_h(reg3, reg2);
+    dst3 = __lasx_xvilvh_h(reg3, reg2);
+    DUP4_ARG3(__lasx_xvpermi_q, dst1, dst0, 0x20, dst1, dst0, 0x31, dst3, dst2,
+              0x20, dst3, dst2, 0x31, reg0, reg1, reg2, reg3);
+    __lasx_xvst(reg0, dst_argb, 0);
+    __lasx_xvst(reg1, dst_argb, 32);
+    __lasx_xvst(reg2, dst_argb, 64);
+    __lasx_xvst(reg3, dst_argb, 96);
+    src_argb1555 += 64;
+    dst_argb += 128;
+  }
+}
+
+void RGB565ToARGBRow_LASX(const uint8_t* src_rgb565,
+                          uint8_t* dst_argb,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, reg2, reg3, dst0, dst1, dst2, dst3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_rgb565, 0);
+    src1 = __lasx_xvld(src_rgb565, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    DUP2_ARG2(__lasx_xvilvl_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst0 = __lasx_xvilvl_h(reg1, reg0);
+    dst1 = __lasx_xvilvh_h(reg1, reg0);
+    DUP2_ARG2(__lasx_xvilvh_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst2 = __lasx_xvilvl_h(reg1, reg0);
+    dst3 = __lasx_xvilvh_h(reg1, reg0);
+    DUP4_ARG3(__lasx_xvpermi_q, dst1, dst0, 0x20, dst1, dst0, 0x31, dst3, dst2,
+              0x20, dst3, dst2, 0x31, reg0, reg1, reg2, reg3);
+    __lasx_xvst(reg0, dst_argb, 0);
+    __lasx_xvst(reg1, dst_argb, 32);
+    __lasx_xvst(reg2, dst_argb, 64);
+    __lasx_xvst(reg3, dst_argb, 96);
+    src_rgb565 += 64;
+    dst_argb += 128;
+  }
+}
+
+void RGB24ToARGBRow_LASX(const uint8_t* src_rgb24,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2;
+  __m256i dst0, dst1, dst2, dst3;
+  __m256i reg0, reg1, reg2, reg3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+                   0x1B1A191817161514};
+  __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+                   0x0706050403020100};
+  __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+                   0x131211100F0E0D0C};
+  __m256i shuf3 = {0x1005040310020100, 0x100B0A0910080706, 0x1005040310020100,
+                   0x100B0A0910080706};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_rgb24, 0);
+    reg1 = __lasx_xvld(src_rgb24, 32);
+    reg2 = __lasx_xvld(src_rgb24, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+              tmp1);
+    tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg3, reg2, 0x20, reg1, reg0,
+              0x31, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_rgb24 += 96;
+    dst_argb += 128;
+  }
+}
+
+void RAWToARGBRow_LASX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2, reg0, reg1, reg2, reg3;
+  __m256i dst0, dst1, dst2, dst3;
+  __m256i alpha = __lasx_xvldi(0xFF);
+  __m256i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514, 0x131211100F0E0D0C,
+                   0x1B1A191817161514};
+  __m256i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100, 0x1F1E1D1C1B1A1918,
+                   0x0706050403020100};
+  __m256i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C, 0x0B0A090807060504,
+                   0x131211100F0E0D0C};
+  __m256i shuf3 = {0x1003040510000102, 0x10090A0B10060708, 0x1003040510000102,
+                   0x10090A0B10060708};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_raw, 0);
+    reg1 = __lasx_xvld(src_raw, 32);
+    reg2 = __lasx_xvld(src_raw, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0,
+              tmp1);
+    tmp2 = __lasx_xvshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lasx_xvshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, reg0, reg1, reg2, reg3);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x20, reg3, reg2, 0x20, reg1, reg0,
+              0x31, reg3, reg2, 0x31, dst0, dst1, dst2, dst3);
+    __lasx_xvst(dst0, dst_argb, 0);
+    __lasx_xvst(dst1, dst_argb, 32);
+    __lasx_xvst(dst2, dst_argb, 64);
+    __lasx_xvst(dst3, dst_argb, 96);
+    src_raw += 96;
+    dst_argb += 128;
+  }
+}
+
+void ARGB1555ToYRow_LASX(const uint8_t* src_argb1555,
+                         uint8_t* dst_y,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, reg2, dst0;
+  __m256i const_66 = __lasx_xvldi(66);
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_25 = __lasx_xvldi(25);
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_argb1555, 0);
+    src1 = __lasx_xvld(src_argb1555, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lasx_xvmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    src_argb1555 += 64;
+    dst_y += 32;
+  }
+}
+
+void ARGB1555ToUVRow_LASX(const uint8_t* src_argb1555,
+                          int src_stride_argb1555,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i reg0, reg1, reg2, reg3, dst0;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb1555, 0, src_argb1555, 32, next_argb1555, 0,
+              next_argb1555, 32, src0, src1, src2, src3);
+    DUP2_ARG2(__lasx_xvpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lasx_xvpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    nexb = __lasx_xvandi_b(tmp2, 0x1F);
+    tmpg = __lasx_xvsrli_b(tmp0, 5);
+    nexg = __lasx_xvsrli_b(tmp2, 5);
+    reg0 = __lasx_xvandi_b(tmp1, 0x03);
+    reg2 = __lasx_xvandi_b(tmp3, 0x03);
+    reg0 = __lasx_xvslli_b(reg0, 3);
+    reg2 = __lasx_xvslli_b(reg2, 3);
+    tmpg = __lasx_xvor_v(tmpg, reg0);
+    nexg = __lasx_xvor_v(nexg, reg2);
+    reg1 = __lasx_xvandi_b(tmp1, 0x7C);
+    reg3 = __lasx_xvandi_b(tmp3, 0x7C);
+    tmpr = __lasx_xvsrli_b(reg1, 2);
+    nexr = __lasx_xvsrli_b(reg3, 2);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvslli_b(tmpg, 3);
+    reg2 = __lasx_xvslli_b(tmpr, 3);
+    tmpb = __lasx_xvsrli_b(tmpb, 2);
+    tmpg = __lasx_xvsrli_b(tmpg, 2);
+    tmpr = __lasx_xvsrli_b(tmpr, 2);
+    tmpb = __lasx_xvor_v(reg0, tmpb);
+    tmpg = __lasx_xvor_v(reg1, tmpg);
+    tmpr = __lasx_xvor_v(reg2, tmpr);
+    reg0 = __lasx_xvslli_b(nexb, 3);
+    reg1 = __lasx_xvslli_b(nexg, 3);
+    reg2 = __lasx_xvslli_b(nexr, 3);
+    nexb = __lasx_xvsrli_b(nexb, 2);
+    nexg = __lasx_xvsrli_b(nexg, 2);
+    nexr = __lasx_xvsrli_b(nexr, 2);
+    nexb = __lasx_xvor_v(reg0, nexb);
+    nexg = __lasx_xvor_v(reg1, nexg);
+    nexr = __lasx_xvor_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    reg0 = __lasx_xvpermi_d(reg0, 0xD8);
+    reg1 = __lasx_xvpermi_d(reg1, 0xD8);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_argb1555 += 64;
+    next_argb1555 += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void RGB565ToYRow_LASX(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1;
+  __m256i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m256i reg0, reg1, dst0;
+  __m256i const_66 = __lasx_xvldi(66);
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_25 = __lasx_xvldi(25);
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lasx_xvld(src_rgb565, 0);
+    src1 = __lasx_xvld(src_rgb565, 32);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lasx_xvmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lasx_xvmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lasx_xvmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    dst0 = __lasx_xvpermi_d(dst0, 0xD8);
+    __lasx_xvst(dst0, dst_y, 0);
+    dst_y += 32;
+    src_rgb565 += 64;
+  }
+}
+
+void RGB565ToUVRow_LASX(const uint8_t* src_rgb565,
+                        int src_stride_rgb565,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  int x;
+  int len = width / 32;
+  const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  __m256i src0, src1, src2, src3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i reg0, reg1, reg2, reg3, dst0;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_rgb565, 0, src_rgb565, 32, next_rgb565, 0,
+              next_rgb565, 32, src0, src1, src2, src3);
+    DUP2_ARG2(__lasx_xvpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lasx_xvpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lasx_xvandi_b(tmp0, 0x1F);
+    tmpr = __lasx_xvandi_b(tmp1, 0xF8);
+    nexb = __lasx_xvandi_b(tmp2, 0x1F);
+    nexr = __lasx_xvandi_b(tmp3, 0xF8);
+    reg1 = __lasx_xvandi_b(tmp1, 0x07);
+    reg3 = __lasx_xvandi_b(tmp3, 0x07);
+    reg0 = __lasx_xvsrli_b(tmp0, 5);
+    reg1 = __lasx_xvslli_b(reg1, 3);
+    reg2 = __lasx_xvsrli_b(tmp2, 5);
+    reg3 = __lasx_xvslli_b(reg3, 3);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    nexg = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvslli_b(tmpb, 3);
+    reg1 = __lasx_xvsrli_b(tmpb, 2);
+    reg2 = __lasx_xvslli_b(nexb, 3);
+    reg3 = __lasx_xvsrli_b(nexb, 2);
+    tmpb = __lasx_xvor_v(reg1, reg0);
+    nexb = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvslli_b(tmpg, 2);
+    reg1 = __lasx_xvsrli_b(tmpg, 4);
+    reg2 = __lasx_xvslli_b(nexg, 2);
+    reg3 = __lasx_xvsrli_b(nexg, 4);
+    tmpg = __lasx_xvor_v(reg1, reg0);
+    nexg = __lasx_xvor_v(reg2, reg3);
+    reg0 = __lasx_xvsrli_b(tmpr, 5);
+    reg2 = __lasx_xvsrli_b(nexr, 5);
+    tmpr = __lasx_xvor_v(tmpr, reg0);
+    nexr = __lasx_xvor_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    reg0 = __lasx_xvpermi_d(reg0, 0xD8);
+    reg1 = __lasx_xvpermi_d(reg1, 0xD8);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    dst_u += 16;
+    dst_v += 16;
+    src_rgb565 += 64;
+    next_rgb565 += 64;
+  }
+}
+
+void RGB24ToYRow_LASX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, dst0;
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_br = {0x4219421942194219, 0x4219421942194219,
+                      0x4219421942194219, 0x4219421942194219};
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+  __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
+                    0x17151412110F0E0C};
+  __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
+                    0x0F0D0C0A09070604};
+  __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
+                    0x001600130010000D};
+  __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
+                    0x000E000B00080005};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_rgb24, 0);
+    reg1 = __lasx_xvld(src_rgb24, 32);
+    reg2 = __lasx_xvld(src_rgb24, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    tmp0 = __lasx_xvshuf_b(src1, src0, shuff0);
+    tmp1 = __lasx_xvshuf_b(src1, src2, shuff1);
+    tmp2 = __lasx_xvshuf_b(src1, src0, shuff2);
+    tmp3 = __lasx_xvshuf_b(src1, src2, shuff3);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmp2, const_129);
+    reg1 = __lasx_xvmaddwev_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvst(dst0, dst_y, 0);
+    dst_y += 32;
+    src_rgb24 += 96;
+  }
+}
+
+void RGB24ToUVRow_LASX(const uint8_t* src_rgb24,
+                       int src_stride_rgb24,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  const uint8_t* next_rgb24 = src_rgb24 + src_stride_rgb24;
+  int len = width / 32;
+  __m256i src0, src1, src2, reg0, reg1, reg2;
+  __m256i nex0, nex1, nex2, dst0, tmp0, tmp1, tmp2;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff0_b = {0x15120F0C09060300, 0x00000000001E1B18,
+                      0x15120F0C09060300, 0x00000000001E1B18};
+  __m256i shuff1_b = {0x0706050403020100, 0x1D1A1714110A0908,
+                      0x0706050403020100, 0x1D1A1714110A0908};
+  __m256i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19,
+                      0x1613100D0A070401, 0x00000000001F1C19};
+  __m256i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908,
+                      0x0706050403020100, 0x1E1B1815120A0908};
+  __m256i shuff0_r = {0x1714110E0B080502, 0x0000000000001D1A,
+                      0x1714110E0B080502, 0x0000000000001D1A};
+  __m256i shuff1_r = {0x0706050403020100, 0x1F1C191613100908,
+                      0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_rgb24, 0, src_rgb24, 32, src_rgb24, 64,
+              next_rgb24, 0, reg0, reg1, reg2, tmp0);
+    DUP2_ARG2(__lasx_xvld, next_rgb24, 32, next_rgb24, 64, tmp1, tmp2);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+              0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+    DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_rgb24 += 96;
+    next_rgb24 += 96;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void RAWToYRow_LASX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, reg2, dst0;
+  __m256i const_129 = __lasx_xvldi(129);
+  __m256i const_br = {0x1942194219421942, 0x1942194219421942,
+                      0x1942194219421942, 0x1942194219421942};
+  __m256i const_1080 = {0x1080108010801080, 0x1080108010801080,
+                        0x1080108010801080, 0x1080108010801080};
+  __m256i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C, 0x0B09080605030200,
+                    0x17151412110F0E0C};
+  __m256i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604, 0x0301001E1D1B1A18,
+                    0x0F0D0C0A09070604};
+  __m256i shuff2 = {0x000A000700040001, 0x001600130010000D, 0x000A000700040001,
+                    0x001600130010000D};
+  __m256i shuff3 = {0x0002001F001C0019, 0x000E000B00080005, 0x0002001F001C0019,
+                    0x000E000B00080005};
+
+  for (x = 0; x < len; x++) {
+    reg0 = __lasx_xvld(src_raw, 0);
+    reg1 = __lasx_xvld(src_raw, 32);
+    reg2 = __lasx_xvld(src_raw, 64);
+    src0 = __lasx_xvpermi_q(reg1, reg0, 0x30);
+    src1 = __lasx_xvpermi_q(reg2, reg0, 0x21);
+    src2 = __lasx_xvpermi_q(reg2, reg1, 0x30);
+    tmp0 = __lasx_xvshuf_b(src1, src0, shuff0);
+    tmp1 = __lasx_xvshuf_b(src1, src2, shuff1);
+    tmp2 = __lasx_xvshuf_b(src1, src0, shuff2);
+    tmp3 = __lasx_xvshuf_b(src1, src2, shuff3);
+    reg0 = __lasx_xvmaddwev_h_bu(const_1080, tmp2, const_129);
+    reg1 = __lasx_xvmaddwev_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvst(dst0, dst_y, 0);
+    dst_y += 32;
+    src_raw += 96;
+  }
+}
+
+void RAWToUVRow_LASX(const uint8_t* src_raw,
+                     int src_stride_raw,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_raw = src_raw + src_stride_raw;
+  int len = width / 32;
+  __m256i src0, src1, src2, reg0, reg1, reg2;
+  __m256i nex0, nex1, nex2, dst0, tmp0, tmp1, tmp2;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_112 = __lasx_xvldi(0x438);
+  __m256i const_74 = __lasx_xvldi(0x425);
+  __m256i const_38 = __lasx_xvldi(0x413);
+  __m256i const_94 = __lasx_xvldi(0x42F);
+  __m256i const_18 = __lasx_xvldi(0x409);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff0_r = {0x15120F0C09060300, 0x00000000001E1B18,
+                      0x15120F0C09060300, 0x00000000001E1B18};
+  __m256i shuff1_r = {0x0706050403020100, 0x1D1A1714110A0908,
+                      0x0706050403020100, 0x1D1A1714110A0908};
+  __m256i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19,
+                      0x1613100D0A070401, 0x00000000001F1C19};
+  __m256i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908,
+                      0x0706050403020100, 0x1E1B1815120A0908};
+  __m256i shuff0_b = {0x1714110E0B080502, 0x0000000000001D1A,
+                      0x1714110E0B080502, 0x0000000000001D1A};
+  __m256i shuff1_b = {0x0706050403020100, 0x1F1C191613100908,
+                      0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_raw, 0, src_raw, 32, src_raw, 64, next_raw, 0,
+              reg0, reg1, reg2, tmp0);
+    DUP2_ARG2(__lasx_xvld, next_raw, 32, next_raw, 64, tmp1, tmp2);
+    DUP4_ARG3(__lasx_xvpermi_q, reg1, reg0, 0x30, reg2, reg0, 0x21, reg2, reg1,
+              0x30, tmp1, tmp0, 0x30, src0, src1, src2, nex0);
+    DUP2_ARG3(__lasx_xvpermi_q, tmp2, tmp0, 0x21, tmp2, tmp1, 0x30, nex1, nex2);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lasx_xvshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, reg0, reg1);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 1);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 2);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    src_raw += 96;
+    next_raw += 96;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
+void NV12ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_vrub, vec_vgug, vec_y, vec_vu;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_vrub = __lasx_xvilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lasx_xvilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_vu = __lasx_xvld(src_uv, 0);
+    vec_vu = __lasx_xvsub_b(vec_vu, const_0x80);
+    vec_vu = __lasx_vext2xv_h_b(vec_vu);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_uv += 16;
+  }
+}
+
+void NV12ToRGB565Row_LASX(const uint8_t* src_y,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_rgb565,
+                          const struct YuvConstants* yuvconstants,
+                          int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_vrub, vec_vgug, vec_y, vec_vu;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_vrub = __lasx_xvilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lasx_xvilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_vu = __lasx_xvld(src_uv, 0);
+    vec_vu = __lasx_xvsub_b(vec_vu, const_0x80);
+    vec_vu = __lasx_vext2xv_h_b(vec_vu);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    out_b = __lasx_xvsrli_h(out_b, 3);
+    out_g = __lasx_xvsrli_h(out_g, 2);
+    out_r = __lasx_xvsrli_h(out_r, 3);
+    out_g = __lasx_xvslli_h(out_g, 5);
+    out_r = __lasx_xvslli_h(out_r, 11);
+    out_r = __lasx_xvor_v(out_r, out_g);
+    out_r = __lasx_xvor_v(out_r, out_b);
+    __lasx_xvst(out_r, dst_rgb565, 0);
+    src_y += 16;
+    src_uv += 16;
+    dst_rgb565 += 32;
+  }
+}
+
+void NV21ToARGBRow_LASX(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m256i vec_yg, vec_yb, vec_ub, vec_vr, vec_ug, vec_vg;
+  __m256i vec_ubvr, vec_ugvg, vec_y, vec_uv;
+  __m256i out_b, out_g, out_r;
+  __m256i const_0x80 = __lasx_xvldi(0x80);
+  __m256i alpha = __lasx_xvldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
+  vec_ubvr = __lasx_xvilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lasx_xvilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lasx_xvld(src_y, 0);
+    vec_uv = __lasx_xvld(src_uv, 0);
+    vec_uv = __lasx_xvsub_b(vec_uv, const_0x80);
+    vec_uv = __lasx_vext2xv_h_b(vec_uv);
+    YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_uv += 16;
+  }
+}
+
+void ARGBToYJRow_LASX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3, dst0;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1;
+  __m256i const_128 = __lasx_xvldi(0x480);
+  __m256i const_150 = __lasx_xvldi(0x96);
+  __m256i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D,
+                      0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
+  __m256i shuff = {0x0000000400000000, 0x0000000500000001, 0x0000000600000002,
+                   0x0000000700000003};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp2 = __lasx_xvpickev_b(src3, src2);
+    tmp3 = __lasx_xvpickod_b(src3, src2);
+    reg0 = __lasx_xvmaddwev_h_bu(const_128, tmp1, const_150);
+    reg1 = __lasx_xvmaddwev_h_bu(const_128, tmp3, const_150);
+    reg0 = __lasx_xvdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lasx_xvdp2add_h_bu(reg1, const_br, tmp2);
+    dst0 = __lasx_xvpickod_b(reg1, reg0);
+    dst0 = __lasx_xvperm_w(dst0, shuff);
+    __lasx_xvst(dst0, dst_y, 0);
+    dst_y += 32;
+    src_argb += 128;
+  }
+}
+
+void ARGBToUVJRow_LASX(const uint8_t* src_argb,
+                       int src_stride_argb,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  const uint8_t* next_argb = src_argb + src_stride_argb;
+  int len = width / 32;
+  __m256i src0, src1, src2, src3;
+  __m256i nex0, nex1, nex2, nex3;
+  __m256i tmp0, tmp1, tmp2, tmp3;
+  __m256i reg0, reg1, dst0;
+  __m256i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m256i const_63 = __lasx_xvldi(0x43F);
+  __m256i const_42 = __lasx_xvldi(0x42A);
+  __m256i const_21 = __lasx_xvldi(0x415);
+  __m256i const_53 = __lasx_xvldi(0x435);
+  __m256i const_10 = __lasx_xvldi(0x40A);
+  __m256i const_8080 = {0x8080808080808080, 0x8080808080808080,
+                        0x8080808080808080, 0x8080808080808080};
+  __m256i shuff = {0x1614060412100200, 0x1E1C0E0C1A180A08, 0x1715070513110301,
+                   0x1F1D0F0D1B190B09};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lasx_xvld, src_argb, 0, src_argb, 32, src_argb, 64, src_argb,
+              96, src0, src1, src2, src3);
+    DUP4_ARG2(__lasx_xvld, next_argb, 0, next_argb, 32, next_argb, 64,
+              next_argb, 96, nex0, nex1, nex2, nex3);
+    tmp0 = __lasx_xvpickev_b(src1, src0);
+    tmp1 = __lasx_xvpickod_b(src1, src0);
+    tmp2 = __lasx_xvpickev_b(src3, src2);
+    tmp3 = __lasx_xvpickod_b(src3, src2);
+    tmpr = __lasx_xvpickod_b(tmp2, tmp0);
+    tmpb = __lasx_xvpickev_b(tmp2, tmp0);
+    tmpg = __lasx_xvpickev_b(tmp3, tmp1);
+    tmp0 = __lasx_xvpickev_b(nex1, nex0);
+    tmp1 = __lasx_xvpickod_b(nex1, nex0);
+    tmp2 = __lasx_xvpickev_b(nex3, nex2);
+    tmp3 = __lasx_xvpickod_b(nex3, nex2);
+    nexr = __lasx_xvpickod_b(tmp2, tmp0);
+    nexb = __lasx_xvpickev_b(tmp2, tmp0);
+    nexg = __lasx_xvpickev_b(tmp3, tmp1);
+    tmp0 = __lasx_xvaddwev_h_bu(tmpb, nexb);
+    tmp1 = __lasx_xvaddwod_h_bu(tmpb, nexb);
+    tmp2 = __lasx_xvaddwev_h_bu(tmpg, nexg);
+    tmp3 = __lasx_xvaddwod_h_bu(tmpg, nexg);
+    reg0 = __lasx_xvaddwev_h_bu(tmpr, nexr);
+    reg1 = __lasx_xvaddwod_h_bu(tmpr, nexr);
+    tmpb = __lasx_xvavgr_hu(tmp0, tmp1);
+    tmpg = __lasx_xvavgr_hu(tmp2, tmp3);
+    tmpr = __lasx_xvavgr_hu(reg0, reg1);
+    reg0 = __lasx_xvmadd_h(const_8080, const_63, tmpb);
+    reg1 = __lasx_xvmadd_h(const_8080, const_63, tmpr);
+    reg0 = __lasx_xvmsub_h(reg0, const_42, tmpg);
+    reg1 = __lasx_xvmsub_h(reg1, const_53, tmpg);
+    reg0 = __lasx_xvmsub_h(reg0, const_21, tmpr);
+    reg1 = __lasx_xvmsub_h(reg1, const_10, tmpb);
+    dst0 = __lasx_xvpackod_b(reg1, reg0);
+    tmp0 = __lasx_xvpermi_d(dst0, 0x44);
+    tmp1 = __lasx_xvpermi_d(dst0, 0xEE);
+    dst0 = __lasx_xvshuf_b(tmp1, tmp0, shuff);
+    __lasx_xvstelm_d(dst0, dst_u, 0, 0);
+    __lasx_xvstelm_d(dst0, dst_v, 0, 2);
+    __lasx_xvstelm_d(dst0, dst_u, 8, 1);
+    __lasx_xvstelm_d(dst0, dst_v, 8, 3);
+    dst_u += 16;
+    dst_v += 16;
+    src_argb += 128;
+    next_argb += 128;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LASX) && defined(__loongarch_asx)
diff --git a/files/source/row_lsx.cc b/files/source/row_lsx.cc
new file mode 100644
index 00000000..3e8b901a
--- /dev/null
+++ b/files/source/row_lsx.cc
@@ -0,0 +1,1829 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Fill YUV -> RGB conversion constants into vectors
+#define YUVTORGB_SETUP(yuvconst, vr, ub, vg, ug, yg, yb) \
+  {                                                      \
+    ub = __lsx_vreplgr2vr_h(yuvconst->kUVToB[0]);        \
+    vr = __lsx_vreplgr2vr_h(yuvconst->kUVToR[1]);        \
+    ug = __lsx_vreplgr2vr_h(yuvconst->kUVToG[0]);        \
+    vg = __lsx_vreplgr2vr_h(yuvconst->kUVToG[1]);        \
+    yg = __lsx_vreplgr2vr_h(yuvconst->kYToRgb[0]);       \
+    yb = __lsx_vreplgr2vr_w(yuvconst->kYBiasToRgb[0]);   \
+  }
+
+// Convert 8 pixels of YUV420 to RGB.
+#define YUVTORGB(in_y, in_vu, vrub, vgug, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    __m128i y_ev, y_od, u_l, v_l;                                      \
+    __m128i tmp0, tmp1, tmp2, tmp3;                                    \
+                                                                       \
+    tmp0 = __lsx_vilvl_b(in_y, in_y);                                  \
+    y_ev = __lsx_vmulwev_w_hu_h(tmp0, yg);                             \
+    y_od = __lsx_vmulwod_w_hu_h(tmp0, yg);                             \
+    y_ev = __lsx_vsrai_w(y_ev, 16);                                    \
+    y_od = __lsx_vsrai_w(y_od, 16);                                    \
+    y_ev = __lsx_vadd_w(y_ev, yb);                                     \
+    y_od = __lsx_vadd_w(y_od, yb);                                     \
+    in_vu = __lsx_vilvl_b(zero, in_vu);                                \
+    in_vu = __lsx_vsub_h(in_vu, const_80);                             \
+    u_l = __lsx_vmulwev_w_h(in_vu, vrub);                              \
+    v_l = __lsx_vmulwod_w_h(in_vu, vrub);                              \
+    tmp0 = __lsx_vadd_w(y_ev, u_l);                                    \
+    tmp1 = __lsx_vadd_w(y_od, u_l);                                    \
+    tmp2 = __lsx_vadd_w(y_ev, v_l);                                    \
+    tmp3 = __lsx_vadd_w(y_od, v_l);                                    \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                     \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                     \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                     \
+    tmp3 = __lsx_vsrai_w(tmp3, 6);                                     \
+    tmp0 = __lsx_vclip255_w(tmp0);                                     \
+    tmp1 = __lsx_vclip255_w(tmp1);                                     \
+    tmp2 = __lsx_vclip255_w(tmp2);                                     \
+    tmp3 = __lsx_vclip255_w(tmp3);                                     \
+    out_b = __lsx_vpackev_h(tmp1, tmp0);                               \
+    out_r = __lsx_vpackev_h(tmp3, tmp2);                               \
+    tmp0 = __lsx_vdp2_w_h(in_vu, vgug);                                \
+    tmp1 = __lsx_vsub_w(y_ev, tmp0);                                   \
+    tmp2 = __lsx_vsub_w(y_od, tmp0);                                   \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                     \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                     \
+    tmp1 = __lsx_vclip255_w(tmp1);                                     \
+    tmp2 = __lsx_vclip255_w(tmp2);                                     \
+    out_g = __lsx_vpackev_h(tmp2, tmp1);                               \
+  }
+
+// Convert I444 pixels of YUV420 to RGB.
+#define I444TORGB(in_yy, in_u, in_v, ub, vr, ugvg, yg, yb, out_b, out_g, \
+                  out_r)                                                 \
+  {                                                                      \
+    __m128i y_ev, y_od, u_ev, v_ev, u_od, v_od;                          \
+    __m128i tmp0, tmp1, tmp2, tmp3;                                      \
+                                                                         \
+    y_ev = __lsx_vmulwev_w_hu_h(in_yy, yg);                              \
+    y_od = __lsx_vmulwod_w_hu_h(in_yy, yg);                              \
+    y_ev = __lsx_vsrai_w(y_ev, 16);                                      \
+    y_od = __lsx_vsrai_w(y_od, 16);                                      \
+    y_ev = __lsx_vadd_w(y_ev, yb);                                       \
+    y_od = __lsx_vadd_w(y_od, yb);                                       \
+    in_u = __lsx_vsub_h(in_u, const_80);                                 \
+    in_v = __lsx_vsub_h(in_v, const_80);                                 \
+    u_ev = __lsx_vmulwev_w_h(in_u, ub);                                  \
+    u_od = __lsx_vmulwod_w_h(in_u, ub);                                  \
+    v_ev = __lsx_vmulwev_w_h(in_v, vr);                                  \
+    v_od = __lsx_vmulwod_w_h(in_v, vr);                                  \
+    tmp0 = __lsx_vadd_w(y_ev, u_ev);                                     \
+    tmp1 = __lsx_vadd_w(y_od, u_od);                                     \
+    tmp2 = __lsx_vadd_w(y_ev, v_ev);                                     \
+    tmp3 = __lsx_vadd_w(y_od, v_od);                                     \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                       \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                       \
+    tmp2 = __lsx_vsrai_w(tmp2, 6);                                       \
+    tmp3 = __lsx_vsrai_w(tmp3, 6);                                       \
+    tmp0 = __lsx_vclip255_w(tmp0);                                       \
+    tmp1 = __lsx_vclip255_w(tmp1);                                       \
+    tmp2 = __lsx_vclip255_w(tmp2);                                       \
+    tmp3 = __lsx_vclip255_w(tmp3);                                       \
+    out_b = __lsx_vpackev_h(tmp1, tmp0);                                 \
+    out_r = __lsx_vpackev_h(tmp3, tmp2);                                 \
+    u_ev = __lsx_vpackev_h(in_u, in_v);                                  \
+    u_od = __lsx_vpackod_h(in_u, in_v);                                  \
+    v_ev = __lsx_vdp2_w_h(u_ev, ugvg);                                   \
+    v_od = __lsx_vdp2_w_h(u_od, ugvg);                                   \
+    tmp0 = __lsx_vsub_w(y_ev, v_ev);                                     \
+    tmp1 = __lsx_vsub_w(y_od, v_od);                                     \
+    tmp0 = __lsx_vsrai_w(tmp0, 6);                                       \
+    tmp1 = __lsx_vsrai_w(tmp1, 6);                                       \
+    tmp0 = __lsx_vclip255_w(tmp0);                                       \
+    tmp1 = __lsx_vclip255_w(tmp1);                                       \
+    out_g = __lsx_vpackev_h(tmp1, tmp0);                                 \
+  }
+
+// Pack and Store 8 ARGB values.
+#define STOREARGB(in_a, in_r, in_g, in_b, pdst_argb) \
+  {                                                  \
+    __m128i temp0, temp1;                            \
+    __m128i dst0, dst1;                              \
+                                                     \
+    temp0 = __lsx_vpackev_b(in_g, in_b);             \
+    temp1 = __lsx_vpackev_b(in_a, in_r);             \
+    dst0 = __lsx_vilvl_h(temp1, temp0);              \
+    dst1 = __lsx_vilvh_h(temp1, temp0);              \
+    __lsx_vst(dst0, pdst_argb, 0);                   \
+    __lsx_vst(dst1, pdst_argb, 16);                  \
+    pdst_argb += 32;                                 \
+  }
+
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _dst0) \
+  {                                                              \
+    __m128i _tmp0, _tmp1, _tmp2, _tmp3;                          \
+    __m128i _reg0, _reg1;                                        \
+    _tmp0 = __lsx_vaddwev_h_bu(_tmpb, _nexb);                    \
+    _tmp1 = __lsx_vaddwod_h_bu(_tmpb, _nexb);                    \
+    _tmp2 = __lsx_vaddwev_h_bu(_tmpg, _nexg);                    \
+    _tmp3 = __lsx_vaddwod_h_bu(_tmpg, _nexg);                    \
+    _reg0 = __lsx_vaddwev_h_bu(_tmpr, _nexr);                    \
+    _reg1 = __lsx_vaddwod_h_bu(_tmpr, _nexr);                    \
+    _tmpb = __lsx_vavgr_hu(_tmp0, _tmp1);                        \
+    _tmpg = __lsx_vavgr_hu(_tmp2, _tmp3);                        \
+    _tmpr = __lsx_vavgr_hu(_reg0, _reg1);                        \
+    _reg0 = __lsx_vmadd_h(const_8080, const_112, _tmpb);         \
+    _reg1 = __lsx_vmadd_h(const_8080, const_112, _tmpr);         \
+    _reg0 = __lsx_vmsub_h(_reg0, const_74, _tmpg);               \
+    _reg1 = __lsx_vmsub_h(_reg1, const_94, _tmpg);               \
+    _reg0 = __lsx_vmsub_h(_reg0, const_38, _tmpr);               \
+    _reg1 = __lsx_vmsub_h(_reg1, const_18, _tmpb);               \
+    _dst0 = __lsx_vpickod_b(_reg1, _reg0);                       \
+  }
+
+void ARGB4444ToARGBRow_LSX(const uint8_t* src_argb4444,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb4444, 0);
+    src1 = __lsx_vld(src_argb4444, 16);
+    tmp0 = __lsx_vandi_b(src0, 0x0F);
+    tmp1 = __lsx_vandi_b(src0, 0xF0);
+    tmp2 = __lsx_vandi_b(src1, 0x0F);
+    tmp3 = __lsx_vandi_b(src1, 0xF0);
+    reg0 = __lsx_vslli_b(tmp0, 4);
+    reg2 = __lsx_vslli_b(tmp2, 4);
+    reg1 = __lsx_vsrli_b(tmp1, 4);
+    reg3 = __lsx_vsrli_b(tmp3, 4);
+    DUP4_ARG2(__lsx_vor_v, tmp0, reg0, tmp1, reg1, tmp2, reg2, tmp3, reg3, tmp0,
+              tmp1, tmp2, tmp3);
+    dst0 = __lsx_vilvl_b(tmp1, tmp0);
+    dst2 = __lsx_vilvl_b(tmp3, tmp2);
+    dst1 = __lsx_vilvh_b(tmp1, tmp0);
+    dst3 = __lsx_vilvh_b(tmp3, tmp2);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_argb4444 += 32;
+  }
+}
+
+void ARGB1555ToARGBRow_LSX(const uint8_t* src_argb1555,
+                           uint8_t* dst_argb,
+                           int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr, tmpa;
+  __m128i reg0, reg1, reg2;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb1555, 0);
+    src1 = __lsx_vld(src_argb1555, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    tmpa = __lsx_vsrli_b(tmp1, 7);
+    tmpa = __lsx_vneg_b(tmpa);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    DUP2_ARG2(__lsx_vilvl_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    dst0 = __lsx_vilvl_h(reg1, reg0);
+    dst1 = __lsx_vilvh_h(reg1, reg0);
+    DUP2_ARG2(__lsx_vilvh_b, tmpg, tmpb, tmpa, tmpr, reg0, reg1);
+    dst2 = __lsx_vilvl_h(reg1, reg0);
+    dst3 = __lsx_vilvh_h(reg1, reg0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_argb1555 += 32;
+  }
+}
+
+void RGB565ToARGBRow_LSX(const uint8_t* src_rgb565,
+                         uint8_t* dst_argb,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb565, 0);
+    src1 = __lsx_vld(src_rgb565, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    DUP2_ARG2(__lsx_vilvl_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst0 = __lsx_vilvl_h(reg1, reg0);
+    dst1 = __lsx_vilvh_h(reg1, reg0);
+    DUP2_ARG2(__lsx_vilvh_b, tmpg, tmpb, alpha, tmpr, reg0, reg1);
+    dst2 = __lsx_vilvl_h(reg1, reg0);
+    dst3 = __lsx_vilvh_h(reg1, reg0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_rgb565 += 32;
+  }
+}
+
+void RGB24ToARGBRow_LSX(const uint8_t* src_rgb24,
+                        uint8_t* dst_argb,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514};
+  __m128i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100};
+  __m128i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C};
+  __m128i shuf3 = {0x1005040310020100, 0x100B0A0910080706};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb24, 0);
+    src1 = __lsx_vld(src_rgb24, 16);
+    src2 = __lsx_vld(src_rgb24, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+    tmp2 = __lsx_vshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_rgb24 += 48;
+  }
+}
+
+void RAWToARGBRow_LSX(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuf0 = {0x131211100F0E0D0C, 0x1B1A191817161514};
+  __m128i shuf1 = {0x1F1E1D1C1B1A1918, 0x0706050403020100};
+  __m128i shuf2 = {0x0B0A090807060504, 0x131211100F0E0D0C};
+  __m128i shuf3 = {0x1003040510000102, 0x10090A0B10060708};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_raw, 0);
+    src1 = __lsx_vld(src_raw, 16);
+    src2 = __lsx_vld(src_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src2, shuf1, tmp0, tmp1);
+    tmp2 = __lsx_vshuf_b(src1, src2, shuf2);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, src0, shuf3, alpha, tmp0, shuf3, alpha,
+              tmp1, shuf3, alpha, tmp2, shuf3, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_raw += 48;
+  }
+}
+
+void ARGB1555ToYRow_LSX(const uint8_t* src_argb1555,
+                        uint8_t* dst_y,
+                        int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, reg2, dst0;
+  __m128i const_66 = __lsx_vldi(66);
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_25 = __lsx_vldi(25);
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_argb1555, 0);
+    src1 = __lsx_vld(src_argb1555, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lsx_vpackod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_argb1555 += 32;
+  }
+}
+
+void ARGB1555ToUVRow_LSX(const uint8_t* src_argb1555,
+                         int src_stride_argb1555,
+                         uint8_t* dst_u,
+                         uint8_t* dst_v,
+                         int width) {
+  int x;
+  int len = width / 16;
+  const uint8_t* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb1555, 0, src_argb1555, 16, next_argb1555, 0,
+              next_argb1555, 16, src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    nexb = __lsx_vandi_b(tmp2, 0x1F);
+    tmpg = __lsx_vsrli_b(tmp0, 5);
+    nexg = __lsx_vsrli_b(tmp2, 5);
+    reg0 = __lsx_vandi_b(tmp1, 0x03);
+    reg2 = __lsx_vandi_b(tmp3, 0x03);
+    reg0 = __lsx_vslli_b(reg0, 3);
+    reg2 = __lsx_vslli_b(reg2, 3);
+    tmpg = __lsx_vor_v(tmpg, reg0);
+    nexg = __lsx_vor_v(nexg, reg2);
+    reg1 = __lsx_vandi_b(tmp1, 0x7C);
+    reg3 = __lsx_vandi_b(tmp3, 0x7C);
+    tmpr = __lsx_vsrli_b(reg1, 2);
+    nexr = __lsx_vsrli_b(reg3, 2);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vslli_b(tmpg, 3);
+    reg2 = __lsx_vslli_b(tmpr, 3);
+    tmpb = __lsx_vsrli_b(tmpb, 2);
+    tmpg = __lsx_vsrli_b(tmpg, 2);
+    tmpr = __lsx_vsrli_b(tmpr, 2);
+    tmpb = __lsx_vor_v(reg0, tmpb);
+    tmpg = __lsx_vor_v(reg1, tmpg);
+    tmpr = __lsx_vor_v(reg2, tmpr);
+    reg0 = __lsx_vslli_b(nexb, 3);
+    reg1 = __lsx_vslli_b(nexg, 3);
+    reg2 = __lsx_vslli_b(nexr, 3);
+    nexb = __lsx_vsrli_b(nexb, 2);
+    nexg = __lsx_vsrli_b(nexg, 2);
+    nexr = __lsx_vsrli_b(nexr, 2);
+    nexb = __lsx_vor_v(reg0, nexb);
+    nexg = __lsx_vor_v(reg1, nexg);
+    nexr = __lsx_vor_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_argb1555 += 32;
+    next_argb1555 += 32;
+  }
+}
+
+void RGB565ToYRow_LSX(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1;
+  __m128i tmp0, tmp1, tmpb, tmpg, tmpr;
+  __m128i reg0, reg1, dst0;
+  __m128i const_66 = __lsx_vldi(66);
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_25 = __lsx_vldi(25);
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb565, 0);
+    src1 = __lsx_vld(src_rgb565, 16);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmpb, const_25);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmpb, const_25);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpg, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpg, const_129);
+    reg0 = __lsx_vmaddwev_h_bu(reg0, tmpr, const_66);
+    reg1 = __lsx_vmaddwod_h_bu(reg1, tmpr, const_66);
+    dst0 = __lsx_vpackod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_rgb565 += 32;
+  }
+}
+
+void RGB565ToUVRow_LSX(const uint8_t* src_rgb565,
+                       int src_stride_rgb565,
+                       uint8_t* dst_u,
+                       uint8_t* dst_v,
+                       int width) {
+  int x;
+  int len = width / 16;
+  const uint8_t* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_rgb565, 0, src_rgb565, 16, next_rgb565, 0,
+              next_rgb565, 16, src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    tmpb = __lsx_vandi_b(tmp0, 0x1F);
+    tmpr = __lsx_vandi_b(tmp1, 0xF8);
+    nexb = __lsx_vandi_b(tmp2, 0x1F);
+    nexr = __lsx_vandi_b(tmp3, 0xF8);
+    reg1 = __lsx_vandi_b(tmp1, 0x07);
+    reg3 = __lsx_vandi_b(tmp3, 0x07);
+    reg0 = __lsx_vsrli_b(tmp0, 5);
+    reg1 = __lsx_vslli_b(reg1, 3);
+    reg2 = __lsx_vsrli_b(tmp2, 5);
+    reg3 = __lsx_vslli_b(reg3, 3);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    nexg = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vslli_b(tmpb, 3);
+    reg1 = __lsx_vsrli_b(tmpb, 2);
+    reg2 = __lsx_vslli_b(nexb, 3);
+    reg3 = __lsx_vsrli_b(nexb, 2);
+    tmpb = __lsx_vor_v(reg1, reg0);
+    nexb = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vslli_b(tmpg, 2);
+    reg1 = __lsx_vsrli_b(tmpg, 4);
+    reg2 = __lsx_vslli_b(nexg, 2);
+    reg3 = __lsx_vsrli_b(nexg, 4);
+    tmpg = __lsx_vor_v(reg1, reg0);
+    nexg = __lsx_vor_v(reg2, reg3);
+    reg0 = __lsx_vsrli_b(tmpr, 5);
+    reg2 = __lsx_vsrli_b(nexr, 5);
+    tmpr = __lsx_vor_v(tmpr, reg0);
+    nexr = __lsx_vor_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgb565 += 32;
+    next_rgb565 += 32;
+  }
+}
+
+void RGB24ToYRow_LSX(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, dst0;
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_br = {0x4219421942194219, 0x4219421942194219};
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+  __m128i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C};
+  __m128i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604};
+  __m128i shuff2 = {0x000A000700040001, 0x001600130010000D};
+  __m128i shuff3 = {0x0002001F001C0019, 0x000E000B00080005};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb24, 0);
+    src1 = __lsx_vld(src_rgb24, 16);
+    src2 = __lsx_vld(src_rgb24, 32);
+    tmp0 = __lsx_vshuf_b(src1, src0, shuff0);
+    tmp1 = __lsx_vshuf_b(src1, src2, shuff1);
+    tmp2 = __lsx_vshuf_b(src1, src0, shuff2);
+    tmp3 = __lsx_vshuf_b(src1, src2, shuff3);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmp2, const_129);
+    reg1 = __lsx_vmaddwev_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp1);
+    dst0 = __lsx_vpickod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_rgb24 += 48;
+  }
+}
+
+void RGB24ToUVRow_LSX(const uint8_t* src_rgb24,
+                      int src_stride_rgb24,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* next_rgb24 = src_rgb24 + src_stride_rgb24;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i nex0, nex1, nex2, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+  __m128i shuff0_b = {0x15120F0C09060300, 0x00000000001E1B18};
+  __m128i shuff1_b = {0x0706050403020100, 0x1D1A1714110A0908};
+  __m128i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19};
+  __m128i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908};
+  __m128i shuff0_r = {0x1714110E0B080502, 0x0000000000001D1A};
+  __m128i shuff1_r = {0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_rgb24, 0);
+    src1 = __lsx_vld(src_rgb24, 16);
+    src2 = __lsx_vld(src_rgb24, 32);
+    nex0 = __lsx_vld(next_rgb24, 0);
+    nex1 = __lsx_vld(next_rgb24, 16);
+    nex2 = __lsx_vld(next_rgb24, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgb24 += 48;
+    next_rgb24 += 48;
+  }
+}
+
+void RAWToYRow_LSX(const uint8_t* src_raw, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, dst0;
+  __m128i const_129 = __lsx_vldi(129);
+  __m128i const_br = {0x1942194219421942, 0x1942194219421942};
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+  __m128i shuff0 = {0x0B09080605030200, 0x17151412110F0E0C};
+  __m128i shuff1 = {0x0301001E1D1B1A18, 0x0F0D0C0A09070604};
+  __m128i shuff2 = {0x000A000700040001, 0x001600130010000D};
+  __m128i shuff3 = {0x0002001F001C0019, 0x000E000B00080005};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_raw, 0);
+    src1 = __lsx_vld(src_raw, 16);
+    src2 = __lsx_vld(src_raw, 32);
+    tmp0 = __lsx_vshuf_b(src1, src0, shuff0);
+    tmp1 = __lsx_vshuf_b(src1, src2, shuff1);
+    tmp2 = __lsx_vshuf_b(src1, src0, shuff2);
+    tmp3 = __lsx_vshuf_b(src1, src2, shuff3);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmp2, const_129);
+    reg1 = __lsx_vmaddwev_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp1);
+    dst0 = __lsx_vsrlni_b_h(reg1, reg0, 8);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_raw += 48;
+  }
+}
+
+void RAWToUVRow_LSX(const uint8_t* src_raw,
+                    int src_stride_raw,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  const uint8_t* next_raw = src_raw + src_stride_raw;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i nex0, nex1, nex2, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+  __m128i shuff0_r = {0x15120F0C09060300, 0x00000000001E1B18};
+  __m128i shuff1_r = {0x0706050403020100, 0x1D1A1714110A0908};
+  __m128i shuff0_g = {0x1613100D0A070401, 0x00000000001F1C19};
+  __m128i shuff1_g = {0x0706050403020100, 0x1E1B1815120A0908};
+  __m128i shuff0_b = {0x1714110E0B080502, 0x0000000000001D1A};
+  __m128i shuff1_b = {0x0706050403020100, 0x1F1C191613100908};
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_raw, 0);
+    src1 = __lsx_vld(src_raw, 16);
+    src2 = __lsx_vld(src_raw, 32);
+    nex0 = __lsx_vld(next_raw, 0);
+    nex1 = __lsx_vld(next_raw, 16);
+    nex2 = __lsx_vld(next_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_b, nex1, nex0, shuff0_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_g, nex1, nex0, shuff0_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0_r, nex1, nex0, shuff0_r, tmpr,
+              nexr);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpb, shuff1_b, nex2, nexb, shuff1_b, tmpb,
+              nexb);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpg, shuff1_g, nex2, nexg, shuff1_g, tmpg,
+              nexg);
+    DUP2_ARG3(__lsx_vshuf_b, src2, tmpr, shuff1_r, nex2, nexr, shuff1_r, tmpr,
+              nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_raw += 48;
+    next_raw += 48;
+  }
+}
+
+void NV12ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_vu = __lsx_vld(src_uv, 0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 8;
+    src_uv += 8;
+  }
+}
+
+void NV12ToRGB565Row_LSX(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_vu = __lsx_vld(src_uv, 0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    out_b = __lsx_vsrli_h(out_b, 3);
+    out_g = __lsx_vsrli_h(out_g, 2);
+    out_r = __lsx_vsrli_h(out_r, 3);
+    out_g = __lsx_vslli_h(out_g, 5);
+    out_r = __lsx_vslli_h(out_r, 11);
+    out_r = __lsx_vor_v(out_r, out_g);
+    out_r = __lsx_vor_v(out_r, out_b);
+    __lsx_vst(out_r, dst_rgb565, 0);
+    src_y += 8;
+    src_uv += 8;
+    dst_rgb565 += 16;
+  }
+}
+
+void NV21ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i vec_y, vec_uv;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_ubvr, vec_ugvg;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_ubvr = __lsx_vilvl_h(vec_ub, vec_vr);
+  vec_ugvg = __lsx_vilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_uv = __lsx_vld(src_vu, 0);
+    YUVTORGB(vec_y, vec_uv, vec_ubvr, vec_ugvg, vec_yg, vec_yb, out_r, out_g,
+             out_b);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 8;
+    src_vu += 8;
+  }
+}
+
+void SobelRow_LSX(const uint8_t* src_sobelx,
+                  const uint8_t* src_sobely,
+                  uint8_t* dst_argb,
+                  int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, tmp0;
+  __m128i out0, out1, out2, out3;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i shuff0 = {0x1001010110000000, 0x1003030310020202};
+  __m128i shuff1 = __lsx_vaddi_bu(shuff0, 0x04);
+  __m128i shuff2 = __lsx_vaddi_bu(shuff1, 0x04);
+  __m128i shuff3 = __lsx_vaddi_bu(shuff2, 0x04);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_sobelx, 0);
+    src1 = __lsx_vld(src_sobely, 0);
+    tmp0 = __lsx_vsadd_bu(src0, src1);
+    DUP4_ARG3(__lsx_vshuf_b, alpha, tmp0, shuff0, alpha, tmp0, shuff1, alpha,
+              tmp0, shuff2, alpha, tmp0, shuff3, out0, out1, out2, out3);
+    __lsx_vst(out0, dst_argb, 0);
+    __lsx_vst(out1, dst_argb, 16);
+    __lsx_vst(out2, dst_argb, 32);
+    __lsx_vst(out3, dst_argb, 48);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void SobelToPlaneRow_LSX(const uint8_t* src_sobelx,
+                         const uint8_t* src_sobely,
+                         uint8_t* dst_y,
+                         int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_sobelx, 0, src_sobelx, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, src_sobely, 0, src_sobely, 16, src2, src3);
+    dst0 = __lsx_vsadd_bu(src0, src2);
+    dst1 = __lsx_vsadd_bu(src1, src3);
+    __lsx_vst(dst0, dst_y, 0);
+    __lsx_vst(dst1, dst_y, 16);
+    src_sobelx += 32;
+    src_sobely += 32;
+    dst_y += 32;
+  }
+}
+
+void SobelXYRow_LSX(const uint8_t* src_sobelx,
+                    const uint8_t* src_sobely,
+                    uint8_t* dst_argb,
+                    int width) {
+  int x;
+  int len = width / 16;
+  __m128i src_r, src_b, src_g;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    src_r = __lsx_vld(src_sobelx, 0);
+    src_b = __lsx_vld(src_sobely, 0);
+    src_g = __lsx_vsadd_bu(src_r, src_b);
+    tmp0 = __lsx_vilvl_b(src_g, src_b);
+    tmp1 = __lsx_vilvh_b(src_g, src_b);
+    tmp2 = __lsx_vilvl_b(alpha, src_r);
+    tmp3 = __lsx_vilvh_b(alpha, src_r);
+    dst0 = __lsx_vilvl_h(tmp2, tmp0);
+    dst1 = __lsx_vilvh_h(tmp2, tmp0);
+    dst2 = __lsx_vilvl_h(tmp3, tmp1);
+    dst3 = __lsx_vilvh_h(tmp3, tmp1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    src_sobelx += 16;
+    src_sobely += 16;
+    dst_argb += 64;
+  }
+}
+
+void ARGBToYJRow_LSX(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1;
+  __m128i const_128 = __lsx_vldi(0x480);
+  __m128i const_150 = __lsx_vldi(0x96);
+  __m128i const_br = {0x4D1D4D1D4D1D4D1D, 0x4D1D4D1D4D1D4D1D};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb, 32, src_argb, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    reg0 = __lsx_vmaddwev_h_bu(const_128, tmp1, const_150);
+    reg1 = __lsx_vmaddwev_h_bu(const_128, tmp3, const_150);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp2);
+    dst0 = __lsx_vpickod_b(reg1, reg0);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_argb += 64;
+  }
+}
+
+void BGRAToYRow_LSX(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1;
+  __m128i const_129 = __lsx_vldi(0x81);
+  __m128i const_br = {0x1942194219421942, 0x1942194219421942};
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_bgra, 0, src_bgra, 16, src_bgra, 32, src_bgra, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    reg0 = __lsx_vmaddwod_h_bu(const_1080, tmp1, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp2);
+    dst0 = __lsx_vsrlni_b_h(reg1, reg0, 8);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_bgra += 64;
+  }
+}
+
+void BGRAToUVRow_LSX(const uint8_t* src_bgra,
+                     int src_stride_bgra,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_bgra = src_bgra + src_stride_bgra;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_bgra, 0, src_bgra, 16, src_bgra, 32, src_bgra, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_bgra, 0, next_bgra, 16, next_bgra, 32, next_bgra,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    tmpb = __lsx_vpickod_b(tmp2, tmp0);
+    tmpr = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickod_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickod_b(nex1, nex0);
+    tmp1 = __lsx_vpickev_b(nex1, nex0);
+    tmp2 = __lsx_vpickod_b(nex3, nex2);
+    tmp3 = __lsx_vpickev_b(nex3, nex2);
+    nexb = __lsx_vpickod_b(tmp2, tmp0);
+    nexr = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickod_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_bgra += 64;
+    next_bgra += 64;
+  }
+}
+
+void ABGRToYRow_LSX(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1;
+  __m128i const_129 = __lsx_vldi(0x81);
+  __m128i const_br = {0x1942194219421942, 0x1942194219421942};
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_abgr, 0, src_abgr, 16, src_abgr, 32, src_abgr, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    reg0 = __lsx_vmaddwev_h_bu(const_1080, tmp1, const_129);
+    reg1 = __lsx_vmaddwev_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp2);
+    dst0 = __lsx_vsrlni_b_h(reg1, reg0, 8);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_abgr += 64;
+  }
+}
+
+void ABGRToUVRow_LSX(const uint8_t* src_abgr,
+                     int src_stride_abgr,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_abgr = src_abgr + src_stride_abgr;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_abgr, 0, src_abgr, 16, src_abgr, 32, src_abgr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_abgr, 0, next_abgr, 16, next_abgr, 32, next_abgr,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    tmpb = __lsx_vpickod_b(tmp2, tmp0);
+    tmpr = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickev_b(nex1, nex0);
+    tmp1 = __lsx_vpickod_b(nex1, nex0);
+    tmp2 = __lsx_vpickev_b(nex3, nex2);
+    tmp3 = __lsx_vpickod_b(nex3, nex2);
+    nexb = __lsx_vpickod_b(tmp2, tmp0);
+    nexr = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickev_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_abgr += 64;
+    next_abgr += 64;
+  }
+}
+
+void RGBAToYRow_LSX(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1;
+  __m128i const_129 = __lsx_vldi(0x81);
+  __m128i const_br = {0x4219421942194219, 0x4219421942194219};
+  __m128i const_1080 = {0x1080108010801080, 0x1080108010801080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_rgba, 0, src_rgba, 16, src_rgba, 32, src_rgba, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    reg0 = __lsx_vmaddwod_h_bu(const_1080, tmp1, const_129);
+    reg1 = __lsx_vmaddwod_h_bu(const_1080, tmp3, const_129);
+    reg0 = __lsx_vdp2add_h_bu(reg0, const_br, tmp0);
+    reg1 = __lsx_vdp2add_h_bu(reg1, const_br, tmp2);
+    dst0 = __lsx_vsrlni_b_h(reg1, reg0, 8);
+    __lsx_vst(dst0, dst_y, 0);
+    dst_y += 16;
+    src_rgba += 64;
+  }
+}
+
+void RGBAToUVRow_LSX(const uint8_t* src_rgba,
+                     int src_stride_rgba,
+                     uint8_t* dst_u,
+                     uint8_t* dst_v,
+                     int width) {
+  int x;
+  const uint8_t* next_rgba = src_rgba + src_stride_rgba;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_112 = __lsx_vldi(0x438);
+  __m128i const_74 = __lsx_vldi(0x425);
+  __m128i const_38 = __lsx_vldi(0x413);
+  __m128i const_94 = __lsx_vldi(0x42F);
+  __m128i const_18 = __lsx_vldi(0x409);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_rgba, 0, src_rgba, 16, src_rgba, 32, src_rgba, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_rgba, 0, next_rgba, 16, next_rgba, 32, next_rgba,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickev_b(src1, src0);
+    tmp2 = __lsx_vpickod_b(src3, src2);
+    tmp3 = __lsx_vpickev_b(src3, src2);
+    tmpr = __lsx_vpickod_b(tmp2, tmp0);
+    tmpb = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickod_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickod_b(nex1, nex0);
+    tmp1 = __lsx_vpickev_b(nex1, nex0);
+    tmp2 = __lsx_vpickod_b(nex3, nex2);
+    tmp3 = __lsx_vpickev_b(nex3, nex2);
+    nexr = __lsx_vpickod_b(tmp2, tmp0);
+    nexb = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickod_b(tmp3, tmp1);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_rgba += 64;
+    next_rgba += 64;
+  }
+}
+
+void ARGBToUVJRow_LSX(const uint8_t* src_argb,
+                      int src_stride_argb,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  int x;
+  const uint8_t* next_argb = src_argb + src_stride_argb;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3;
+  __m128i nex0, nex1, nex2, nex3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i reg0, reg1, dst0;
+  __m128i tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  __m128i const_63 = __lsx_vldi(0x43F);
+  __m128i const_42 = __lsx_vldi(0x42A);
+  __m128i const_21 = __lsx_vldi(0x415);
+  __m128i const_53 = __lsx_vldi(0x435);
+  __m128i const_10 = __lsx_vldi(0x40A);
+  __m128i const_8080 = {0x8080808080808080, 0x8080808080808080};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb, 32, src_argb, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, next_argb, 0, next_argb, 16, next_argb, 32, next_argb,
+              48, nex0, nex1, nex2, nex3);
+    tmp0 = __lsx_vpickev_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src1, src0);
+    tmp2 = __lsx_vpickev_b(src3, src2);
+    tmp3 = __lsx_vpickod_b(src3, src2);
+    tmpr = __lsx_vpickod_b(tmp2, tmp0);
+    tmpb = __lsx_vpickev_b(tmp2, tmp0);
+    tmpg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vpickev_b(nex1, nex0);
+    tmp1 = __lsx_vpickod_b(nex1, nex0);
+    tmp2 = __lsx_vpickev_b(nex3, nex2);
+    tmp3 = __lsx_vpickod_b(nex3, nex2);
+    nexr = __lsx_vpickod_b(tmp2, tmp0);
+    nexb = __lsx_vpickev_b(tmp2, tmp0);
+    nexg = __lsx_vpickev_b(tmp3, tmp1);
+    tmp0 = __lsx_vaddwev_h_bu(tmpb, nexb);
+    tmp1 = __lsx_vaddwod_h_bu(tmpb, nexb);
+    tmp2 = __lsx_vaddwev_h_bu(tmpg, nexg);
+    tmp3 = __lsx_vaddwod_h_bu(tmpg, nexg);
+    reg0 = __lsx_vaddwev_h_bu(tmpr, nexr);
+    reg1 = __lsx_vaddwod_h_bu(tmpr, nexr);
+    tmpb = __lsx_vavgr_hu(tmp0, tmp1);
+    tmpg = __lsx_vavgr_hu(tmp2, tmp3);
+    tmpr = __lsx_vavgr_hu(reg0, reg1);
+    reg0 = __lsx_vmadd_h(const_8080, const_63, tmpb);
+    reg1 = __lsx_vmadd_h(const_8080, const_63, tmpr);
+    reg0 = __lsx_vmsub_h(reg0, const_42, tmpg);
+    reg1 = __lsx_vmsub_h(reg1, const_53, tmpg);
+    reg0 = __lsx_vmsub_h(reg0, const_21, tmpr);
+    reg1 = __lsx_vmsub_h(reg1, const_10, tmpb);
+    dst0 = __lsx_vpickod_b(reg1, reg0);
+    __lsx_vstelm_d(dst0, dst_u, 0, 0);
+    __lsx_vstelm_d(dst0, dst_v, 0, 1);
+    dst_u += 8;
+    dst_v += 8;
+    src_argb += 64;
+    next_argb += 64;
+  }
+}
+
+void I444ToARGBRow_LSX(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, vec_u, vec_v, out_b, out_g, out_r;
+  __m128i vec_yl, vec_yh, vec_ul, vec_vl, vec_uh, vec_vh;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb, vec_ugvg;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i zero = __lsx_vldi(0);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_ugvg = __lsx_vilvl_h(vec_ug, vec_vg);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_u = __lsx_vld(src_u, 0);
+    vec_v = __lsx_vld(src_v, 0);
+    vec_yl = __lsx_vilvl_b(vec_y, vec_y);
+    vec_ul = __lsx_vilvl_b(zero, vec_u);
+    vec_vl = __lsx_vilvl_b(zero, vec_v);
+    I444TORGB(vec_yl, vec_ul, vec_vl, vec_ub, vec_vr, vec_ugvg, vec_yg, vec_yb,
+              out_b, out_g, out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    vec_yh = __lsx_vilvh_b(vec_y, vec_y);
+    vec_uh = __lsx_vilvh_b(zero, vec_u);
+    vec_vh = __lsx_vilvh_b(zero, vec_v);
+    I444TORGB(vec_yh, vec_uh, vec_vh, vec_ub, vec_vr, vec_ugvg, vec_yg, vec_yb,
+              out_b, out_g, out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_y += 16;
+    src_u += 16;
+    src_v += 16;
+  }
+}
+
+void I400ToARGBRow_LSX(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, vec_yl, vec_yh, out0;
+  __m128i y_ev, y_od, dst0, dst1, dst2, dst3;
+  __m128i temp0, temp1;
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i vec_yg = __lsx_vreplgr2vr_h(yuvconstants->kYToRgb[0]);
+  __m128i vec_yb = __lsx_vreplgr2vr_w(yuvconstants->kYBiasToRgb[0]);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    vec_yl = __lsx_vilvl_b(vec_y, vec_y);
+    y_ev = __lsx_vmulwev_w_hu_h(vec_yl, vec_yg);
+    y_od = __lsx_vmulwod_w_hu_h(vec_yl, vec_yg);
+    y_ev = __lsx_vsrai_w(y_ev, 16);
+    y_od = __lsx_vsrai_w(y_od, 16);
+    y_ev = __lsx_vadd_w(y_ev, vec_yb);
+    y_od = __lsx_vadd_w(y_od, vec_yb);
+    y_ev = __lsx_vsrai_w(y_ev, 6);
+    y_od = __lsx_vsrai_w(y_od, 6);
+    y_ev = __lsx_vclip255_w(y_ev);
+    y_od = __lsx_vclip255_w(y_od);
+    out0 = __lsx_vpackev_h(y_od, y_ev);
+    temp0 = __lsx_vpackev_b(out0, out0);
+    temp1 = __lsx_vpackev_b(alpha, out0);
+    dst0 = __lsx_vilvl_h(temp1, temp0);
+    dst1 = __lsx_vilvh_h(temp1, temp0);
+    vec_yh = __lsx_vilvh_b(vec_y, vec_y);
+    y_ev = __lsx_vmulwev_w_hu_h(vec_yh, vec_yg);
+    y_od = __lsx_vmulwod_w_hu_h(vec_yh, vec_yg);
+    y_ev = __lsx_vsrai_w(y_ev, 16);
+    y_od = __lsx_vsrai_w(y_od, 16);
+    y_ev = __lsx_vadd_w(y_ev, vec_yb);
+    y_od = __lsx_vadd_w(y_od, vec_yb);
+    y_ev = __lsx_vsrai_w(y_ev, 6);
+    y_od = __lsx_vsrai_w(y_od, 6);
+    y_ev = __lsx_vclip255_w(y_ev);
+    y_od = __lsx_vclip255_w(y_od);
+    out0 = __lsx_vpackev_h(y_od, y_ev);
+    temp0 = __lsx_vpackev_b(out0, out0);
+    temp1 = __lsx_vpackev_b(alpha, out0);
+    dst2 = __lsx_vilvl_h(temp1, temp0);
+    dst3 = __lsx_vilvh_h(temp1, temp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_y += 16;
+  }
+}
+
+void J400ToARGBRow_LSX(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  int x;
+  int len = width / 16;
+  __m128i vec_y, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  for (x = 0; x < len; x++) {
+    vec_y = __lsx_vld(src_y, 0);
+    tmp0 = __lsx_vilvl_b(vec_y, vec_y);
+    tmp1 = __lsx_vilvh_b(vec_y, vec_y);
+    tmp2 = __lsx_vilvl_b(alpha, vec_y);
+    tmp3 = __lsx_vilvh_b(alpha, vec_y);
+    dst0 = __lsx_vilvl_h(tmp2, tmp0);
+    dst1 = __lsx_vilvh_h(tmp2, tmp0);
+    dst2 = __lsx_vilvl_h(tmp3, tmp1);
+    dst3 = __lsx_vilvh_h(tmp3, tmp1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+    src_y += 16;
+  }
+}
+
+void YUY2ToARGBRow_LSX(const uint8_t* src_yuy2,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_yuy2, 0);
+    vec_y = __lsx_vpickev_b(src0, src0);
+    vec_vu = __lsx_vpickod_b(src0, src0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_yuy2 += 16;
+  }
+}
+
+void UYVYToARGBRow_LSX(const uint8_t* src_uyvy,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, vec_y, vec_vu;
+  __m128i vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb;
+  __m128i vec_vrub, vec_vgug;
+  __m128i out_b, out_g, out_r;
+  __m128i const_80 = __lsx_vldi(0x480);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+
+  YUVTORGB_SETUP(yuvconstants, vec_vr, vec_ub, vec_vg, vec_ug, vec_yg, vec_yb);
+  vec_vrub = __lsx_vilvl_h(vec_vr, vec_ub);
+  vec_vgug = __lsx_vilvl_h(vec_vg, vec_ug);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_uyvy, 0);
+    vec_y = __lsx_vpickod_b(src0, src0);
+    vec_vu = __lsx_vpickev_b(src0, src0);
+    YUVTORGB(vec_y, vec_vu, vec_vrub, vec_vgug, vec_yg, vec_yb, out_b, out_g,
+             out_r);
+    STOREARGB(alpha, out_r, out_g, out_b, dst_argb);
+    src_uyvy += 16;
+  }
+}
+
+void InterpolateRow_LSX(uint8_t* dst_ptr,
+                        const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        int width,
+                        int32_t source_y_fraction) {
+  int x;
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8_t* nex_ptr = src_ptr + src_stride;
+  uint16_t y_fractions;
+  int len = width / 32;
+  __m128i src0, src1, nex0, nex1;
+  __m128i dst0, dst1, y_frac;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+  __m128i const_128 = __lsx_vldi(0x480);
+
+  if (y1_fraction == 0) {
+    for (x = 0; x < len; x++) {
+      DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+      __lsx_vst(src0, dst_ptr, 0);
+      __lsx_vst(src1, dst_ptr, 16);
+      src_ptr += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  if (y1_fraction == 128) {
+    for (x = 0; x < len; x++) {
+      DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+      DUP2_ARG2(__lsx_vld, nex_ptr, 0, nex_ptr, 16, nex0, nex1);
+      dst0 = __lsx_vavgr_bu(src0, nex0);
+      dst1 = __lsx_vavgr_bu(src1, nex1);
+      __lsx_vst(dst0, dst_ptr, 0);
+      __lsx_vst(dst1, dst_ptr, 16);
+      src_ptr += 32;
+      nex_ptr += 32;
+      dst_ptr += 32;
+    }
+    return;
+  }
+
+  y_fractions = (uint16_t)(y0_fraction + (y1_fraction << 8));
+  y_frac = __lsx_vreplgr2vr_h(y_fractions);
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, nex_ptr, 0, nex_ptr, 16, nex0, nex1);
+    tmp0 = __lsx_vilvl_b(nex0, src0);
+    tmp1 = __lsx_vilvh_b(nex0, src0);
+    tmp2 = __lsx_vilvl_b(nex1, src1);
+    tmp3 = __lsx_vilvh_b(nex1, src1);
+    tmp0 = __lsx_vdp2add_h_bu(const_128, tmp0, y_frac);
+    tmp1 = __lsx_vdp2add_h_bu(const_128, tmp1, y_frac);
+    tmp2 = __lsx_vdp2add_h_bu(const_128, tmp2, y_frac);
+    tmp3 = __lsx_vdp2add_h_bu(const_128, tmp3, y_frac);
+    dst0 = __lsx_vsrlni_b_h(tmp1, tmp0, 8);
+    dst1 = __lsx_vsrlni_b_h(tmp3, tmp2, 8);
+    __lsx_vst(dst0, dst_ptr, 0);
+    __lsx_vst(dst1, dst_ptr, 16);
+    src_ptr += 32;
+    nex_ptr += 32;
+    dst_ptr += 32;
+  }
+}
+
+void ARGBSetRow_LSX(uint8_t* dst_argb, uint32_t v32, int width) {
+  int x;
+  int len = width / 4;
+  __m128i dst0 = __lsx_vreplgr2vr_w(v32);
+
+  for (x = 0; x < len; x++) {
+    __lsx_vst(dst0, dst_argb, 0);
+    dst_argb += 16;
+  }
+}
+
+void RAWToRGB24Row_LSX(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2;
+  __m128i dst0, dst1, dst2;
+  __m128i shuf0 = {0x0708030405000102, 0x110C0D0E090A0B06};
+  __m128i shuf1 = {0x1516171213140F10, 0x1F1E1B1C1D18191A};
+  __m128i shuf2 = {0x090405060102031E, 0x0D0E0F0A0B0C0708};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_raw, 0, src_raw, 16, src0, src1);
+    src2 = __lsx_vld(src_raw, 32);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuf0, src1, src0, shuf1, dst0, dst1);
+    dst2 = __lsx_vshuf_b(src1, src2, shuf2);
+    dst1 = __lsx_vinsgr2vr_b(dst1, src_raw[32], 0x0E);
+    __lsx_vst(dst0, dst_rgb24, 0);
+    __lsx_vst(dst1, dst_rgb24, 16);
+    __lsx_vst(dst2, dst_rgb24, 32);
+    dst_rgb24 += 48;
+    src_raw += 48;
+  }
+}
+
+void MergeUVRow_LSX(const uint8_t* src_u,
+                    const uint8_t* src_v,
+                    uint8_t* dst_uv,
+                    int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_u, 0, src_v, 0, src0, src1);
+    dst0 = __lsx_vilvl_b(src1, src0);
+    dst1 = __lsx_vilvh_b(src1, src0);
+    __lsx_vst(dst0, dst_uv, 0);
+    __lsx_vst(dst1, dst_uv, 16);
+    src_u += 16;
+    src_v += 16;
+    dst_uv += 32;
+  }
+}
+
+void ARGBExtractAlphaRow_LSX(const uint8_t* src_argb,
+                             uint8_t* dst_a,
+                             int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb, 32, src_argb, 48,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vpickod_b(src1, src0);
+    tmp1 = __lsx_vpickod_b(src3, src2);
+    dst0 = __lsx_vpickod_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst_a, 0);
+    src_argb += 64;
+    dst_a += 16;
+  }
+}
+
+void ARGBBlendRow_LSX(const uint8_t* src_argb,
+                      const uint8_t* src_argb1,
+                      uint8_t* dst_argb,
+                      int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, dst0, dst1;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i a0, a1, a2, a3;
+  __m128i const_256 = __lsx_vldi(0x500);
+  __m128i zero = __lsx_vldi(0);
+  __m128i alpha = __lsx_vldi(0xFF);
+  __m128i control = {0xFF000000FF000000, 0xFF000000FF000000};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src_argb1, 0, src_argb1, 16,
+              src0, src1, src2, src3);
+    tmp0 = __lsx_vshuf4i_b(src0, 0xFF);
+    tmp1 = __lsx_vshuf4i_b(src1, 0xFF);
+    a0 = __lsx_vilvl_b(zero, tmp0);
+    a1 = __lsx_vilvh_b(zero, tmp0);
+    a2 = __lsx_vilvl_b(zero, tmp1);
+    a3 = __lsx_vilvh_b(zero, tmp1);
+    reg0 = __lsx_vilvl_b(zero, src2);
+    reg1 = __lsx_vilvh_b(zero, src2);
+    reg2 = __lsx_vilvl_b(zero, src3);
+    reg3 = __lsx_vilvh_b(zero, src3);
+    DUP4_ARG2(__lsx_vsub_h, const_256, a0, const_256, a1, const_256, a2,
+              const_256, a3, a0, a1, a2, a3);
+    DUP4_ARG2(__lsx_vmul_h, a0, reg0, a1, reg1, a2, reg2, a3, reg3, reg0, reg1,
+              reg2, reg3);
+    DUP2_ARG3(__lsx_vsrani_b_h, reg1, reg0, 8, reg3, reg2, 8, dst0, dst1);
+    dst0 = __lsx_vsadd_bu(dst0, src0);
+    dst1 = __lsx_vsadd_bu(dst1, src1);
+    dst0 = __lsx_vbitsel_v(dst0, alpha, control);
+    dst1 = __lsx_vbitsel_v(dst1, alpha, control);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    src_argb += 32;
+    src_argb1 += 32;
+    dst_argb += 32;
+  }
+}
+
+void ARGBQuantizeRow_LSX(uint8_t* dst_argb,
+                         int scale,
+                         int interval_size,
+                         int interval_offset,
+                         int width) {
+  int x;
+  int len = width / 16;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec_size = __lsx_vreplgr2vr_b(interval_size);
+  __m128i vec_offset = __lsx_vreplgr2vr_b(interval_offset);
+  __m128i vec_scale = __lsx_vreplgr2vr_w(scale);
+  __m128i zero = __lsx_vldi(0);
+  __m128i control = {0xFF000000FF000000, 0xFF000000FF000000};
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, dst_argb, 0, dst_argb, 16, dst_argb, 32, dst_argb, 48,
+              src0, src1, src2, src3);
+    reg0 = __lsx_vilvl_b(zero, src0);
+    reg1 = __lsx_vilvh_b(zero, src0);
+    reg2 = __lsx_vilvl_b(zero, src1);
+    reg3 = __lsx_vilvh_b(zero, src1);
+    reg4 = __lsx_vilvl_b(zero, src2);
+    reg5 = __lsx_vilvh_b(zero, src2);
+    reg6 = __lsx_vilvl_b(zero, src3);
+    reg7 = __lsx_vilvh_b(zero, src3);
+    tmp0 = __lsx_vilvl_h(zero, reg0);
+    tmp1 = __lsx_vilvh_h(zero, reg0);
+    tmp2 = __lsx_vilvl_h(zero, reg1);
+    tmp3 = __lsx_vilvh_h(zero, reg1);
+    tmp4 = __lsx_vilvl_h(zero, reg2);
+    tmp5 = __lsx_vilvh_h(zero, reg2);
+    tmp6 = __lsx_vilvl_h(zero, reg3);
+    tmp7 = __lsx_vilvh_h(zero, reg3);
+    DUP4_ARG2(__lsx_vmul_w, tmp0, vec_scale, tmp1, vec_scale, tmp2, vec_scale,
+              tmp3, vec_scale, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vmul_w, tmp4, vec_scale, tmp5, vec_scale, tmp6, vec_scale,
+              tmp7, vec_scale, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vsrani_h_w, tmp1, tmp0, 16, tmp3, tmp2, 16, tmp5, tmp4, 16,
+              tmp7, tmp6, 16, reg0, reg1, reg2, reg3);
+    dst0 = __lsx_vpickev_b(reg1, reg0);
+    dst1 = __lsx_vpickev_b(reg3, reg2);
+    tmp0 = __lsx_vilvl_h(zero, reg4);
+    tmp1 = __lsx_vilvh_h(zero, reg4);
+    tmp2 = __lsx_vilvl_h(zero, reg5);
+    tmp3 = __lsx_vilvh_h(zero, reg5);
+    tmp4 = __lsx_vilvl_h(zero, reg6);
+    tmp5 = __lsx_vilvh_h(zero, reg6);
+    tmp6 = __lsx_vilvl_h(zero, reg7);
+    tmp7 = __lsx_vilvh_h(zero, reg7);
+    DUP4_ARG2(__lsx_vmul_w, tmp0, vec_scale, tmp1, vec_scale, tmp2, vec_scale,
+              tmp3, vec_scale, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vmul_w, tmp4, vec_scale, tmp5, vec_scale, tmp6, vec_scale,
+              tmp7, vec_scale, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vsrani_h_w, tmp1, tmp0, 16, tmp3, tmp2, 16, tmp5, tmp4, 16,
+              tmp7, tmp6, 16, reg0, reg1, reg2, reg3);
+    dst2 = __lsx_vpickev_b(reg1, reg0);
+    dst3 = __lsx_vpickev_b(reg3, reg2);
+    DUP4_ARG2(__lsx_vmul_b, dst0, vec_size, dst1, vec_size, dst2, vec_size,
+              dst3, vec_size, dst0, dst1, dst2, dst3);
+    DUP4_ARG2(__lsx_vadd_b, dst0, vec_offset, dst1, vec_offset, dst2,
+              vec_offset, dst3, vec_offset, dst0, dst1, dst2, dst3);
+    DUP4_ARG3(__lsx_vbitsel_v, dst0, src0, control, dst1, src1, control, dst2,
+              src2, control, dst3, src3, control, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    __lsx_vst(dst2, dst_argb, 32);
+    __lsx_vst(dst3, dst_argb, 48);
+    dst_argb += 64;
+  }
+}
+
+void ARGBColorMatrixRow_LSX(const uint8_t* src_argb,
+                            uint8_t* dst_argb,
+                            const int8_t* matrix_argb,
+                            int width) {
+  int x;
+  int len = width / 8;
+  __m128i src0, src1, tmp0, tmp1, dst0, dst1;
+  __m128i tmp_b, tmp_g, tmp_r, tmp_a;
+  __m128i reg_b, reg_g, reg_r, reg_a;
+  __m128i matrix_b = __lsx_vldrepl_w(matrix_argb, 0);
+  __m128i matrix_g = __lsx_vldrepl_w(matrix_argb, 4);
+  __m128i matrix_r = __lsx_vldrepl_w(matrix_argb, 8);
+  __m128i matrix_a = __lsx_vldrepl_w(matrix_argb, 12);
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    DUP4_ARG2(__lsx_vdp2_h_bu_b, src0, matrix_b, src0, matrix_g, src0, matrix_r,
+              src0, matrix_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vdp2_h_bu_b, src1, matrix_b, src1, matrix_g, src1, matrix_r,
+              src1, matrix_a, reg_b, reg_g, reg_r, reg_a);
+    DUP4_ARG2(__lsx_vhaddw_w_h, tmp_b, tmp_b, tmp_g, tmp_g, tmp_r, tmp_r, tmp_a,
+              tmp_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vhaddw_w_h, reg_b, reg_b, reg_g, reg_g, reg_r, reg_r, reg_a,
+              reg_a, reg_b, reg_g, reg_r, reg_a);
+    DUP4_ARG2(__lsx_vsrai_w, tmp_b, 6, tmp_g, 6, tmp_r, 6, tmp_a, 6, tmp_b,
+              tmp_g, tmp_r, tmp_a);
+    DUP4_ARG2(__lsx_vsrai_w, reg_b, 6, reg_g, 6, reg_r, 6, reg_a, 6, reg_b,
+              reg_g, reg_r, reg_a);
+    DUP4_ARG1(__lsx_vclip255_w, tmp_b, tmp_g, tmp_r, tmp_a, tmp_b, tmp_g, tmp_r,
+              tmp_a)
+    DUP4_ARG1(__lsx_vclip255_w, reg_b, reg_g, reg_r, reg_a, reg_b, reg_g, reg_r,
+              reg_a)
+    DUP4_ARG2(__lsx_vpickev_h, reg_b, tmp_b, reg_g, tmp_g, reg_r, tmp_r, reg_a,
+              tmp_a, tmp_b, tmp_g, tmp_r, tmp_a);
+    tmp0 = __lsx_vpackev_b(tmp_g, tmp_b);
+    tmp1 = __lsx_vpackev_b(tmp_a, tmp_r);
+    dst0 = __lsx_vilvl_h(tmp1, tmp0);
+    dst1 = __lsx_vilvh_h(tmp1, tmp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    src_argb += 32;
+    dst_argb += 32;
+  }
+}
+
+void SplitUVRow_LSX(const uint8_t* src_uv,
+                    uint8_t* dst_u,
+                    uint8_t* dst_v,
+                    int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_uv, 0, src_uv, 16, src_uv, 32, src_uv, 48, src0,
+              src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, dst0, dst1);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, dst2, dst3);
+    __lsx_vst(dst0, dst_u, 0);
+    __lsx_vst(dst1, dst_u, 16);
+    __lsx_vst(dst2, dst_v, 0);
+    __lsx_vst(dst3, dst_v, 16);
+    src_uv += 64;
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void SetRow_LSX(uint8_t* dst, uint8_t v8, int width) {
+  int x;
+  int len = width / 16;
+  __m128i dst0 = __lsx_vreplgr2vr_b(v8);
+
+  for (x = 0; x < len; x++) {
+    __lsx_vst(dst0, dst, 0);
+    dst += 16;
+  }
+}
+
+void MirrorSplitUVRow_LSX(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
+  int x;
+  int len = width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2, dst3;
+  __m128i shuff0 = {0x10121416181A1C1E, 0x00020406080A0C0E};
+  __m128i shuff1 = {0x11131517191B1D1F, 0x01030507090B0D0F};
+
+  src_uv += (width << 1);
+  for (x = 0; x < len; x++) {
+    src_uv -= 64;
+    DUP4_ARG2(__lsx_vld, src_uv, 0, src_uv, 16, src_uv, 32, src_uv, 48, src2,
+              src3, src0, src1);
+    DUP4_ARG3(__lsx_vshuf_b, src1, src0, shuff1, src3, src2, shuff1, src1, src0,
+              shuff0, src3, src2, shuff0, dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst_v, 0);
+    __lsx_vst(dst1, dst_v, 16);
+    __lsx_vst(dst2, dst_u, 0);
+    __lsx_vst(dst3, dst_u, 16);
+    dst_u += 32;
+    dst_v += 32;
+  }
+}
+
+void HalfFloatRow_LSX(const uint16_t* src,
+                      uint16_t* dst,
+                      float scale,
+                      int width) {
+  int x;
+  int len = width / 32;
+  float mult = 1.9259299444e-34f * scale;
+  __m128i src0, src1, src2, src3, dst0, dst1, dst2, dst3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128 vec_mult = (__m128)__lsx_vldrepl_w(&mult, 0);
+  __m128i zero = __lsx_vldi(0);
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src, 0, src, 16, src, 32, src, 48, src0, src1, src2,
+              src3);
+    DUP4_ARG2(__lsx_vilvl_h, zero, src0, zero, src1, zero, src2, zero, src3,
+              tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vilvh_h, zero, src0, zero, src1, zero, src2, zero, src3,
+              tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG1(__lsx_vffint_s_wu, tmp0, tmp2, tmp4, tmp6, reg0, reg2, reg4,
+              reg6);
+    DUP4_ARG1(__lsx_vffint_s_wu, tmp1, tmp3, tmp5, tmp7, reg1, reg3, reg5,
+              reg7);
+    DUP4_ARG2(__lsx_vfmul_s, reg0, vec_mult, reg1, vec_mult, reg2, vec_mult,
+              reg3, vec_mult, reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vfmul_s, reg4, vec_mult, reg5, vec_mult, reg6, vec_mult,
+              reg7, vec_mult, reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vsrli_w, (v4u32)reg0, 13, (v4u32)reg1, 13, (v4u32)reg2, 13,
+              (v4u32)reg3, 13, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrli_w, (v4u32)reg4, 13, (v4u32)reg5, 13, (v4u32)reg6, 13,
+              (v4u32)reg7, 13, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG2(__lsx_vpickev_h, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
+              dst0, dst1, dst2, dst3);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    __lsx_vst(dst2, dst, 32);
+    __lsx_vst(dst3, dst, 48);
+    src += 32;
+    dst += 32;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/files/source/row_mmi.cc b/files/source/row_mmi.cc
index d8726d09..362fd1cf 100644
--- a/files/source/row_mmi.cc
+++ b/files/source/row_mmi.cc
@@ -21,6 +21,8 @@ extern "C" {
 // This module is for Mips MMI.
 #if !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
 
+// clang-format off
+
 void RGB24ToARGBRow_MMI(const uint8_t* src_rgb24,
                         uint8_t* dst_argb,
                         int width) {
@@ -603,7 +605,7 @@ void ARGBToARGB4444Row_MMI(const uint8_t* src_argb,
       : "memory");
 }
 
-void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ARGBToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -611,8 +613,8 @@ void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -624,8 +626,8 @@ void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x08(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -637,8 +639,8 @@ void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x17(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x10(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -650,8 +652,8 @@ void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x1f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x18(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -669,35 +671,38 @@ void ARGBToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x20          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
+void ARGBToUVRow_MMI(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x0026004a00700002;
-  const uint64_t mask_v = 0x00020070005e0012;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -707,15 +712,16 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
       "pinsrh_3   %[dest0_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x08(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -725,7 +731,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -741,8 +748,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x10(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x17(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -752,15 +759,16 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
       "pinsrh_3   %[dest1_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -770,7 +778,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -786,8 +795,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x20(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x27(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -797,15 +806,16 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
       "pinsrh_3   %[dest2_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x28(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -815,7 +825,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -831,8 +842,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x30(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x37(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -842,15 +853,16 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
       "pinsrh_3   %[dest3_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x38(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x3f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -860,7 +872,8 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -888,7 +901,7 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x40              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -898,16 +911,17 @@ void ARGBToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void BGRAToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -915,8 +929,8 @@ void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -928,8 +942,8 @@ void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x08(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -941,8 +955,8 @@ void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x17(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x10(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -954,8 +968,8 @@ void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x1f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x18(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -973,35 +987,38 @@ void BGRAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x20          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
+void BGRAToUVRow_MMI(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x00020070004a0026;
-  const uint64_t mask_v = 0x0012005e00700002;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002f00380002;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1011,15 +1028,16 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
       "pinsrh_0   %[dest0_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x08(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1029,7 +1047,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -1045,8 +1064,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x10(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x17(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1056,15 +1075,16 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
       "pinsrh_0   %[dest1_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1074,7 +1094,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -1090,8 +1111,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x20(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x27(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1101,15 +1122,16 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
       "pinsrh_0   %[dest2_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x28(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1119,7 +1141,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]        \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -1135,8 +1158,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x30(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x37(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1146,15 +1169,16 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
       "pinsrh_0   %[dest3_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x38(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x3f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1164,7 +1188,8 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsrl       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_0   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -1192,7 +1217,7 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x40              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -1202,16 +1227,17 @@ void BGRAToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ABGRToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -1219,8 +1245,8 @@ void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1232,8 +1258,8 @@ void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x08(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1245,8 +1271,8 @@ void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x17(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x10(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1258,8 +1284,8 @@ void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x1f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x18(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1277,35 +1303,38 @@ void ABGRToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x20          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
+void ABGRToUVRow_MMI(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x00020070004a0026;
-  const uint64_t mask_v = 0x0012005e00700002;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002F00380002;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1315,15 +1344,16 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest0_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x08(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1333,7 +1363,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1349,8 +1380,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x10(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x17(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1360,15 +1391,16 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest1_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1378,7 +1410,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1394,8 +1427,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x20(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x27(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1405,15 +1438,16 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest2_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x28(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1423,7 +1457,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1439,8 +1474,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x30(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x37(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1450,15 +1485,16 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest3_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x38(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x3f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1468,7 +1504,8 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1496,7 +1533,7 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x40              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -1506,16 +1543,17 @@ void ABGRToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RGBAToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -1523,8 +1561,8 @@ void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1536,8 +1574,8 @@ void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x08(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x08(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1549,8 +1587,8 @@ void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x17(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x10(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x17(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x10(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1562,8 +1600,8 @@ void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x1f(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x18(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x1f(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x18(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1581,35 +1619,38 @@ void RGBAToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x20          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x20          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
+void RGBAToUVRow_MMI(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x0026004a00700002;
-  const uint64_t mask_v = 0x00020070005e0012;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1619,15 +1660,16 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[dest0_u],      %[src0],           %[value]          \n\t"
       "dsrl       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x08(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1637,7 +1679,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1653,8 +1696,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x10(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x17(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1664,15 +1707,16 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[dest1_u],      %[src0],           %[value]          \n\t"
       "dsrl       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1682,7 +1726,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1698,8 +1743,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x20(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x27(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1709,15 +1754,16 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[dest2_u],      %[src0],           %[value]          \n\t"
       "dsrl       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x28(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1727,7 +1773,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1743,8 +1790,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x30(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x37(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1754,15 +1801,16 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[dest3_u],      %[src0],           %[value]          \n\t"
       "dsrl       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x38(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x3f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1772,7 +1820,8 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "paddh      %[src0],         %[src0],           %[src_lo]         \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_0   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsrl       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_3   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -1800,7 +1849,7 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x40              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -1810,16 +1859,17 @@ void RGBAToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RGB24ToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -1827,8 +1877,8 @@ void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1841,8 +1891,8 @@ void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0d(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x06(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0d(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x06(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1855,8 +1905,8 @@ void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x13(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x0c(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x13(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x0c(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1869,8 +1919,8 @@ void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x19(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x12(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x19(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x12(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -1889,35 +1939,38 @@ void RGB24ToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x18          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x18          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
+void RGB24ToUVRow_MMI(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x0026004a00700002;
-  const uint64_t mask_v = 0x00020070005e0012;
+  const uint64_t mask_u = 0x0013002500380002;
+  const uint64_t mask_v = 0x00020038002f0009;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1929,15 +1982,16 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
       "pinsrh_3   %[dest0_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x06(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0d(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x06(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0d(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x06(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0d(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1949,7 +2003,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -1965,8 +2020,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x0c(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x13(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x0c(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x13(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x0c(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x13(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1978,15 +2033,16 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
       "pinsrh_3   %[dest1_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x12(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x19(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x12(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x19(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x12(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x19(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -1998,7 +2054,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -2014,8 +2071,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2027,15 +2084,16 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
       "pinsrh_3   %[dest2_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x1e(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x25(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x1e(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x25(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x1e(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x25(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2047,7 +2105,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -2063,8 +2122,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x24(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2b(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x24(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2b(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x24(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2b(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2076,15 +2135,16 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
       "pinsrh_3   %[dest3_v],      %[src0],           %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x2a(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x31(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x2a(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x31(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x2a(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x31(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2096,7 +2156,8 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
       "pinsrh_3   %[src_hi],       %[src0],           %[value]          \n\t"
@@ -2124,7 +2185,7 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x30              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x30              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -2134,16 +2195,17 @@ void RGB24ToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RAWToYRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest0, dest1, dest2, dest3;
   const uint64_t value = 0x1080;
@@ -2151,8 +2213,8 @@ void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
 
   __asm__ volatile(
       "1:                                                           \n\t"
-      "gsldlc1    %[src],          0x07(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x00(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x07(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x00(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -2165,8 +2227,8 @@ void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest0],        %[dest0],          %[src]        \n\t"
       "psrlw      %[dest0],        %[dest0],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x0d(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x06(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x0d(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x06(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -2179,8 +2241,8 @@ void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest1],        %[dest1],          %[src]        \n\t"
       "psrlw      %[dest1],        %[dest1],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x13(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x0c(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x13(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x0c(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -2193,8 +2255,8 @@ void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "paddw      %[dest2],        %[dest2],          %[src]        \n\t"
       "psrlw      %[dest2],        %[dest2],          %[eight]      \n\t"
 
-      "gsldlc1    %[src],          0x19(%[src_argb0])               \n\t"
-      "gsldrc1    %[src],          0x12(%[src_argb0])               \n\t"
+      "gsldlc1    %[src],          0x19(%[src_argb])               \n\t"
+      "gsldrc1    %[src],          0x12(%[src_argb])               \n\t"
       "punpcklbh  %[src_lo],       %[src],            %[zero]       \n\t"
       "pinsrh_3   %[src_lo],       %[src_lo],         %[value]      \n\t"
       "pmaddhw    %[src_lo],       %[src_lo],         %[mask]       \n\t"
@@ -2213,35 +2275,38 @@ void RAWToYRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
       "gssdlc1    %[dest0],        0x07(%[dst_y])                   \n\t"
       "gssdrc1    %[dest0],        0x00(%[dst_y])                   \n\t"
 
-      "daddiu     %[src_argb0],    %[src_argb0],      0x18          \n\t"
+      "daddiu     %[src_argb],    %[src_argb],      0x18          \n\t"
       "daddiu     %[dst_y],        %[dst_y],          0x08          \n\t"
       "daddi      %[width],        %[width],         -0x08          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src] "=&f"(src), [src_hi] "=&f"(src_hi), [src_lo] "=&f"(src_lo),
         [dest0] "=&f"(dest0), [dest1] "=&f"(dest1), [dest2] "=&f"(dest2),
         [dest3] "=&f"(dest3)
-      : [src_argb0] "r"(src_argb0), [dst_y] "r"(dst_y), [width] "r"(width),
+      : [src_argb] "r"(src_argb), [dst_y] "r"(dst_y), [width] "r"(width),
         [mask] "f"(mask), [value] "f"(value), [eight] "f"(0x08),
         [zero] "f"(0x00)
       : "memory");
 }
 
-void RAWToUVRow_MMI(const uint8_t* src_rgb0,
+void RAWToUVRow_MMI(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width) {
   uint64_t src_rgb1;
-  uint64_t ftmp[12];
+  uint64_t ftmp[13];
+  uint64_t tmp[1];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x00020070004a0026;
-  const uint64_t mask_v = 0x0012005e00700002;
+  const uint64_t mask_u = 0x0002003800250013;
+  const uint64_t mask_v = 0x0009002f00380002;
 
   __asm__ volatile(
+      "dli        %[tmp0],         0x0001000100010001                   \n\t"
+      "dmtc1      %[tmp0],         %[ftmp12]                            \n\t"
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2253,15 +2318,16 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest0_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest0_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest0_v],      %[dest0_v],        %[value]          \n\t"
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x06(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0d(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x06(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0d(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x06(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0d(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2273,7 +2339,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -2289,8 +2356,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x0c(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x13(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x0c(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x13(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x0c(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x13(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2302,15 +2369,16 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest1_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest1_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest1_v],      %[dest1_v],        %[value]          \n\t"
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x12(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x19(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x12(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x19(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x12(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x19(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2322,7 +2390,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -2338,8 +2407,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2351,15 +2420,16 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest2_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest2_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest2_v],      %[dest2_v],        %[value]          \n\t"
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x1e(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x25(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x1e(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x25(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x1e(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x25(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2371,7 +2441,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -2387,8 +2458,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src0],           %[src1]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x24(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2b(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x24(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2b(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x24(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2b(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2400,15 +2471,16 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[dest3_u],      %[src0],           %[value]          \n\t"
       "dsll       %[dest3_v],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest3_v],      %[dest3_v],        %[value]          \n\t"
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x2a(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x31(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x2a(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x31(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x2a(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x31(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
@@ -2420,7 +2492,8 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "dsll       %[src1],         %[src1],           %[eight]          \n\t"
       "punpckhbh  %[src_hi],       %[src1],           %[zero]           \n\t"
       "paddh      %[src0],         %[src0],           %[src_hi]         \n\t"
-      "psrlh      %[src0],         %[src0],           %[two]            \n\t"
+      "paddh      %[src0],         %[src0],           %[ftmp12]         \n\t"
+      "psrlh      %[src0],         %[src0],           %[one]            \n\t"
       "pinsrh_3   %[src_lo],       %[src0],           %[value]          \n\t"
       "dsll       %[src_hi],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_hi],       %[src_hi],         %[value]          \n\t"
@@ -2448,7 +2521,7 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x30              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x30              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -2458,23 +2531,24 @@ void RAWToUVRow_MMI(const uint8_t* src_rgb0,
         [dest0_u] "=&f"(ftmp[4]), [dest0_v] "=&f"(ftmp[5]),
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
-        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+        [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11]),
+        [ftmp12] "=&f"(ftmp[12]), [tmp0] "=&r"(tmp[0])
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08), [one] "f"(0x01),
         [sixteen] "f"(0x10)
       : "memory");
 }
 
-void ARGBToYJRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ARGBToYJRow_MMI(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   uint64_t src, src_hi, src_lo;
   uint64_t dest, dest0, dest1, dest2, dest3;
   uint64_t tmp0, tmp1;
-  const uint64_t shift = 0x07;
-  const uint64_t value = 0x0040;
+  const uint64_t shift = 0x08;
+  const uint64_t value = 0x80;
   const uint64_t mask0 = 0x0;
-  const uint64_t mask1 = 0x00010026004B000FULL;
+  const uint64_t mask1 = 0x0001004D0096001DULL;
 
   __asm__ volatile(
       "1:                                                           \n\t"
@@ -2544,13 +2618,13 @@ void ARGBToYJRow_MMI(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
         [src_lo] "=&f"(src_lo), [dest0] "=&f"(dest0), [dest1] "=&f"(dest1),
         [dest2] "=&f"(dest2), [dest3] "=&f"(dest3), [tmp0] "=&f"(tmp0),
         [tmp1] "=&f"(tmp1)
-      : [src_ptr] "r"(src_argb0), [dst_ptr] "r"(dst_y), [mask0] "f"(mask0),
+      : [src_ptr] "r"(src_argb), [dst_ptr] "r"(dst_y), [mask0] "f"(mask0),
         [mask1] "f"(mask1), [shift] "f"(shift), [value] "f"(value),
         [width] "r"(width)
       : "memory");
 }
 
-void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
+void ARGBToUVJRow_MMI(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
@@ -2558,22 +2632,22 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
   uint64_t src_rgb1;
   uint64_t ftmp[12];
   const uint64_t value = 0x4040;
-  const uint64_t mask_u = 0x002b0054007f0002;
-  const uint64_t mask_v = 0x0002007f006b0014;
+  const uint64_t mask_u = 0x0015002a003f0002;
+  const uint64_t mask_v = 0x0002003f0035000a;
 
   __asm__ volatile(
       "1:                                                               \n\t"
-      "daddu      %[src_rgb1],     %[src_rgb0],       %[src_stride_rgb] \n\t"
-      "gsldrc1    %[src0],         0x00(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x07(%[src_rgb0])                    \n\t"
+      "daddu      %[src_rgb1],     %[src_rgb],       %[src_stride_rgb] \n\t"
+      "gsldrc1    %[src0],         0x00(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x07(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x00(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x07(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[dest0_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest0_u],      %[dest0_u],        %[value]          \n\t"
@@ -2581,16 +2655,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "pmaddhw    %[dest0_u],      %[dest0_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest0_v],      %[dest0_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x08(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x0f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x08(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x0f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x08(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x0f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
@@ -2607,16 +2681,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest0_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest0_v],      %[dest0_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x10(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x17(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x10(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x17(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x10(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x17(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[dest1_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest1_u],      %[dest1_u],        %[value]          \n\t"
@@ -2624,16 +2698,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "pmaddhw    %[dest1_u],      %[dest1_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest1_v],      %[dest1_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x18(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x1f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x18(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x1f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x18(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x1f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
@@ -2650,16 +2724,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest1_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest1_v],      %[dest1_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x20(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x27(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x20(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x27(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x20(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x27(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[dest2_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest2_u],      %[dest2_u],        %[value]          \n\t"
@@ -2667,16 +2741,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "pmaddhw    %[dest2_u],      %[dest2_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest2_v],      %[dest2_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x28(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x2f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x28(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x2f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x28(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x2f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
@@ -2693,16 +2767,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "psubw      %[dest2_v],      %[src1],           %[src0]           \n\t"
       "psraw      %[dest2_v],      %[dest2_v],        %[eight]          \n\t"
 
-      "gsldrc1    %[src0],         0x30(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x37(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x30(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x37(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x30(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x37(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[dest3_u],      %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[dest3_u],      %[dest3_u],        %[value]          \n\t"
@@ -2710,16 +2784,16 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "pmaddhw    %[dest3_u],      %[dest3_u],        %[mask_u]         \n\t"
       "pmaddhw    %[dest3_v],      %[dest3_v],        %[mask_v]         \n\t"
 
-      "gsldrc1    %[src0],         0x38(%[src_rgb0])                    \n\t"
-      "gsldlc1    %[src0],         0x3f(%[src_rgb0])                    \n\t"
+      "gsldrc1    %[src0],         0x38(%[src_rgb])                    \n\t"
+      "gsldlc1    %[src0],         0x3f(%[src_rgb])                    \n\t"
       "gsldrc1    %[src1],         0x38(%[src_rgb1])                    \n\t"
       "gsldlc1    %[src1],         0x3f(%[src_rgb1])                    \n\t"
       "punpcklbh  %[src_lo],       %[src0],           %[zero]           \n\t"
       "punpckhbh  %[src_hi],       %[src0],           %[zero]           \n\t"
       "punpcklbh  %[src0],         %[src1],           %[zero]           \n\t"
       "punpckhbh  %[src1],         %[src1],           %[zero]           \n\t"
-      "pavgh      %[src0],         %[src_lo],         %[src0]           \n\t"
-      "pavgh      %[src1],         %[src_hi],         %[src1]           \n\t"
+      "paddh      %[src0],         %[src_lo],         %[src0]           \n\t"
+      "paddh      %[src1],         %[src_hi],         %[src1]           \n\t"
       "pavgh      %[src0],         %[src0],           %[src1]           \n\t"
       "dsll       %[src_lo],       %[src0],           %[sixteen]        \n\t"
       "pinsrh_0   %[src_lo],       %[src_lo],         %[value]          \n\t"
@@ -2748,7 +2822,7 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
       "gssdlc1    %[dest0_v],      0x07(%[dst_v])                       \n\t"
       "gssdrc1    %[dest0_v],      0x00(%[dst_v])                       \n\t"
 
-      "daddiu     %[src_rgb0],     %[src_rgb0],       0x40              \n\t"
+      "daddiu     %[src_rgb],     %[src_rgb],       0x40              \n\t"
       "daddiu     %[dst_u],        %[dst_u],          0x08              \n\t"
       "daddiu     %[dst_v],        %[dst_v],          0x08              \n\t"
       "daddi      %[width],        %[width],         -0x10              \n\t"
@@ -2759,10 +2833,10 @@ void ARGBToUVJRow_MMI(const uint8_t* src_rgb0,
         [dest1_u] "=&f"(ftmp[6]), [dest1_v] "=&f"(ftmp[7]),
         [dest2_u] "=&f"(ftmp[8]), [dest2_v] "=&f"(ftmp[9]),
         [dest3_u] "=&f"(ftmp[10]), [dest3_v] "=&f"(ftmp[11])
-      : [src_rgb0] "r"(src_rgb0), [src_stride_rgb] "r"(src_stride_rgb),
+      : [src_rgb] "r"(src_rgb), [src_stride_rgb] "r"(src_stride_rgb),
         [dst_u] "r"(dst_u), [dst_v] "r"(dst_v), [width] "r"(width),
         [mask_u] "f"(mask_u), [mask_v] "f"(mask_v), [value] "f"(value),
-        [zero] "f"(0x00), [eight] "f"(0x08), [two] "f"(0x02),
+        [zero] "f"(0x00), [eight] "f"(0x08),
         [sixteen] "f"(0x10)
       : "memory");
 }
@@ -4052,10 +4126,10 @@ void ARGBGrayRow_MMI(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   uint64_t tmp0, tmp1;
   const uint64_t mask0 = 0x0;
   const uint64_t mask1 = 0x01;
-  const uint64_t mask2 = 0x00400026004B000FULL;
+  const uint64_t mask2 = 0x0080004D0096001DULL;
   const uint64_t mask3 = 0xFF000000FF000000ULL;
   const uint64_t mask4 = ~mask3;
-  const uint64_t shift = 0x07;
+  const uint64_t shift = 0x08;
 
   __asm__ volatile(
       "1:                                                           \n\t"
@@ -4312,7 +4386,7 @@ void ARGBShadeRow_MMI(const uint8_t* src_argb,
       : "memory");
 }
 
-void ARGBMultiplyRow_MMI(const uint8_t* src_argb0,
+void ARGBMultiplyRow_MMI(const uint8_t* src_argb,
                          const uint8_t* src_argb1,
                          uint8_t* dst_argb,
                          int width) {
@@ -4348,12 +4422,12 @@ void ARGBMultiplyRow_MMI(const uint8_t* src_argb0,
         [src1_hi] "=&f"(src1_hi), [src1_lo] "=&f"(src1_lo),
         [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo), [src0] "=&f"(src0),
         [src1] "=&f"(src1), [dest] "=&f"(dest)
-      : [src0_ptr] "r"(src_argb0), [src1_ptr] "r"(src_argb1),
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
         [dst_ptr] "r"(dst_argb), [width] "r"(width), [mask] "f"(mask)
       : "memory");
 }
 
-void ARGBAddRow_MMI(const uint8_t* src_argb0,
+void ARGBAddRow_MMI(const uint8_t* src_argb,
                     const uint8_t* src_argb1,
                     uint8_t* dst_argb,
                     int width) {
@@ -4375,12 +4449,12 @@ void ARGBAddRow_MMI(const uint8_t* src_argb0,
       "daddi      %[width],        %[width],         -0x02          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
-      : [src0_ptr] "r"(src_argb0), [src1_ptr] "r"(src_argb1),
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
         [dst_ptr] "r"(dst_argb), [width] "r"(width)
       : "memory");
 }
 
-void ARGBSubtractRow_MMI(const uint8_t* src_argb0,
+void ARGBSubtractRow_MMI(const uint8_t* src_argb,
                          const uint8_t* src_argb1,
                          uint8_t* dst_argb,
                          int width) {
@@ -4402,7 +4476,7 @@ void ARGBSubtractRow_MMI(const uint8_t* src_argb0,
       "daddi      %[width],        %[width],         -0x02          \n\t"
       "bnez       %[width],        1b                               \n\t"
       : [src0] "=&f"(src0), [src1] "=&f"(src1), [dest] "=&f"(dest)
-      : [src0_ptr] "r"(src_argb0), [src1_ptr] "r"(src_argb1),
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
         [dst_ptr] "r"(dst_argb), [width] "r"(width)
       : "memory");
 }
@@ -4778,7 +4852,9 @@ void J400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* dst_argb, int width) {
       : "memory");
 }
 
-void I400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* rgb_buf, int width) {
+// TODO - respect YuvConstants
+void I400ToARGBRow_MMI(const uint8_t* src_y, uint8_t* rgb_buf,
+                       const struct YuvConstants*, int width) {
   uint64_t src, src_lo, src_hi, dest, dest_lo, dest_hi;
   const uint64_t mask0 = 0x0;
   const uint64_t mask1 = 0x55;
@@ -4912,10 +4988,10 @@ void MirrorRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
       : "memory");
 }
 
-void MirrorUVRow_MMI(const uint8_t* src_uv,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width) {
+void MirrorSplitUVRow_MMI(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
   uint64_t src0, src1, dest0, dest1;
   const uint64_t mask0 = 0x00ff00ff00ff00ffULL;
   const uint64_t mask1 = 0x1b;
@@ -5476,10 +5552,10 @@ void UYVYToYRow_MMI(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
       : "memory");
 }
 
-// Blend src_argb0 over src_argb1 and store to dst_argb.
-// dst_argb may be src_argb0 or src_argb1.
+// Blend src_argb over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb or src_argb1.
 // This code mimics the SSSE3 version for better testability.
-void ARGBBlendRow_MMI(const uint8_t* src_argb0,
+void ARGBBlendRow_MMI(const uint8_t* src_argb,
                       const uint8_t* src_argb1,
                       uint8_t* dst_argb,
                       int width) {
@@ -5532,7 +5608,7 @@ void ARGBBlendRow_MMI(const uint8_t* src_argb0,
         [dest] "=&f"(dest), [src0_hi] "=&f"(src0_hi), [src0_lo] "=&f"(src0_lo),
         [src1_hi] "=&f"(src1_hi), [src1_lo] "=&f"(src1_lo),
         [dest_hi] "=&f"(dest_hi), [dest_lo] "=&f"(dest_lo)
-      : [src0_ptr] "r"(src_argb0), [src1_ptr] "r"(src_argb1),
+      : [src0_ptr] "r"(src_argb), [src1_ptr] "r"(src_argb1),
         [dst_ptr] "r"(dst_argb), [mask0] "f"(mask0), [mask1] "f"(mask1),
         [mask2] "f"(mask2), [mask3] "f"(mask3), [mask4] "f"(mask4),
         [shift] "f"(shift), [width] "r"(width)
@@ -6034,6 +6110,1730 @@ void ARGBCopyYToAlphaRow_MMI(const uint8_t* src, uint8_t* dst, int width) {
       : "memory");
 }
 
+void I444ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+  __asm__ volatile (
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"//yg
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"//bb
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"//ub
+    "or         %[ub],           %[ub],             %[mask]       \n\t"//must sign extension
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"//bg
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"//ug
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"//vg
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"//br
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"//vr
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"//sign extension
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"//u
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"//v
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"//u*ug
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"//v*vg
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"//v*vr
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"//ffffgggg
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//gbgbgbgb
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//frfrfrfr
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x04          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+// Also used for 420
+void I422ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"//yg
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"//bb
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"//ub
+    "or         %[ub],           %[ub],             %[mask]       \n\t"//must sign extension
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"//bg
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"//ug
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"//vg
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"//br
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"//vr
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"//sign extension
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"//v
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"//u*ug
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"//v*vg
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"//v*vr
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"//ffffgggg
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//gbgbgbgb
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"//frfrfrfr
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"//frgbfrgb
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+// 10 bit YUV to ARGB
+void I210ToARGBRow_MMI(const uint16_t* src_y,
+                       const uint16_t* src_u,
+                       const uint16_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[y_ptr])                   \n\t"
+    "gsldrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "psllh      %[y],            %[y],              %[six]        \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "punpcklhw  %[u],            %[u],              %[u]          \n\t"
+    "psrah      %[u],            %[u],              %[two]        \n\t"
+    "punpcklhw  %[v],            %[v],              %[v]          \n\t"
+    "psrah      %[v],            %[v],              %[two]        \n\t"
+    "pminsh     %[u],            %[u],              %[mask1]      \n\t"
+    "pminsh     %[v],            %[v],              %[mask1]      \n\t"
+
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "packushb   %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[alpha]      \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x08          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x04          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),
+      [u]"=&f"(u),                         [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [alpha]"f"(-1),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask),                     [two]"f"(0x02),
+      [mask1]"f"(0x00ff00ff00ff00ff)
+    : "memory"
+  );
+}
+
+void I422AlphaToARGBRow_MMI(const uint8_t* src_y,
+                            const uint8_t* src_u,
+                            const uint8_t* src_v,
+                            const uint8_t* src_a,
+                            uint8_t* rgb_buf,
+                            const struct YuvConstants* yuvconstants,
+                            int width) {
+  uint64_t y,u,v,a;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+    "gslwlc1    %[a],            0x03(%[a_ptr])                   \n\t"
+    "gslwrc1    %[a],            0x00(%[a_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"//rrrrbbbb
+    "packushb   %[g_vec0],       %[g_vec0],         %[a]          \n\t"
+    "punpcklwd  %[g_vec0],       %[g_vec0],         %[a]          \n\t"//aaaagggg
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[g_vec1],       0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[a_ptr],        %[a_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),                         [a]"=&f"(a),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [a_ptr]"r"(src_a),                   [zero]"f"(0x00),
+      [six]"f"(0x6),                       [five]"f"(0x55),
+      [mask]"f"(mask)
+    : "memory"
+  );
+}
+
+void I422ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_u,
+                        const uint8_t* src_v,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y,u,v;
+  uint64_t b_vec[2],g_vec[2],r_vec[2];
+  uint64_t mask = 0xff00ff00ff00ff00ULL;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec0],       %[y],              %[bb]         \n\t"
+    "pmullh     %[b_vec1],       %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec0],       %[b_vec0],         %[b_vec1]     \n\t"
+    "psrah      %[b_vec0],       %[b_vec0],         %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec0],       %[y],              %[bg]         \n\t"
+    "pmullh     %[g_vec1],       %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "pmullh     %[g_vec1],       %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec0],       %[g_vec0],         %[g_vec1]     \n\t"
+    "psrah      %[g_vec0],       %[g_vec0],         %[six]        \n\t"
+
+    "paddsh     %[r_vec0],       %[y],              %[br]         \n\t"
+    "pmullh     %[r_vec1],       %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec0],       %[r_vec0],         %[r_vec1]     \n\t"
+    "psrah      %[r_vec0],       %[r_vec0],         %[six]        \n\t"
+
+    "packushb   %[r_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "packushb   %[g_vec0],       %[g_vec0],         %[zero]       \n\t"
+    "punpcklbh  %[b_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpckhbh  %[r_vec0],       %[r_vec0],         %[g_vec0]     \n\t"
+    "punpcklhw  %[g_vec0],       %[b_vec0],         %[r_vec0]     \n\t"
+    "punpckhhw  %[g_vec1],       %[b_vec0],         %[r_vec0]     \n\t"
+
+    "punpckhwd  %[r_vec0],       %[g_vec0],         %[g_vec0]     \n\t"
+    "psllw      %[r_vec1],       %[r_vec0],         %[lmove1]     \n\t"
+    "or         %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "psrlw      %[r_vec1],       %[r_vec0],         %[rmove1]     \n\t"
+    "pextrh     %[r_vec1],       %[r_vec1],         %[zero]       \n\t"
+    "pinsrh_2   %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "pextrh     %[r_vec1],       %[g_vec1],         %[zero]       \n\t"
+    "pinsrh_3   %[g_vec0],       %[g_vec0],         %[r_vec1]     \n\t"
+    "pextrh     %[r_vec1],       %[g_vec1],         %[one]        \n\t"
+    "punpckhwd  %[g_vec1],       %[g_vec1],         %[g_vec1]     \n\t"
+    "psllw      %[g_vec1],       %[g_vec1],         %[rmove1]     \n\t"
+    "or         %[g_vec1],       %[g_vec1],         %[r_vec1]     \n\t"
+    "gssdlc1    %[g_vec0],       0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec0],       0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[g_vec1],       0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[g_vec1],       0x08(%[rgbbuf_ptr])              \n\t"
+
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0c          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec0]"=&f"(b_vec[0]),             [b_vec1]"=&f"(b_vec[1]),
+      [g_vec0]"=&f"(g_vec[0]),             [g_vec1]"=&f"(g_vec[1]),
+      [r_vec0]"=&f"(r_vec[0]),             [r_vec1]"=&f"(r_vec[1]),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask]"f"(mask),
+      [lmove1]"f"(0x18),                   [rmove1]"f"(0x8),
+      [one]"f"(0x1)
+    : "memory"
+  );
+}
+
+void I422ToARGB4444Row_MMI(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb4444,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask]       \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask]       \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"//y*0x0101
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"//y1
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"//u
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "and        %[g_vec],        %[g_vec],          %[mask1]      \n\t"
+    "psrlw      %[g_vec],        %[g_vec],          %[four]       \n\t"
+    "psrlw      %[r_vec],        %[g_vec],          %[four]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpcklbh  %[r_vec],        %[alpha],          %[zero]       \n\t"
+    "and        %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "and        %[b_vec],        %[b_vec],          %[mask1]      \n\t"
+    "psrlw      %[b_vec],        %[b_vec],          %[four]       \n\t"
+    "psrlw      %[r_vec],        %[b_vec],          %[four]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpcklbh  %[r_vec],        %[alpha],          %[zero]       \n\t"
+    "and        %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_argb4444])            \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_argb4444])            \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_argb4444], %[dst_argb4444],   0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_argb4444]"r"(dst_argb4444),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask]"f"(0xff00ff00ff00ff00),
+      [four]"f"(0x4),                      [mask1]"f"(0xf0f0f0f0f0f0f0f0),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void I422ToARGB1555Row_MMI(const uint8_t* src_y,
+                           const uint8_t* src_u,
+                           const uint8_t* src_v,
+                           uint8_t* dst_argb1555,
+                           const struct YuvConstants* yuvconstants,
+                           int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlw      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "or         %[g_vec],        %[g_vec],          %[mask3]      \n\t"
+
+    "psrlw      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "or         %[b_vec],        %[b_vec],          %[mask3]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_argb1555])            \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_argb1555])            \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_argb1555], %[dst_argb1555],   0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_argb1555]"r"(dst_argb1555),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [mask3]"f"(0x800000008000),
+      [lmove5]"f"(0x5)
+    : "memory"
+  );
+}
+
+void I422ToRGB565Row_MMI(const uint8_t* src_y,
+                         const uint8_t* src_u,
+                         const uint8_t* src_v,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    //u3|u2|u1|u0 --> u1|u1|u0|u0
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    //v3|v2|v1|v0 --> v1|v1|v0|v0
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[lmove5]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_rgb565])             \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_rgb565])             \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[dst_rgb565],   %[dst_rgb565],     0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [dst_rgb565]"r"(dst_rgb565),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [seven]"f"(0x7),
+      [lmove5]"f"(0x5)
+    : "memory"
+  );
+}
+
+void NV12ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_uv,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void NV21ToARGBRow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_vu,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[vu_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[vu_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[ushu]       \n\t"
+    "pshufh     %[u],            %[u],              %[vshu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[vu_ptr],       %[vu_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [vu_ptr]"r"(src_vu),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void NV12ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_uv,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpckhwd  %[r_vec],        %[g_vec],          %[g_vec]      \n\t"
+    "psllw      %[temp],         %[r_vec],          %[lmove1]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrlw      %[temp],         %[r_vec],          %[rmove1]     \n\t"
+    "pextrh     %[temp],         %[temp],           %[zero]       \n\t"
+    "pinsrh_2   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[zero]       \n\t"
+    "pinsrh_3   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[one]        \n\t"
+    "punpckhwd  %[b_vec],        %[b_vec],          %[b_vec]      \n\t"
+    "psllw      %[b_vec],        %[b_vec],          %[rmove1]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[b_vec],        0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0C          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [lmove1]"f"(0x18),
+      [one]"f"(0x1),                       [rmove1]"f"(0x8)
+    : "memory"
+  );
+}
+
+void NV21ToRGB24Row_MMI(const uint8_t* src_y,
+                        const uint8_t* src_vu,
+                        uint8_t* rgb_buf,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[vu_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[vu_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[ushu]       \n\t"
+    "pshufh     %[u],            %[u],              %[vshu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpckhwd  %[r_vec],        %[g_vec],          %[g_vec]      \n\t"
+    "psllw      %[temp],         %[r_vec],          %[lmove1]     \n\t"
+    "or         %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrlw      %[temp],         %[r_vec],          %[rmove1]     \n\t"
+    "pextrh     %[temp],         %[temp],           %[zero]       \n\t"
+    "pinsrh_2   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[zero]       \n\t"
+    "pinsrh_3   %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pextrh     %[temp],         %[b_vec],          %[one]        \n\t"
+    "punpckhwd  %[b_vec],        %[b_vec],          %[b_vec]      \n\t"
+    "psllw      %[b_vec],        %[b_vec],          %[rmove1]     \n\t"
+    "or         %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gsswlc1    %[b_vec],        0x0b(%[rgbbuf_ptr])              \n\t"
+    "gsswrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[vu_ptr],       %[vu_ptr],         0x04          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x0C          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [vu_ptr]"r"(src_vu),
+      [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [lmove1]"f"(0x18),                   [rmove1]"f"(0x8),
+      [one]"f"(0x1)
+    : "memory"
+  );
+}
+
+void NV12ToRGB565Row_MMI(const uint8_t* src_y,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_rgb565,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[uv_ptr])                  \n\t"
+    "gslwrc1    %[u],            0x00(%[uv_ptr])                  \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "pshufh     %[v],            %[u],              %[vshu]       \n\t"
+    "pshufh     %[u],            %[u],              %[ushu]       \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[g_vec],          %[three]      \n\t"
+    "and        %[g_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psubb      %[y],            %[eight],          %[three]      \n\t"//5
+    "psllw      %[r_vec],        %[r_vec],          %[y]          \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[g_vec],        %[g_vec],          %[r_vec]      \n\t"
+
+    "psrlh      %[temp],         %[b_vec],          %[three]      \n\t"
+    "and        %[b_vec],        %[temp],           %[mask2]      \n\t"
+    "psrlw      %[temp],         %[temp],           %[seven]      \n\t"
+    "psrlw      %[r_vec],        %[mask1],          %[eight]      \n\t"
+    "and        %[r_vec],        %[temp],           %[r_vec]      \n\t"
+    "psubb      %[y],            %[eight],          %[three]      \n\t"//5
+    "psllw      %[r_vec],        %[r_vec],          %[y]          \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "paddb      %[r_vec],        %[three],          %[six]        \n\t"
+    "psrlw      %[temp],         %[temp],           %[r_vec]      \n\t"
+    "and        %[r_vec],        %[temp],           %[mask2]      \n\t"
+    "paddb      %[temp],         %[three],          %[eight]      \n\t"
+    "psllw      %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "or         %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "punpcklhw  %[r_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[g_vec],          %[b_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[r_vec],          %[b_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[dst_rgb565])             \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[dst_rgb565])             \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+	"daddiu     %[uv_ptr],       %[uv_ptr],         0x04          \n\t"
+    "daddiu     %[dst_rgb565],   %[dst_rgb565],     0x08          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [uv_ptr]"r"(src_uv),
+      [dst_rgb565]"r"(dst_rgb565),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [three]"f"(0x3),                     [mask2]"f"(0x1f0000001f),
+      [eight]"f"(0x8),                     [seven]"f"(0x7)
+    : "memory"
+  );
+}
+
+void YUY2ToARGBRow_MMI(const uint8_t* src_yuy2,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[yuy2_ptr])                \n\t"
+    "gsldrc1    %[y],            0x00(%[yuy2_ptr])                \n\t"
+    "psrlh      %[temp],         %[y],              %[eight]      \n\t"
+    "pshufh     %[u],            %[temp],           %[ushu]       \n\t"
+    "pshufh     %[v],            %[temp],           %[vshu]       \n\t"
+
+    "psrlh      %[temp],         %[mask1],          %[eight]      \n\t"
+    "and        %[y],            %[y],              %[temp]       \n\t"
+    "psllh      %[temp],         %[y],              %[eight]      \n\t"
+    "or         %[y],            %[y],              %[temp]       \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[b_vec],        0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[yuy2_ptr],     %[yuy2_ptr],       0x08          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [yuy2_ptr]"r"(src_yuy2),             [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [eight]"f"(0x8)
+    : "memory"
+  );
+}
+
+void UYVYToARGBRow_MMI(const uint8_t* src_uyvy,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gsldlc1    %[y],            0x07(%[uyvy_ptr])                \n\t"
+    "gsldrc1    %[y],            0x00(%[uyvy_ptr])                \n\t"
+    "psrlh      %[temp],         %[mask1],          %[eight]      \n\t"
+    "and        %[temp],         %[y],              %[temp]       \n\t"
+    "pshufh     %[u],            %[temp],           %[ushu]       \n\t"
+    "pshufh     %[v],            %[temp],           %[vshu]       \n\t"
+
+    "psrlh      %[y],            %[y],              %[eight]      \n\t"
+    "psllh      %[temp],         %[y],              %[eight]      \n\t"
+    "or         %[y],            %[y],              %[temp]       \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[g_vec],          %[alpha]      \n\t"
+    "punpcklbh  %[b_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[r_vec],          %[g_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],        0x07(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[g_vec],        0x00(%[rgbbuf_ptr])              \n\t"
+    "gssdlc1    %[b_vec],        0x0f(%[rgbbuf_ptr])              \n\t"
+    "gssdrc1    %[b_vec],        0x08(%[rgbbuf_ptr])              \n\t"
+
+    "daddiu     %[uyvy_ptr],     %[uyvy_ptr],       0x08          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [uyvy_ptr]"r"(src_uyvy),             [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [ushu]"f"(0xA0),                     [vshu]"f"(0xf5),
+      [alpha]"f"(-1),                      [eight]"f"(0x8)
+    : "memory"
+  );
+}
+
+void I422ToRGBARow_MMI(const uint8_t* src_y,
+                       const uint8_t* src_u,
+                       const uint8_t* src_v,
+                       uint8_t* rgb_buf,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
+  uint64_t y, u, v;
+  uint64_t b_vec, g_vec, r_vec, temp;
+  uint64_t ub,ug,vg,vr,bb,bg,br,yg;
+
+  __asm__ volatile(
+    "ldc1       %[yg],           0xc0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[bb],           0x60(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ub],           0x00(%[yuvcons_ptr])             \n\t"
+    "or         %[ub],           %[ub],             %[mask1]      \n\t"
+    "ldc1       %[bg],           0x80(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[ug],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[ug],           %[ug],             %[zero]       \n\t"
+    "pshufh     %[ug],           %[ug],             %[zero]       \n\t"
+    "ldc1       %[vg],           0x20(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vg],           %[vg],             %[zero]       \n\t"
+    "pshufh     %[vg],           %[vg],             %[five]       \n\t"
+    "ldc1       %[br],           0xa0(%[yuvcons_ptr])             \n\t"
+    "ldc1       %[vr],           0x40(%[yuvcons_ptr])             \n\t"
+    "punpcklbh  %[vr],           %[vr],             %[zero]       \n\t"
+    "pshufh     %[vr],           %[vr],             %[five]       \n\t"
+    "or         %[vr],           %[vr],             %[mask1]      \n\t"
+
+    "1:                                                           \n\t"
+    "gslwlc1    %[y],            0x03(%[y_ptr])                   \n\t"
+    "gslwrc1    %[y],            0x00(%[y_ptr])                   \n\t"
+    "gslwlc1    %[u],            0x03(%[u_ptr])                   \n\t"
+    "gslwrc1    %[u],            0x00(%[u_ptr])                   \n\t"
+    "gslwlc1    %[v],            0x03(%[v_ptr])                   \n\t"
+    "gslwrc1    %[v],            0x00(%[v_ptr])                   \n\t"
+
+    "punpcklbh  %[y],            %[y],              %[y]          \n\t"
+    "pmulhuh    %[y],            %[y],              %[yg]         \n\t"
+
+    "punpcklbh  %[u],            %[u],              %[u]          \n\t"
+    "punpcklbh  %[u],            %[u],              %[zero]       \n\t"
+    "paddsh     %[b_vec],        %[y],              %[bb]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ub]         \n\t"
+    "psubsh     %[b_vec],        %[b_vec],          %[temp]       \n\t"
+    "psrah      %[b_vec],        %[b_vec],          %[six]        \n\t"
+
+    "punpcklbh  %[v],            %[v],              %[v]          \n\t"
+    "punpcklbh  %[v],            %[v],              %[zero]       \n\t"
+    "paddsh     %[g_vec],        %[y],              %[bg]         \n\t"
+    "pmullh     %[temp],         %[u],              %[ug]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "pmullh     %[temp],         %[v],              %[vg]         \n\t"
+    "psubsh     %[g_vec],        %[g_vec],          %[temp]       \n\t"
+    "psrah      %[g_vec],        %[g_vec],          %[six]        \n\t"
+
+    "paddsh     %[r_vec],        %[y],              %[br]         \n\t"
+    "pmullh     %[temp],         %[v],              %[vr]         \n\t"
+    "psubsh     %[r_vec],        %[r_vec],          %[temp]       \n\t"
+    "psrah      %[r_vec],        %[r_vec],          %[six]        \n\t"
+
+    "packushb   %[r_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "packushb   %[g_vec],        %[g_vec],          %[zero]       \n\t"
+    "punpcklwd  %[g_vec],        %[alpha],          %[g_vec]      \n\t"
+    "punpcklbh  %[b_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpckhbh  %[r_vec],        %[g_vec],          %[r_vec]      \n\t"
+    "punpcklhw  %[g_vec],        %[b_vec],          %[r_vec]      \n\t"
+    "punpckhhw  %[b_vec],        %[b_vec],          %[r_vec]      \n\t"
+
+    "gssdlc1    %[g_vec],       0x07(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[g_vec],       0x00(%[rgbbuf_ptr])               \n\t"
+    "gssdlc1    %[b_vec],       0x0f(%[rgbbuf_ptr])               \n\t"
+    "gssdrc1    %[b_vec],       0x08(%[rgbbuf_ptr])               \n\t"
+
+    "daddiu     %[y_ptr],        %[y_ptr],          0x04          \n\t"
+    "daddiu     %[u_ptr],        %[u_ptr],          0x02          \n\t"
+    "daddiu     %[v_ptr],        %[v_ptr],          0x02          \n\t"
+    "daddiu     %[rgbbuf_ptr],   %[rgbbuf_ptr],     0x10          \n\t"
+    "daddi      %[width],        %[width],          -0x04         \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [y]"=&f"(y),                         [u]"=&f"(u),
+      [v]"=&f"(v),
+      [b_vec]"=&f"(b_vec),                 [g_vec]"=&f"(g_vec),
+      [r_vec]"=&f"(r_vec),                 [temp]"=&f"(temp),
+      [ub]"=&f"(ub),                       [ug]"=&f"(ug),
+      [vg]"=&f"(vg),                       [vr]"=&f"(vr),
+      [bb]"=&f"(bb),                       [bg]"=&f"(bg),
+      [br]"=&f"(br),                       [yg]"=&f"(yg)
+    : [y_ptr]"r"(src_y),                   [u_ptr]"r"(src_u),
+      [v_ptr]"r"(src_v),                   [rgbbuf_ptr]"r"(rgb_buf),
+      [yuvcons_ptr]"r"(yuvconstants),      [width]"r"(width),
+      [zero]"f"(0x00),                     [five]"f"(0x55),
+      [six]"f"(0x6),                       [mask1]"f"(0xff00ff00ff00ff00),
+      [alpha]"f"(-1)
+    : "memory"
+  );
+}
+
+void ARGBSetRow_MMI(uint8_t* dst_argb, uint32_t v32, int width) {
+  __asm__ volatile (
+    "punpcklwd  %[v32],          %[v32],            %[v32]        \n\t"
+    "1:                                                           \n\t"
+    "gssdlc1    %[v32],          0x07(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[v32],          0x00(%[dst_ptr])                 \n\t"
+    "gssdlc1    %[v32],          0x0f(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[v32],          0x08(%[dst_ptr])                 \n\t"
+
+    "daddi      %[width],        %[width],         -0x04          \n\t"
+    "daddiu     %[dst_ptr],      %[dst_ptr],        0x10          \n\t"
+    "bnez       %[width],        1b                               \n\t"
+    : [v32]"+&f"(v32)
+    : [dst_ptr]"r"(dst_argb),           [width]"r"(width)
+    : "memory"
+  );
+}
+// clang-format on
+
+// 10 bit YUV to ARGB
 #endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
 
 #ifdef __cplusplus
diff --git a/files/source/row_msa.cc b/files/source/row_msa.cc
index 5c0239a3..b7d5bb5e 100644
--- a/files/source/row_msa.cc
+++ b/files/source/row_msa.cc
@@ -24,16 +24,14 @@ extern "C" {
 #define ALPHA_VAL (-1)
 
 // Fill YUV -> RGB conversion constants into vectors
-#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, bb, bg, br, yg) \
-  {                                                              \
-    ub = __msa_fill_w(yuvconst->kUVToB[0]);                      \
-    vr = __msa_fill_w(yuvconst->kUVToR[1]);                      \
-    ug = __msa_fill_w(yuvconst->kUVToG[0]);                      \
-    vg = __msa_fill_w(yuvconst->kUVToG[1]);                      \
-    bb = __msa_fill_w(yuvconst->kUVBiasB[0]);                    \
-    bg = __msa_fill_w(yuvconst->kUVBiasG[0]);                    \
-    br = __msa_fill_w(yuvconst->kUVBiasR[0]);                    \
-    yg = __msa_fill_w(yuvconst->kYToRgb[0]);                     \
+#define YUVTORGB_SETUP(yuvconst, ub, vr, ug, vg, yg, yb) \
+  {                                                      \
+    ub = __msa_fill_w(yuvconst->kUVToB[0]);              \
+    vr = __msa_fill_w(yuvconst->kUVToR[1]);              \
+    ug = __msa_fill_w(yuvconst->kUVToG[0]);              \
+    vg = __msa_fill_w(yuvconst->kUVToG[1]);              \
+    yg = __msa_fill_w(yuvconst->kYToRgb[0]);             \
+    yb = __msa_fill_w(yuvconst->kYBiasToRgb[0]);         \
   }
 
 // Load YUV 422 pixel data
@@ -70,54 +68,52 @@ extern "C" {
   }
 
 // Convert 8 pixels of YUV 420 to RGB.
-#define YUVTORGB(in_y, in_uv, ubvr, ugvg, bb, bg, br, yg, out_b, out_g, out_r) \
-  {                                                                            \
-    v8i16 vec0_m, vec1_m;                                                      \
-    v4i32 reg0_m, reg1_m, reg2_m, reg3_m, reg4_m;                              \
-    v4i32 reg5_m, reg6_m, reg7_m;                                              \
-    v16i8 zero_m = {0};                                                        \
-                                                                               \
-    vec0_m = (v8i16)__msa_ilvr_b((v16i8)in_y, (v16i8)in_y);                    \
-    vec1_m = (v8i16)__msa_ilvr_b((v16i8)zero_m, (v16i8)in_uv);                 \
-    reg0_m = (v4i32)__msa_ilvr_h((v8i16)zero_m, (v8i16)vec0_m);                \
-    reg1_m = (v4i32)__msa_ilvl_h((v8i16)zero_m, (v8i16)vec0_m);                \
-    reg2_m = (v4i32)__msa_ilvr_h((v8i16)zero_m, (v8i16)vec1_m);                \
-    reg3_m = (v4i32)__msa_ilvl_h((v8i16)zero_m, (v8i16)vec1_m);                \
-    reg0_m *= yg;                                                              \
-    reg1_m *= yg;                                                              \
-    reg2_m *= ubvr;                                                            \
-    reg3_m *= ubvr;                                                            \
-    reg0_m = __msa_srai_w(reg0_m, 16);                                         \
-    reg1_m = __msa_srai_w(reg1_m, 16);                                         \
-    reg4_m = __msa_dotp_s_w((v8i16)vec1_m, (v8i16)ugvg);                       \
-    reg5_m = __msa_ilvev_w(reg2_m, reg2_m);                                    \
-    reg6_m = __msa_ilvev_w(reg3_m, reg3_m);                                    \
-    reg7_m = __msa_ilvr_w(reg4_m, reg4_m);                                     \
-    reg2_m = __msa_ilvod_w(reg2_m, reg2_m);                                    \
-    reg3_m = __msa_ilvod_w(reg3_m, reg3_m);                                    \
-    reg4_m = __msa_ilvl_w(reg4_m, reg4_m);                                     \
-    reg5_m = reg0_m - reg5_m;                                                  \
-    reg6_m = reg1_m - reg6_m;                                                  \
-    reg2_m = reg0_m - reg2_m;                                                  \
-    reg3_m = reg1_m - reg3_m;                                                  \
-    reg7_m = reg0_m - reg7_m;                                                  \
-    reg4_m = reg1_m - reg4_m;                                                  \
-    reg5_m += bb;                                                              \
-    reg6_m += bb;                                                              \
-    reg7_m += bg;                                                              \
-    reg4_m += bg;                                                              \
-    reg2_m += br;                                                              \
-    reg3_m += br;                                                              \
-    reg5_m = __msa_srai_w(reg5_m, 6);                                          \
-    reg6_m = __msa_srai_w(reg6_m, 6);                                          \
-    reg7_m = __msa_srai_w(reg7_m, 6);                                          \
-    reg4_m = __msa_srai_w(reg4_m, 6);                                          \
-    reg2_m = __msa_srai_w(reg2_m, 6);                                          \
-    reg3_m = __msa_srai_w(reg3_m, 6);                                          \
-    CLIP_0TO255(reg5_m, reg6_m, reg7_m, reg4_m, reg2_m, reg3_m);               \
-    out_b = __msa_pckev_h((v8i16)reg6_m, (v8i16)reg5_m);                       \
-    out_g = __msa_pckev_h((v8i16)reg4_m, (v8i16)reg7_m);                       \
-    out_r = __msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);                       \
+#define YUVTORGB(in_y, in_uv, ubvr, ugvg, yg, yb, out_b, out_g, out_r) \
+  {                                                                    \
+    v8i16 vec0_m, vec1_m;                                              \
+    v4i32 reg0_m, reg1_m, reg2_m, reg3_m, reg4_m;                      \
+    v4i32 reg5_m, reg6_m, reg7_m;                                      \
+    v16i8 temp_m, zero_m = {0};                                        \
+                                                                       \
+    vec0_m = (v8i16)__msa_ilvr_b((v16i8)in_y, (v16i8)in_y);            \
+    vec1_m = (v8i16)__msa_ilvr_b((v16i8)zero_m, (v16i8)in_uv);         \
+    reg0_m = (v4i32)__msa_ilvr_h((v8i16)zero_m, (v8i16)vec0_m);        \
+    reg1_m = (v4i32)__msa_ilvl_h((v8i16)zero_m, (v8i16)vec0_m);        \
+    vec1_m = (v8i16)__msa_subv_h(vec1_m, const_0x80);                  \
+    temp_m = (v16i8)__msa_clti_s_h(vec1_m, 0);                         \
+    reg2_m = (v4i32)__msa_ilvr_h((v8i16)temp_m, (v8i16)vec1_m);        \
+    reg3_m = (v4i32)__msa_ilvl_h((v8i16)temp_m, (v8i16)vec1_m);        \
+    reg0_m *= yg;                                                      \
+    reg1_m *= yg;                                                      \
+    reg2_m *= ubvr;                                                    \
+    reg3_m *= ubvr;                                                    \
+    reg0_m = __msa_srai_w(reg0_m, 16);                                 \
+    reg1_m = __msa_srai_w(reg1_m, 16);                                 \
+    reg0_m += yb;                                                      \
+    reg1_m += yb;                                                      \
+    reg4_m = __msa_dotp_s_w((v8i16)vec1_m, (v8i16)ugvg);               \
+    reg5_m = __msa_ilvev_w(reg2_m, reg2_m);                            \
+    reg6_m = __msa_ilvev_w(reg3_m, reg3_m);                            \
+    reg7_m = __msa_ilvr_w(reg4_m, reg4_m);                             \
+    reg2_m = __msa_ilvod_w(reg2_m, reg2_m);                            \
+    reg3_m = __msa_ilvod_w(reg3_m, reg3_m);                            \
+    reg4_m = __msa_ilvl_w(reg4_m, reg4_m);                             \
+    reg5_m = reg0_m + reg5_m;                                          \
+    reg6_m = reg1_m + reg6_m;                                          \
+    reg2_m = reg0_m + reg2_m;                                          \
+    reg3_m = reg1_m + reg3_m;                                          \
+    reg7_m = reg0_m - reg7_m;                                          \
+    reg4_m = reg1_m - reg4_m;                                          \
+    reg5_m = __msa_srai_w(reg5_m, 6);                                  \
+    reg6_m = __msa_srai_w(reg6_m, 6);                                  \
+    reg7_m = __msa_srai_w(reg7_m, 6);                                  \
+    reg4_m = __msa_srai_w(reg4_m, 6);                                  \
+    reg2_m = __msa_srai_w(reg2_m, 6);                                  \
+    reg3_m = __msa_srai_w(reg3_m, 6);                                  \
+    CLIP_0TO255(reg5_m, reg6_m, reg7_m, reg4_m, reg2_m, reg3_m);       \
+    out_b = __msa_pckev_h((v8i16)reg6_m, (v8i16)reg5_m);               \
+    out_g = __msa_pckev_h((v8i16)reg4_m, (v8i16)reg7_m);               \
+    out_r = __msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);               \
   }
 
 // Pack and Store 8 ARGB values.
@@ -155,11 +151,10 @@ extern "C" {
   }
 
 // Loads current and next row of ARGB input and averages it to calculate U and V
-#define READ_ARGB(s_ptr, t_ptr, argb0, argb1, argb2, argb3)               \
+#define READ_ARGB(s_ptr, t_ptr, argb0, argb1, argb2, argb3, const_0x0101) \
   {                                                                       \
     v16u8 src0_m, src1_m, src2_m, src3_m, src4_m, src5_m, src6_m, src7_m; \
     v16u8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m; \
-    v16u8 vec8_m, vec9_m;                                                 \
     v8u16 reg0_m, reg1_m, reg2_m, reg3_m, reg4_m, reg5_m, reg6_m, reg7_m; \
     v8u16 reg8_m, reg9_m;                                                 \
                                                                           \
@@ -195,81 +190,81 @@ extern "C" {
     reg1_m = (v8u16)__msa_pckev_d((v2i64)reg7_m, (v2i64)reg3_m);          \
     reg0_m += (v8u16)__msa_pckod_d((v2i64)reg6_m, (v2i64)reg2_m);         \
     reg1_m += (v8u16)__msa_pckod_d((v2i64)reg7_m, (v2i64)reg3_m);         \
-    reg8_m = (v8u16)__msa_srai_h((v8i16)reg8_m, 2);                       \
-    reg9_m = (v8u16)__msa_srai_h((v8i16)reg9_m, 2);                       \
-    reg0_m = (v8u16)__msa_srai_h((v8i16)reg0_m, 2);                       \
-    reg1_m = (v8u16)__msa_srai_h((v8i16)reg1_m, 2);                       \
-    argb0 = (v16u8)__msa_pckev_b((v16i8)reg9_m, (v16i8)reg8_m);           \
-    argb1 = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m);           \
-    src0_m = (v16u8)__msa_ld_b((void*)s, 64);                             \
-    src1_m = (v16u8)__msa_ld_b((void*)s, 80);                             \
-    src2_m = (v16u8)__msa_ld_b((void*)s, 96);                             \
-    src3_m = (v16u8)__msa_ld_b((void*)s, 112);                            \
-    src4_m = (v16u8)__msa_ld_b((void*)t, 64);                             \
-    src5_m = (v16u8)__msa_ld_b((void*)t, 80);                             \
-    src6_m = (v16u8)__msa_ld_b((void*)t, 96);                             \
-    src7_m = (v16u8)__msa_ld_b((void*)t, 112);                            \
-    vec2_m = (v16u8)__msa_ilvr_b((v16i8)src0_m, (v16i8)src4_m);           \
-    vec3_m = (v16u8)__msa_ilvr_b((v16i8)src1_m, (v16i8)src5_m);           \
-    vec4_m = (v16u8)__msa_ilvr_b((v16i8)src2_m, (v16i8)src6_m);           \
-    vec5_m = (v16u8)__msa_ilvr_b((v16i8)src3_m, (v16i8)src7_m);           \
-    vec6_m = (v16u8)__msa_ilvl_b((v16i8)src0_m, (v16i8)src4_m);           \
-    vec7_m = (v16u8)__msa_ilvl_b((v16i8)src1_m, (v16i8)src5_m);           \
-    vec8_m = (v16u8)__msa_ilvl_b((v16i8)src2_m, (v16i8)src6_m);           \
-    vec9_m = (v16u8)__msa_ilvl_b((v16i8)src3_m, (v16i8)src7_m);           \
-    reg0_m = __msa_hadd_u_h(vec2_m, vec2_m);                              \
-    reg1_m = __msa_hadd_u_h(vec3_m, vec3_m);                              \
-    reg2_m = __msa_hadd_u_h(vec4_m, vec4_m);                              \
-    reg3_m = __msa_hadd_u_h(vec5_m, vec5_m);                              \
-    reg4_m = __msa_hadd_u_h(vec6_m, vec6_m);                              \
-    reg5_m = __msa_hadd_u_h(vec7_m, vec7_m);                              \
-    reg6_m = __msa_hadd_u_h(vec8_m, vec8_m);                              \
-    reg7_m = __msa_hadd_u_h(vec9_m, vec9_m);                              \
-    reg8_m = (v8u16)__msa_pckev_d((v2i64)reg4_m, (v2i64)reg0_m);          \
-    reg9_m = (v8u16)__msa_pckev_d((v2i64)reg5_m, (v2i64)reg1_m);          \
-    reg8_m += (v8u16)__msa_pckod_d((v2i64)reg4_m, (v2i64)reg0_m);         \
-    reg9_m += (v8u16)__msa_pckod_d((v2i64)reg5_m, (v2i64)reg1_m);         \
-    reg0_m = (v8u16)__msa_pckev_d((v2i64)reg6_m, (v2i64)reg2_m);          \
-    reg1_m = (v8u16)__msa_pckev_d((v2i64)reg7_m, (v2i64)reg3_m);          \
-    reg0_m += (v8u16)__msa_pckod_d((v2i64)reg6_m, (v2i64)reg2_m);         \
-    reg1_m += (v8u16)__msa_pckod_d((v2i64)reg7_m, (v2i64)reg3_m);         \
-    reg8_m = (v8u16)__msa_srai_h((v8i16)reg8_m, 2);                       \
-    reg9_m = (v8u16)__msa_srai_h((v8i16)reg9_m, 2);                       \
-    reg0_m = (v8u16)__msa_srai_h((v8i16)reg0_m, 2);                       \
-    reg1_m = (v8u16)__msa_srai_h((v8i16)reg1_m, 2);                       \
-    argb2 = (v16u8)__msa_pckev_b((v16i8)reg9_m, (v16i8)reg8_m);           \
-    argb3 = (v16u8)__msa_pckev_b((v16i8)reg1_m, (v16i8)reg0_m);           \
+    reg8_m += const_0x0101;                                               \
+    reg9_m += const_0x0101;                                               \
+    reg0_m += const_0x0101;                                               \
+    reg1_m += const_0x0101;                                               \
+    argb0 = (v8u16)__msa_srai_h((v8i16)reg8_m, 1);                        \
+    argb1 = (v8u16)__msa_srai_h((v8i16)reg9_m, 1);                        \
+    argb2 = (v8u16)__msa_srai_h((v8i16)reg0_m, 1);                        \
+    argb3 = (v8u16)__msa_srai_h((v8i16)reg1_m, 1);                        \
   }
 
-// Takes ARGB input and calculates U and V.
 #define ARGBTOUV(argb0, argb1, argb2, argb3, const0, const1, const2, const3, \
-                 shf0, shf1, shf2, shf3, v_out, u_out)                       \
+                 shf0, shf1, shf2, shf3, shift, u_out, v_out)                \
   {                                                                          \
-    v16u8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
-    v8u16 reg0_m, reg1_m, reg2_m, reg3_m;                                    \
+    v8u16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
+    v4u32 reg0_m, reg1_m, reg2_m, reg3_m;                                    \
                                                                              \
-    vec0_m = (v16u8)__msa_vshf_b(shf0, (v16i8)argb1, (v16i8)argb0);          \
-    vec1_m = (v16u8)__msa_vshf_b(shf0, (v16i8)argb3, (v16i8)argb2);          \
-    vec2_m = (v16u8)__msa_vshf_b(shf1, (v16i8)argb1, (v16i8)argb0);          \
-    vec3_m = (v16u8)__msa_vshf_b(shf1, (v16i8)argb3, (v16i8)argb2);          \
-    vec4_m = (v16u8)__msa_vshf_b(shf2, (v16i8)argb1, (v16i8)argb0);          \
-    vec5_m = (v16u8)__msa_vshf_b(shf2, (v16i8)argb3, (v16i8)argb2);          \
-    vec6_m = (v16u8)__msa_vshf_b(shf3, (v16i8)argb1, (v16i8)argb0);          \
-    vec7_m = (v16u8)__msa_vshf_b(shf3, (v16i8)argb3, (v16i8)argb2);          \
-    reg0_m = __msa_dotp_u_h(vec0_m, const1);                                 \
-    reg1_m = __msa_dotp_u_h(vec1_m, const1);                                 \
-    reg2_m = __msa_dotp_u_h(vec4_m, const1);                                 \
-    reg3_m = __msa_dotp_u_h(vec5_m, const1);                                 \
-    reg0_m += const3;                                                        \
-    reg1_m += const3;                                                        \
-    reg2_m += const3;                                                        \
-    reg3_m += const3;                                                        \
-    reg0_m -= __msa_dotp_u_h(vec2_m, const0);                                \
-    reg1_m -= __msa_dotp_u_h(vec3_m, const0);                                \
-    reg2_m -= __msa_dotp_u_h(vec6_m, const2);                                \
-    reg3_m -= __msa_dotp_u_h(vec7_m, const2);                                \
-    v_out = (v16u8)__msa_pckod_b((v16i8)reg1_m, (v16i8)reg0_m);              \
-    u_out = (v16u8)__msa_pckod_b((v16i8)reg3_m, (v16i8)reg2_m);              \
+    vec0_m = (v8u16)__msa_vshf_h(shf0, (v16i8)argb1, (v16i8)argb0);          \
+    vec1_m = (v8u16)__msa_vshf_h(shf0, (v16i8)argb3, (v16i8)argb2);          \
+    vec2_m = (v8u16)__msa_vshf_h(shf1, (v16i8)argb1, (v16i8)argb0);          \
+    vec3_m = (v8u16)__msa_vshf_h(shf1, (v16i8)argb3, (v16i8)argb2);          \
+    vec4_m = (v8u16)__msa_vshf_h(shf2, (v16i8)argb1, (v16i8)argb0);          \
+    vec5_m = (v8u16)__msa_vshf_h(shf2, (v16i8)argb3, (v16i8)argb2);          \
+    vec6_m = (v8u16)__msa_vshf_h(shf3, (v16i8)argb1, (v16i8)argb0);          \
+    vec7_m = (v8u16)__msa_vshf_h(shf3, (v16i8)argb3, (v16i8)argb2);          \
+    reg0_m = __msa_dotp_u_w(vec0_m, const0);                                 \
+    reg1_m = __msa_dotp_u_w(vec1_m, const0);                                 \
+    reg2_m = __msa_dotp_u_w(vec4_m, const0);                                 \
+    reg3_m = __msa_dotp_u_w(vec5_m, const0);                                 \
+    reg0_m += const1;                                                        \
+    reg1_m += const1;                                                        \
+    reg2_m += const1;                                                        \
+    reg3_m += const1;                                                        \
+    reg0_m -= (v4u32)__msa_dotp_u_w(vec2_m, const2);                         \
+    reg1_m -= (v4u32)__msa_dotp_u_w(vec3_m, const2);                         \
+    reg2_m -= (v4u32)__msa_dotp_u_w(vec6_m, const3);                         \
+    reg3_m -= (v4u32)__msa_dotp_u_w(vec7_m, const3);                         \
+    reg0_m = __msa_srl_w(reg0_m, shift);                                     \
+    reg1_m = __msa_srl_w(reg1_m, shift);                                     \
+    reg2_m = __msa_srl_w(reg2_m, shift);                                     \
+    reg3_m = __msa_srl_w(reg3_m, shift);                                     \
+    u_out = (v8u16)__msa_pckev_h((v8i16)reg1_m, (v8i16)reg0_m);              \
+    v_out = (v8u16)__msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);              \
+  }
+
+// Takes ARGB input and calculates U and V.
+#define ARGBTOUV_H(argb0, argb1, argb2, argb3, const0, const1, const2, const3, \
+                   shf0, shf1, shf2, shf3, v_out, u_out)                       \
+  {                                                                            \
+    v8u16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;      \
+    v4u32 reg0_m, reg1_m, reg2_m, reg3_m;                                      \
+                                                                               \
+    vec0_m = __msa_vshf_h(shf0, (v16i8)argb1, (v16i8)argb0);                   \
+    vec1_m = __msa_vshf_h(shf0, (v16i8)argb3, (v16i8)argb2);                   \
+    vec2_m = __msa_vshf_h(shf1, (v16i8)argb1, (v16i8)argb0);                   \
+    vec3_m = __msa_vshf_h(shf1, (v16i8)argb3, (v16i8)argb2);                   \
+    vec4_m = __msa_vshf_h(shf2, (v16i8)argb1, (v16i8)argb0);                   \
+    vec5_m = __msa_vshf_h(shf2, (v16i8)argb3, (v16i8)argb2);                   \
+    vec6_m = __msa_vshf_h(shf3, (v16i8)argb1, (v16i8)argb0);                   \
+    vec7_m = __msa_vshf_h(shf3, (v16i8)argb3, (v16i8)argb2);                   \
+    reg0_m = __msa_dotp_u_w(vec0_m, const1);                                   \
+    reg1_m = __msa_dotp_u_w(vec1_m, const1);                                   \
+    reg2_m = __msa_dotp_u_w(vec4_m, const1);                                   \
+    reg3_m = __msa_dotp_u_w(vec5_m, const1);                                   \
+    reg0_m += (v4u32)const3;                                                   \
+    reg1_m += (v4u32)const3;                                                   \
+    reg2_m += (v4u32)const3;                                                   \
+    reg3_m += (v4u32)const3;                                                   \
+    reg0_m -= __msa_dotp_u_w(vec2_m, const0);                                  \
+    reg1_m -= __msa_dotp_u_w(vec3_m, const0);                                  \
+    reg2_m -= __msa_dotp_u_w(vec6_m, const2);                                  \
+    reg3_m -= __msa_dotp_u_w(vec7_m, const2);                                  \
+    u_out = (v16u8)__msa_pckev_h((v8i16)reg3_m, (v8i16)reg2_m);                \
+    v_out = (v16u8)__msa_pckev_h((v8i16)reg1_m, (v8i16)reg0_m);                \
+    u_out = (v16u8)__msa_pckod_b((v16i8)u_out, (v16i8)u_out);                  \
+    v_out = (v16u8)__msa_pckod_b((v16i8)v_out, (v16i8)v_out);                  \
   }
 
 // Load I444 pixel data
@@ -285,6 +280,34 @@ extern "C" {
     out_v = (v16u8)__msa_insert_d(zero_m, 0, (int64_t)v_m);   \
   }
 
+#define RGBTOUV(_tmpb, _tmpg, _tmpr, _nexb, _nexg, _nexr, _dst0) \
+  {                                                              \
+    v16u8 _tmp0, _tmp1, _tmp2, _tmp3, _tmp4, _tmp5;              \
+    v8i16 _reg0, _reg1, _reg2, _reg3, _reg4, _reg5;              \
+    _tmp0 = (v16u8)__msa_ilvev_b(_tmpb, _nexb);                  \
+    _tmp1 = (v16u8)__msa_ilvod_b(_tmpb, _nexb);                  \
+    _tmp2 = (v16u8)__msa_ilvev_b(_tmpg, _nexg);                  \
+    _tmp3 = (v16u8)__msa_ilvod_b(_tmpg, _nexg);                  \
+    _tmp4 = (v16u8)__msa_ilvev_b(_tmpr, _nexr);                  \
+    _tmp5 = (v16u8)__msa_ilvod_b(_tmpr, _nexr);                  \
+    _reg0 = (v8i16)__msa_hadd_u_h(_tmp0, _tmp0);                 \
+    _reg1 = (v8i16)__msa_hadd_u_h(_tmp1, _tmp1);                 \
+    _reg2 = (v8i16)__msa_hadd_u_h(_tmp2, _tmp2);                 \
+    _reg3 = (v8i16)__msa_hadd_u_h(_tmp3, _tmp3);                 \
+    _reg4 = (v8i16)__msa_hadd_u_h(_tmp4, _tmp4);                 \
+    _reg5 = (v8i16)__msa_hadd_u_h(_tmp5, _tmp5);                 \
+    _reg0 = (v8i16)__msa_aver_u_h(_reg0, _reg1);                 \
+    _reg2 = (v8i16)__msa_aver_u_h(_reg2, _reg3);                 \
+    _reg4 = (v8i16)__msa_aver_u_h(_reg4, _reg5);                 \
+    _reg1 = const_8080 + const_112 * _reg0;                      \
+    _reg3 = const_8080 + const_112 * _reg4;                      \
+    _reg1 = (v8i16)__msa_msubv_h(_reg1, const_74, _reg2);        \
+    _reg3 = (v8i16)__msa_msubv_h(_reg3, const_94, _reg2);        \
+    _reg1 = (v8i16)__msa_msubv_h(_reg1, const_38, _reg4);        \
+    _reg3 = (v8i16)__msa_msubv_h(_reg3, const_18, _reg0);        \
+    _dst0 = (v16u8)__msa_pckod_b(_reg3, _reg1);                  \
+  }
+
 void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width) {
   int x;
   v16u8 src0, src1, src2, src3;
@@ -302,6 +325,20 @@ void MirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width) {
   }
 }
 
+void MirrorUVRow_MSA(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  int x;
+  v8u16 src, dst;
+  v8u16 shuffler = {7, 6, 5, 4, 3, 2, 1, 0};
+  src_uv += (width - 8) << 1;
+  for (x = 0; x < width; x += 8) {
+    src = LD_UH(src_uv);
+    dst = __msa_vshf_h(shuffler, src, src);
+    ST_UH(dst, dst_uv);
+    src_uv -= 16;
+    dst_uv += 16;
+  }
+}
+
 void ARGBMirrorRow_MSA(const uint8_t* src, uint8_t* dst, int width) {
   int x;
   v16u8 src0, src1, src2, src3;
@@ -376,20 +413,19 @@ void I422ToARGBRow_MSA(const uint8_t* src_y,
   int x;
   v16u8 src0, src1, src2;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
   for (x = 0; x < width; x += 8) {
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
     src_y += 8;
     src_u += 4;
@@ -407,20 +443,19 @@ void I422ToRGBARow_MSA(const uint8_t* src_y,
   int x;
   v16u8 src0, src1, src2;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
   for (x = 0; x < width; x += 8) {
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     STOREARGB(alpha, vec0, vec1, vec2, dst_argb);
     src_y += 8;
     src_u += 4;
@@ -440,12 +475,12 @@ void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
   int64_t data_a;
   v16u8 src0, src1, src2, src3;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v4i32 zero = {0};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -454,8 +489,7 @@ void I422AlphaToARGBRow_MSA(const uint8_t* src_y,
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
     src3 = (v16u8)__msa_insert_d((v2i64)zero, 0, data_a);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     src3 = (v16u8)__msa_ilvr_b((v16i8)src3, (v16i8)src3);
     STOREARGB(vec0, vec1, vec2, src3, dst_argb);
     src_y += 8;
@@ -476,17 +510,17 @@ void I422ToRGB24Row_MSA(const uint8_t* src_y,
   int64_t data_u, data_v;
   v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2;
   v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 reg0, reg1, reg2, reg3;
   v2i64 zero = {0};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
   v16i8 shuffler0 = {0, 1, 16, 2, 3, 17, 4, 5, 18, 6, 7, 19, 8, 9, 20, 10};
   v16i8 shuffler1 = {0, 21, 1, 2, 22, 3, 4, 23, 5, 6, 24, 7, 8, 25, 9, 10};
   v16i8 shuffler2 = {26, 6,  7,  27, 8,  9,  28, 10,
                      11, 29, 12, 13, 30, 14, 15, 31};
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -499,10 +533,8 @@ void I422ToRGB24Row_MSA(const uint8_t* src_y,
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
     src3 = (v16u8)__msa_sldi_b((v16i8)src0, (v16i8)src0, 8);
     src4 = (v16u8)__msa_sldi_b((v16i8)src1, (v16i8)src1, 8);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
-    YUVTORGB(src3, src4, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec3, vec4, vec5);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    YUVTORGB(src3, src4, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec3, vec4, vec5);
     reg0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
     reg2 = (v16u8)__msa_ilvev_b((v16i8)vec4, (v16i8)vec3);
     reg3 = (v16u8)__msa_pckev_b((v16i8)vec5, (v16i8)vec2);
@@ -529,24 +561,23 @@ void I422ToRGB565Row_MSA(const uint8_t* src_y,
   int x;
   v16u8 src0, src1, src2, dst0;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
   for (x = 0; x < width; x += 8) {
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec2, vec1);
-    vec0 = __msa_srai_h(vec0, 3);
-    vec1 = __msa_srai_h(vec1, 3);
-    vec2 = __msa_srai_h(vec2, 2);
-    vec1 = __msa_slli_h(vec1, 11);
-    vec2 = __msa_slli_h(vec2, 5);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    vec0 = __msa_srli_h(vec0, 3);
+    vec1 = __msa_srli_h(vec1, 2);
+    vec2 = __msa_srli_h(vec2, 3);
+    vec2 = __msa_slli_h(vec2, 11);
+    vec1 = __msa_slli_h(vec1, 5);
     vec0 |= vec1;
     dst0 = (v16u8)(vec2 | vec0);
     ST_UB(dst0, dst_rgb565);
@@ -568,25 +599,24 @@ void I422ToARGB4444Row_MSA(const uint8_t* src_y,
   v16u8 src0, src1, src2, dst0;
   v8i16 vec0, vec1, vec2;
   v8u16 reg0, reg1, reg2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v8u16 const_0xF000 = (v8u16)__msa_fill_h(0xF000);
+  v8u16 mask = (v8u16)__msa_fill_h(0x00F0);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
   for (x = 0; x < width; x += 8) {
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
-    reg0 = (v8u16)__msa_srai_h(vec0, 4);
-    reg1 = (v8u16)__msa_srai_h(vec1, 4);
-    reg2 = (v8u16)__msa_srai_h(vec2, 4);
-    reg1 = (v8u16)__msa_slli_h((v8i16)reg1, 4);
-    reg2 = (v8u16)__msa_slli_h((v8i16)reg2, 8);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    reg0 = (v8u16)__msa_srli_h(vec0, 4);
+    reg2 = (v8u16)__msa_srli_h(vec2, 4);
+    reg1 = (v8u16)__msa_and_v(vec1, mask);
+    reg2 = (v8u16)__msa_slli_h(reg2, 8);
     reg1 |= const_0xF000;
     reg0 |= reg2;
     dst0 = (v16u8)(reg1 | reg0);
@@ -608,23 +638,22 @@ void I422ToARGB1555Row_MSA(const uint8_t* src_y,
   v16u8 src0, src1, src2, dst0;
   v8i16 vec0, vec1, vec2;
   v8u16 reg0, reg1, reg2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v8u16 const_0x8000 = (v8u16)__msa_fill_h(0x8000);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
   for (x = 0; x < width; x += 8) {
     READYUV422(src_y, src_u, src_v, src0, src1, src2);
     src1 = (v16u8)__msa_ilvr_b((v16i8)src2, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
-    reg0 = (v8u16)__msa_srai_h(vec0, 3);
-    reg1 = (v8u16)__msa_srai_h(vec1, 3);
-    reg2 = (v8u16)__msa_srai_h(vec2, 3);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
+    reg0 = (v8u16)__msa_srli_h(vec0, 3);
+    reg1 = (v8u16)__msa_srli_h(vec1, 3);
+    reg2 = (v8u16)__msa_srli_h(vec2, 3);
     reg1 = (v8u16)__msa_slli_h((v8i16)reg1, 5);
     reg2 = (v8u16)__msa_slli_h((v8i16)reg2, 10);
     reg1 |= const_0x8000;
@@ -768,7 +797,7 @@ void UYVYToUV422Row_MSA(const uint8_t* src_uyvy,
   }
 }
 
-void ARGBToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ARGBToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, src3, vec0, vec1, vec2, vec3, dst0;
   v8u16 reg0, reg1, reg2, reg3, reg4, reg5;
@@ -779,10 +808,10 @@ void ARGBToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 48);
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb, 48);
     vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
     vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
     vec2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
@@ -809,38 +838,39 @@ void ARGBToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
     reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 8);
     dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 64;
+    src_argb += 64;
     dst_y += 16;
   }
 }
 
-void ARGBToUVRow_MSA(const uint8_t* src_argb0,
+void ARGBToUVRow_MSA(const uint8_t* src_argb,
                      int src_stride_argb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   int x;
-  const uint8_t* src_argb0_next = src_argb0 + src_stride_argb;
+  const uint8_t* src_argb_next = src_argb + src_stride_argb;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
   v8u16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, reg8, reg9;
   v16u8 dst0, dst1;
-  v8u16 const_0x70 = (v8u16)__msa_ldi_h(0x70);
-  v8u16 const_0x4A = (v8u16)__msa_ldi_h(0x4A);
-  v8u16 const_0x26 = (v8u16)__msa_ldi_h(0x26);
-  v8u16 const_0x5E = (v8u16)__msa_ldi_h(0x5E);
-  v8u16 const_0x12 = (v8u16)__msa_ldi_h(0x12);
+  v8u16 const_0x70 = (v8u16)__msa_ldi_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_ldi_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_ldi_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_ldi_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_ldi_h(0x09);
   v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
 
   for (x = 0; x < width; x += 32) {
-    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 48);
-    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 64);
-    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 80);
-    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 96);
-    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb0, 112);
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb, 48);
+    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb, 64);
+    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb, 80);
+    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb, 96);
+    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb, 112);
     vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
     vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
     vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4);
@@ -861,14 +891,14 @@ void ARGBToUVRow_MSA(const uint8_t* src_argb0,
     reg3 = __msa_hadd_u_h(vec5, vec5);
     reg4 = __msa_hadd_u_h(vec0, vec0);
     reg5 = __msa_hadd_u_h(vec1, vec1);
-    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 0);
-    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 16);
-    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 32);
-    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 48);
-    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 64);
-    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 80);
-    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 96);
-    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb0_next, 112);
+    src0 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 0);
+    src1 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 16);
+    src2 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 32);
+    src3 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 48);
+    src4 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 64);
+    src5 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 80);
+    src6 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 96);
+    src7 = (v16u8)__msa_ld_b((v16u8*)src_argb_next, 112);
     vec0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
     vec1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
     vec2 = (v16u8)__msa_pckev_b((v16i8)src5, (v16i8)src4);
@@ -889,12 +919,18 @@ void ARGBToUVRow_MSA(const uint8_t* src_argb0,
     reg3 += __msa_hadd_u_h(vec5, vec5);
     reg4 += __msa_hadd_u_h(vec0, vec0);
     reg5 += __msa_hadd_u_h(vec1, vec1);
-    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 2);
-    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 2);
-    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 2);
-    reg3 = (v8u16)__msa_srai_h((v8i16)reg3, 2);
-    reg4 = (v8u16)__msa_srai_h((v8i16)reg4, 2);
-    reg5 = (v8u16)__msa_srai_h((v8i16)reg5, 2);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg4 += const_0x0001;
+    reg5 += const_0x0001;
+    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 1);
+    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 1);
+    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 1);
+    reg3 = (v8u16)__msa_srai_h((v8i16)reg3, 1);
+    reg4 = (v8u16)__msa_srai_h((v8i16)reg4, 1);
+    reg5 = (v8u16)__msa_srai_h((v8i16)reg5, 1);
     reg6 = reg0 * const_0x70;
     reg7 = reg1 * const_0x70;
     reg8 = reg2 * const_0x4A;
@@ -925,8 +961,8 @@ void ARGBToUVRow_MSA(const uint8_t* src_argb0,
     dst1 = (v16u8)__msa_pckev_b((v16i8)reg5, (v16i8)reg4);
     ST_UB(dst0, dst_u);
     ST_UB(dst1, dst_v);
-    src_argb0 += 128;
-    src_argb0_next += 128;
+    src_argb += 128;
+    src_argb_next += 128;
     dst_u += 16;
     dst_v += 16;
   }
@@ -1153,7 +1189,7 @@ void ARGBToUV444Row_MSA(const uint8_t* src_argb,
   }
 }
 
-void ARGBMultiplyRow_MSA(const uint8_t* src_argb0,
+void ARGBMultiplyRow_MSA(const uint8_t* src_argb,
                          const uint8_t* src_argb1,
                          uint8_t* dst_argb,
                          int width) {
@@ -1164,7 +1200,7 @@ void ARGBMultiplyRow_MSA(const uint8_t* src_argb0,
   v8i16 zero = {0};
 
   for (x = 0; x < width; x += 4) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
     src1 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
     vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
     vec1 = (v8u16)__msa_ilvl_b((v16i8)src0, (v16i8)src0);
@@ -1186,13 +1222,13 @@ void ARGBMultiplyRow_MSA(const uint8_t* src_argb0,
     vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
     dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
     ST_UB(dst0, dst_argb);
-    src_argb0 += 16;
+    src_argb += 16;
     src_argb1 += 16;
     dst_argb += 16;
   }
 }
 
-void ARGBAddRow_MSA(const uint8_t* src_argb0,
+void ARGBAddRow_MSA(const uint8_t* src_argb,
                     const uint8_t* src_argb1,
                     uint8_t* dst_argb,
                     int width) {
@@ -1200,20 +1236,20 @@ void ARGBAddRow_MSA(const uint8_t* src_argb0,
   v16u8 src0, src1, src2, src3, dst0, dst1;
 
   for (x = 0; x < width; x += 8) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
     src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
     src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
     dst0 = __msa_adds_u_b(src0, src2);
     dst1 = __msa_adds_u_b(src1, src3);
     ST_UB2(dst0, dst1, dst_argb, 16);
-    src_argb0 += 32;
+    src_argb += 32;
     src_argb1 += 32;
     dst_argb += 32;
   }
 }
 
-void ARGBSubtractRow_MSA(const uint8_t* src_argb0,
+void ARGBSubtractRow_MSA(const uint8_t* src_argb,
                          const uint8_t* src_argb1,
                          uint8_t* dst_argb,
                          int width) {
@@ -1221,14 +1257,14 @@ void ARGBSubtractRow_MSA(const uint8_t* src_argb0,
   v16u8 src0, src1, src2, src3, dst0, dst1;
 
   for (x = 0; x < width; x += 8) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
     src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
     src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
     dst0 = __msa_subs_u_b(src0, src2);
     dst1 = __msa_subs_u_b(src1, src3);
     ST_UB2(dst0, dst1, dst_argb, 16);
-    src_argb0 += 32;
+    src_argb += 32;
     src_argb1 += 32;
     dst_argb += 32;
   }
@@ -1412,17 +1448,17 @@ void ARGBGrayRow_MSA(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   int x;
   v16u8 src0, src1, vec0, vec1, dst0, dst1;
   v8u16 reg0;
-  v16u8 const_0x26 = (v16u8)__msa_ldi_h(0x26);
-  v16u8 const_0x4B0F = (v16u8)__msa_fill_h(0x4B0F);
+  v16u8 const_0x4D = (v16u8)__msa_ldi_h(0x4D);
+  v16u8 const_0x961D = (v16u8)__msa_fill_h(0x961D);
 
   for (x = 0; x < width; x += 8) {
     src0 = (v16u8)__msa_ld_b((v16u8*)src_argb, 0);
     src1 = (v16u8)__msa_ld_b((v16u8*)src_argb, 16);
     vec0 = (v16u8)__msa_pckev_h((v8i16)src1, (v8i16)src0);
     vec1 = (v16u8)__msa_pckod_h((v8i16)src1, (v8i16)src0);
-    reg0 = __msa_dotp_u_h(vec0, const_0x4B0F);
-    reg0 = __msa_dpadd_u_h(reg0, vec1, const_0x26);
-    reg0 = (v8u16)__msa_srari_h((v8i16)reg0, 7);
+    reg0 = __msa_dotp_u_h(vec0, const_0x961D);
+    reg0 = __msa_dpadd_u_h(reg0, vec1, const_0x4D);
+    reg0 = (v8u16)__msa_srari_h((v8i16)reg0, 8);
     vec0 = (v16u8)__msa_ilvev_b((v16i8)reg0, (v16i8)reg0);
     vec1 = (v16u8)__msa_ilvod_b((v16i8)vec1, (v16i8)vec0);
     dst0 = (v16u8)__msa_ilvr_b((v16i8)vec1, (v16i8)vec0);
@@ -1656,56 +1692,51 @@ void ARGB1555ToYRow_MSA(const uint8_t* src_argb1555,
                         uint8_t* dst_y,
                         int width) {
   int x;
-  v8u16 src0, src1, vec0, vec1, vec2, vec3, vec4, vec5;
-  v8u16 reg0, reg1, reg2, reg3, reg4, reg5;
-  v16u8 dst0;
-  v8u16 const_0x19 = (v8u16)__msa_ldi_h(0x19);
-  v8u16 const_0x81 = (v8u16)__msa_ldi_h(0x81);
-  v8u16 const_0x42 = (v8u16)__msa_ldi_h(0x42);
-  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
-  v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
+  v16u8 src0, src1, tmp0, tmp1, tmpb, tmpg, tmpr;
+  v16u8 reg0, reg1, reg2, dst;
+  v8i16 tmpr_l, tmpr_r, tmpg_l, tmpg_r, tmpb_l, tmpb_r;
+  v8i16 res0, res1;
+  v8i16 const_66 = (v8i16)__msa_ldi_h(66);
+  v8i16 const_129 = (v8i16)__msa_ldi_h(129);
+  v8i16 const_25 = (v8i16)__msa_ldi_h(25);
+  v8u16 const_1080 = (v8u16)__msa_fill_h(0x1080);
+  v16u8 zero = (v16u8)__msa_ldi_b(0);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v8u16)__msa_ld_b((void*)src_argb1555, 0);
-    src1 = (v8u16)__msa_ld_b((void*)src_argb1555, 16);
-    vec0 = src0 & const_0x1F;
-    vec1 = src1 & const_0x1F;
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 5);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 5);
-    vec2 = src0 & const_0x1F;
-    vec3 = src1 & const_0x1F;
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 5);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 5);
-    vec4 = src0 & const_0x1F;
-    vec5 = src1 & const_0x1F;
-    reg0 = (v8u16)__msa_slli_h((v8i16)vec0, 3);
-    reg1 = (v8u16)__msa_slli_h((v8i16)vec1, 3);
-    reg0 |= (v8u16)__msa_srai_h((v8i16)vec0, 2);
-    reg1 |= (v8u16)__msa_srai_h((v8i16)vec1, 2);
-    reg2 = (v8u16)__msa_slli_h((v8i16)vec2, 3);
-    reg3 = (v8u16)__msa_slli_h((v8i16)vec3, 3);
-    reg2 |= (v8u16)__msa_srai_h((v8i16)vec2, 2);
-    reg3 |= (v8u16)__msa_srai_h((v8i16)vec3, 2);
-    reg4 = (v8u16)__msa_slli_h((v8i16)vec4, 3);
-    reg5 = (v8u16)__msa_slli_h((v8i16)vec5, 3);
-    reg4 |= (v8u16)__msa_srai_h((v8i16)vec4, 2);
-    reg5 |= (v8u16)__msa_srai_h((v8i16)vec5, 2);
-    reg0 *= const_0x19;
-    reg1 *= const_0x19;
-    reg2 *= const_0x81;
-    reg3 *= const_0x81;
-    reg4 *= const_0x42;
-    reg5 *= const_0x42;
-    reg0 += reg2;
-    reg1 += reg3;
-    reg0 += reg4;
-    reg1 += reg5;
-    reg0 += const_0x1080;
-    reg1 += const_0x1080;
-    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 8);
-    reg1 = (v8u16)__msa_srai_h((v8i16)reg1, 8);
-    dst0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
-    ST_UB(dst0, dst_y);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb1555, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb1555, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpg = (v16u8)__msa_srli_b(tmp0, 5);
+    reg0 = (v16u8)__msa_andi_b(tmp1, 0x03);
+    reg0 = (v16u8)__msa_slli_b(reg0, 3);
+    tmpg = (v16u8)__msa_or_v(tmpg, reg0);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x7C);
+    tmpr = (v16u8)__msa_srli_b(reg1, 2);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_slli_b(tmpg, 3);
+    reg2 = (v16u8)__msa_slli_b(tmpr, 3);
+    tmpb = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpg = (v16u8)__msa_srli_b(tmpg, 2);
+    tmpr = (v16u8)__msa_srli_b(tmpr, 2);
+    tmpb = (v16u8)__msa_or_v(reg0, tmpb);
+    tmpg = (v16u8)__msa_or_v(reg1, tmpg);
+    tmpr = (v16u8)__msa_or_v(reg2, tmpr);
+    tmpb_r = (v8i16)__msa_ilvr_b(zero, tmpb);
+    tmpb_l = (v8i16)__msa_ilvl_b(zero, tmpb);
+    tmpg_r = (v8i16)__msa_ilvr_b(zero, tmpg);
+    tmpg_l = (v8i16)__msa_ilvl_b(zero, tmpg);
+    tmpr_r = (v8i16)__msa_ilvr_b(zero, tmpr);
+    tmpr_l = (v8i16)__msa_ilvl_b(zero, tmpr);
+    res0 = const_1080 + const_25 * tmpb_r;
+    res1 = const_1080 + const_25 * tmpb_l;
+    res0 += const_129 * tmpg_r;
+    res1 += const_129 * tmpg_l;
+    res0 += const_66 * tmpr_r;
+    res1 += const_66 * tmpr_l;
+    dst = (v16u8)__msa_pckod_b(res1, res0);
+    ST_UB(dst, dst_y);
     src_argb1555 += 32;
     dst_y += 16;
   }
@@ -1713,68 +1744,55 @@ void ARGB1555ToYRow_MSA(const uint8_t* src_argb1555,
 
 void RGB565ToYRow_MSA(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
   int x;
-  v8u16 src0, src1, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
-  v8u16 reg0, reg1, reg2, reg3, reg4, reg5;
-  v4u32 res0, res1, res2, res3;
-  v16u8 dst0;
-  v4u32 const_0x810019 = (v4u32)__msa_fill_w(0x810019);
-  v4u32 const_0x010042 = (v4u32)__msa_fill_w(0x010042);
-  v8i16 const_0x1080 = __msa_fill_h(0x1080);
-  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
-  v8u16 const_0x7E0 = (v8u16)__msa_fill_h(0x7E0);
-  v8u16 const_0xF800 = (v8u16)__msa_fill_h(0xF800);
+  v16u8 src0, src1, tmp0, tmp1, tmpb, tmpg, tmpr;
+  v16u8 reg0, reg1, dst;
+  v8i16 tmpr_l, tmpr_r, tmpg_l, tmpg_r, tmpb_l, tmpb_r;
+  v8i16 res0, res1;
+  v8i16 const_66 = (v8i16)__msa_ldi_h(66);
+  v8i16 const_129 = (v8i16)__msa_ldi_h(129);
+  v8i16 const_25 = (v8i16)__msa_ldi_h(25);
+  v8i16 const_1080 = (v8i16)__msa_fill_h(0x1080);
+  v16u8 zero = __msa_ldi_b(0);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v8u16)__msa_ld_b((void*)src_rgb565, 0);
-    src1 = (v8u16)__msa_ld_b((void*)src_rgb565, 16);
-    vec0 = src0 & const_0x1F;
-    vec1 = src0 & const_0x7E0;
-    vec2 = src0 & const_0xF800;
-    vec3 = src1 & const_0x1F;
-    vec4 = src1 & const_0x7E0;
-    vec5 = src1 & const_0xF800;
-    reg0 = (v8u16)__msa_slli_h((v8i16)vec0, 3);
-    reg1 = (v8u16)__msa_srli_h((v8i16)vec1, 3);
-    reg2 = (v8u16)__msa_srli_h((v8i16)vec2, 8);
-    reg3 = (v8u16)__msa_slli_h((v8i16)vec3, 3);
-    reg4 = (v8u16)__msa_srli_h((v8i16)vec4, 3);
-    reg5 = (v8u16)__msa_srli_h((v8i16)vec5, 8);
-    reg0 |= (v8u16)__msa_srli_h((v8i16)vec0, 2);
-    reg1 |= (v8u16)__msa_srli_h((v8i16)vec1, 9);
-    reg2 |= (v8u16)__msa_srli_h((v8i16)vec2, 13);
-    reg3 |= (v8u16)__msa_srli_h((v8i16)vec3, 2);
-    reg4 |= (v8u16)__msa_srli_h((v8i16)vec4, 9);
-    reg5 |= (v8u16)__msa_srli_h((v8i16)vec5, 13);
-    vec0 = (v8u16)__msa_ilvr_h((v8i16)reg1, (v8i16)reg0);
-    vec1 = (v8u16)__msa_ilvl_h((v8i16)reg1, (v8i16)reg0);
-    vec2 = (v8u16)__msa_ilvr_h((v8i16)reg4, (v8i16)reg3);
-    vec3 = (v8u16)__msa_ilvl_h((v8i16)reg4, (v8i16)reg3);
-    vec4 = (v8u16)__msa_ilvr_h(const_0x1080, (v8i16)reg2);
-    vec5 = (v8u16)__msa_ilvl_h(const_0x1080, (v8i16)reg2);
-    vec6 = (v8u16)__msa_ilvr_h(const_0x1080, (v8i16)reg5);
-    vec7 = (v8u16)__msa_ilvl_h(const_0x1080, (v8i16)reg5);
-    res0 = __msa_dotp_u_w(vec0, (v8u16)const_0x810019);
-    res1 = __msa_dotp_u_w(vec1, (v8u16)const_0x810019);
-    res2 = __msa_dotp_u_w(vec2, (v8u16)const_0x810019);
-    res3 = __msa_dotp_u_w(vec3, (v8u16)const_0x810019);
-    res0 = __msa_dpadd_u_w(res0, vec4, (v8u16)const_0x010042);
-    res1 = __msa_dpadd_u_w(res1, vec5, (v8u16)const_0x010042);
-    res2 = __msa_dpadd_u_w(res2, vec6, (v8u16)const_0x010042);
-    res3 = __msa_dpadd_u_w(res3, vec7, (v8u16)const_0x010042);
-    res0 = (v4u32)__msa_srai_w((v4i32)res0, 8);
-    res1 = (v4u32)__msa_srai_w((v4i32)res1, 8);
-    res2 = (v4u32)__msa_srai_w((v4i32)res2, 8);
-    res3 = (v4u32)__msa_srai_w((v4i32)res3, 8);
-    vec0 = (v8u16)__msa_pckev_h((v8i16)res1, (v8i16)res0);
-    vec1 = (v8u16)__msa_pckev_h((v8i16)res3, (v8i16)res2);
-    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
-    ST_UB(dst0, dst_y);
+    src0 = (v16u8)__msa_ld_b((void*)src_rgb565, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_rgb565, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpr = (v16u8)__msa_andi_b(tmp1, 0xF8);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x07);
+    reg0 = (v16u8)__msa_srli_b(tmp0, 5);
+    reg1 = (v16u8)__msa_slli_b(reg1, 3);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpb = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_slli_b(tmpg, 2);
+    reg1 = (v16u8)__msa_srli_b(tmpg, 4);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    reg0 = (v16u8)__msa_srli_b(tmpr, 5);
+    tmpr = (v16u8)__msa_or_v(tmpr, reg0);
+    tmpb_r = (v8i16)__msa_ilvr_b(zero, tmpb);
+    tmpb_l = (v8i16)__msa_ilvl_b(zero, tmpb);
+    tmpg_r = (v8i16)__msa_ilvr_b(zero, tmpg);
+    tmpg_l = (v8i16)__msa_ilvl_b(zero, tmpg);
+    tmpr_r = (v8i16)__msa_ilvr_b(zero, tmpr);
+    tmpr_l = (v8i16)__msa_ilvl_b(zero, tmpr);
+    res0 = const_1080 + const_25 * tmpb_r;
+    res1 = const_1080 + const_25 * tmpb_l;
+    res0 += const_129 * tmpg_r;
+    res1 += const_129 * tmpg_l;
+    res0 += const_66 * tmpr_r;
+    res1 += const_66 * tmpr_l;
+    dst = (v16u8)__msa_pckod_b(res1, res0);
+    ST_UB(dst, dst_y);
     src_rgb565 += 32;
     dst_y += 16;
   }
 }
 
-void RGB24ToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RGB24ToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0;
   v8u16 vec0, vec1, vec2, vec3;
@@ -1789,9 +1807,9 @@ void RGB24ToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v16i8 zero = {0};
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
     reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0);
     reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
     reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1);
@@ -1810,12 +1828,12 @@ void RGB24ToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
     vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
     dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 48;
+    src_argb += 48;
     dst_y += 16;
   }
 }
 
-void RAWToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RAWToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, reg0, reg1, reg2, reg3, dst0;
   v8u16 vec0, vec1, vec2, vec3;
@@ -1830,9 +1848,9 @@ void RAWToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v16i8 zero = {0};
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
     reg0 = (v16u8)__msa_vshf_b(mask0, zero, (v16i8)src0);
     reg1 = (v16u8)__msa_vshf_b(mask1, (v16i8)src1, (v16i8)src0);
     reg2 = (v16u8)__msa_vshf_b(mask2, (v16i8)src2, (v16i8)src1);
@@ -1851,7 +1869,7 @@ void RAWToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
     vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
     dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 48;
+    src_argb += 48;
     dst_y += 16;
   }
 }
@@ -1865,69 +1883,61 @@ void ARGB1555ToUVRow_MSA(const uint8_t* src_argb1555,
   const uint16_t* s = (const uint16_t*)src_argb1555;
   const uint16_t* t = (const uint16_t*)(src_argb1555 + src_stride_argb1555);
   int64_t res0, res1;
-  v8u16 src0, src1, src2, src3, reg0, reg1, reg2, reg3;
-  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6;
-  v16u8 dst0;
-  v8u16 const_0x70 = (v8u16)__msa_ldi_h(0x70);
-  v8u16 const_0x4A = (v8u16)__msa_ldi_h(0x4A);
-  v8u16 const_0x26 = (v8u16)__msa_ldi_h(0x26);
-  v8u16 const_0x5E = (v8u16)__msa_ldi_h(0x5E);
-  v8u16 const_0x12 = (v8u16)__msa_ldi_h(0x12);
-  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
-  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
+  v16u8 src0, src1, src2, src3, dst;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 reg0, reg1, reg2, reg3;
+  v16u8 tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  v8i16 const_112 = (v8i16)__msa_ldi_h(0x38);
+  v8i16 const_74 = (v8i16)__msa_ldi_h(0x25);
+  v8i16 const_38 = (v8i16)__msa_ldi_h(0x13);
+  v8i16 const_94 = (v8i16)__msa_ldi_h(0x2F);
+  v8i16 const_18 = (v8i16)__msa_ldi_h(0x09);
+  v8u16 const_8080 = (v8u16)__msa_fill_h(0x8080);
 
   for (x = 0; x < width; x += 16) {
     src0 = (v8u16)__msa_ld_b((void*)s, 0);
     src1 = (v8u16)__msa_ld_b((void*)s, 16);
     src2 = (v8u16)__msa_ld_b((void*)t, 0);
     src3 = (v8u16)__msa_ld_b((void*)t, 16);
-    vec0 = src0 & const_0x1F;
-    vec1 = src1 & const_0x1F;
-    vec0 += src2 & const_0x1F;
-    vec1 += src3 & const_0x1F;
-    vec0 = (v8u16)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 5);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 5);
-    src2 = (v8u16)__msa_srai_h((v8i16)src2, 5);
-    src3 = (v8u16)__msa_srai_h((v8i16)src3, 5);
-    vec2 = src0 & const_0x1F;
-    vec3 = src1 & const_0x1F;
-    vec2 += src2 & const_0x1F;
-    vec3 += src3 & const_0x1F;
-    vec2 = (v8u16)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 5);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 5);
-    src2 = (v8u16)__msa_srai_h((v8i16)src2, 5);
-    src3 = (v8u16)__msa_srai_h((v8i16)src3, 5);
-    vec4 = src0 & const_0x1F;
-    vec5 = src1 & const_0x1F;
-    vec4 += src2 & const_0x1F;
-    vec5 += src3 & const_0x1F;
-    vec4 = (v8u16)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
-    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
-    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
-    vec4 = __msa_hadd_u_h((v16u8)vec4, (v16u8)vec4);
-    vec6 = (v8u16)__msa_slli_h((v8i16)vec0, 1);
-    vec6 |= (v8u16)__msa_srai_h((v8i16)vec0, 6);
-    vec0 = (v8u16)__msa_slli_h((v8i16)vec2, 1);
-    vec0 |= (v8u16)__msa_srai_h((v8i16)vec2, 6);
-    vec2 = (v8u16)__msa_slli_h((v8i16)vec4, 1);
-    vec2 |= (v8u16)__msa_srai_h((v8i16)vec4, 6);
-    reg0 = vec6 * const_0x70;
-    reg1 = vec0 * const_0x4A;
-    reg2 = vec2 * const_0x70;
-    reg3 = vec0 * const_0x5E;
-    reg0 += const_0x8080;
-    reg1 += vec2 * const_0x26;
-    reg2 += const_0x8080;
-    reg3 += vec6 * const_0x12;
-    reg0 -= reg1;
-    reg2 -= reg3;
-    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 8);
-    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 8);
-    dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0);
-    res0 = __msa_copy_u_d((v2i64)dst0, 0);
-    res1 = __msa_copy_u_d((v2i64)dst0, 1);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmp2 = (v16u8)__msa_pckev_b(src3, src2);
+    tmp3 = (v16u8)__msa_pckod_b(src3, src2);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    nexb = (v16u8)__msa_andi_b(tmp2, 0x1F);
+    tmpg = (v16u8)__msa_srli_b(tmp0, 5);
+    nexg = (v16u8)__msa_srli_b(tmp2, 5);
+    reg0 = (v16u8)__msa_andi_b(tmp1, 0x03);
+    reg2 = (v16u8)__msa_andi_b(tmp3, 0x03);
+    reg0 = (v16u8)__msa_slli_b(reg0, 3);
+    reg2 = (v16u8)__msa_slli_b(reg2, 3);
+    tmpg = (v16u8)__msa_or_v(tmpg, reg0);
+    nexg = (v16u8)__msa_or_v(nexg, reg2);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x7C);
+    reg3 = (v16u8)__msa_andi_b(tmp3, 0x7C);
+    tmpr = (v16u8)__msa_srli_b(reg1, 2);
+    nexr = (v16u8)__msa_srli_b(reg3, 2);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_slli_b(tmpg, 3);
+    reg2 = (v16u8)__msa_slli_b(tmpr, 3);
+    tmpb = (v16u8)__msa_srli_b(tmpb, 2);
+    tmpg = (v16u8)__msa_srli_b(tmpg, 2);
+    tmpr = (v16u8)__msa_srli_b(tmpr, 2);
+    tmpb = (v16u8)__msa_or_v(reg0, tmpb);
+    tmpg = (v16u8)__msa_or_v(reg1, tmpg);
+    tmpr = (v16u8)__msa_or_v(reg2, tmpr);
+    reg0 = (v16u8)__msa_slli_b(nexb, 3);
+    reg1 = (v16u8)__msa_slli_b(nexg, 3);
+    reg2 = (v16u8)__msa_slli_b(nexr, 3);
+    nexb = (v16u8)__msa_srli_b(nexb, 2);
+    nexg = (v16u8)__msa_srli_b(nexg, 2);
+    nexr = (v16u8)__msa_srli_b(nexr, 2);
+    nexb = (v16u8)__msa_or_v(reg0, nexb);
+    nexg = (v16u8)__msa_or_v(reg1, nexg);
+    nexr = (v16u8)__msa_or_v(reg2, nexr);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst);
+    res0 = __msa_copy_u_d((v2i64)dst, 0);
+    res1 = __msa_copy_u_d((v2i64)dst, 1);
     SD(res0, dst_u);
     SD(res1, dst_v);
     s += 16;
@@ -1946,68 +1956,57 @@ void RGB565ToUVRow_MSA(const uint8_t* src_rgb565,
   const uint16_t* s = (const uint16_t*)src_rgb565;
   const uint16_t* t = (const uint16_t*)(src_rgb565 + src_stride_rgb565);
   int64_t res0, res1;
-  v8u16 src0, src1, src2, src3, reg0, reg1, reg2, reg3;
-  v8u16 vec0, vec1, vec2, vec3, vec4, vec5;
-  v16u8 dst0;
-  v8u16 const_0x70 = (v8u16)__msa_ldi_h(0x70);
-  v8u16 const_0x4A = (v8u16)__msa_ldi_h(0x4A);
-  v8u16 const_0x26 = (v8u16)__msa_ldi_h(0x26);
-  v8u16 const_0x5E = (v8u16)__msa_ldi_h(0x5E);
-  v8u16 const_0x12 = (v8u16)__msa_ldi_h(0x12);
-  v8u16 const_32896 = (v8u16)__msa_fill_h(0x8080);
-  v8u16 const_0x1F = (v8u16)__msa_ldi_h(0x1F);
-  v8u16 const_0x3F = (v8u16)__msa_fill_h(0x3F);
+  v16u8 src0, src1, src2, src3, dst;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 reg0, reg1, reg2, reg3;
+  v16u8 tmpb, tmpg, tmpr, nexb, nexg, nexr;
+  v8i16 const_112 = (v8i16)__msa_ldi_h(0x38);
+  v8i16 const_74 = (v8i16)__msa_ldi_h(0x25);
+  v8i16 const_38 = (v8i16)__msa_ldi_h(0x13);
+  v8i16 const_94 = (v8i16)__msa_ldi_h(0x2F);
+  v8i16 const_18 = (v8i16)__msa_ldi_h(0x09);
+  v8u16 const_8080 = (v8u16)__msa_fill_h(0x8080);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v8u16)__msa_ld_b((void*)s, 0);
-    src1 = (v8u16)__msa_ld_b((void*)s, 16);
-    src2 = (v8u16)__msa_ld_b((void*)t, 0);
-    src3 = (v8u16)__msa_ld_b((void*)t, 16);
-    vec0 = src0 & const_0x1F;
-    vec1 = src1 & const_0x1F;
-    vec0 += src2 & const_0x1F;
-    vec1 += src3 & const_0x1F;
-    vec0 = (v8u16)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 5);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 5);
-    src2 = (v8u16)__msa_srai_h((v8i16)src2, 5);
-    src3 = (v8u16)__msa_srai_h((v8i16)src3, 5);
-    vec2 = src0 & const_0x3F;
-    vec3 = src1 & const_0x3F;
-    vec2 += src2 & const_0x3F;
-    vec3 += src3 & const_0x3F;
-    vec1 = (v8u16)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
-    src0 = (v8u16)__msa_srai_h((v8i16)src0, 6);
-    src1 = (v8u16)__msa_srai_h((v8i16)src1, 6);
-    src2 = (v8u16)__msa_srai_h((v8i16)src2, 6);
-    src3 = (v8u16)__msa_srai_h((v8i16)src3, 6);
-    vec4 = src0 & const_0x1F;
-    vec5 = src1 & const_0x1F;
-    vec4 += src2 & const_0x1F;
-    vec5 += src3 & const_0x1F;
-    vec2 = (v8u16)__msa_pckev_b((v16i8)vec5, (v16i8)vec4);
-    vec0 = __msa_hadd_u_h((v16u8)vec0, (v16u8)vec0);
-    vec1 = __msa_hadd_u_h((v16u8)vec1, (v16u8)vec1);
-    vec2 = __msa_hadd_u_h((v16u8)vec2, (v16u8)vec2);
-    vec3 = (v8u16)__msa_slli_h((v8i16)vec0, 1);
-    vec3 |= (v8u16)__msa_srai_h((v8i16)vec0, 6);
-    vec4 = (v8u16)__msa_slli_h((v8i16)vec2, 1);
-    vec4 |= (v8u16)__msa_srai_h((v8i16)vec2, 6);
-    reg0 = vec3 * const_0x70;
-    reg1 = vec1 * const_0x4A;
-    reg2 = vec4 * const_0x70;
-    reg3 = vec1 * const_0x5E;
-    reg0 += const_32896;
-    reg1 += vec4 * const_0x26;
-    reg2 += const_32896;
-    reg3 += vec3 * const_0x12;
-    reg0 -= reg1;
-    reg2 -= reg3;
-    reg0 = (v8u16)__msa_srai_h((v8i16)reg0, 8);
-    reg2 = (v8u16)__msa_srai_h((v8i16)reg2, 8);
-    dst0 = (v16u8)__msa_pckev_b((v16i8)reg2, (v16i8)reg0);
-    res0 = __msa_copy_u_d((v2i64)dst0, 0);
-    res1 = __msa_copy_u_d((v2i64)dst0, 1);
+    src0 = (v16u8)__msa_ld_b((void*)s, 0);
+    src1 = (v16u8)__msa_ld_b((void*)s, 16);
+    src2 = (v16u8)__msa_ld_b((void*)t, 0);
+    src3 = (v16u8)__msa_ld_b((void*)t, 16);
+    tmp0 = (v16u8)__msa_pckev_b(src1, src0);
+    tmp1 = (v16u8)__msa_pckod_b(src1, src0);
+    tmp2 = (v16u8)__msa_pckev_b(src3, src2);
+    tmp3 = (v16u8)__msa_pckod_b(src3, src2);
+    tmpb = (v16u8)__msa_andi_b(tmp0, 0x1F);
+    tmpr = (v16u8)__msa_andi_b(tmp1, 0xF8);
+    nexb = (v16u8)__msa_andi_b(tmp2, 0x1F);
+    nexr = (v16u8)__msa_andi_b(tmp3, 0xF8);
+    reg1 = (v16u8)__msa_andi_b(tmp1, 0x07);
+    reg3 = (v16u8)__msa_andi_b(tmp3, 0x07);
+    reg0 = (v16u8)__msa_srli_b(tmp0, 5);
+    reg1 = (v16u8)__msa_slli_b(reg1, 3);
+    reg2 = (v16u8)__msa_srli_b(tmp2, 5);
+    reg3 = (v16u8)__msa_slli_b(reg3, 3);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    nexg = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_slli_b(tmpb, 3);
+    reg1 = (v16u8)__msa_srli_b(tmpb, 2);
+    reg2 = (v16u8)__msa_slli_b(nexb, 3);
+    reg3 = (v16u8)__msa_srli_b(nexb, 2);
+    tmpb = (v16u8)__msa_or_v(reg1, reg0);
+    nexb = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_slli_b(tmpg, 2);
+    reg1 = (v16u8)__msa_srli_b(tmpg, 4);
+    reg2 = (v16u8)__msa_slli_b(nexg, 2);
+    reg3 = (v16u8)__msa_srli_b(nexg, 4);
+    tmpg = (v16u8)__msa_or_v(reg1, reg0);
+    nexg = (v16u8)__msa_or_v(reg2, reg3);
+    reg0 = (v16u8)__msa_srli_b(tmpr, 5);
+    reg2 = (v16u8)__msa_srli_b(nexr, 5);
+    tmpr = (v16u8)__msa_or_v(tmpr, reg0);
+    nexr = (v16u8)__msa_or_v(nexr, reg2);
+    RGBTOUV(tmpb, tmpg, tmpr, nexb, nexg, nexr, dst);
+    res0 = __msa_copy_u_d((v2i64)dst, 0);
+    res1 = __msa_copy_u_d((v2i64)dst, 1);
     SD(res0, dst_u);
     SD(res1, dst_v);
     s += 16;
@@ -2017,26 +2016,27 @@ void RGB565ToUVRow_MSA(const uint8_t* src_rgb565,
   }
 }
 
-void RGB24ToUVRow_MSA(const uint8_t* src_rgb0,
+void RGB24ToUVRow_MSA(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
   int64_t res0, res1;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v16u8 inp0, inp1, inp2, inp3, inp4, inp5;
   v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
   v8i16 reg0, reg1, reg2, reg3;
   v16u8 dst0;
-  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x70);
-  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x4A);
-  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x26);
-  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x5E);
-  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x12);
+  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x09);
   v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
   v16i8 mask = {0, 1, 2, 16, 3, 4, 5, 17, 6, 7, 8, 18, 9, 10, 11, 19};
   v16i8 zero = {0};
 
@@ -2085,10 +2085,14 @@ void RGB24ToUVRow_MSA(const uint8_t* src_rgb0,
     reg1 += (v8i16)__msa_pckod_d((v2i64)vec3, (v2i64)vec2);
     reg2 += (v8i16)__msa_pckod_d((v2i64)vec5, (v2i64)vec4);
     reg3 += (v8i16)__msa_pckod_d((v2i64)vec7, (v2i64)vec6);
-    reg0 = __msa_srai_h((v8i16)reg0, 2);
-    reg1 = __msa_srai_h((v8i16)reg1, 2);
-    reg2 = __msa_srai_h((v8i16)reg2, 2);
-    reg3 = __msa_srai_h((v8i16)reg3, 2);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg0 = __msa_srai_h((v8i16)reg0, 1);
+    reg1 = __msa_srai_h((v8i16)reg1, 1);
+    reg2 = __msa_srai_h((v8i16)reg2, 1);
+    reg3 = __msa_srai_h((v8i16)reg3, 1);
     vec4 = (v8u16)__msa_pckev_h(reg1, reg0);
     vec5 = (v8u16)__msa_pckev_h(reg3, reg2);
     vec6 = (v8u16)__msa_pckod_h(reg1, reg0);
@@ -2122,26 +2126,27 @@ void RGB24ToUVRow_MSA(const uint8_t* src_rgb0,
   }
 }
 
-void RAWToUVRow_MSA(const uint8_t* src_rgb0,
+void RAWToUVRow_MSA(const uint8_t* src_rgb,
                     int src_stride_rgb,
                     uint8_t* dst_u,
                     uint8_t* dst_v,
                     int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
   int64_t res0, res1;
   v16u8 inp0, inp1, inp2, inp3, inp4, inp5;
   v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
   v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
   v8i16 reg0, reg1, reg2, reg3;
   v16u8 dst0;
-  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x70);
-  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x4A);
-  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x26);
-  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x5E);
-  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x12);
+  v8u16 const_0x70 = (v8u16)__msa_fill_h(0x38);
+  v8u16 const_0x4A = (v8u16)__msa_fill_h(0x25);
+  v8u16 const_0x26 = (v8u16)__msa_fill_h(0x13);
+  v8u16 const_0x5E = (v8u16)__msa_fill_h(0x2f);
+  v8u16 const_0x12 = (v8u16)__msa_fill_h(0x09);
   v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
   v16i8 mask = {0, 1, 2, 16, 3, 4, 5, 17, 6, 7, 8, 18, 9, 10, 11, 19};
   v16i8 zero = {0};
 
@@ -2190,10 +2195,14 @@ void RAWToUVRow_MSA(const uint8_t* src_rgb0,
     reg1 += (v8i16)__msa_pckod_d((v2i64)vec3, (v2i64)vec2);
     reg2 += (v8i16)__msa_pckod_d((v2i64)vec5, (v2i64)vec4);
     reg3 += (v8i16)__msa_pckod_d((v2i64)vec7, (v2i64)vec6);
-    reg0 = __msa_srai_h(reg0, 2);
-    reg1 = __msa_srai_h(reg1, 2);
-    reg2 = __msa_srai_h(reg2, 2);
-    reg3 = __msa_srai_h(reg3, 2);
+    reg0 += const_0x0001;
+    reg1 += const_0x0001;
+    reg2 += const_0x0001;
+    reg3 += const_0x0001;
+    reg0 = __msa_srai_h(reg0, 1);
+    reg1 = __msa_srai_h(reg1, 1);
+    reg2 = __msa_srai_h(reg2, 1);
+    reg3 = __msa_srai_h(reg3, 1);
     vec4 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
     vec5 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
     vec6 = (v8u16)__msa_pckod_h((v8i16)reg1, (v8i16)reg0);
@@ -2236,13 +2245,13 @@ void NV12ToARGBRow_MSA(const uint8_t* src_y,
   uint64_t val0, val1;
   v16u8 src0, src1, res0, res1, dst0, dst1;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 zero = {0};
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -2251,8 +2260,7 @@ void NV12ToARGBRow_MSA(const uint8_t* src_y,
     val1 = LD(src_uv);
     src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
     src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0);
     res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1);
     dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0);
@@ -2273,12 +2281,12 @@ void NV12ToRGB565Row_MSA(const uint8_t* src_y,
   uint64_t val0, val1;
   v16u8 src0, src1, dst0;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
   v16u8 zero = {0};
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -2287,8 +2295,7 @@ void NV12ToRGB565Row_MSA(const uint8_t* src_y,
     val1 = LD(src_uv);
     src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
     src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     vec0 = vec0 >> 3;
     vec1 = (vec1 >> 2) << 5;
     vec2 = (vec2 >> 3) << 11;
@@ -2309,14 +2316,14 @@ void NV21ToARGBRow_MSA(const uint8_t* src_y,
   uint64_t val0, val1;
   v16u8 src0, src1, res0, res1, dst0, dst1;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
   v16u8 zero = {0};
   v16i8 shuffler = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -2326,8 +2333,7 @@ void NV21ToARGBRow_MSA(const uint8_t* src_y,
     src0 = (v16u8)__msa_insert_d((v2i64)zero, 0, val0);
     src1 = (v16u8)__msa_insert_d((v2i64)zero, 0, val1);
     src1 = (v16u8)__msa_vshf_b(shuffler, (v16i8)src1, (v16i8)src1);
-    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src0, src1, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     res0 = (v16u8)__msa_ilvev_b((v16i8)vec2, (v16i8)vec0);
     res1 = (v16u8)__msa_ilvev_b((v16i8)alpha, (v16i8)vec1);
     dst0 = (v16u8)__msa_ilvr_b((v16i8)res1, (v16i8)res0);
@@ -2416,27 +2422,27 @@ void SobelXYRow_MSA(const uint8_t* src_sobelx,
   }
 }
 
-void ARGBToYJRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ARGBToYJRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, src3, dst0;
-  v16u8 const_0x4B0F = (v16u8)__msa_fill_h(0x4B0F);
-  v16u8 const_0x26 = (v16u8)__msa_fill_h(0x26);
-  v8u16 const_0x40 = (v8u16)__msa_fill_h(0x40);
+  v16u8 const_0x961D = (v16u8)__msa_fill_h(0x961D);
+  v16u8 const_0x4D = (v16u8)__msa_fill_h(0x4D);
+  v8u16 const_0x80 = (v8u16)__msa_fill_h(0x80);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((void*)src_argb0, 48);
-    ARGBTOY(src0, src1, src2, src3, const_0x4B0F, const_0x26, const_0x40, 7,
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
+    ARGBTOY(src0, src1, src2, src3, const_0x961D, const_0x4D, const_0x80, 8,
             dst0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 64;
+    src_argb += 64;
     dst_y += 16;
   }
 }
 
-void BGRAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void BGRAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, src3, dst0;
   v16u8 const_0x4200 = (v16u8)__msa_fill_h(0x4200);
@@ -2444,19 +2450,19 @@ void BGRAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((void*)src_argb0, 48);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
     ARGBTOY(src0, src1, src2, src3, const_0x4200, const_0x1981, const_0x1080, 8,
             dst0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 64;
+    src_argb += 64;
     dst_y += 16;
   }
 }
 
-void ABGRToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void ABGRToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, src3, dst0;
   v16u8 const_0x8142 = (v16u8)__msa_fill_h(0x8142);
@@ -2464,19 +2470,19 @@ void ABGRToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((void*)src_argb0, 48);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
     ARGBTOY(src0, src1, src2, src3, const_0x8142, const_0x19, const_0x1080, 8,
             dst0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 64;
+    src_argb += 64;
     dst_y += 16;
   }
 }
 
-void RGBAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
+void RGBAToYRow_MSA(const uint8_t* src_argb, uint8_t* dst_y, int width) {
   int x;
   v16u8 src0, src1, src2, src3, dst0;
   v16u8 const_0x1900 = (v16u8)__msa_fill_h(0x1900);
@@ -2484,81 +2490,143 @@ void RGBAToYRow_MSA(const uint8_t* src_argb0, uint8_t* dst_y, int width) {
   v8u16 const_0x1080 = (v8u16)__msa_fill_h(0x1080);
 
   for (x = 0; x < width; x += 16) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
-    src2 = (v16u8)__msa_ld_b((void*)src_argb0, 32);
-    src3 = (v16u8)__msa_ld_b((void*)src_argb0, 48);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((void*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((void*)src_argb, 48);
     ARGBTOY(src0, src1, src2, src3, const_0x1900, const_0x4281, const_0x1080, 8,
             dst0);
     ST_UB(dst0, dst_y);
-    src_argb0 += 64;
+    src_argb += 64;
     dst_y += 16;
   }
 }
 
-void ARGBToUVJRow_MSA(const uint8_t* src_rgb0,
+void ARGBToUVJRow_MSA(const uint8_t* src_rgb,
                       int src_stride_rgb,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
-  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
-  v16u8 vec0, vec1, vec2, vec3;
-  v16u8 dst0, dst1;
-  v16i8 shuffler0 = {0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};
-  v16i8 shuffler1 = {2,  3,  6,  7,  10, 11, 14, 15,
-                     18, 19, 22, 23, 26, 27, 30, 31};
-  v16i8 shuffler2 = {0, 3, 4, 7, 8, 11, 12, 15, 16, 19, 20, 23, 24, 27, 28, 31};
-  v16i8 shuffler3 = {1, 2, 5, 6, 9, 10, 13, 14, 17, 18, 21, 22, 25, 26, 29, 30};
-  v16u8 const_0x7F = (v16u8)__msa_fill_h(0x7F);
-  v16u8 const_0x6B14 = (v16u8)__msa_fill_h(0x6B14);
-  v16u8 const_0x2B54 = (v16u8)__msa_fill_h(0x2B54);
-  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  v8u16 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 dst0, dst1, dst2, dst3;
+  v16u8 zero = {0};
+  v8i16 shuffler0 = {0, 3, 4, 7, 8, 11, 12, 15};
+  v8i16 shuffler1 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8i16 shuffler2 = {2, 3, 6, 7, 10, 11, 14, 15};
+  v8i16 shuffler3 = {0, 1, 4, 5, 8, 9, 12, 13};
+  v8u16 const_0x0000003f = (v8u16)__msa_fill_w(0x0000003f);
+  v4u32 const_0x00008080 = (v8u16)__msa_fill_w(0x00008080);
+  v8u16 const_0x0015002a = (v8u16)__msa_fill_w(0x0015002a);
+  v8u16 const_0x0035000a = (v8u16)__msa_fill_w(0x0035000a);
+  v4i32 shift = __msa_fill_w(0x00000008);
 
   for (x = 0; x < width; x += 32) {
-    src0 = (v16u8)__msa_ld_b((void*)s, 0);
-    src1 = (v16u8)__msa_ld_b((void*)s, 16);
-    src2 = (v16u8)__msa_ld_b((void*)s, 32);
-    src3 = (v16u8)__msa_ld_b((void*)s, 48);
-    src4 = (v16u8)__msa_ld_b((void*)t, 0);
-    src5 = (v16u8)__msa_ld_b((void*)t, 16);
-    src6 = (v16u8)__msa_ld_b((void*)t, 32);
-    src7 = (v16u8)__msa_ld_b((void*)t, 48);
-    src0 = __msa_aver_u_b(src0, src4);
-    src1 = __msa_aver_u_b(src1, src5);
-    src2 = __msa_aver_u_b(src2, src6);
-    src3 = __msa_aver_u_b(src3, src7);
-    src4 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0);
-    src5 = (v16u8)__msa_pckev_w((v4i32)src3, (v4i32)src2);
-    src6 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
-    src7 = (v16u8)__msa_pckod_w((v4i32)src3, (v4i32)src2);
-    vec0 = __msa_aver_u_b(src4, src6);
-    vec1 = __msa_aver_u_b(src5, src7);
-    src0 = (v16u8)__msa_ld_b((void*)s, 64);
-    src1 = (v16u8)__msa_ld_b((void*)s, 80);
-    src2 = (v16u8)__msa_ld_b((void*)s, 96);
-    src3 = (v16u8)__msa_ld_b((void*)s, 112);
-    src4 = (v16u8)__msa_ld_b((void*)t, 64);
-    src5 = (v16u8)__msa_ld_b((void*)t, 80);
-    src6 = (v16u8)__msa_ld_b((void*)t, 96);
-    src7 = (v16u8)__msa_ld_b((void*)t, 112);
-    src0 = __msa_aver_u_b(src0, src4);
-    src1 = __msa_aver_u_b(src1, src5);
-    src2 = __msa_aver_u_b(src2, src6);
-    src3 = __msa_aver_u_b(src3, src7);
-    src4 = (v16u8)__msa_pckev_w((v4i32)src1, (v4i32)src0);
-    src5 = (v16u8)__msa_pckev_w((v4i32)src3, (v4i32)src2);
-    src6 = (v16u8)__msa_pckod_w((v4i32)src1, (v4i32)src0);
-    src7 = (v16u8)__msa_pckod_w((v4i32)src3, (v4i32)src2);
-    vec2 = __msa_aver_u_b(src4, src6);
-    vec3 = __msa_aver_u_b(src5, src7);
-    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x6B14, const_0x7F, const_0x2B54,
-             const_0x8080, shuffler1, shuffler0, shuffler2, shuffler3, dst0,
-             dst1);
-    ST_UB(dst0, dst_v);
-    ST_UB(dst1, dst_u);
+    src1 = __msa_ld_b((void*)s, 0);
+    src3 = __msa_ld_b((void*)s, 16);
+    src5 = __msa_ld_b((void*)t, 0);
+    src7 = __msa_ld_b((void*)t, 16);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec0 = __msa_aver_u_h(src4, src5);
+    vec1 = __msa_aver_u_h(src6, src7);
+
+    src1 = __msa_ld_b((void*)s, 32);
+    src3 = __msa_ld_b((void*)s, 48);
+    src5 = __msa_ld_b((void*)t, 32);
+    src7 = __msa_ld_b((void*)t, 48);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec2 = __msa_aver_u_h(src4, src5);
+    vec3 = __msa_aver_u_h(src6, src7);
+    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x0000003f, const_0x00008080,
+             const_0x0015002a, const_0x0035000a, shuffler0, shuffler1,
+             shuffler2, shuffler3, shift, dst0, dst1);
+
+    src1 = __msa_ld_b((void*)s, 64);
+    src3 = __msa_ld_b((void*)s, 80);
+    src5 = __msa_ld_b((void*)t, 64);
+    src7 = __msa_ld_b((void*)t, 80);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec0 = __msa_aver_u_h(src4, src5);
+    vec1 = __msa_aver_u_h(src6, src7);
+
+    src1 = __msa_ld_b((void*)s, 96);
+    src3 = __msa_ld_b((void*)s, 112);
+    src5 = __msa_ld_b((void*)t, 96);
+    src7 = __msa_ld_b((void*)t, 112);
+    src0 = __msa_ilvr_b(zero, src1);
+    src1 = __msa_ilvl_b(zero, src1);
+    src2 = __msa_ilvr_b(zero, src3);
+    src3 = __msa_ilvl_b(zero, src3);
+    src4 = __msa_ilvr_b(zero, src5);
+    src5 = __msa_ilvl_b(zero, src5);
+    src6 = __msa_ilvr_b(zero, src7);
+    src7 = __msa_ilvl_b(zero, src7);
+    src0 += src4;
+    src1 += src5;
+    src2 += src6;
+    src3 += src7;
+    src4 = __msa_ilvev_d(src1, src0);
+    src5 = __msa_ilvod_d(src1, src0);
+    src6 = __msa_ilvev_d(src3, src2);
+    src7 = __msa_ilvod_d(src3, src2);
+    vec2 = __msa_aver_u_h(src4, src5);
+    vec3 = __msa_aver_u_h(src6, src7);
+    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x0000003f, const_0x00008080,
+             const_0x0015002a, const_0x0035000a, shuffler0, shuffler1,
+             shuffler2, shuffler3, shift, dst2, dst3);
+
+    dst0 = (v8u16)__msa_pckev_b(dst2, dst0);
+    dst1 = (v8u16)__msa_pckev_b(dst3, dst1);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
     s += 128;
     t += 128;
     dst_v += 16;
@@ -2566,103 +2634,108 @@ void ARGBToUVJRow_MSA(const uint8_t* src_rgb0,
   }
 }
 
-void BGRAToUVRow_MSA(const uint8_t* src_rgb0,
+void BGRAToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
-  v16u8 dst0, dst1, vec0, vec1, vec2, vec3;
-  v16i8 shuffler0 = {0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};
-  v16i8 shuffler1 = {2,  3,  6,  7,  10, 11, 14, 15,
-                     18, 19, 22, 23, 26, 27, 30, 31};
-  v16i8 shuffler2 = {0, 3, 4, 7, 8, 11, 12, 15, 16, 19, 20, 23, 24, 27, 28, 31};
-  v16i8 shuffler3 = {2, 1, 6, 5, 10, 9, 14, 13, 18, 17, 22, 21, 26, 25, 30, 29};
-  v16u8 const_0x125E = (v16u8)__msa_fill_h(0x125E);
-  v16u8 const_0x7000 = (v16u8)__msa_fill_h(0x7000);
-  v16u8 const_0x264A = (v16u8)__msa_fill_h(0x264A);
-  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
+  v16u8 dst0, dst1;
+  v8i16 shuffler0 = {1, unused, 5, unused, 9, unused, 13, unused};
+  v8i16 shuffler1 = {2, 3, 6, 7, 10, 11, 14, 15};
+  v8i16 shuffler2 = {3, unused, 7, unused, 11, unused, 15, unused};
+  v8i16 shuffler3 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
 
-  for (x = 0; x < width; x += 32) {
-    READ_ARGB(s, t, vec0, vec1, vec2, vec3);
-    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x125E, const_0x7000, const_0x264A,
-             const_0x8080, shuffler0, shuffler1, shuffler2, shuffler3, dst0,
-             dst1);
-    ST_UB(dst0, dst_v);
-    ST_UB(dst1, dst_u);
-    s += 128;
-    t += 128;
-    dst_v += 16;
-    dst_u += 16;
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
   }
 }
 
-void ABGRToUVRow_MSA(const uint8_t* src_rgb0,
+void ABGRToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
-  v16u8 src0, src1, src2, src3;
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
   v16u8 dst0, dst1;
-  v16i8 shuffler0 = {0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};
-  v16i8 shuffler1 = {2,  3,  6,  7,  10, 11, 14, 15,
-                     18, 19, 22, 23, 26, 27, 30, 31};
-  v16i8 shuffler2 = {0, 3, 4, 7, 8, 11, 12, 15, 16, 19, 20, 23, 24, 27, 28, 31};
-  v16i8 shuffler3 = {1, 2, 5, 6, 9, 10, 13, 14, 17, 18, 21, 22, 25, 26, 29, 30};
-  v16u8 const_0x4A26 = (v16u8)__msa_fill_h(0x4A26);
-  v16u8 const_0x0070 = (v16u8)__msa_fill_h(0x0070);
-  v16u8 const_0x125E = (v16u8)__msa_fill_h(0x125E);
-  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  v8i16 shuffler0 = {0, unused, 4, unused, 8, unused, 12, unused};
+  v8i16 shuffler1 = {1, 2, 5, 6, 9, 10, 13, 14};
+  v8i16 shuffler2 = {2, unused, 6, unused, 10, unused, 14, unused};
+  v8i16 shuffler3 = {0, 1, 4, 5, 8, 9, 12, 13};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
 
-  for (x = 0; x < width; x += 32) {
-    READ_ARGB(s, t, src0, src1, src2, src3);
-    ARGBTOUV(src0, src1, src2, src3, const_0x4A26, const_0x0070, const_0x125E,
-             const_0x8080, shuffler1, shuffler0, shuffler2, shuffler3, dst0,
-             dst1);
-    ST_UB(dst0, dst_u);
-    ST_UB(dst1, dst_v);
-    s += 128;
-    t += 128;
-    dst_u += 16;
-    dst_v += 16;
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
   }
 }
 
-void RGBAToUVRow_MSA(const uint8_t* src_rgb0,
+void RGBAToUVRow_MSA(const uint8_t* src_rgb,
                      int src_stride_rgb,
                      uint8_t* dst_u,
                      uint8_t* dst_v,
                      int width) {
   int x;
-  const uint8_t* s = src_rgb0;
-  const uint8_t* t = src_rgb0 + src_stride_rgb;
-  v16u8 dst0, dst1, vec0, vec1, vec2, vec3;
-  v16i8 shuffler0 = {0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};
-  v16i8 shuffler1 = {2,  3,  6,  7,  10, 11, 14, 15,
-                     18, 19, 22, 23, 26, 27, 30, 31};
-  v16i8 shuffler2 = {0, 3, 4, 7, 8, 11, 12, 15, 16, 19, 20, 23, 24, 27, 28, 31};
-  v16i8 shuffler3 = {2, 1, 6, 5, 10, 9, 14, 13, 18, 17, 22, 21, 26, 25, 30, 29};
-  v16u8 const_0x125E = (v16u8)__msa_fill_h(0x264A);
-  v16u8 const_0x7000 = (v16u8)__msa_fill_h(0x7000);
-  v16u8 const_0x264A = (v16u8)__msa_fill_h(0x125E);
-  v8u16 const_0x8080 = (v8u16)__msa_fill_h(0x8080);
+  const uint8_t* s = src_rgb;
+  const uint8_t* t = src_rgb + src_stride_rgb;
+  const uint8_t unused = 0xf;
+  v8u16 src0, src1, src2, src3;
+  v16u8 dst0, dst1;
+  v8i16 shuffler0 = {3, unused, 7, unused, 11, unused, 15, unused};
+  v8i16 shuffler1 = {2, 1, 6, 5, 10, 9, 14, 13};
+  v8i16 shuffler2 = {1, unused, 5, unused, 9, unused, 13, unused};
+  v8i16 shuffler3 = {3, 2, 7, 6, 11, 10, 15, 14};
+  v8u16 const_0x09002f = (v8u16)__msa_fill_w(0x09002f);
+  v8u16 const_0x000038 = (v8u16)__msa_fill_w(0x0038);
+  v8u16 const_0x250013 = (v8u16)__msa_fill_w(0x250013);
+  v4u32 const_0x008080 = (v4u32)__msa_fill_w(0x8080);
+  v8u16 const_0x0001 = (v8u16)__msa_fill_h(0x0001);
 
-  for (x = 0; x < width; x += 32) {
-    READ_ARGB(s, t, vec0, vec1, vec2, vec3);
-    ARGBTOUV(vec0, vec1, vec2, vec3, const_0x125E, const_0x7000, const_0x264A,
-             const_0x8080, shuffler0, shuffler1, shuffler2, shuffler3, dst0,
-             dst1);
-    ST_UB(dst0, dst_u);
-    ST_UB(dst1, dst_v);
-    s += 128;
-    t += 128;
-    dst_u += 16;
-    dst_v += 16;
+  for (x = 0; x < width; x += 16) {
+    READ_ARGB(s, t, src0, src1, src2, src3, const_0x0001);
+    ARGBTOUV_H(src0, src1, src2, src3, const_0x09002f, const_0x000038,
+               const_0x250013, const_0x008080, shuffler0, shuffler1, shuffler2,
+               shuffler3, dst0, dst1);
+    *((uint64_t*)dst_v) = __msa_copy_u_d((v2i64)dst0, 0);
+    *((uint64_t*)dst_u) = __msa_copy_u_d((v2i64)dst1, 0);
+    s += 64;
+    t += 64;
+    dst_u += 8;
+    dst_v += 8;
   }
 }
 
@@ -2674,54 +2747,57 @@ void I444ToARGBRow_MSA(const uint8_t* src_y,
                        int width) {
   int x;
   v16u8 src0, src1, src2, dst0, dst1;
-  v8u16 vec0, vec1, vec2;
+  v8i16 vec0, vec1, vec2;
   v4i32 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, reg8, reg9;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
   v8i16 zero = {0};
+  v4i32 const_0x80 = __msa_fill_w(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
 
   for (x = 0; x < width; x += 8) {
     READI444(src_y, src_u, src_v, src0, src1, src2);
-    vec0 = (v8u16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
+    vec0 = (v8i16)__msa_ilvr_b((v16i8)src0, (v16i8)src0);
     reg0 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
     reg1 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
     reg0 *= vec_yg;
     reg1 *= vec_yg;
     reg0 = __msa_srai_w(reg0, 16);
     reg1 = __msa_srai_w(reg1, 16);
-    reg4 = reg0 + vec_br;
-    reg5 = reg1 + vec_br;
-    reg2 = reg0 + vec_bg;
-    reg3 = reg1 + vec_bg;
-    reg0 += vec_bb;
-    reg1 += vec_bb;
+    reg0 += vec_yb;
+    reg1 += vec_yb;
     vec0 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src1);
     vec1 = (v8u16)__msa_ilvr_b((v16i8)zero, (v16i8)src2);
     reg6 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec0);
     reg7 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec0);
     reg8 = (v4i32)__msa_ilvr_h((v8i16)zero, (v8i16)vec1);
     reg9 = (v4i32)__msa_ilvl_h((v8i16)zero, (v8i16)vec1);
-    reg0 -= reg6 * vec_ub;
-    reg1 -= reg7 * vec_ub;
-    reg2 -= reg6 * vec_ug;
-    reg3 -= reg7 * vec_ug;
-    reg4 -= reg8 * vec_vr;
-    reg5 -= reg9 * vec_vr;
-    reg2 -= reg8 * vec_vg;
-    reg3 -= reg9 * vec_vg;
-    reg0 = __msa_srai_w(reg0, 6);
-    reg1 = __msa_srai_w(reg1, 6);
-    reg2 = __msa_srai_w(reg2, 6);
-    reg3 = __msa_srai_w(reg3, 6);
-    reg4 = __msa_srai_w(reg4, 6);
-    reg5 = __msa_srai_w(reg5, 6);
+    reg6 -= const_0x80;
+    reg7 -= const_0x80;
+    reg8 -= const_0x80;
+    reg9 -= const_0x80;
+    tmp0 = reg0 + reg6 * vec_ub;
+    tmp1 = reg1 + reg7 * vec_ub;
+    tmp2 = reg0 + reg8 * vec_vr;
+    tmp3 = reg1 + reg9 * vec_vr;
+    tmp4 = reg6 * vec_ug;
+    tmp5 = reg7 * vec_ug;
+    tmp4 += reg8 * vec_vg;
+    tmp5 += reg9 * vec_vg;
+    tmp4 = reg0 - tmp4;
+    tmp5 = reg1 - tmp5;
+    reg0 = __msa_srai_w(tmp0, 6);
+    reg1 = __msa_srai_w(tmp1, 6);
+    reg2 = __msa_srai_w(tmp2, 6);
+    reg3 = __msa_srai_w(tmp3, 6);
+    reg4 = __msa_srai_w(tmp4, 6);
+    reg5 = __msa_srai_w(tmp5, 6);
     CLIP_0TO255(reg0, reg1, reg2, reg3, reg4, reg5);
     vec0 = (v8u16)__msa_pckev_h((v8i16)reg1, (v8i16)reg0);
-    vec1 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
-    vec2 = (v8u16)__msa_pckev_h((v8i16)reg5, (v8i16)reg4);
+    vec1 = (v8u16)__msa_pckev_h((v8i16)reg5, (v8i16)reg4);
+    vec2 = (v8u16)__msa_pckev_h((v8i16)reg3, (v8i16)reg2);
     vec0 = (v8u16)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
     vec1 = (v8u16)__msa_ilvev_b((v16i8)alpha, (v16i8)vec2);
     dst0 = (v16u8)__msa_ilvr_h((v8i16)vec1, (v8i16)vec0);
@@ -2734,13 +2810,24 @@ void I444ToARGBRow_MSA(const uint8_t* src_y,
   }
 }
 
-void I400ToARGBRow_MSA(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+// TODO - respect YuvConstants
+void I400ToARGBRow_MSA(const uint8_t* src_y,
+                       uint8_t* dst_argb,
+                       const struct YuvConstants* yuvconstants,
+                       int width) {
   int x;
+#if defined(__aarch64__) || defined(__arm__)
+  int ygb = yuvconstants->kUVBiasBGR[3];
+  int yg = yuvconstants->kYToRgb[1];
+#else
+  int ygb = yuvconstants->kYBiasToRgb[0];
+  int yg = yuvconstants->kYToRgb[0];
+#endif
   v16u8 src0, res0, res1, res2, res3, res4, dst0, dst1, dst2, dst3;
   v8i16 vec0, vec1;
   v4i32 reg0, reg1, reg2, reg3;
-  v4i32 vec_yg = __msa_fill_w(0x4A35);
-  v8i16 vec_ygb = __msa_fill_h(0xFB78);
+  v4i32 vec_yg = __msa_fill_w(yg);
+  v8i16 vec_ygb = __msa_fill_h(ygb);
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
   v8i16 max = __msa_ldi_h(0xFF);
   v8i16 zero = {0};
@@ -2814,12 +2901,12 @@ void YUY2ToARGBRow_MSA(const uint8_t* src_yuy2,
   int x;
   v16u8 src0, src1, src2;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -2827,8 +2914,7 @@ void YUY2ToARGBRow_MSA(const uint8_t* src_yuy2,
     src0 = (v16u8)__msa_ld_b((void*)src_yuy2, 0);
     src1 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0);
     src2 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0);
-    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
     src_yuy2 += 16;
     dst_argb += 32;
@@ -2842,12 +2928,12 @@ void UYVYToARGBRow_MSA(const uint8_t* src_uyvy,
   int x;
   v16u8 src0, src1, src2;
   v8i16 vec0, vec1, vec2;
-  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg, vec_br, vec_yg;
+  v4i32 vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb;
   v4i32 vec_ubvr, vec_ugvg;
+  v8i16 const_0x80 = __msa_ldi_h(0x80);
   v16u8 alpha = (v16u8)__msa_ldi_b(ALPHA_VAL);
 
-  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_bb, vec_bg,
-                 vec_br, vec_yg);
+  YUVTORGB_SETUP(yuvconstants, vec_ub, vec_vr, vec_ug, vec_vg, vec_yg, vec_yb);
   vec_ubvr = __msa_ilvr_w(vec_vr, vec_ub);
   vec_ugvg = (v4i32)__msa_ilvev_h((v8i16)vec_vg, (v8i16)vec_ug);
 
@@ -2855,8 +2941,7 @@ void UYVYToARGBRow_MSA(const uint8_t* src_uyvy,
     src0 = (v16u8)__msa_ld_b((void*)src_uyvy, 0);
     src1 = (v16u8)__msa_pckod_b((v16i8)src0, (v16i8)src0);
     src2 = (v16u8)__msa_pckev_b((v16i8)src0, (v16i8)src0);
-    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_bb, vec_bg, vec_br, vec_yg,
-             vec0, vec1, vec2);
+    YUVTORGB(src1, src2, vec_ubvr, vec_ugvg, vec_yg, vec_yb, vec0, vec1, vec2);
     STOREARGB(vec0, vec1, vec2, alpha, dst_argb);
     src_uyvy += 16;
     dst_argb += 32;
@@ -3001,12 +3086,12 @@ void ARGBExtractAlphaRow_MSA(const uint8_t* src_argb,
   }
 }
 
-void ARGBBlendRow_MSA(const uint8_t* src_argb0,
+void ARGBBlendRow_MSA(const uint8_t* src_argb,
                       const uint8_t* src_argb1,
                       uint8_t* dst_argb,
                       int width) {
   int x;
-  v16u8 src0, src1, src2, src3, dst0, dst1;
+  v16u8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
   v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
   v8u16 vec8, vec9, vec10, vec11, vec12, vec13;
   v8u16 const_256 = (v8u16)__msa_ldi_h(256);
@@ -3015,8 +3100,8 @@ void ARGBBlendRow_MSA(const uint8_t* src_argb0,
   v16i8 zero = {0};
 
   for (x = 0; x < width; x += 8) {
-    src0 = (v16u8)__msa_ld_b((void*)src_argb0, 0);
-    src1 = (v16u8)__msa_ld_b((void*)src_argb0, 16);
+    src0 = (v16u8)__msa_ld_b((void*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((void*)src_argb, 16);
     src2 = (v16u8)__msa_ld_b((void*)src_argb1, 0);
     src3 = (v16u8)__msa_ld_b((void*)src_argb1, 16);
     vec0 = (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
@@ -3051,16 +3136,16 @@ void ARGBBlendRow_MSA(const uint8_t* src_argb0,
     vec9 = (v8u16)__msa_srai_h((v8i16)vec9, 8);
     vec10 = (v8u16)__msa_srai_h((v8i16)vec10, 8);
     vec11 = (v8u16)__msa_srai_h((v8i16)vec11, 8);
-    vec0 += vec8;
-    vec1 += vec9;
-    vec2 += vec10;
-    vec3 += vec11;
     dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
     dst1 = (v16u8)__msa_pckev_b((v16i8)vec3, (v16i8)vec2);
+    dst2 = (v16u8)__msa_pckev_b((v16i8)vec9, (v16i8)vec8);
+    dst3 = (v16u8)__msa_pckev_b((v16i8)vec11, (v16i8)vec10);
+    dst0 = (v16u8)__msa_adds_u_b(dst0, dst2);
+    dst1 = (v16u8)__msa_adds_u_b(dst1, dst3);
     dst0 = __msa_bmnz_v(dst0, const_255, mask);
     dst1 = __msa_bmnz_v(dst1, const_255, mask);
     ST_UB2(dst0, dst1, dst_argb, 16);
-    src_argb0 += 32;
+    src_argb += 32;
     src_argb1 += 32;
     dst_argb += 32;
   }
@@ -3082,7 +3167,7 @@ void ARGBQuantizeRow_MSA(uint8_t* dst_argb,
   v16i8 mask = {0, 1, 2, 19, 4, 5, 6, 23, 8, 9, 10, 27, 12, 13, 14, 31};
   v16i8 zero = {0};
 
-  for (x = 0; x < width; x += 8) {
+  for (x = 0; x < width; x += 16) {
     src0 = (v16u8)__msa_ld_b((void*)dst_argb, 0);
     src1 = (v16u8)__msa_ld_b((void*)dst_argb, 16);
     src2 = (v16u8)__msa_ld_b((void*)dst_argb, 32);
@@ -3315,10 +3400,10 @@ void SetRow_MSA(uint8_t* dst, uint8_t v8, int width) {
   }
 }
 
-void MirrorUVRow_MSA(const uint8_t* src_uv,
-                     uint8_t* dst_u,
-                     uint8_t* dst_v,
-                     int width) {
+void MirrorSplitUVRow_MSA(const uint8_t* src_uv,
+                          uint8_t* dst_u,
+                          uint8_t* dst_v,
+                          int width) {
   int x;
   v16u8 src0, src1, src2, src3;
   v16u8 dst0, dst1, dst2, dst3;
diff --git a/files/source/row_neon.cc b/files/source/row_neon.cc
index a12fa790..804ff839 100644
--- a/files/source/row_neon.cc
+++ b/files/source/row_neon.cc
@@ -10,8 +10,6 @@
 
 #include "libyuv/row.h"
 
-#include <stdio.h>
-
 #ifdef __cplusplus
 namespace libyuv {
 extern "C" {
@@ -21,90 +19,118 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
     !defined(__aarch64__)
 
+// d8-d15, r4-r11,r14(lr) need to be preserved if used. r13(sp),r15(pc) are
+// reserved.
+
+// q0: Y uint16x8_t
+// d2: U uint8x8_t
+// d3: V uint8x8_t
+
 // Read 8 Y, 4 U and 4 V from 422
 #define READYUV422                               \
-  "vld1.8     {d0}, [%0]!                    \n" \
-  "vld1.32    {d2[0]}, [%1]!                 \n" \
-  "vld1.32    {d2[1]}, [%2]!                 \n"
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vld1.32    {d2[0]}, [%[src_u]]!           \n" \
+  "vld1.32    {d2[1]}, [%[src_v]]!           \n" \
+  "vmov.u8    d1, d0                         \n" \
+  "vmovl.u8   q1, d2                         \n" \
+  "vzip.u8    d0, d1                         \n" \
+  "vsli.u16   q1, q1, #8                     \n"
 
 // Read 8 Y, 8 U and 8 V from 444
 #define READYUV444                               \
-  "vld1.8     {d0}, [%0]!                    \n" \
-  "vld1.8     {d2}, [%1]!                    \n" \
-  "vld1.8     {d3}, [%2]!                    \n" \
-  "vpaddl.u8  q1, q1                         \n" \
-  "vrshrn.u16 d2, q1, #1                     \n"
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vld1.8     {d2}, [%[src_u]]!              \n" \
+  "vmovl.u8   q0, d0                         \n" \
+  "vld1.8     {d3}, [%[src_v]]!              \n" \
+  "vsli.u16   q0, q0, #8                     \n"
 
 // Read 8 Y, and set 4 U and 4 V to 128
 #define READYUV400                               \
-  "vld1.8     {d0}, [%0]!                    \n" \
-  "vmov.u8    d2, #128                       \n"
+  "vld1.8     {d0}, [%[src_y]]!              \n" \
+  "vmov.u8    q1, #128                       \n" \
+  "vmovl.u8   q0, d0                         \n" \
+  "vsli.u16   q0, q0, #8                     \n"
 
 // Read 8 Y and 4 UV from NV12
-#define READNV12                                                               \
-  "vld1.8     {d0}, [%0]!                    \n"                               \
-  "vld1.8     {d2}, [%1]!                    \n"                               \
-  "vmov.u8    d3, d2                         \n" /* split odd/even uv apart */ \
-  "vuzp.u8    d2, d3                         \n"                               \
-  "vtrn.u32   d2, d3                         \n"
+#define READNV12                                                              \
+  "vld1.8     {d0}, [%[src_y]]!              \n"                              \
+  "vld1.8     {d2}, [%[src_uv]]!             \n"                              \
+  "vmov.u8    d1, d0                         \n"                              \
+  "vmov.u8    d3, d2                         \n"                              \
+  "vzip.u8    d0, d1                         \n"                              \
+  "vsli.u16   d2, d2, #8                     \n" /* Duplicate low byte (U) */ \
+  "vsri.u16   d3, d3, #8                     \n" /* Duplicate high byte (V) */
 
 // Read 8 Y and 4 VU from NV21
 #define READNV21                                                               \
-  "vld1.8     {d0}, [%0]!                    \n"                               \
-  "vld1.8     {d2}, [%1]!                    \n"                               \
-  "vmov.u8    d3, d2                         \n" /* split odd/even uv apart */ \
-  "vuzp.u8    d3, d2                         \n"                               \
-  "vtrn.u32   d2, d3                         \n"
+  "vld1.8     {d0}, [%[src_y]]!              \n"                               \
+  "vld1.8     {d2}, [%[src_vu]]!             \n"                               \
+  "vmov.u8    d1, d0                         \n"                               \
+  "vmov.u8    d3, d2                         \n"                               \
+  "vzip.u8    d0, d1                         \n"                               \
+  "vsri.u16   d2, d2, #8                     \n" /* Duplicate high byte (U) */ \
+  "vsli.u16   d3, d3, #8                     \n" /* Duplicate low byte (V) */
 
 // Read 8 YUY2
 #define READYUY2                                 \
-  "vld2.8     {d0, d2}, [%0]!                \n" \
+  "vld2.8     {d0, d2}, [%[src_yuy2]]!       \n" \
+  "vmovl.u8   q0, d0                         \n" \
   "vmov.u8    d3, d2                         \n" \
-  "vuzp.u8    d2, d3                         \n" \
-  "vtrn.u32   d2, d3                         \n"
+  "vsli.u16   q0, q0, #8                     \n" \
+  "vsli.u16   d2, d2, #8                     \n" \
+  "vsri.u16   d3, d3, #8                     \n"
 
 // Read 8 UYVY
 #define READUYVY                                 \
-  "vld2.8     {d2, d3}, [%0]!                \n" \
-  "vmov.u8    d0, d3                         \n" \
+  "vld2.8     {d2, d3}, [%[src_uyvy]]!       \n" \
+  "vmovl.u8   q0, d3                         \n" \
   "vmov.u8    d3, d2                         \n" \
-  "vuzp.u8    d2, d3                         \n" \
-  "vtrn.u32   d2, d3                         \n"
-
-#define YUVTORGB_SETUP                             \
-  "vld1.8     {d24}, [%[kUVToRB]]            \n"   \
-  "vld1.8     {d25}, [%[kUVToG]]             \n"   \
-  "vld1.16    {d26[], d27[]}, [%[kUVBiasBGR]]! \n" \
-  "vld1.16    {d8[], d9[]}, [%[kUVBiasBGR]]!   \n" \
-  "vld1.16    {d28[], d29[]}, [%[kUVBiasBGR]]  \n" \
-  "vld1.32    {d30[], d31[]}, [%[kYToRgb]]     \n"
-
-#define YUVTORGB                                                              \
-  "vmull.u8   q8, d2, d24                    \n" /* u/v B/R component      */ \
-  "vmull.u8   q9, d2, d25                    \n" /* u/v G component        */ \
-  "vmovl.u8   q0, d0                         \n" /* Y                      */ \
-  "vmovl.s16  q10, d1                        \n"                              \
-  "vmovl.s16  q0, d0                         \n"                              \
-  "vmul.s32   q10, q10, q15                  \n"                              \
-  "vmul.s32   q0, q0, q15                    \n"                              \
-  "vqshrun.s32 d0, q0, #16                   \n"                              \
-  "vqshrun.s32 d1, q10, #16                  \n" /* Y                      */ \
-  "vadd.s16   d18, d19                       \n"                              \
-  "vshll.u16  q1, d16, #16                   \n" /* Replicate u * UB       */ \
-  "vshll.u16  q10, d17, #16                  \n" /* Replicate v * VR       */ \
-  "vshll.u16  q3, d18, #16                   \n" /* Replicate (v*VG + u*UG)*/ \
-  "vaddw.u16  q1, q1, d16                    \n"                              \
-  "vaddw.u16  q10, q10, d17                  \n"                              \
-  "vaddw.u16  q3, q3, d18                    \n"                              \
-  "vqadd.s16  q8, q0, q13                    \n" /* B */                      \
-  "vqadd.s16  q9, q0, q14                    \n" /* R */                      \
-  "vqadd.s16  q0, q0, q4                     \n" /* G */                      \
-  "vqadd.s16  q8, q8, q1                     \n" /* B */                      \
-  "vqadd.s16  q9, q9, q10                    \n" /* R */                      \
-  "vqsub.s16  q0, q0, q3                     \n" /* G */                      \
-  "vqshrun.s16 d20, q8, #6                   \n" /* B */                      \
-  "vqshrun.s16 d22, q9, #6                   \n" /* R */                      \
-  "vqshrun.s16 d21, q0, #6                   \n" /* G */
+  "vsli.u16   q0, q0, #8                     \n" \
+  "vsli.u16   d2, d2, #8                     \n" \
+  "vsri.u16   d3, d3, #8                     \n"
+
+#define YUVTORGB_SETUP                                        \
+  "vld4.8     {d26[], d27[], d28[], d29[]}, [%[kUVCoeff]] \n" \
+  "vld1.16    {d31[]}, [%[kRGBCoeffBias]]!   \n"              \
+  "vld1.16    {d20[], d21[]}, [%[kRGBCoeffBias]]! \n"         \
+  "vld1.16    {d22[], d23[]}, [%[kRGBCoeffBias]]! \n"         \
+  "vld1.16    {d24[], d25[]}, [%[kRGBCoeffBias]] \n"
+
+// q0: B uint16x8_t
+// q1: G uint16x8_t
+// q2: R uint16x8_t
+
+// Convert from YUV to 2.14 fixed point RGB
+#define YUVTORGB                                           \
+  "vmull.u16  q2, d1, d31                    \n"           \
+  "vmull.u8   q8, d3, d29                    \n" /* DGV */ \
+  "vmull.u16  q0, d0, d31                    \n"           \
+  "vmlal.u8   q8, d2, d28                    \n" /* DG */  \
+  "vqshrn.u32 d0, q0, #16                    \n"           \
+  "vqshrn.u32 d1, q2, #16                    \n" /* Y */   \
+  "vmull.u8   q9, d2, d26                    \n" /* DB */  \
+  "vmull.u8   q2, d3, d27                    \n" /* DR */  \
+  "vadd.u16   q4, q0, q11                    \n" /* G */   \
+  "vadd.u16   q2, q0, q2                     \n" /* R */   \
+  "vadd.u16   q0, q0, q9                     \n" /* B */   \
+  "vqsub.u16  q1, q4, q8                     \n" /* G */   \
+  "vqsub.u16  q0, q0, q10                    \n" /* B */   \
+  "vqsub.u16  q2, q2, q12                    \n" /* R */
+
+// Convert from 2.14 fixed point RGB To 8 bit RGB
+#define RGBTORGB8                                        \
+  "vqshrn.u16 d4, q2, #6                     \n" /* R */ \
+  "vqshrn.u16 d2, q1, #6                     \n" /* G */ \
+  "vqshrn.u16 d0, q0, #6                     \n" /* B */
+
+#define YUVTORGB_REGS \
+  "q0", "q1", "q2", "q4", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "d31"
+
+#define STORERGBA                                \
+  "vmov.u8    d1, d0                         \n" \
+  "vmov.u8    d3, d4                         \n" \
+  "vmov.u8    d0, d6                         \n" \
+  "vst4.8     {d0, d1, d2, d3}, [%[dst_rgba]]! \n"
 
 void I444ToARGBRow_NEON(const uint8_t* src_y,
                         const uint8_t* src_u,
@@ -114,22 +140,20 @@ void I444ToARGBRow_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "vmov.u8    d23, #255                      \n"
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV444 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),     // %0
-        "+r"(src_u),     // %1
-        "+r"(src_v),     // %2
-        "+r"(dst_argb),  // %3
-        "+r"(width)      // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void I422ToARGBRow_NEON(const uint8_t* src_y,
@@ -140,22 +164,46 @@ void I422ToARGBRow_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "vmov.u8    d23, #255                      \n"
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),     // %0
-        "+r"(src_u),     // %1
-        "+r"(src_v),     // %2
-        "+r"(dst_argb),  // %3
-        "+r"(width)      // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
+}
+
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "vld1.8      {d6}, [%[src_a]]!             \n"
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
@@ -168,22 +216,20 @@ void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
   asm volatile(
       YUVTORGB_SETUP
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %5, %5, #8                     \n"
-      "vld1.8     {d23}, [%3]!                   \n"
-      "vst4.8     {d20, d21, d22, d23}, [%4]!    \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),     // %0
-        "+r"(src_u),     // %1
-        "+r"(src_v),     // %2
-        "+r"(src_a),     // %3
-        "+r"(dst_argb),  // %4
-        "+r"(width)      // %5
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "vld1.8      {d6}, [%[src_a]]!             \n"
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void I422ToRGBARow_NEON(const uint8_t* src_y,
@@ -194,22 +240,18 @@ void I422ToRGBARow_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vmov.u8    d19, #255                      \n"  // YUVTORGB modified d19
-      "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),     // %0
-        "+r"(src_u),     // %1
-        "+r"(src_v),     // %2
-        "+r"(dst_rgba),  // %3
-        "+r"(width)      // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      RGBTORGB8 "subs        %[width], %[width], #8        \n" STORERGBA
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgba] "+r"(dst_rgba),                         // %[dst_rgba]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void I422ToRGB24Row_NEON(const uint8_t* src_y,
@@ -220,29 +262,28 @@ void I422ToRGB24Row_NEON(const uint8_t* src_y,
                          int width) {
   asm volatile(
       YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vst3.8     {d20, d21, d22}, [%3]!         \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),      // %0
-        "+r"(src_u),      // %1
-        "+r"(src_v),      // %2
-        "+r"(dst_rgb24),  // %3
-        "+r"(width)       // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 #define ARGBTORGB565                                                        \
-  "vshll.u8    q0, d22, #8                   \n" /* R                    */ \
-  "vshll.u8    q8, d21, #8                   \n" /* G                    */ \
-  "vshll.u8    q9, d20, #8                   \n" /* B                    */ \
-  "vsri.16     q0, q8, #5                    \n" /* RG                   */ \
-  "vsri.16     q0, q9, #11                   \n" /* RGB                  */
+  "vshll.u8    q2, d4, #8                    \n" /* R                    */ \
+  "vshll.u8    q1, d2, #8                    \n" /* G                    */ \
+  "vshll.u8    q0, d0, #8                    \n" /* B                    */ \
+  "vsri.16     q2, q1, #5                    \n" /* RG                   */ \
+  "vsri.16     q2, q0, #11                   \n" /* RGB                  */
 
 void I422ToRGB565Row_NEON(const uint8_t* src_y,
                           const uint8_t* src_u,
@@ -252,31 +293,29 @@ void I422ToRGB565Row_NEON(const uint8_t* src_y,
                           int width) {
   asm volatile(
       YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n" ARGBTORGB565
-      "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
-      "bgt        1b                             \n"
-      : "+r"(src_y),       // %0
-        "+r"(src_u),       // %1
-        "+r"(src_v),       // %2
-        "+r"(dst_rgb565),  // %3
-        "+r"(width)        // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      RGBTORGB8 "subs        %[width], %[width], #8        \n" ARGBTORGB565
+      "vst1.8      {q2}, [%[dst_rgb565]]!        \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 #define ARGBTOARGB1555                                                      \
-  "vshll.u8    q0, d23, #8                   \n" /* A                    */ \
-  "vshll.u8    q8, d22, #8                   \n" /* R                    */ \
-  "vshll.u8    q9, d21, #8                   \n" /* G                    */ \
-  "vshll.u8    q10, d20, #8                  \n" /* B                    */ \
-  "vsri.16     q0, q8, #1                    \n" /* AR                   */ \
-  "vsri.16     q0, q9, #6                    \n" /* ARG                  */ \
-  "vsri.16     q0, q10, #11                  \n" /* ARGB                 */
+  "vshll.u8    q3, d6, #8                    \n" /* A                    */ \
+  "vshll.u8    q2, d4, #8                    \n" /* R                    */ \
+  "vshll.u8    q1, d2, #8                    \n" /* G                    */ \
+  "vshll.u8    q0, d0, #8                    \n" /* B                    */ \
+  "vsri.16     q3, q2, #1                    \n" /* AR                   */ \
+  "vsri.16     q3, q1, #6                    \n" /* ARG                  */ \
+  "vsri.16     q3, q0, #11                   \n" /* ARGB                 */
 
 void I422ToARGB1555Row_NEON(const uint8_t* src_y,
                             const uint8_t* src_u,
@@ -287,30 +326,28 @@ void I422ToARGB1555Row_NEON(const uint8_t* src_y,
   asm volatile(
       YUVTORGB_SETUP
       "1:                                        \n" READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vmov.u8    d23, #255                      \n" ARGBTOARGB1555
-      "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels
-      "bgt        1b                             \n"
-      : "+r"(src_y),         // %0
-        "+r"(src_u),         // %1
-        "+r"(src_v),         // %2
-        "+r"(dst_argb1555),  // %3
-        "+r"(width)          // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vmov.u8     d6, #0xff                     \n" ARGBTOARGB1555
+      "vst1.8      {q3}, [%[dst_argb1555]]!      \n"  // store 8 pixels RGB1555.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb1555] "+r"(dst_argb1555),                 // %[dst_argb1555]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "q3");
 }
 
 #define ARGBTOARGB4444                                                      \
-  "vshr.u8    d20, d20, #4                   \n" /* B                    */ \
-  "vbic.32    d21, d21, d4                   \n" /* G                    */ \
-  "vshr.u8    d22, d22, #4                   \n" /* R                    */ \
-  "vbic.32    d23, d23, d4                   \n" /* A                    */ \
-  "vorr       d0, d20, d21                   \n" /* BG                   */ \
-  "vorr       d1, d22, d23                   \n" /* RA                   */ \
+  "vshr.u8    d0, d0, #4                     \n" /* B                    */ \
+  "vbic.32    d2, d2, d7                     \n" /* G                    */ \
+  "vshr.u8    d4, d4, #4                     \n" /* R                    */ \
+  "vbic.32    d6, d6, d7                     \n" /* A                    */ \
+  "vorr       d0, d0, d2                     \n" /* BG                   */ \
+  "vorr       d1, d4, d6                     \n" /* RA                   */ \
   "vzip.u8    d0, d1                         \n" /* BGRA                 */
 
 void I422ToARGB4444Row_NEON(const uint8_t* src_y,
@@ -321,56 +358,53 @@ void I422ToARGB4444Row_NEON(const uint8_t* src_y,
                             int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "vmov.u8    d4, #0x0f                      \n"  // vbic bits to clear
-      "1:                                        \n"
-
-      READYUV422 YUVTORGB
-      "subs       %4, %4, #8                     \n"
-      "vmov.u8    d23, #255                      \n" ARGBTOARGB4444
-      "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels
-      "bgt        1b                             \n"
-      : "+r"(src_y),         // %0
-        "+r"(src_u),         // %1
-        "+r"(src_v),         // %2
-        "+r"(dst_argb4444),  // %3
-        "+r"(width)          // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      "vmov.u8     d6, #255                      \n"
+      "vmov.u8     d7, #0x0f                     \n"  // vbic bits to clear
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n" ARGBTOARGB4444
+      "vst1.8      {q0}, [%[dst_argb4444]]!      \n"  // store 8 pixels
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb4444] "+r"(dst_argb4444),                 // %[dst_argb4444]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "q3");
 }
 
-void I400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "vmov.u8    d23, #255                      \n"
+      "vmov.u8     d6, #255                      \n"
       "1:                                        \n" READYUV400 YUVTORGB
-      "subs       %2, %2, #8                     \n"
-      "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),     // %0
-        "+r"(dst_argb),  // %1
-        "+r"(width)      // %2
-      : [kUVToRB] "r"(&kYuvI601Constants.kUVToRB),
-        [kUVToG] "r"(&kYuvI601Constants.kUVToG),
-        [kUVBiasBGR] "r"(&kYuvI601Constants.kUVBiasBGR),
-        [kYToRgb] "r"(&kYuvI601Constants.kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+          RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
   asm volatile(
-      "vmov.u8    d23, #255                      \n"
+      "vmov.u8     d23, #255                     \n"
       "1:                                        \n"
-      "vld1.8     {d20}, [%0]!                   \n"
-      "vmov       d21, d20                       \n"
-      "vmov       d22, d20                       \n"
-      "subs       %2, %2, #8                     \n"
-      "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-      "bgt        1b                             \n"
+      "vld1.8      {d20}, [%0]!                  \n"
+      "vmov        d21, d20                      \n"
+      "vmov        d22, d20                      \n"
+      "subs        %2, %2, #8                    \n"
+      "vst4.8      {d20, d21, d22, d23}, [%1]!   \n"
+      "bgt         1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -383,22 +417,20 @@ void NV12ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile(YUVTORGB_SETUP
-               "vmov.u8    d23, #255                      \n"
-               "1:                                        \n" READNV12 YUVTORGB
-               "subs       %3, %3, #8                     \n"
-               "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-               "bgt        1b                             \n"
-               : "+r"(src_y),     // %0
-                 "+r"(src_uv),    // %1
-                 "+r"(dst_argb),  // %2
-                 "+r"(width)      // %3
-               : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-                 [kUVToG] "r"(&yuvconstants->kUVToG),
-                 [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-                 [kYToRgb] "r"(&yuvconstants->kYToRgb)
-               : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9",
-                 "q10", "q11", "q12", "q13", "q14", "q15");
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void NV21ToARGBRow_NEON(const uint8_t* src_y,
@@ -406,22 +438,20 @@ void NV21ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile(YUVTORGB_SETUP
-               "vmov.u8    d23, #255                      \n"
-               "1:                                        \n" READNV21 YUVTORGB
-               "subs       %3, %3, #8                     \n"
-               "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
-               "bgt        1b                             \n"
-               : "+r"(src_y),     // %0
-                 "+r"(src_vu),    // %1
-                 "+r"(dst_argb),  // %2
-                 "+r"(width)      // %3
-               : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-                 [kUVToG] "r"(&yuvconstants->kUVToG),
-                 [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-                 [kYToRgb] "r"(&yuvconstants->kYToRgb)
-               : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9",
-                 "q10", "q11", "q12", "q13", "q14", "q15");
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV21 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_vu] "+r"(src_vu),                             // %[src_vu]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void NV12ToRGB24Row_NEON(const uint8_t* src_y,
@@ -430,25 +460,19 @@ void NV12ToRGB24Row_NEON(const uint8_t* src_y,
                          const struct YuvConstants* yuvconstants,
                          int width) {
   asm volatile(
-
       YUVTORGB_SETUP
-
-      "1:                                        \n"
-
-      READNV12 YUVTORGB
-      "subs       %3, %3, #8                     \n"
-      "vst3.8     {d20, d21, d22}, [%2]!         \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),      // %0
-        "+r"(src_uv),     // %1
-        "+r"(dst_rgb24),  // %2
-        "+r"(width)       // %3
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 void NV21ToRGB24Row_NEON(const uint8_t* src_y,
@@ -457,25 +481,19 @@ void NV21ToRGB24Row_NEON(const uint8_t* src_y,
                          const struct YuvConstants* yuvconstants,
                          int width) {
   asm volatile(
-
       YUVTORGB_SETUP
-
-      "1:                                        \n"
-
-      READNV21 YUVTORGB
-      "subs       %3, %3, #8                     \n"
-      "vst3.8     {d20, d21, d22}, [%2]!         \n"
-      "bgt        1b                             \n"
-      : "+r"(src_y),      // %0
-        "+r"(src_vu),     // %1
-        "+r"(dst_rgb24),  // %2
-        "+r"(width)       // %3
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV21 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst3.8      {d0, d2, d4}, [%[dst_rgb24]]! \n"
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_vu] "+r"(src_vu),                             // %[src_vu]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 void NV12ToRGB565Row_NEON(const uint8_t* src_y,
@@ -485,62 +503,56 @@ void NV12ToRGB565Row_NEON(const uint8_t* src_y,
                           int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "1:                                        \n" READNV12 YUVTORGB
-      "subs       %3, %3, #8                     \n" ARGBTORGB565
-      "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
-      "bgt        1b                             \n"
-      : "+r"(src_y),       // %0
-        "+r"(src_uv),      // %1
-        "+r"(dst_rgb565),  // %2
-        "+r"(width)        // %3
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9", "q10", "q11",
-        "q12", "q13", "q14", "q15");
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n" ARGBTORGB565
+      "vst1.8      {q2}, [%[dst_rgb565]]!        \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_uv] "+r"(src_uv),                             // %[src_uv]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 void YUY2ToARGBRow_NEON(const uint8_t* src_yuy2,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile(YUVTORGB_SETUP
-               "vmov.u8    d23, #255                      \n"
-               "1:                                        \n" READYUY2 YUVTORGB
-               "subs       %2, %2, #8                     \n"
-               "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-               "bgt        1b                             \n"
-               : "+r"(src_yuy2),  // %0
-                 "+r"(dst_argb),  // %1
-                 "+r"(width)      // %2
-               : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-                 [kUVToG] "r"(&yuvconstants->kUVToG),
-                 [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-                 [kYToRgb] "r"(&yuvconstants->kYToRgb)
-               : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9",
-                 "q10", "q11", "q12", "q13", "q14", "q15");
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READYUY2 YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_yuy2] "+r"(src_yuy2),                         // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 void UYVYToARGBRow_NEON(const uint8_t* src_uyvy,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile(YUVTORGB_SETUP
-               "vmov.u8    d23, #255                      \n"
-               "1:                                        \n" READUYVY YUVTORGB
-               "subs       %2, %2, #8                     \n"
-               "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
-               "bgt        1b                             \n"
-               : "+r"(src_uyvy),  // %0
-                 "+r"(dst_argb),  // %1
-                 "+r"(width)      // %2
-               : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-                 [kUVToG] "r"(&yuvconstants->kUVToG),
-                 [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-                 [kYToRgb] "r"(&yuvconstants->kYToRgb)
-               : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q8", "q9",
-                 "q10", "q11", "q12", "q13", "q14", "q15");
+  asm volatile(
+      YUVTORGB_SETUP
+      "vmov.u8     d6, #255                      \n"
+      "1:                                        \n" READUYVY YUVTORGB RGBTORGB8
+      "subs        %[width], %[width], #8        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%[dst_argb]]! \n"
+      "bgt         1b                            \n"
+      : [src_uyvy] "+r"(src_uyvy),                         // %[src_uyvy]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "d6");
 }
 
 // Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
@@ -550,11 +562,11 @@ void SplitUVRow_NEON(const uint8_t* src_uv,
                      int width) {
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop
-      "vst1.8     {q0}, [%1]!                    \n"  // store U
-      "vst1.8     {q1}, [%2]!                    \n"  // store V
-      "bgt        1b                             \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pairs of UV
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%1]!                   \n"  // store U
+      "vst1.8      {q1}, [%2]!                   \n"  // store V
+      "bgt         1b                            \n"
       : "+r"(src_uv),               // %0
         "+r"(dst_u),                // %1
         "+r"(dst_v),                // %2
@@ -564,6 +576,52 @@ void SplitUVRow_NEON(const uint8_t* src_uv,
   );
 }
 
+// Reads 16 byte Y's from tile and writes out 16 Y's.
+// MM21 Y tiles are 16x32 so src_tile_stride = 512 bytes
+// MM21 UV tiles are 8x16 so src_tile_stride = 256 bytes
+// width measured in bytes so 8 UV = 16.
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0], %3                \n"  // load 16 bytes
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "pld         [%0, 1792]                    \n"
+      "vst1.16     {q0}, [%1]!                   \n"  // store 16 bytes
+      "bgt         1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "q0"  // Clobber List
+  );
+}
+
+// Read 16 bytes of UV, detile, and write 8 bytes of U and 8 bytes of V.
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld2.8      {d0, d1}, [%0], %4            \n"
+      "subs        %3, %3, #16                   \n"
+      "pld         [%0, 1792]                    \n"
+      "vst1.8      {d0}, [%1]!                   \n"
+      "vst1.8      {d1}, [%2]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3
+      : "r"(src_tile_stride)        // %4
+      : "cc", "memory", "d0", "d1"  // Clobber List
+  );
+}
+
 // Reads 16 U's and V's and writes out 16 pairs of UV.
 void MergeUVRow_NEON(const uint8_t* src_u,
                      const uint8_t* src_v,
@@ -571,11 +629,11 @@ void MergeUVRow_NEON(const uint8_t* src_u,
                      int width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load U
-      "vld1.8     {q1}, [%1]!                    \n"  // load V
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop
-      "vst2.8     {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load U
+      "vld1.8      {q1}, [%1]!                   \n"  // load V
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vst2.8      {q0, q1}, [%2]!               \n"  // store 16 pairs of UV
+      "bgt         1b                            \n"
       : "+r"(src_u),                // %0
         "+r"(src_v),                // %1
         "+r"(dst_uv),               // %2
@@ -593,13 +651,13 @@ void SplitRGBRow_NEON(const uint8_t* src_rgb,
                       int width) {
   asm volatile(
       "1:                                        \n"
-      "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB
-      "vld3.8     {d1, d3, d5}, [%0]!            \n"  // next 8 RGB
-      "subs       %4, %4, #16                    \n"  // 16 processed per loop
-      "vst1.8     {q0}, [%1]!                    \n"  // store R
-      "vst1.8     {q1}, [%2]!                    \n"  // store G
-      "vst1.8     {q2}, [%3]!                    \n"  // store B
-      "bgt        1b                             \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // next 8 RGB
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%1]!                   \n"  // store R
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%3]!                   \n"  // store B
+      "bgt         1b                            \n"
       : "+r"(src_rgb),                    // %0
         "+r"(dst_r),                      // %1
         "+r"(dst_g),                      // %2
@@ -618,13 +676,13 @@ void MergeRGBRow_NEON(const uint8_t* src_r,
                       int width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load R
-      "vld1.8     {q1}, [%1]!                    \n"  // load G
-      "vld1.8     {q2}, [%2]!                    \n"  // load B
-      "subs       %4, %4, #16                    \n"  // 16 processed per loop
-      "vst3.8     {d0, d2, d4}, [%3]!            \n"  // store 8 RGB
-      "vst3.8     {d1, d3, d5}, [%3]!            \n"  // next 8 RGB
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q2}, [%2]!                   \n"  // load B
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst3.8      {d0, d2, d4}, [%3]!           \n"  // store 8 RGB
+      "vst3.8      {d1, d3, d5}, [%3]!           \n"  // next 8 RGB
+      "bgt         1b                            \n"
       : "+r"(src_r),                      // %0
         "+r"(src_g),                      // %1
         "+r"(src_b),                      // %2
@@ -635,14 +693,341 @@ void MergeRGBRow_NEON(const uint8_t* src_r,
   );
 }
 
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b, dst_a.
+void SplitARGBRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // next 8 ARGB
+      "subs        %5, %5, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%3]!                   \n"  // store B
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%1]!                   \n"  // store R
+      "vst1.8      {q3}, [%4]!                   \n"  // store A
+      "bgt         1b                            \n"
+      : "+r"(src_argb),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(dst_a),                            // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's and B's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q0}, [%2]!                   \n"  // load B
+      "vld1.8      {q3}, [%3]!                   \n"  // load A
+      "subs        %5, %5, #16                   \n"  // 16 processed per loop
+      "vst4.8      {d0, d2, d4, d6}, [%4]!       \n"  // store 8 ARGB
+      "vst4.8      {d1, d3, d5, d7}, [%4]!       \n"  // next 8 ARGB
+      "bgt         1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(src_a),                            // %3
+        "+r"(dst_argb),                         // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b.
+void SplitXRGBRow_NEON(const uint8_t* src_argb,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // next 8 ARGB
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q0}, [%3]!                   \n"  // store B
+      "vst1.8      {q1}, [%2]!                   \n"  // store G
+      "vst1.8      {q2}, [%1]!                   \n"  // store R
+      "bgt         1b                            \n"
+      : "+r"(src_argb),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "vmov.u8     q3, #255                      \n"  // load A(255)
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load R
+      "vld1.8      {q1}, [%1]!                   \n"  // load G
+      "vld1.8      {q0}, [%2]!                   \n"  // load B
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop
+      "vst4.8      {d0, d2, d4, d6}, [%3]!       \n"  // store 8 ARGB
+      "vst4.8      {d1, d3, d5, d7}, [%3]!       \n"  // next 8 ARGB
+      "bgt         1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(dst_argb),                         // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = 10 - depth;
+  asm volatile(
+      "vmov.u32    q14, #1023                    \n"
+      "vdup.32     q15, %5                       \n"
+      "1:                                        \n"
+      "vld1.16     {d4}, [%2]!                   \n"  // B
+      "vld1.16     {d2}, [%1]!                   \n"  // G
+      "vld1.16     {d0}, [%0]!                   \n"  // R
+      "vmovl.u16   q2, d4                        \n"  // B
+      "vmovl.u16   q1, d2                        \n"  // G
+      "vmovl.u16   q0, d0                        \n"  // R
+      "vshl.u32    q2, q2, q15                   \n"  // 000B
+      "vshl.u32    q1, q1, q15                   \n"
+      "vshl.u32    q0, q0, q15                   \n"
+      "vmin.u32    q2, q2, q14                   \n"
+      "vmin.u32    q1, q1, q14                   \n"
+      "vmin.u32    q0, q0, q14                   \n"
+      "vsli.u32    q2, q1, #10                   \n"  // 00GB
+      "vsli.u32    q2, q0, #20                   \n"  // 0RGB
+      "vorr.u32    q2, #0xc0000000               \n"  // ARGB (AR30)
+      "subs        %4, %4, #4                    \n"
+      "vst1.8      {q2}, [%3]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "q0", "q1", "q2", "q14", "q15");
+}
+
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width) {
+  asm volatile(
+      "vmov.u32    q14, #1023                    \n"
+      "1:                                        \n"
+      "vld1.16     {d4}, [%2]!                   \n"  // B
+      "vld1.16     {d2}, [%1]!                   \n"  // G
+      "vld1.16     {d0}, [%0]!                   \n"  // R
+      "vmovl.u16   q2, d4                        \n"  // 000B
+      "vmovl.u16   q1, d2                        \n"  // G
+      "vmovl.u16   q0, d0                        \n"  // R
+      "vmin.u32    q2, q2, q14                   \n"
+      "vmin.u32    q1, q1, q14                   \n"
+      "vmin.u32    q0, q0, q14                   \n"
+      "vsli.u32    q2, q1, #10                   \n"  // 00GB
+      "vsli.u32    q2, q0, #20                   \n"  // 0RGB
+      "vorr.u32    q2, #0xc0000000               \n"  // ARGB (AR30)
+      "subs        %4, %4, #4                    \n"
+      "vst1.8      {q2}, [%3]!                   \n"
+      "bgt         1b                            \n"
+      "3:                                        \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q14");
+}
+
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "vdup.u16    q15, %6                       \n"
+      "vdup.u16    q14, %7                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vld1.16     {q3}, [%3]!                   \n"  // A
+      "vmin.u16    q2, q2, q14                   \n"
+      "vmin.u16    q1, q1, q14                   \n"
+      "vmin.u16    q0, q0, q14                   \n"
+      "vmin.u16    q3, q3, q14                   \n"
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vshl.u16    q3, q3, q15                   \n"
+      "subs        %5, %5, #8                    \n"
+      "vst4.16     {d0, d2, d4, d6}, [%4]!       \n"
+      "vst4.16     {d1, d3, d5, d7}, [%4]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+        "+r"(width)      // %5
+      : "r"(shift),      // %6
+        "r"(mask)        // %7
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "vmov.u8     q3, #0xff                     \n"  // A (0xffff)
+      "vdup.u16    q15, %5                       \n"
+      "vdup.u16    q14, %6                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vmin.u16    q2, q2, q14                   \n"
+      "vmin.u16    q1, q1, q14                   \n"
+      "vmin.u16    q0, q0, q14                   \n"
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "subs        %4, %4, #8                    \n"
+      "vst4.16     {d0, d2, d4, d6}, [%3]!       \n"
+      "vst4.16     {d1, d3, d5, d7}, [%3]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar64),  // %3
+        "+r"(width)      // %4
+      : "r"(shift),      // %5
+        "r"(mask)        // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "vdup.16     q15, %6                       \n"
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vld1.16     {q3}, [%3]!                   \n"  // A
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vshl.u16    q3, q3, q15                   \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d1, q1                        \n"
+      "vqmovn.u16  d2, q2                        \n"
+      "vqmovn.u16  d3, q3                        \n"
+      "subs        %5, %5, #8                    \n"
+      "vst4.8      {d0, d1, d2, d3}, [%4]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      : "r"(shift)       // %6
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q15");
+}
+
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "vdup.16     q15, %5                       \n"
+      "vmov.u8     d6, #0xff                     \n"  // A (0xff)
+      "1:                                        \n"
+      "vld1.16     {q2}, [%0]!                   \n"  // R
+      "vld1.16     {q1}, [%1]!                   \n"  // G
+      "vld1.16     {q0}, [%2]!                   \n"  // B
+      "vshl.u16    q2, q2, q15                   \n"
+      "vshl.u16    q1, q1, q15                   \n"
+      "vshl.u16    q0, q0, q15                   \n"
+      "vqmovn.u16  d5, q2                        \n"
+      "vqmovn.u16  d4, q1                        \n"
+      "vqmovn.u16  d3, q0                        \n"
+      "subs        %4, %4, #8                    \n"
+      "vst4.u8     {d3, d4, d5, d6}, [%3]!       \n"
+      "bgt         1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "q0", "q1", "q2", "d6", "q15");
+}
+
 // Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
 void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
-      "subs       %2, %2, #32                    \n"  // 32 processed per loop
-      "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
-      "bgt        1b                             \n"
+      "vld1.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 32
+      "subs        %2, %2, #32                   \n"  // 32 processed per loop
+      "vst1.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 32
+      "bgt         1b                            \n"
       : "+r"(src),                  // %0
         "+r"(dst),                  // %1
         "+r"(width)                 // %2  // Output registers
@@ -654,11 +1039,11 @@ void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
 // SetRow writes 'width' bytes using an 8 bit value repeated.
 void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
   asm volatile(
-      "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
+      "vdup.8      q0, %2                        \n"  // duplicate 16 bytes
       "1:                                        \n"
-      "subs      %1, %1, #16                     \n"  // 16 bytes per loop
-      "vst1.8    {q0}, [%0]!                     \n"  // store
-      "bgt       1b                              \n"
+      "subs        %1, %1, #16                   \n"  // 16 bytes per loop
+      "vst1.8      {q0}, [%0]!                   \n"  // store
+      "bgt         1b                            \n"
       : "+r"(dst),   // %0
         "+r"(width)  // %1
       : "r"(v8)      // %2
@@ -668,11 +1053,11 @@ void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
 // ARGBSetRow writes 'width' pixels using an 32 bit value repeated.
 void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width) {
   asm volatile(
-      "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+      "vdup.u32    q0, %2                        \n"  // duplicate 4 ints
       "1:                                        \n"
-      "subs      %1, %1, #4                      \n"  // 4 pixels per loop
-      "vst1.8    {q0}, [%0]!                     \n"  // store
-      "bgt       1b                              \n"
+      "subs        %1, %1, #4                    \n"  // 4 pixels per loop
+      "vst1.8      {q0}, [%0]!                   \n"  // store
+      "bgt         1b                            \n"
       : "+r"(dst),   // %0
         "+r"(width)  // %1
       : "r"(v32)     // %2
@@ -682,41 +1067,62 @@ void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width) {
 void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       // Start at end of source row.
-      "mov        r3, #-16                       \n"
-      "add        %0, %0, %2                     \n"
-      "sub        %0, #16                        \n"
+      "add         %0, %0, %2                    \n"
+      "sub         %0, %0, #32                   \n"  // 32 bytes per loop
 
       "1:                                        \n"
-      "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
-      "subs       %2, #16                        \n"  // 16 pixels per loop.
-      "vrev64.8   q0, q0                         \n"
-      "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
-      "vst1.8     {d0}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q1, q2}, [%0], %3            \n"  // src -= 32
+      "subs        %2, #32                       \n"  // 32 pixels per loop.
+      "vrev64.8    q0, q2                        \n"
+      "vrev64.8    q1, q1                        \n"
+      "vswp        d0, d1                        \n"
+      "vswp        d2, d3                        \n"
+      "vst1.8      {q0, q1}, [%1]!               \n"  // dst += 32
+      "bgt         1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
+      : "r"(-32)     // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "mov         r12, #-16                     \n"
+      "add         %0, %0, %2, lsl #1            \n"
+      "sub         %0, #16                       \n"
+
+      "1:                                        \n"
+      "vld2.8      {d0, d1}, [%0], r12           \n"  // src -= 16
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    q0, q0                        \n"
+      "vst2.8      {d0, d1}, [%1]!               \n"  // dst += 16
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_uv),  // %1
+        "+r"(width)    // %2
       :
-      : "cc", "memory", "r3", "q0");
+      : "cc", "memory", "r12", "q0");
 }
 
-void MirrorUVRow_NEON(const uint8_t* src_uv,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width) {
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
   asm volatile(
       // Start at end of source row.
-      "mov        r12, #-16                      \n"
-      "add        %0, %0, %3, lsl #1             \n"
-      "sub        %0, #16                        \n"
+      "mov         r12, #-16                     \n"
+      "add         %0, %0, %3, lsl #1            \n"
+      "sub         %0, #16                       \n"
 
       "1:                                        \n"
-      "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
-      "subs       %3, #8                         \n"  // 8 pixels per loop.
-      "vrev64.8   q0, q0                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
-      "vst1.8     {d1}, [%2]!                    \n"
-      "bgt        1b                             \n"
+      "vld2.8      {d0, d1}, [%0], r12           \n"  // src -= 16
+      "subs        %3, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    q0, q0                        \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // dst += 8
+      "vst1.8      {d1}, [%2]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_uv),  // %0
         "+r"(dst_u),   // %1
         "+r"(dst_v),   // %2
@@ -725,37 +1131,57 @@ void MirrorUVRow_NEON(const uint8_t* src_uv,
       : "cc", "memory", "r12", "q0");
 }
 
-void ARGBMirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   asm volatile(
-      // Start at end of source row.
-      "mov        r3, #-16                       \n"
-      "add        %0, %0, %2, lsl #2             \n"
-      "sub        %0, #16                        \n"
+      "add         %0, %0, %2, lsl #2            \n"
+      "sub         %0, #32                       \n"
 
       "1:                                        \n"
-      "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
-      "subs       %2, #4                         \n"  // 4 pixels per loop.
-      "vrev64.32  q0, q0                         \n"
-      "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
-      "vst1.8     {d0}, [%1]!                    \n"
-      "bgt        1b                             \n"
-      : "+r"(src),   // %0
-        "+r"(dst),   // %1
-        "+r"(width)  // %2
-      :
-      : "cc", "memory", "r3", "q0");
+      "vld4.8      {d0, d1, d2, d3}, [%0], %3    \n"  // src -= 32
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    d0, d0                        \n"
+      "vrev64.8    d1, d1                        \n"
+      "vrev64.8    d2, d2                        \n"
+      "vrev64.8    d3, d3                        \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // dst += 32
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      : "r"(-32)         // %3
+      : "cc", "memory", "d0", "d1", "d2", "d3");
+}
+
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  src_rgb24 += width * 3 - 24;
+  asm volatile(
+      "1:                                        \n"
+      "vld3.8      {d0, d1, d2}, [%0], %3        \n"  // src -= 24
+      "subs        %2, #8                        \n"  // 8 pixels per loop.
+      "vrev64.8    d0, d0                        \n"
+      "vrev64.8    d1, d1                        \n"
+      "vrev64.8    d2, d2                        \n"
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"  // dst += 24
+      "bgt         1b                            \n"
+      : "+r"(src_rgb24),  // %0
+        "+r"(dst_rgb24),  // %1
+        "+r"(width)       // %2
+      : "r"(-24)          // %3
+      : "cc", "memory", "d0", "d1", "d2");
 }
 
 void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
                          uint8_t* dst_argb,
                          int width) {
   asm volatile(
-      "vmov.u8    d4, #255                       \n"  // Alpha
+      "vmov.u8     d4, #255                      \n"  // Alpha
       "1:                                        \n"
-      "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RGB24.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst4.8      {d1, d2, d3, d4}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_rgb24),  // %0
         "+r"(dst_argb),   // %1
         "+r"(width)       // %2
@@ -766,13 +1192,13 @@ void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
 
 void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   asm volatile(
-      "vmov.u8    d4, #255                       \n"  // Alpha
+      "vmov.u8     d4, #255                      \n"  // Alpha
       "1:                                        \n"
-      "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vswp.u8    d1, d3                         \n"  // swap R, B
-      "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst4.8      {d1, d2, d3, d4}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_raw),   // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -781,15 +1207,31 @@ void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   );
 }
 
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "vmov.u8     d0, #255                      \n"  // Alpha
+      "1:                                        \n"
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of RGBA.
+      "bgt         1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_rgba),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
 void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
   asm volatile(
       "1:                                        \n"
-      "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vswp.u8    d1, d3                         \n"  // swap R, B
-      "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of
+      "vld3.8      {d1, d2, d3}, [%0]!           \n"  // load 8 pixels of RAW.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst3.8      {d1, d2, d3}, [%1]!           \n"  // store 8 pixels of
                                                       // RGB24.
-      "bgt        1b                             \n"
+      "bgt         1b                            \n"
       : "+r"(src_raw),    // %0
         "+r"(dst_rgb24),  // %1
         "+r"(width)       // %2
@@ -814,13 +1256,13 @@ void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
-      "vmov.u8    d3, #255                       \n"  // Alpha
+      "vmov.u8     d3, #255                      \n"  // Alpha
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       RGB565TOARGB
-      "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_rgb565),  // %0
         "+r"(dst_argb),    // %1
         "+r"(width)        // %2
@@ -860,13 +1302,13 @@ void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
                             uint8_t* dst_argb,
                             int width) {
   asm volatile(
-      "vmov.u8    d3, #255                       \n"  // Alpha
+      "vmov.u8     d3, #255                      \n"  // Alpha
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGB1555TOARGB
-      "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_argb1555),  // %0
         "+r"(dst_argb),      // %1
         "+r"(width)          // %2
@@ -889,13 +1331,13 @@ void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
                             uint8_t* dst_argb,
                             int width) {
   asm volatile(
-      "vmov.u8    d3, #255                       \n"  // Alpha
+      "vmov.u8     d3, #255                      \n"  // Alpha
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGB4444TOARGB
-      "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_argb4444),  // %0
         "+r"(dst_argb),      // %1
         "+r"(width)          // %2
@@ -909,27 +1351,28 @@ void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of
-                                                      // RGB24.
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of ARGB.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst3.8      {d0, d2, d4}, [%1]!           \n"  // store 16 RGB24 pixels.
+      "vst3.8      {d1, d3, d5}, [%1]!           \n"
+      "bgt         1b                            \n"
       : "+r"(src_argb),   // %0
         "+r"(dst_rgb24),  // %1
         "+r"(width)       // %2
       :
-      : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+      : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 
 void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vswp.u8    d1, d3                         \n"  // swap R, B
-      "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
-      "bgt        1b                             \n"
+      "vld4.8      {d1, d2, d3, d4}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vswp.u8     d1, d3                        \n"  // swap R, B
+      "vst3.8      {d1, d2, d3}, [%1]!           \n"  // store 8 pixels of RAW.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_raw),   // %1
         "+r"(width)      // %2
@@ -941,10 +1384,10 @@ void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
 void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop.
-      "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
-      "bgt        1b                             \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pixels of YUY2.
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 16 pixels of Y.
+      "bgt         1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -956,10 +1399,10 @@ void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
 void UYVYToYRow_NEON(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop.
-      "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
-      "bgt        1b                             \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 16 pixels of UYVY.
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vst1.8      {q1}, [%1]!                   \n"  // store 16 pixels of Y.
+      "bgt         1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -974,11 +1417,11 @@ void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
-      "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
-      "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
-      "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of YUY2.
+      "subs        %3, %3, #16                   \n"  // 16 pixels = 8 UVs.
+      "vst1.8      {d1}, [%1]!                   \n"  // store 8 U.
+      "vst1.8      {d3}, [%2]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -994,11 +1437,11 @@ void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
-      "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
-      "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of UYVY.
+      "subs        %3, %3, #16                   \n"  // 16 pixels = 8 UVs.
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 U.
+      "vst1.8      {d2}, [%2]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1014,16 +1457,16 @@ void YUY2ToUVRow_NEON(const uint8_t* src_yuy2,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // stride + src_yuy2
+      "add         %1, %0, %1                    \n"  // stride + src_yuy2
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
-      "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
-      "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
-      "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
-      "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
-      "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
-      "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of YUY2.
+      "subs        %4, %4, #16                   \n"  // 16 pixels = 8 UVs.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load next row YUY2.
+      "vrhadd.u8   d1, d1, d5                    \n"  // average rows of U
+      "vrhadd.u8   d3, d3, d7                    \n"  // average rows of V
+      "vst1.8      {d1}, [%2]!                   \n"  // store 8 U.
+      "vst1.8      {d3}, [%3]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
       : "+r"(src_yuy2),     // %0
         "+r"(stride_yuy2),  // %1
         "+r"(dst_u),        // %2
@@ -1041,16 +1484,16 @@ void UYVYToUVRow_NEON(const uint8_t* src_uyvy,
                       uint8_t* dst_v,
                       int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // stride + src_uyvy
+      "add         %1, %0, %1                    \n"  // stride + src_uyvy
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
-      "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
-      "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
-      "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
-      "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
-      "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
-      "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 16 pixels of UYVY.
+      "subs        %4, %4, #16                   \n"  // 16 pixels = 8 UVs.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load next row UYVY.
+      "vrhadd.u8   d0, d0, d4                    \n"  // average rows of U
+      "vrhadd.u8   d2, d2, d6                    \n"  // average rows of V
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 U.
+      "vst1.8      {d2}, [%3]!                   \n"  // store 8 V.
+      "bgt         1b                            \n"
       : "+r"(src_uyvy),     // %0
         "+r"(stride_uyvy),  // %1
         "+r"(dst_u),        // %2
@@ -1068,14 +1511,14 @@ void ARGBShuffleRow_NEON(const uint8_t* src_argb,
                          const uint8_t* shuffler,
                          int width) {
   asm volatile(
-      "vld1.8     {q2}, [%3]                     \n"  // shuffler
+      "vld1.8      {q2}, [%3]                    \n"  // shuffler
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
-      "subs       %2, %2, #4                     \n"  // 4 processed per loop
-      "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
-      "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
-      "vst1.8     {q1}, [%1]!                    \n"  // store 4.
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 4 pixels.
+      "subs        %2, %2, #4                    \n"  // 4 processed per loop
+      "vtbl.8      d2, {d0, d1}, d4              \n"  // look up 2 first pixels
+      "vtbl.8      d3, {d0, d1}, d5              \n"  // look up 2 next pixels
+      "vst1.8      {q1}, [%1]!                   \n"  // store 4.
+      "bgt         1b                            \n"
       : "+r"(src_argb),                   // %0
         "+r"(dst_argb),                   // %1
         "+r"(width)                       // %2
@@ -1091,12 +1534,12 @@ void I422ToYUY2Row_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
-      "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
-      "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
-      "subs       %4, %4, #16                    \n"  // 16 pixels
-      "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
-      "bgt        1b                             \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 16 Ys
+      "vld1.8      {d1}, [%1]!                   \n"  // load 8 Us
+      "vld1.8      {d3}, [%2]!                   \n"  // load 8 Vs
+      "subs        %4, %4, #16                   \n"  // 16 pixels
+      "vst4.8      {d0, d1, d2, d3}, [%3]!       \n"  // Store 8 YUY2/16 pixels.
+      "bgt         1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -1113,12 +1556,12 @@ void I422ToUYVYRow_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
-      "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
-      "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
-      "subs       %4, %4, #16                    \n"  // 16 pixels
-      "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
-      "bgt        1b                             \n"
+      "vld2.8      {d1, d3}, [%0]!               \n"  // load 16 Ys
+      "vld1.8      {d0}, [%1]!                   \n"  // load 8 Us
+      "vld1.8      {d2}, [%2]!                   \n"  // load 8 Vs
+      "subs        %4, %4, #16                   \n"  // 16 pixels
+      "vst4.8      {d0, d1, d2, d3}, [%3]!       \n"  // Store 8 UYVY/16 pixels.
+      "bgt         1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -1133,16 +1576,16 @@ void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
                           int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGBTORGB565
-      "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
-      "bgt        1b                             \n"
+      "vst1.8      {q2}, [%1]!                   \n"  // store 8 pixels RGB565.
+      "bgt         1b                            \n"
       : "+r"(src_argb),    // %0
         "+r"(dst_rgb565),  // %1
         "+r"(width)        // %2
       :
-      : "cc", "memory", "q0", "q8", "q9", "q10", "q11");
+      : "cc", "memory", "q0", "q1", "q2", "d6");
 }
 
 void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
@@ -1150,21 +1593,21 @@ void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
                                 const uint32_t dither4,
                                 int width) {
   asm volatile(
-      "vdup.32    d2, %2                         \n"  // dither4
+      "vdup.32     d7, %2                        \n"  // dither4
       "1:                                        \n"
-      "vld4.8     {d20, d21, d22, d23}, [%1]!    \n"  // load 8 pixels of ARGB.
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vqadd.u8   d20, d20, d2                   \n"
-      "vqadd.u8   d21, d21, d2                   \n"
-      "vqadd.u8   d22, d22, d2                   \n"  // add for dither
+      "vld4.8      {d0, d2, d4, d6}, [%1]!       \n"  // load 8 pixels of ARGB.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d0, d0, d7                    \n"
+      "vqadd.u8    d2, d2, d7                    \n"
+      "vqadd.u8    d4, d4, d7                    \n"  // add for dither
       ARGBTORGB565
-      "vst1.8     {q0}, [%0]!                    \n"  // store 8 RGB565.
-      "bgt        1b                             \n"
+      "vst1.8      {q2}, [%0]!                   \n"  // store 8 RGB565.
+      "bgt         1b                            \n"
       : "+r"(dst_rgb)   // %0
       : "r"(src_argb),  // %1
         "r"(dither4),   // %2
         "r"(width)      // %3
-      : "cc", "memory", "q0", "q1", "q8", "q9", "q10", "q11");
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
@@ -1172,58 +1615,35 @@ void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
                             int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGBTOARGB1555
-      "vst1.8     {q0}, [%1]!                    \n"  // store 8 ARGB1555.
-      "bgt        1b                             \n"
+      "vst1.8      {q3}, [%1]!                   \n"  // store 8 ARGB1555.
+      "bgt         1b                            \n"
       : "+r"(src_argb),      // %0
         "+r"(dst_argb1555),  // %1
         "+r"(width)          // %2
       :
-      : "cc", "memory", "q0", "q8", "q9", "q10", "q11");
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
                             uint8_t* dst_argb4444,
                             int width) {
   asm volatile(
-      "vmov.u8    d4, #0x0f                      \n"  // bits to clear with
+      "vmov.u8     d7, #0x0f                     \n"  // bits to clear with
                                                       // vbic.
       "1:                                        \n"
-      "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGBTOARGB4444
-      "vst1.8     {q0}, [%1]!                    \n"  // store 8 ARGB4444.
-      "bgt        1b                             \n"
+      "vst1.8      {q0}, [%1]!                   \n"  // store 8 ARGB4444.
+      "bgt         1b                            \n"
       : "+r"(src_argb),      // %0
         "+r"(dst_argb4444),  // %1
         "+r"(width)          // %2
       :
-      : "cc", "memory", "q0", "q8", "q9", "q10", "q11");
-}
-
-void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
-  asm volatile(
-      "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-      "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-      "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-      "vmov.u8    d27, #16                       \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d27                        \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_argb),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
-      : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
 void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
@@ -1231,11 +1651,11 @@ void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
                               int width) {
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels
-      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop
-      "vst1.8     {q3}, [%1]!                    \n"  // store 16 A's.
-      "bgt       1b                              \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q3}, [%1]!                   \n"  // store 16 A's.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_a),     // %1
         "+r"(width)      // %2
@@ -1244,59 +1664,36 @@ void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
   );
 }
 
-void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
-  asm volatile(
-      "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-      "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-      "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_argb),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
-      : "cc", "memory", "q0", "q1", "q2", "q12", "q13");
-}
-
 // 8x1 pixels.
 void ARGBToUV444Row_NEON(const uint8_t* src_argb,
                          uint8_t* dst_u,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "vmov.u8    d24, #112                      \n"  // UB / VR 0.875
+      "vmov.u8     d24, #112                     \n"  // UB / VR 0.875
                                                       // coefficient
-      "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
-      "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
-      "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
-      "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
-      "vmov.u16   q15, #0x8080                   \n"  // 128.5
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlsl.u8   q2, d1, d25                    \n"  // G
-      "vmlsl.u8   q2, d2, d26                    \n"  // R
-      "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
-
-      "vmull.u8   q3, d2, d24                    \n"  // R
-      "vmlsl.u8   q3, d1, d28                    \n"  // G
-      "vmlsl.u8   q3, d0, d27                    \n"  // B
-      "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
-
-      "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
-      "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
-
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
-      "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
-      "bgt        1b                             \n"
+      "vmov.u8     d25, #74                      \n"  // UG -0.5781 coefficient
+      "vmov.u8     d26, #38                      \n"  // UR -0.2969 coefficient
+      "vmov.u8     d27, #18                      \n"  // VB -0.1406 coefficient
+      "vmov.u8     d28, #94                      \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlsl.u8    q2, d1, d25                   \n"  // G
+      "vmlsl.u8    q2, d2, d26                   \n"  // R
+
+      "vmull.u8    q3, d2, d24                   \n"  // R
+      "vmlsl.u8    q3, d1, d28                   \n"  // G
+      "vmlsl.u8    q3, d0, d27                   \n"  // B
+
+      "vaddhn.u16  d0, q2, q15                   \n"  // +128 -> unsigned
+      "vaddhn.u16  d1, q3, q15                   \n"  // +128 -> unsigned
+
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%2]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1312,13 +1709,11 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
   "vmul.s16   q8, " #QB ", q10               \n" /* B                    */ \
   "vmls.s16   q8, " #QG ", q11               \n" /* G                    */ \
   "vmls.s16   q8, " #QR ", q12               \n" /* R                    */ \
-  "vadd.u16   q8, q8, q15                    \n" /* +128 -> unsigned     */ \
   "vmul.s16   q9, " #QR ", q10               \n" /* R                    */ \
   "vmls.s16   q9, " #QG ", q14               \n" /* G                    */ \
   "vmls.s16   q9, " #QB ", q13               \n" /* B                    */ \
-  "vadd.u16   q9, q9, q15                    \n" /* +128 -> unsigned     */ \
-  "vqshrn.u16  d0, q8, #8                    \n" /* 16 bit to 8 bit U    */ \
-  "vqshrn.u16  d1, q9, #8                    \n" /* 16 bit to 8 bit V    */
+  "vaddhn.u16 d0, q8, q15                    \n" /* +128 -> unsigned     */ \
+  "vaddhn.u16 d1, q9, q15                    \n" /* +128 -> unsigned     */
 // clang-format on
 
 // TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
@@ -1328,34 +1723,34 @@ void ARGBToUVRow_NEON(const uint8_t* src_argb,
                       uint8_t* dst_v,
                       int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ARGB pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ARGB pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_stride_argb),  // %1
     "+r"(dst_u),     // %2
@@ -1374,34 +1769,34 @@ void ARGBToUVJRow_NEON(const uint8_t* src_argb,
                        uint8_t* dst_v,
                        int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_argb
-    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
-    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
-    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
-    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
-    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ARGB pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ARGB pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_stride_argb),  // %1
     "+r"(dst_u),     // %2
@@ -1413,40 +1808,132 @@ void ARGBToUVJRow_NEON(const uint8_t* src_argb,
   );
 }
 
+// TODO(fbarchard): Subsample match C code.
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  asm volatile (
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgb24
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB24 pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RGB24 pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RGB24 pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RGB24 pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  asm volatile (
+      "add         %1, %0, %1                    \n"  // src_stride + src_raw
+      "vmov.s16    q10, #127 / 2                 \n"  // UB / VR 0.500 coefficient
+      "vmov.s16    q11, #84 / 2                  \n"  // UG -0.33126 coefficient
+      "vmov.s16    q12, #43 / 2                  \n"  // UR -0.16874 coefficient
+      "vmov.s16    q13, #20 / 2                  \n"  // VB -0.08131 coefficient
+      "vmov.s16    q14, #107 / 2                 \n"  // VG -0.41869 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RAW pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RAW pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RAW pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RAW pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+    RGBTOUV(q2, q1, q0)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
 void BGRAToUVRow_NEON(const uint8_t* src_bgra,
                       int src_stride_bgra,
                       uint8_t* dst_u,
                       uint8_t* dst_v,
                       int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
-    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
-    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
-    "vrshr.u16  q2, q2, #1                     \n"
-    "vrshr.u16  q3, q3, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_bgra
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 BGRA pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 BGRA pixels.
+      "vpaddl.u8   q3, q3                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more BGRA pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 BGRA pixels.
+      "vpadal.u8   q3, q7                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q1, q1, #1                    \n"  // 2x average
+      "vrshr.u16   q2, q2, #1                    \n"
+      "vrshr.u16   q3, q3, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q3, q2, q1)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_bgra),  // %0
     "+r"(src_stride_bgra),  // %1
     "+r"(dst_u),     // %2
@@ -1464,34 +1951,34 @@ void ABGRToUVRow_NEON(const uint8_t* src_abgr,
                       uint8_t* dst_v,
                       int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_abgr
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ABGR pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ABGR pixels.
+      "vpaddl.u8   q2, q2                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q0, q0                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more ABGR pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 ABGR pixels.
+      "vpadal.u8   q2, q6                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q2, q1, q0)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_abgr),  // %0
     "+r"(src_stride_abgr),  // %1
     "+r"(dst_u),     // %2
@@ -1509,34 +1996,34 @@ void RGBAToUVRow_NEON(const uint8_t* src_rgba,
                       uint8_t* dst_v,
                       int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
-    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
-    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
-    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
-    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
-    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgba
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 RGBA pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 RGBA pixels.
+      "vpaddl.u8   q0, q1                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q2                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q3                        \n"  // R 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more RGBA pixels.
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 RGBA pixels.
+      "vpadal.u8   q0, q5                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q6                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q7                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_rgba),  // %0
     "+r"(src_stride_rgba),  // %1
     "+r"(dst_u),     // %2
@@ -1554,34 +2041,34 @@ void RGB24ToUVRow_NEON(const uint8_t* src_rgb24,
                        uint8_t* dst_v,
                        int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
-    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
-    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_rgb24
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RGB24 pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RGB24 pixels.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RGB24 pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RGB24 pixels.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q6                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q0, q1, q2)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_rgb24),  // %0
     "+r"(src_stride_rgb24),  // %1
     "+r"(dst_u),     // %2
@@ -1599,34 +2086,34 @@ void RAWToUVRow_NEON(const uint8_t* src_raw,
                      uint8_t* dst_v,
                      int width) {
   asm volatile (
-    "add        %1, %0, %1                     \n"  // src_stride + src_raw
-    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
-    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-    "vmov.u16   q15, #0x8080                   \n"  // 128.5
-    "1:                                        \n"
-    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
-    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
-    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
-    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
-    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
-    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
-    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
-    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
-    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
-
-    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
-    "vrshr.u16  q1, q1, #1                     \n"
-    "vrshr.u16  q2, q2, #1                     \n"
-
-    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+      "add         %1, %0, %1                    \n"  // src_stride + src_raw
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875 coefficient
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
+      "1:                                        \n"
+      "vld3.8      {d0, d2, d4}, [%0]!           \n"  // load 8 RAW pixels.
+      "vld3.8      {d1, d3, d5}, [%0]!           \n"  // load next 8 RAW pixels.
+      "vpaddl.u8   q2, q2                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q0, q0                        \n"  // R 16 bytes -> 8 shorts.
+      "vld3.8      {d8, d10, d12}, [%1]!         \n"  // load 8 more RAW pixels.
+      "vld3.8      {d9, d11, d13}, [%1]!         \n"  // load last 8 RAW pixels.
+      "vpadal.u8   q2, q6                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // R 16 bytes -> 8 shorts.
+
+      "vrshr.u16   q0, q0, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q1, #1                    \n"
+      "vrshr.u16   q2, q2, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
     RGBTOUV(q2, q1, q0)
-    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-    "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
   : "+r"(src_raw),  // %0
     "+r"(src_stride_raw),  // %1
     "+r"(dst_u),     // %2
@@ -1645,55 +2132,55 @@ void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
                         uint8_t* dst_v,
                         int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // src_stride + src_argb
-      "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
                                                       // coefficient
-      "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-      "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-      "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-      "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-      "vmov.u16   q15, #0x8080                   \n"  // 128.5
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
       RGB565TOARGB
-      "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 RGB565 pixels.
       RGB565TOARGB
-      "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
 
-      "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 RGB565 pixels.
       RGB565TOARGB
-      "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 RGB565 pixels.
       RGB565TOARGB
-      "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-      "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-      "vrshr.u16  q5, q5, #1                     \n"
-      "vrshr.u16  q6, q6, #1                     \n"
-
-      "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-      "vmul.s16   q8, q4, q10                    \n"  // B
-      "vmls.s16   q8, q5, q11                    \n"  // G
-      "vmls.s16   q8, q6, q12                    \n"  // R
-      "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-      "vmul.s16   q9, q6, q10                    \n"  // R
-      "vmls.s16   q9, q5, q14                    \n"  // G
-      "vmls.s16   q9, q4, q13                    \n"  // B
-      "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q4, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q5, q5, #1                    \n"
+      "vrshr.u16   q6, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      "vmul.s16    q8, q4, q10                   \n"  // B
+      "vmls.s16    q8, q5, q11                   \n"  // G
+      "vmls.s16    q8, q6, q12                   \n"  // R
+      "vadd.u16    q8, q8, q15                   \n"  // +128 -> unsigned
+      "vmul.s16    q9, q6, q10                   \n"  // R
+      "vmls.s16    q9, q5, q14                   \n"  // G
+      "vmls.s16    q9, q4, q13                   \n"  // B
+      "vadd.u16    q9, q9, q15                   \n"  // +128 -> unsigned
       "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
       "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-      "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-      "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-      "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
       : "+r"(src_rgb565),         // %0
         "+r"(src_stride_rgb565),  // %1
         "+r"(dst_u),              // %2
@@ -1711,55 +2198,55 @@ void ARGB1555ToUVRow_NEON(const uint8_t* src_argb1555,
                           uint8_t* dst_v,
                           int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // src_stride + src_argb
-      "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
                                                       // coefficient
-      "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-      "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-      "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-      "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-      "vmov.u16   q15, #0x8080                   \n"  // 128.5
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
       RGB555TOARGB
-      "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 ARGB1555 pixels.
       RGB555TOARGB
-      "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
 
-      "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 ARGB1555 pixels.
       RGB555TOARGB
-      "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 ARGB1555 pixels.
       RGB555TOARGB
-      "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-      "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-      "vrshr.u16  q5, q5, #1                     \n"
-      "vrshr.u16  q6, q6, #1                     \n"
-
-      "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-      "vmul.s16   q8, q4, q10                    \n"  // B
-      "vmls.s16   q8, q5, q11                    \n"  // G
-      "vmls.s16   q8, q6, q12                    \n"  // R
-      "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-      "vmul.s16   q9, q6, q10                    \n"  // R
-      "vmls.s16   q9, q5, q14                    \n"  // G
-      "vmls.s16   q9, q4, q13                    \n"  // B
-      "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q4, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q5, q5, #1                    \n"
+      "vrshr.u16   q6, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      "vmul.s16    q8, q4, q10                   \n"  // B
+      "vmls.s16    q8, q5, q11                   \n"  // G
+      "vmls.s16    q8, q6, q12                   \n"  // R
+      "vadd.u16    q8, q8, q15                   \n"  // +128 -> unsigned
+      "vmul.s16    q9, q6, q10                   \n"  // R
+      "vmls.s16    q9, q5, q14                   \n"  // G
+      "vmls.s16    q9, q4, q13                   \n"  // B
+      "vadd.u16    q9, q9, q15                   \n"  // +128 -> unsigned
       "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
       "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-      "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-      "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-      "bgt        1b                             \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
       : "+r"(src_argb1555),         // %0
         "+r"(src_stride_argb1555),  // %1
         "+r"(dst_u),                // %2
@@ -1777,55 +2264,46 @@ void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
                           uint8_t* dst_v,
                           int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // src_stride + src_argb
-      "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875
+      "add         %1, %0, %1                    \n"  // src_stride + src_argb
+      "vmov.s16    q10, #112 / 2                 \n"  // UB / VR 0.875
                                                       // coefficient
-      "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
-      "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
-      "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
-      "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
-      "vmov.u16   q15, #0x8080                   \n"  // 128.5
+      "vmov.s16    q11, #74 / 2                  \n"  // UG -0.5781 coefficient
+      "vmov.s16    q12, #38 / 2                  \n"  // UR -0.2969 coefficient
+      "vmov.s16    q13, #18 / 2                  \n"  // VB -0.1406 coefficient
+      "vmov.s16    q14, #94 / 2                  \n"  // VG -0.7344 coefficient
+      "vmov.u16    q15, #0x8080                  \n"  // 128.5
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+      "vpaddl.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%0]!                   \n"  // next 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+      "vpaddl.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpaddl.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpaddl.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
 
-      "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+      "vld1.8      {q0}, [%1]!                   \n"  // load 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
-      "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+      "vpadal.u8   d8, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d10, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d12, d2                       \n"  // R 8 bytes -> 4 shorts.
+      "vld1.8      {q0}, [%1]!                   \n"  // next 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
-      "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
-      "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
-
-      "vrshr.u16  q4, q4, #1                     \n"  // 2x average
-      "vrshr.u16  q5, q5, #1                     \n"
-      "vrshr.u16  q6, q6, #1                     \n"
-
-      "subs       %4, %4, #16                    \n"  // 16 processed per loop.
-      "vmul.s16   q8, q4, q10                    \n"  // B
-      "vmls.s16   q8, q5, q11                    \n"  // G
-      "vmls.s16   q8, q6, q12                    \n"  // R
-      "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
-      "vmul.s16   q9, q6, q10                    \n"  // R
-      "vmls.s16   q9, q5, q14                    \n"  // G
-      "vmls.s16   q9, q4, q13                    \n"  // B
-      "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
-      "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
-      "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
-      "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
-      "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
-      "bgt        1b                             \n"
+      "vpadal.u8   d9, d0                        \n"  // B 8 bytes -> 4 shorts.
+      "vpadal.u8   d11, d1                       \n"  // G 8 bytes -> 4 shorts.
+      "vpadal.u8   d13, d2                       \n"  // R 8 bytes -> 4 shorts.
+
+      "vrshr.u16   q0, q4, #1                    \n"  // 2x average
+      "vrshr.u16   q1, q5, #1                    \n"
+      "vrshr.u16   q2, q6, #1                    \n"
+
+      "subs        %4, %4, #16                   \n"  // 16 processed per loop.
+      RGBTOUV(q0, q1, q2)
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 pixels U.
+      "vst1.8      {d1}, [%3]!                   \n"  // store 8 pixels V.
+      "bgt         1b                            \n"
       : "+r"(src_argb4444),         // %0
         "+r"(src_stride_argb4444),  // %1
         "+r"(dst_u),                // %2
@@ -1838,21 +2316,21 @@ void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
 
 void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
   asm volatile(
-      "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-      "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-      "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-      "vmov.u8    d27, #16                       \n"  // Add 16 constant
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 RGB565 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       RGB565TOARGB
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d27                        \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
       : "+r"(src_rgb565),  // %0
         "+r"(dst_y),       // %1
         "+r"(width)        // %2
@@ -1864,21 +2342,21 @@ void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
                          uint8_t* dst_y,
                          int width) {
   asm volatile(
-      "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-      "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-      "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-      "vmov.u8    d27, #16                       \n"  // Add 16 constant
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB1555 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGB1555TOARGB
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d27                        \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
       : "+r"(src_argb1555),  // %0
         "+r"(dst_y),         // %1
         "+r"(width)          // %2
@@ -1890,21 +2368,21 @@ void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
                          uint8_t* dst_y,
                          int width) {
   asm volatile(
-      "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-      "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-      "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-      "vmov.u8    d27, #16                       \n"  // Add 16 constant
+      "vmov.u8     d24, #25                      \n"  // B * 0.1016 coefficient
+      "vmov.u8     d25, #129                     \n"  // G * 0.5078 coefficient
+      "vmov.u8     d26, #66                      \n"  // R * 0.2578 coefficient
+      "vmov.u8     d27, #16                      \n"  // Add 16 constant
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+      "vld1.8      {q0}, [%0]!                   \n"  // load 8 ARGB4444 pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
       ARGB4444TOARGB
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d27                        \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit Y
+      "vqadd.u8    d0, d27                       \n"
+      "vst1.8      {d0}, [%1]!                   \n"  // store 8 pixels Y.
+      "bgt         1b                            \n"
       : "+r"(src_argb4444),  // %0
         "+r"(dst_y),         // %1
         "+r"(width)          // %2
@@ -1912,119 +2390,276 @@ void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
       : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13");
 }
 
-void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
   asm volatile(
-      "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-      "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-      "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-      "vmov.u8    d7, #16                        \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q8, d1, d4                     \n"  // R
-      "vmlal.u8   q8, d2, d5                     \n"  // G
-      "vmlal.u8   q8, d3, d6                     \n"  // B
-      "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d7                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_bgra),  // %0
-        "+r"(dst_y),     // %1
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q2}, [%0]!                   \n"
+      "vmov.u8     q1, q0                        \n"
+      "vmov.u8     q3, q2                        \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst2.8      {q0, q1}, [%1]!               \n"  // store 4 pixels
+      "vst2.8      {q2, q3}, [%1]!               \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_argb),  // %0
+        "+r"(dst_ar64),  // %1
         "+r"(width)      // %2
       :
-      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+      : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
-void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
   asm volatile(
-      "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-      "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-      "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-      "vmov.u8    d7, #16                        \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q8, d0, d4                     \n"  // R
-      "vmlal.u8   q8, d1, d5                     \n"  // G
-      "vmlal.u8   q8, d2, d6                     \n"  // B
-      "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d7                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_abgr),  // %0
-        "+r"(dst_y),     // %1
+      "vld1.8      {q4}, [%3]                    \n"  // shuffler
+
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"
+      "vld1.8      {q2}, [%0]!                   \n"
+      "vtbl.8      d2, {d0, d1}, d8              \n"
+      "vtbl.8      d3, {d0, d1}, d9              \n"
+      "vtbl.8      d6, {d4, d5}, d8              \n"
+      "vtbl.8      d7, {d4, d5}, d9              \n"
+      "vmov.u8     q0, q1                        \n"
+      "vmov.u8     q2, q3                        \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst2.8      {q0, q1}, [%1]!               \n"  // store 4 pixels
+      "vst2.8      {q2, q3}, [%1]!               \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_ab64),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleARGBToABGR)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vld1.16     {q2}, [%0]!                   \n"
+      "vld1.16     {q3}, [%0]!                   \n"
+      "vshrn.u16   d0, q0, #8                    \n"
+      "vshrn.u16   d1, q1, #8                    \n"
+      "vshrn.u16   d4, q2, #8                    \n"
+      "vshrn.u16   d5, q3, #8                    \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 4 pixels
+      "vst1.8      {q2}, [%1]!                   \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ar64),  // %0
+        "+r"(dst_argb),  // %1
         "+r"(width)      // %2
       :
-      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+static const uvec8 kShuffleAB64ToARGB = {5, 3, 1, 7, 13, 11, 9, 15};
+
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "vld1.8      {d8}, [%3]                    \n"  // shuffler
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vld1.16     {q2}, [%0]!                   \n"
+      "vld1.16     {q3}, [%0]!                   \n"
+      "vtbl.8      d0, {d0, d1}, d8              \n"
+      "vtbl.8      d1, {d2, d3}, d8              \n"
+      "vtbl.8      d4, {d4, d5}, d8              \n"
+      "vtbl.8      d5, {d6, d7}, d8              \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vst1.8      {q0}, [%1]!                   \n"  // store 4 pixels
+      "vst1.8      {q2}, [%1]!                   \n"  // store 4 pixels
+      "bgt         1b                            \n"
+      : "+r"(src_ab64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAB64ToARGB)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of ARGB
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d0, d20                   \n"  // B
+      "vmull.u8    q9, d1, d20                   \n"
+      "vmlal.u8    q8, d2, d21                   \n"  // G
+      "vmlal.u8    q9, d3, d21                   \n"
+      "vmlal.u8    q8, d4, d22                   \n"  // R
+      "vmlal.u8    q9, d5, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
+}
+
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
+  asm volatile(
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 16 pixels of RGBA
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d2, d20                   \n"  // B
+      "vmull.u8    q9, d3, d20                   \n"
+      "vmlal.u8    q8, d4, d21                   \n"  // G
+      "vmlal.u8    q9, d5, d21                   \n"
+      "vmlal.u8    q8, d6, d22                   \n"  // R
+      "vmlal.u8    q9, d7, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_rgba),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
 }
 
 void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+                          uint8_t* dst_y,
+                          int width,
+                          const struct RgbConstants* rgbconstants) {
   asm volatile(
-      "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-      "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-      "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-      "vmov.u8    d7, #16                        \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q8, d1, d4                     \n"  // B
-      "vmlal.u8   q8, d2, d5                     \n"  // G
-      "vmlal.u8   q8, d3, d6                     \n"  // R
-      "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d7                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_rgba),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
-      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+      "vld1.8      {d0}, [%3]                    \n"  // load rgbconstants
+      "vdup.u8     d20, d0[0]                    \n"
+      "vdup.u8     d21, d0[1]                    \n"
+      "vdup.u8     d22, d0[2]                    \n"
+      "vdup.u16    q12, d0[2]                    \n"
+      "1:                                        \n"
+      "vld3.8      {d2, d4, d6}, [%0]!           \n"  // load 16 pixels of
+                                                      // RGB24.
+      "vld3.8      {d3, d5, d7}, [%0]!           \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop.
+      "vmull.u8    q8, d2, d20                   \n"  // B
+      "vmull.u8    q9, d3, d20                   \n"
+      "vmlal.u8    q8, d4, d21                   \n"  // G
+      "vmlal.u8    q9, d5, d21                   \n"
+      "vmlal.u8    q8, d6, d22                   \n"  // R
+      "vmlal.u8    q9, d7, d22                   \n"
+      "vaddhn.u16  d0, q8, q12                   \n"  // 16 bit to 8 bit Y
+      "vaddhn.u16  d1, q9, q12                   \n"
+      "vst1.8      {d0, d1}, [%1]!               \n"  // store 16 pixels Y.
+      "bgt         1b                            \n"
+      : "+r"(src_rgb),     // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "d20", "d21", "d22",
+        "q12");
+}
+
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
 }
 
 void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
-  asm volatile(
-      "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
-      "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-      "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
-      "vmov.u8    d7, #16                        \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q8, d0, d4                     \n"  // B
-      "vmlal.u8   q8, d1, d5                     \n"  // G
-      "vmlal.u8   q8, d2, d6                     \n"  // R
-      "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d7                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_rgb24),  // %0
-        "+r"(dst_y),      // %1
-        "+r"(width)       // %2
-      :
-      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+  RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
 }
 
 void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
-  asm volatile(
-      "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
-      "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
-      "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
-      "vmov.u8    d7, #16                        \n"  // Add 16 constant
-      "1:                                        \n"
-      "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q8, d0, d4                     \n"  // B
-      "vmlal.u8   q8, d1, d5                     \n"  // G
-      "vmlal.u8   q8, d2, d6                     \n"  // R
-      "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
-      "vqadd.u8   d0, d7                         \n"
-      "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
-      "bgt        1b                             \n"
-      : "+r"(src_raw),  // %0
-        "+r"(dst_y),    // %1
-        "+r"(width)     // %2
-      :
-      : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8");
+  RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
 }
 
 // Bilinear filter 16x2 -> 16x1
@@ -2035,46 +2670,46 @@ void InterpolateRow_NEON(uint8_t* dst_ptr,
                          int source_y_fraction) {
   int y1_fraction = source_y_fraction;
   asm volatile(
-      "cmp        %4, #0                         \n"
-      "beq        100f                           \n"
-      "add        %2, %1                         \n"
-      "cmp        %4, #128                       \n"
-      "beq        50f                            \n"
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "add         %2, %1                        \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
 
-      "vdup.8     d5, %4                         \n"
-      "rsb        %4, #256                       \n"
-      "vdup.8     d4, %4                         \n"
+      "vdup.8      d5, %4                        \n"
+      "rsb         %4, #256                      \n"
+      "vdup.8      d4, %4                        \n"
       // General purpose row blend.
       "1:                                        \n"
-      "vld1.8     {q0}, [%1]!                    \n"
-      "vld1.8     {q1}, [%2]!                    \n"
-      "subs       %3, %3, #16                    \n"
-      "vmull.u8   q13, d0, d4                    \n"
-      "vmull.u8   q14, d1, d4                    \n"
-      "vmlal.u8   q13, d2, d5                    \n"
-      "vmlal.u8   q14, d3, d5                    \n"
-      "vrshrn.u16 d0, q13, #8                    \n"
-      "vrshrn.u16 d1, q14, #8                    \n"
-      "vst1.8     {q0}, [%0]!                    \n"
-      "bgt        1b                             \n"
-      "b          99f                            \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vmull.u8    q13, d0, d4                   \n"
+      "vmull.u8    q14, d1, d4                   \n"
+      "vmlal.u8    q13, d2, d5                   \n"
+      "vmlal.u8    q14, d3, d5                   \n"
+      "vrshrn.u16  d0, q13, #8                   \n"
+      "vrshrn.u16  d1, q14, #8                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
 
       // Blend 50 / 50.
       "50:                                       \n"
-      "vld1.8     {q0}, [%1]!                    \n"
-      "vld1.8     {q1}, [%2]!                    \n"
-      "subs       %3, %3, #16                    \n"
-      "vrhadd.u8  q0, q1                         \n"
-      "vst1.8     {q0}, [%0]!                    \n"
-      "bgt        50b                            \n"
-      "b          99f                            \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       "100:                                      \n"
-      "vld1.8     {q0}, [%1]!                    \n"
-      "subs       %3, %3, #16                    \n"
-      "vst1.8     {q0}, [%0]!                    \n"
-      "bgt        100b                           \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
 
       "99:                                       \n"
       : "+r"(dst_ptr),     // %0
@@ -2086,59 +2721,119 @@ void InterpolateRow_NEON(uint8_t* dst_ptr,
       : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14");
 }
 
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int dst_width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+
+  asm volatile(
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
+
+      "vdup.16     d17, %4                       \n"
+      "vdup.16     d16, %5                       \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "vld1.16     {q1}, [%2]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vmull.u16   q2, d0, d16                   \n"
+      "vmull.u16   q3, d1, d16                   \n"
+      "vmlal.u16   q2, d2, d17                   \n"
+      "vmlal.u16   q3, d3, d17                   \n"
+      "vrshrn.u32  d0, q2, #8                    \n"
+      "vrshrn.u32  d1, q3, #8                    \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "vld1.16     {q1}, [%2]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vrhadd.u16  q0, q1                        \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "vld1.16     {q0}, [%1]!                   \n"
+      "subs        %3, %3, #8                    \n"
+      "vst1.16     {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction)   // %5
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q8");
+}
+
 // dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
-void ARGBBlendRow_NEON(const uint8_t* src_argb0,
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width) {
   asm volatile(
-      "subs       %3, #8                         \n"
-      "blt        89f                            \n"
+      "subs        %3, #8                        \n"
+      "blt         89f                           \n"
       // Blend 8 pixels.
       "8:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
-      "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q10, d4, d3                    \n"  // db * a
-      "vmull.u8   q11, d5, d3                    \n"  // dg * a
-      "vmull.u8   q12, d6, d3                    \n"  // dr * a
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 pixels of ARGB0.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 pixels of ARGB1.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q10, d4, d3                   \n"  // db * a
+      "vmull.u8    q11, d5, d3                   \n"  // dg * a
+      "vmull.u8    q12, d6, d3                   \n"  // dr * a
       "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
       "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
       "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
-      "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
-      "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
-      "vqadd.u8   q0, q0, q2                     \n"  // + sbg
-      "vqadd.u8   d2, d2, d6                     \n"  // + sr
-      "vmov.u8    d3, #255                       \n"  // a = 255
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
-      "bge        8b                             \n"
+      "vqsub.u8    q2, q2, q10                   \n"  // dbg - dbg * a / 256
+      "vqsub.u8    d6, d6, d22                   \n"  // dr - dr * a / 256
+      "vqadd.u8    q0, q0, q2                    \n"  // + sbg
+      "vqadd.u8    d2, d2, d6                    \n"  // + sr
+      "vmov.u8     d3, #255                      \n"  // a = 255
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 pixels of ARGB.
+      "bge         8b                            \n"
 
       "89:                                       \n"
-      "adds       %3, #8-1                       \n"
-      "blt        99f                            \n"
+      "adds        %3, #8-1                      \n"
+      "blt         99f                           \n"
 
       // Blend 1 pixels.
       "1:                                        \n"
-      "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
-      "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
-      "subs       %3, %3, #1                     \n"    // 1 processed per loop.
-      "vmull.u8   q10, d4, d3                    \n"    // db * a
-      "vmull.u8   q11, d5, d3                    \n"    // dg * a
-      "vmull.u8   q12, d6, d3                    \n"    // dr * a
-      "vqrshrn.u16 d20, q10, #8                  \n"    // db >>= 8
-      "vqrshrn.u16 d21, q11, #8                  \n"    // dg >>= 8
-      "vqrshrn.u16 d22, q12, #8                  \n"    // dr >>= 8
-      "vqsub.u8   q2, q2, q10                    \n"    // dbg - dbg * a / 256
-      "vqsub.u8   d6, d6, d22                    \n"    // dr - dr * a / 256
-      "vqadd.u8   q0, q0, q2                     \n"    // + sbg
-      "vqadd.u8   d2, d2, d6                     \n"    // + sr
-      "vmov.u8    d3, #255                       \n"    // a = 255
-      "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
-      "bge        1b                             \n"
-
-      "99:                                         \n"
-
-      : "+r"(src_argb0),  // %0
+      "vld4.8      {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+      "vld4.8      {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+      "subs        %3, %3, #1                    \n"  // 1 processed per loop.
+      "vmull.u8    q10, d4, d3                   \n"  // db * a
+      "vmull.u8    q11, d5, d3                   \n"  // dg * a
+      "vmull.u8    q12, d6, d3                   \n"  // dr * a
+      "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+      "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+      "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+      "vqsub.u8    q2, q2, q10                   \n"  // dbg - dbg * a / 256
+      "vqsub.u8    d6, d6, d22                   \n"  // dr - dr * a / 256
+      "vqadd.u8    q0, q0, q2                    \n"  // + sbg
+      "vqadd.u8    d2, d2, d6                    \n"  // + sr
+      "vmov.u8     d3, #255                      \n"  // a = 255
+      "vst4.8      {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+      "bge         1b                            \n"
+
+      "99:                                       \n"
+
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2153,16 +2848,16 @@ void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
   asm volatile(
       // Attenuate 8 pixels.
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q10, d0, d3                    \n"  // b * a
-      "vmull.u8   q11, d1, d3                    \n"  // g * a
-      "vmull.u8   q12, d2, d3                    \n"  // r * a
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q10, d0, d3                   \n"  // b * a
+      "vmull.u8    q11, d1, d3                   \n"  // g * a
+      "vmull.u8    q12, d2, d3                   \n"  // r * a
       "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
       "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
       "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
-      "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2178,32 +2873,32 @@ void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
                           int interval_offset,
                           int width) {
   asm volatile(
-      "vdup.u16   q8, %2                         \n"
-      "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
-      "vdup.u16   q9, %3                         \n"  // interval multiply.
-      "vdup.u16   q10, %4                        \n"  // interval add
+      "vdup.u16    q8, %2                        \n"
+      "vshr.u16    q8, q8, #1                    \n"  // scale >>= 1
+      "vdup.u16    q9, %3                        \n"  // interval multiply.
+      "vdup.u16    q10, %4                       \n"  // interval add
 
       // 8 pixel loop.
       "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
-      "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-      "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
-      "vmovl.u8   q1, d2                         \n"
-      "vmovl.u8   q2, d4                         \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]        \n"  // load 8 pixels of ARGB.
+      "subs        %1, %1, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q0, d0                        \n"  // b (0 .. 255)
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q2, d4                        \n"
       "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
       "vqdmulh.s16 q1, q1, q8                    \n"  // g
       "vqdmulh.s16 q2, q2, q8                    \n"  // r
-      "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
-      "vmul.u16   q1, q1, q9                     \n"  // g
-      "vmul.u16   q2, q2, q9                     \n"  // r
-      "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
-      "vadd.u16   q1, q1, q10                    \n"  // g
-      "vadd.u16   q2, q2, q10                    \n"  // r
-      "vqmovn.u16 d0, q0                         \n"
-      "vqmovn.u16 d2, q1                         \n"
-      "vqmovn.u16 d4, q2                         \n"
-      "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vmul.u16    q0, q0, q9                    \n"  // b * interval_size
+      "vmul.u16    q1, q1, q9                    \n"  // g
+      "vmul.u16    q2, q2, q9                    \n"  // r
+      "vadd.u16    q0, q0, q10                   \n"  // b + interval_offset
+      "vadd.u16    q1, q1, q10                   \n"  // g
+      "vadd.u16    q2, q2, q10                   \n"  // r
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d2, q1                        \n"
+      "vqmovn.u16  d4, q2                        \n"
+      "vst4.8      {d0, d2, d4, d6}, [%0]!       \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(dst_argb),       // %0
         "+r"(width)           // %1
       : "r"(scale),           // %2
@@ -2220,28 +2915,28 @@ void ARGBShadeRow_NEON(const uint8_t* src_argb,
                        int width,
                        uint32_t value) {
   asm volatile(
-      "vdup.u32   q0, %3                         \n"  // duplicate scale value.
-      "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
-      "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+      "vdup.u32    q0, %3                        \n"  // duplicate scale value.
+      "vzip.u8     d0, d1                        \n"  // d0 aarrggbb.
+      "vshr.u16    q0, q0, #1                    \n"  // scale / 2.
 
       // 8 pixel loop.
       "1:                                        \n"
-      "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
-      "vmovl.u8   q11, d22                       \n"
-      "vmovl.u8   q12, d24                       \n"
-      "vmovl.u8   q13, d26                       \n"
+      "vld4.8      {d20, d22, d24, d26}, [%0]!   \n"  // load 8 pixels of ARGB.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q10, d20                      \n"  // b (0 .. 255)
+      "vmovl.u8    q11, d22                      \n"
+      "vmovl.u8    q12, d24                      \n"
+      "vmovl.u8    q13, d26                      \n"
       "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
       "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
       "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
       "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
-      "vqmovn.u16 d20, q10                       \n"
-      "vqmovn.u16 d22, q11                       \n"
-      "vqmovn.u16 d24, q12                       \n"
-      "vqmovn.u16 d26, q13                       \n"
-      "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
-      "bgt        1b                             \n"
+      "vqmovn.u16  d20, q10                      \n"
+      "vqmovn.u16  d22, q11                      \n"
+      "vqmovn.u16  d24, q12                      \n"
+      "vqmovn.u16  d26, q13                      \n"
+      "vst4.8      {d20, d22, d24, d26}, [%1]!   \n"  // store 8 pixels of ARGB.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2251,23 +2946,23 @@ void ARGBShadeRow_NEON(const uint8_t* src_argb,
 
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 // Similar to ARGBToYJ but stores ARGB.
-// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+// C code is (29 * b + 150 * g + 77 * r + 128) >> 8;
 void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   asm volatile(
-      "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-      "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-      "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q2, d0, d24                    \n"  // B
-      "vmlal.u8   q2, d1, d25                    \n"  // G
-      "vmlal.u8   q2, d2, d26                    \n"  // R
-      "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
-      "vmov       d1, d0                         \n"  // G
-      "vmov       d2, d0                         \n"  // R
-      "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
+      "vmov.u8     d24, #29                      \n"  // B * 0.1140 coefficient
+      "vmov.u8     d25, #150                     \n"  // G * 0.5870 coefficient
+      "vmov.u8     d26, #77                      \n"  // R * 0.2990 coefficient
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d24                   \n"  // B
+      "vmlal.u8    q2, d1, d25                   \n"  // G
+      "vmlal.u8    q2, d2, d26                   \n"  // R
+      "vqrshrn.u16 d0, q2, #8                    \n"  // 16 bit to 8 bit B
+      "vmov        d1, d0                        \n"  // G
+      "vmov        d2, d0                        \n"  // R
+      "vst4.8      {d0, d1, d2, d3}, [%1]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2281,32 +2976,32 @@ void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
 //    r = (r * 50 + g * 98 + b * 24) >> 7
 void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width) {
   asm volatile(
-      "vmov.u8    d20, #17                       \n"  // BB coefficient
-      "vmov.u8    d21, #68                       \n"  // BG coefficient
-      "vmov.u8    d22, #35                       \n"  // BR coefficient
-      "vmov.u8    d24, #22                       \n"  // GB coefficient
-      "vmov.u8    d25, #88                       \n"  // GG coefficient
-      "vmov.u8    d26, #45                       \n"  // GR coefficient
-      "vmov.u8    d28, #24                       \n"  // BB coefficient
-      "vmov.u8    d29, #98                       \n"  // BG coefficient
-      "vmov.u8    d30, #50                       \n"  // BR coefficient
-      "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
-      "subs       %1, %1, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
-      "vmlal.u8   q2, d1, d21                    \n"  // G
-      "vmlal.u8   q2, d2, d22                    \n"  // R
-      "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
-      "vmlal.u8   q3, d1, d25                    \n"  // G
-      "vmlal.u8   q3, d2, d26                    \n"  // R
-      "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
-      "vmlal.u8   q8, d1, d29                    \n"  // G
-      "vmlal.u8   q8, d2, d30                    \n"  // R
-      "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
-      "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
-      "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
-      "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
+      "vmov.u8     d20, #17                      \n"  // BB coefficient
+      "vmov.u8     d21, #68                      \n"  // BG coefficient
+      "vmov.u8     d22, #35                      \n"  // BR coefficient
+      "vmov.u8     d24, #22                      \n"  // GB coefficient
+      "vmov.u8     d25, #88                      \n"  // GG coefficient
+      "vmov.u8     d26, #45                      \n"  // GR coefficient
+      "vmov.u8     d28, #24                      \n"  // BB coefficient
+      "vmov.u8     d29, #98                      \n"  // BG coefficient
+      "vmov.u8     d30, #50                      \n"  // BR coefficient
+      "1:                                        \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]        \n"  // load 8 ARGB pixels.
+      "subs        %1, %1, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q2, d0, d20                   \n"  // B to Sepia B
+      "vmlal.u8    q2, d1, d21                   \n"  // G
+      "vmlal.u8    q2, d2, d22                   \n"  // R
+      "vmull.u8    q3, d0, d24                   \n"  // B to Sepia G
+      "vmlal.u8    q3, d1, d25                   \n"  // G
+      "vmlal.u8    q3, d2, d26                   \n"  // R
+      "vmull.u8    q8, d0, d28                   \n"  // B to Sepia R
+      "vmlal.u8    q8, d1, d29                   \n"  // G
+      "vmlal.u8    q8, d2, d30                   \n"  // R
+      "vqshrn.u16  d0, q2, #7                    \n"  // 16 bit to 8 bit B
+      "vqshrn.u16  d1, q3, #7                    \n"  // 16 bit to 8 bit G
+      "vqshrn.u16  d2, q8, #7                    \n"  // 16 bit to 8 bit R
+      "vst4.8      {d0, d1, d2, d3}, [%0]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
       : "+r"(dst_argb),  // %0
         "+r"(width)      // %1
       :
@@ -2322,51 +3017,51 @@ void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
                              const int8_t* matrix_argb,
                              int width) {
   asm volatile(
-      "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
-      "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
-      "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
-
-      "1:                                        \n"
-      "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-      "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
-      "vmovl.u8   q9, d18                        \n"  // g
-      "vmovl.u8   q10, d20                       \n"  // r
-      "vmovl.u8   q11, d22                       \n"  // a
-      "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
-      "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
-      "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
-      "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
-      "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
-      "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
-      "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
-      "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
-      "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-      "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-      "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-      "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-      "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
-      "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
-      "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
-      "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
-      "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-      "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-      "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-      "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
-      "vmul.s16   q4, q11, d0[3]                 \n"  // B += A * Matrix B
-      "vmul.s16   q5, q11, d1[3]                 \n"  // G += A * Matrix G
-      "vmul.s16   q6, q11, d2[3]                 \n"  // R += A * Matrix R
-      "vmul.s16   q7, q11, d3[3]                 \n"  // A += A * Matrix A
-      "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
-      "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
-      "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
-      "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+      "vld1.8      {q2}, [%3]                    \n"  // load 3 ARGB vectors.
+      "vmovl.s8    q0, d4                        \n"  // B,G coefficients s16.
+      "vmovl.s8    q1, d5                        \n"  // R,A coefficients s16.
+
+      "1:                                        \n"
+      "vld4.8      {d16, d18, d20, d22}, [%0]!   \n"  // load 8 ARGB pixels.
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop.
+      "vmovl.u8    q8, d16                       \n"  // b (0 .. 255) 16 bit
+      "vmovl.u8    q9, d18                       \n"  // g
+      "vmovl.u8    q10, d20                      \n"  // r
+      "vmovl.u8    q11, d22                      \n"  // a
+      "vmul.s16    q12, q8, d0[0]                \n"  // B = B * Matrix B
+      "vmul.s16    q13, q8, d1[0]                \n"  // G = B * Matrix G
+      "vmul.s16    q14, q8, d2[0]                \n"  // R = B * Matrix R
+      "vmul.s16    q15, q8, d3[0]                \n"  // A = B * Matrix A
+      "vmul.s16    q4, q9, d0[1]                 \n"  // B += G * Matrix B
+      "vmul.s16    q5, q9, d1[1]                 \n"  // G += G * Matrix G
+      "vmul.s16    q6, q9, d2[1]                 \n"  // R += G * Matrix R
+      "vmul.s16    q7, q9, d3[1]                 \n"  // A += G * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
+      "vmul.s16    q4, q10, d0[2]                \n"  // B += R * Matrix B
+      "vmul.s16    q5, q10, d1[2]                \n"  // G += R * Matrix G
+      "vmul.s16    q6, q10, d2[2]                \n"  // R += R * Matrix R
+      "vmul.s16    q7, q10, d3[2]                \n"  // A += R * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
+      "vmul.s16    q4, q11, d0[3]                \n"  // B += A * Matrix B
+      "vmul.s16    q5, q11, d1[3]                \n"  // G += A * Matrix G
+      "vmul.s16    q6, q11, d2[3]                \n"  // R += A * Matrix R
+      "vmul.s16    q7, q11, d3[3]                \n"  // A += A * Matrix A
+      "vqadd.s16   q12, q12, q4                  \n"  // Accumulate B
+      "vqadd.s16   q13, q13, q5                  \n"  // Accumulate G
+      "vqadd.s16   q14, q14, q6                  \n"  // Accumulate R
+      "vqadd.s16   q15, q15, q7                  \n"  // Accumulate A
       "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
       "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
       "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
       "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
-      "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
+      "vst4.8      {d16, d18, d20, d22}, [%1]!   \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
       : "+r"(src_argb),   // %0
         "+r"(dst_argb),   // %1
         "+r"(width)       // %2
@@ -2376,27 +3071,27 @@ void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
 }
 
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_NEON(const uint8_t* src_argb0,
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vmull.u8   q0, d0, d1                     \n"  // multiply B
-      "vmull.u8   q1, d2, d3                     \n"  // multiply G
-      "vmull.u8   q2, d4, d5                     \n"  // multiply R
-      "vmull.u8   q3, d6, d7                     \n"  // multiply A
-      "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
-      "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
-      "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
-      "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vmull.u8    q0, d0, d1                    \n"  // multiply B
+      "vmull.u8    q1, d2, d3                    \n"  // multiply G
+      "vmull.u8    q2, d4, d5                    \n"  // multiply R
+      "vmull.u8    q3, d6, d7                    \n"  // multiply A
+      "vrshrn.u16  d0, q0, #8                    \n"  // 16 bit to 8 bit B
+      "vrshrn.u16  d1, q1, #8                    \n"  // 16 bit to 8 bit G
+      "vrshrn.u16  d2, q2, #8                    \n"  // 16 bit to 8 bit R
+      "vrshrn.u16  d3, q3, #8                    \n"  // 16 bit to 8 bit A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2405,21 +3100,21 @@ void ARGBMultiplyRow_NEON(const uint8_t* src_argb0,
 }
 
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBAddRow_NEON(const uint8_t* src_argb0,
+void ARGBAddRow_NEON(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vqadd.u8   q0, q0, q2                     \n"  // add B, G
-      "vqadd.u8   q1, q1, q3                     \n"  // add R, A
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    q0, q0, q2                    \n"  // add B, G
+      "vqadd.u8    q1, q1, q3                    \n"  // add R, A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2428,21 +3123,21 @@ void ARGBAddRow_NEON(const uint8_t* src_argb0,
 }
 
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_NEON(const uint8_t* src_argb0,
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
-      "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d4, d5, d6, d7}, [%1]!       \n"  // load 8 more ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqsub.u8    q0, q0, q2                    \n"  // subtract B, G
+      "vqsub.u8    q1, q1, q3                    \n"  // subtract R, A
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2460,17 +3155,17 @@ void SobelRow_NEON(const uint8_t* src_sobelx,
                    uint8_t* dst_argb,
                    int width) {
   asm volatile(
-      "vmov.u8    d3, #255                       \n"  // alpha
+      "vmov.u8     d3, #255                      \n"  // alpha
       // 8 pixel loop.
       "1:                                        \n"
-      "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
-      "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vqadd.u8   d0, d0, d1                     \n"  // add
-      "vmov.u8    d1, d0                         \n"
-      "vmov.u8    d2, d0                         \n"
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
+      "vld1.8      {d0}, [%0]!                   \n"  // load 8 sobelx.
+      "vld1.8      {d1}, [%1]!                   \n"  // load 8 sobely.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d0, d0, d1                    \n"  // add
+      "vmov.u8     d1, d0                        \n"
+      "vmov.u8     d2, d0                        \n"
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -2487,12 +3182,12 @@ void SobelToPlaneRow_NEON(const uint8_t* src_sobelx,
   asm volatile(
       // 16 pixel loop.
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
-      "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop.
-      "vqadd.u8   q0, q0, q1                     \n"  // add
-      "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 sobelx.
+      "vld1.8      {q1}, [%1]!                   \n"  // load 16 sobely.
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vqadd.u8    q0, q0, q1                    \n"  // add
+      "vst1.8      {q0}, [%2]!                   \n"  // store 16 pixels.
+      "bgt         1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_y),       // %2
@@ -2511,15 +3206,15 @@ void SobelXYRow_NEON(const uint8_t* src_sobelx,
                      uint8_t* dst_argb,
                      int width) {
   asm volatile(
-      "vmov.u8    d3, #255                       \n"  // alpha
+      "vmov.u8     d3, #255                      \n"  // alpha
       // 8 pixel loop.
       "1:                                        \n"
-      "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
-      "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vqadd.u8   d1, d0, d2                     \n"  // add
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
-      "bgt        1b                             \n"
+      "vld1.8      {d2}, [%0]!                   \n"  // load 8 sobelx.
+      "vld1.8      {d0}, [%1]!                   \n"  // load 8 sobely.
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vqadd.u8    d1, d0, d2                    \n"  // add
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"  // store 8 ARGB pixels.
+      "bgt         1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -2539,23 +3234,23 @@ void SobelXRow_NEON(const uint8_t* src_y0,
                     int width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {d0}, [%0],%5                  \n"  // top
-      "vld1.8     {d1}, [%0],%6                  \n"
-      "vsubl.u8   q0, d0, d1                     \n"
-      "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
-      "vld1.8     {d3}, [%1],%6                  \n"
-      "vsubl.u8   q1, d2, d3                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vld1.8     {d2}, [%2],%5                  \n"  // bottom
-      "vld1.8     {d3}, [%2],%6                  \n"
-      "subs       %4, %4, #8                     \n"  // 8 pixels
-      "vsubl.u8   q1, d2, d3                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vabs.s16   q0, q0                         \n"
-      "vqmovn.u16 d0, q0                         \n"
-      "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
-      "bgt        1b                             \n"
+      "vld1.8      {d0}, [%0],%5                 \n"  // top
+      "vld1.8      {d1}, [%0],%6                 \n"
+      "vsubl.u8    q0, d0, d1                    \n"
+      "vld1.8      {d2}, [%1],%5                 \n"  // center * 2
+      "vld1.8      {d3}, [%1],%6                 \n"
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vld1.8      {d2}, [%2],%5                 \n"  // bottom
+      "vld1.8      {d3}, [%2],%6                 \n"
+      "subs        %4, %4, #8                    \n"  // 8 pixels
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vabs.s16    q0, q0                        \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vst1.8      {d0}, [%3]!                   \n"  // store 8 sobelx
+      "bgt         1b                            \n"
       : "+r"(src_y0),               // %0
         "+r"(src_y1),               // %1
         "+r"(src_y2),               // %2
@@ -2577,23 +3272,23 @@ void SobelYRow_NEON(const uint8_t* src_y0,
                     int width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {d0}, [%0],%4                  \n"  // left
-      "vld1.8     {d1}, [%1],%4                  \n"
-      "vsubl.u8   q0, d0, d1                     \n"
-      "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
-      "vld1.8     {d3}, [%1],%4                  \n"
-      "vsubl.u8   q1, d2, d3                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vld1.8     {d2}, [%0],%5                  \n"  // right
-      "vld1.8     {d3}, [%1],%5                  \n"
-      "subs       %3, %3, #8                     \n"  // 8 pixels
-      "vsubl.u8   q1, d2, d3                     \n"
-      "vadd.s16   q0, q0, q1                     \n"
-      "vabs.s16   q0, q0                         \n"
-      "vqmovn.u16 d0, q0                         \n"
-      "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
-      "bgt        1b                             \n"
+      "vld1.8      {d0}, [%0],%4                 \n"  // left
+      "vld1.8      {d1}, [%1],%4                 \n"
+      "vsubl.u8    q0, d0, d1                    \n"
+      "vld1.8      {d2}, [%0],%4                 \n"  // center * 2
+      "vld1.8      {d3}, [%1],%4                 \n"
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vld1.8      {d2}, [%0],%5                 \n"  // right
+      "vld1.8      {d3}, [%1],%5                 \n"
+      "subs        %3, %3, #8                    \n"  // 8 pixels
+      "vsubl.u8    q1, d2, d3                    \n"
+      "vadd.s16    q0, q0, q1                    \n"
+      "vabs.s16    q0, q0                        \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vst1.8      {d0}, [%2]!                   \n"  // store 8 sobely
+      "bgt         1b                            \n"
       : "+r"(src_y0),               // %0
         "+r"(src_y1),               // %1
         "+r"(dst_sobely),           // %2
@@ -2615,18 +3310,18 @@ void HalfFloat1Row_NEON(const uint16_t* src,
   asm volatile(
 
       "1:                                        \n"
-      "vld1.8     {q1}, [%0]!                    \n"  // load 8 shorts
-      "subs       %2, %2, #8                     \n"  // 8 pixels per loop
-      "vmovl.u16  q2, d2                         \n"  // 8 int's
-      "vmovl.u16  q3, d3                         \n"
-      "vcvt.f32.u32  q2, q2                      \n"  // 8 floats
-      "vcvt.f32.u32  q3, q3                      \n"
-      "vmul.f32   q2, q2, %y3                    \n"  // adjust exponent
-      "vmul.f32   q3, q3, %y3                    \n"
-      "vqshrn.u32 d2, q2, #13                    \n"  // isolate halffloat
-      "vqshrn.u32 d3, q3, #13                    \n"
-      "vst1.8     {q1}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q1}, [%0]!                   \n"  // load 8 shorts
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u16   q2, d2                        \n"  // 8 int's
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // adjust exponent
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vqshrn.u32  d2, q2, #13                   \n"  // isolate halffloat
+      "vqshrn.u32  d3, q3, #13                   \n"
+      "vst1.8      {q1}, [%1]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src),              // %0
         "+r"(dst),              // %1
         "+r"(width)             // %2
@@ -2641,18 +3336,18 @@ void HalfFloatRow_NEON(const uint16_t* src,
   asm volatile(
 
       "1:                                        \n"
-      "vld1.8     {q1}, [%0]!                    \n"  // load 8 shorts
-      "subs       %2, %2, #8                     \n"  // 8 pixels per loop
-      "vmovl.u16  q2, d2                         \n"  // 8 int's
-      "vmovl.u16  q3, d3                         \n"
-      "vcvt.f32.u32  q2, q2                      \n"  // 8 floats
-      "vcvt.f32.u32  q3, q3                      \n"
-      "vmul.f32   q2, q2, %y3                    \n"  // adjust exponent
-      "vmul.f32   q3, q3, %y3                    \n"
-      "vqshrn.u32 d2, q2, #13                    \n"  // isolate halffloat
-      "vqshrn.u32 d3, q3, #13                    \n"
-      "vst1.8     {q1}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q1}, [%0]!                   \n"  // load 8 shorts
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u16   q2, d2                        \n"  // 8 int's
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // adjust exponent
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vqshrn.u32  d2, q2, #13                   \n"  // isolate halffloat
+      "vqshrn.u32  d3, q3, #13                   \n"
+      "vst1.8      {q1}, [%1]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src),                      // %0
         "+r"(dst),                      // %1
         "+r"(width)                     // %2
@@ -2667,17 +3362,17 @@ void ByteToFloatRow_NEON(const uint8_t* src,
   asm volatile(
 
       "1:                                        \n"
-      "vld1.8     {d2}, [%0]!                    \n"  // load 8 bytes
-      "subs       %2, %2, #8                     \n"  // 8 pixels per loop
-      "vmovl.u8   q1, d2                         \n"  // 8 shorts
-      "vmovl.u16  q2, d2                         \n"  // 8 ints
-      "vmovl.u16  q3, d3                         \n"
-      "vcvt.f32.u32  q2, q2                      \n"  // 8 floats
-      "vcvt.f32.u32  q3, q3                      \n"
-      "vmul.f32   q2, q2, %y3                    \n"  // scale
-      "vmul.f32   q3, q3, %y3                    \n"
-      "vst1.8     {q2, q3}, [%1]!                \n"  // store 8 floats
-      "bgt        1b                             \n"
+      "vld1.8      {d2}, [%0]!                   \n"  // load 8 bytes
+      "subs        %2, %2, #8                    \n"  // 8 pixels per loop
+      "vmovl.u8    q1, d2                        \n"  // 8 shorts
+      "vmovl.u16   q2, d2                        \n"  // 8 ints
+      "vmovl.u16   q3, d3                        \n"
+      "vcvt.f32.u32 q2, q2                       \n"  // 8 floats
+      "vcvt.f32.u32 q3, q3                       \n"
+      "vmul.f32    q2, q2, %y3                   \n"  // scale
+      "vmul.f32    q3, q3, %y3                   \n"
+      "vst1.8      {q2, q3}, [%1]!               \n"  // store 8 floats
+      "bgt         1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -2694,26 +3389,26 @@ void GaussCol_NEON(const uint16_t* src0,
                    uint32_t* dst,
                    int width) {
   asm volatile(
-      "vmov.u16   d6, #4                         \n"  // constant 4
-      "vmov.u16   d7, #6                         \n"  // constant 6
-
-      "1:                                        \n"
-      "vld1.16    {q1}, [%0]!                    \n"  // load 8 samples, 5 rows
-      "vld1.16    {q2}, [%4]!                    \n"
-      "vaddl.u16  q0, d2, d4                     \n"  // * 1
-      "vaddl.u16  q1, d3, d5                     \n"  // * 1
-      "vld1.16    {q2}, [%1]!                    \n"
-      "vmlal.u16  q0, d4, d6                     \n"  // * 4
-      "vmlal.u16  q1, d5, d6                     \n"  // * 4
-      "vld1.16    {q2}, [%2]!                    \n"
-      "vmlal.u16  q0, d4, d7                     \n"  // * 6
-      "vmlal.u16  q1, d5, d7                     \n"  // * 6
-      "vld1.16    {q2}, [%3]!                    \n"
-      "vmlal.u16  q0, d4, d6                     \n"  // * 4
-      "vmlal.u16  q1, d5, d6                     \n"  // * 4
-      "subs       %6, %6, #8                     \n"  // 8 processed per loop
-      "vst1.32    {q0, q1}, [%5]!                \n"  // store 8 samples
-      "bgt        1b                             \n"
+      "vmov.u16    d6, #4                        \n"  // constant 4
+      "vmov.u16    d7, #6                        \n"  // constant 6
+
+      "1:                                        \n"
+      "vld1.16     {q1}, [%0]!                   \n"  // load 8 samples, 5 rows
+      "vld1.16     {q2}, [%4]!                   \n"
+      "vaddl.u16   q0, d2, d4                    \n"  // * 1
+      "vaddl.u16   q1, d3, d5                    \n"  // * 1
+      "vld1.16     {q2}, [%1]!                   \n"
+      "vmlal.u16   q0, d4, d6                    \n"  // * 4
+      "vmlal.u16   q1, d5, d6                    \n"  // * 4
+      "vld1.16     {q2}, [%2]!                   \n"
+      "vmlal.u16   q0, d4, d7                    \n"  // * 6
+      "vmlal.u16   q1, d5, d7                    \n"  // * 6
+      "vld1.16     {q2}, [%3]!                   \n"
+      "vmlal.u16   q0, d4, d6                    \n"  // * 4
+      "vmlal.u16   q1, d5, d6                    \n"  // * 4
+      "subs        %6, %6, #8                    \n"  // 8 processed per loop
+      "vst1.32     {q0, q1}, [%5]!               \n"  // store 8 samples
+      "bgt         1b                            \n"
       : "+r"(src0),  // %0
         "+r"(src1),  // %1
         "+r"(src2),  // %2
@@ -2731,8 +3426,8 @@ void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width) {
   const uint32_t* src2 = src + 2;
   const uint32_t* src3 = src + 3;
   asm volatile(
-      "vmov.u32    q10, #4                        \n"  // constant 4
-      "vmov.u32    q11, #6                        \n"  // constant 6
+      "vmov.u32    q10, #4                       \n"  // constant 4
+      "vmov.u32    q11, #6                       \n"  // constant 6
 
       "1:                                        \n"
       "vld1.32     {q0, q1}, [%0]!               \n"  // load 12 source samples
@@ -2769,17 +3464,17 @@ void NV21ToYUV24Row_NEON(const uint8_t* src_y,
                          uint8_t* dst_yuv24,
                          int width) {
   asm volatile(
-      "1:                                          \n"
-      "vld1.8    {q2}, [%0]!                     \n"  // load 16 Y values
-      "vld2.8    {d0, d2}, [%1]!                 \n"  // load 8 VU values
-      "vmov      d1, d0                          \n"
-      "vzip.u8   d0, d1                          \n"  // VV
-      "vmov      d3, d2                          \n"
-      "vzip.u8   d2, d3                          \n"  // UU
-      "subs      %3, %3, #16                     \n"  // 16 pixels per loop
-      "vst3.8    {d0, d2, d4}, [%2]!             \n"  // store 16 YUV pixels
-      "vst3.8    {d1, d3, d5}, [%2]!             \n"
-      "bgt        1b                             \n"
+      "1:                                        \n"
+      "vld1.8      {q2}, [%0]!                   \n"  // load 16 Y values
+      "vld2.8      {d0, d2}, [%1]!               \n"  // load 8 VU values
+      "vmov        d1, d0                        \n"
+      "vzip.u8     d0, d1                        \n"  // VV
+      "vmov        d3, d2                        \n"
+      "vzip.u8     d2, d3                        \n"  // UU
+      "subs        %3, %3, #16                   \n"  // 16 pixels per loop
+      "vst3.8      {d0, d2, d4}, [%2]!           \n"  // store 16 YUV pixels
+      "vst3.8      {d1, d3, d5}, [%2]!           \n"
+      "bgt         1b                            \n"
       : "+r"(src_y),      // %0
         "+r"(src_vu),     // %1
         "+r"(dst_yuv24),  // %2
@@ -2793,24 +3488,24 @@ void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
                       uint8_t* dst_uv,
                       int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // src_stride + src_AYUV
+      "add         %1, %0, %1                    \n"  // src_stride + src_AYUV
       "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 AYUV pixels.
-      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 AYUV
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV
                                                       // pixels.
-      "vpaddl.u8  q0, q0                         \n"  // V 16 bytes -> 8 shorts.
-      "vpaddl.u8  q1, q1                         \n"  // U 16 bytes -> 8 shorts.
-      "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more AYUV
+      "vpaddl.u8   q0, q0                        \n"  // V 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // U 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more AYUV
                                                       // pixels.
-      "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 AYUV
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 AYUV
                                                       // pixels.
-      "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-      "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
       "vqrshrun.s16 d1, q0, #2                   \n"  // 2x2 average
       "vqrshrun.s16 d0, q1, #2                   \n"
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop.
-      "vst2.8     {d0, d1}, [%2]!                \n"  // store 8 pixels UV.
-      "bgt        1b                             \n"
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store 8 pixels UV.
+      "bgt         1b                            \n"
       : "+r"(src_ayuv),         // %0
         "+r"(src_stride_ayuv),  // %1
         "+r"(dst_uv),           // %2
@@ -2824,24 +3519,24 @@ void AYUVToVURow_NEON(const uint8_t* src_ayuv,
                       uint8_t* dst_vu,
                       int width) {
   asm volatile(
-      "add        %1, %0, %1                     \n"  // src_stride + src_AYUV
+      "add         %1, %0, %1                    \n"  // src_stride + src_AYUV
       "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 AYUV pixels.
-      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 AYUV
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV
                                                       // pixels.
-      "vpaddl.u8  q0, q0                         \n"  // V 16 bytes -> 8 shorts.
-      "vpaddl.u8  q1, q1                         \n"  // U 16 bytes -> 8 shorts.
-      "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more AYUV
+      "vpaddl.u8   q0, q0                        \n"  // V 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // U 16 bytes -> 8 shorts.
+      "vld4.8      {d8, d10, d12, d14}, [%1]!    \n"  // load 8 more AYUV
                                                       // pixels.
-      "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 AYUV
+      "vld4.8      {d9, d11, d13, d15}, [%1]!    \n"  // load last 8 AYUV
                                                       // pixels.
-      "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
-      "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q0, q4                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q5                        \n"  // G 16 bytes -> 8 shorts.
       "vqrshrun.s16 d0, q0, #2                   \n"  // 2x2 average
       "vqrshrun.s16 d1, q1, #2                   \n"
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop.
-      "vst2.8     {d0, d1}, [%2]!                \n"  // store 8 pixels VU.
-      "bgt        1b                             \n"
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop.
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store 8 pixels VU.
+      "bgt         1b                            \n"
       : "+r"(src_ayuv),         // %0
         "+r"(src_stride_ayuv),  // %1
         "+r"(dst_vu),           // %2
@@ -2854,12 +3549,12 @@ void AYUVToVURow_NEON(const uint8_t* src_ayuv,
 // Similar to ARGBExtractAlphaRow_NEON
 void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
   asm volatile(
-      "1:                                          \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 AYUV pixels
-      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 AYUV pixels
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop
-      "vst1.8     {q2}, [%1]!                    \n"  // store 16 Y's.
-      "bgt       1b                              \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 AYUV pixels
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 AYUV pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q2}, [%1]!                   \n"  // store 16 Y's.
+      "bgt         1b                            \n"
       : "+r"(src_ayuv),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -2867,16 +3562,16 @@ void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
       : "cc", "memory", "q0", "q1", "q2", "q3");
 }
 
-// Convert biplanar UV channel of NV12 to NV21
-void UVToVURow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
   asm volatile(
-      "1:                                          \n"
-      "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 UV values
-      "vld2.8     {d1, d3}, [%0]!                \n"
-      "vorr.u8    q2, q0, q0                     \n"  // move U after V
-      "subs       %2, %2, #16                    \n"  // 16 pixels per loop
-      "vst2.8     {q1, q2}, [%1]!                \n"  // store 16 VU pixels
-      "bgt        1b                             \n"
+      "1:                                        \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 16 UV values
+      "vld2.8      {d1, d3}, [%0]!               \n"
+      "vorr.u8     q2, q0, q0                    \n"  // move U after V
+      "subs        %2, %2, #16                   \n"  // 16 pixels per loop
+      "vst2.8      {q1, q2}, [%1]!               \n"  // store 16 VU pixels
+      "bgt         1b                            \n"
       : "+r"(src_uv),  // %0
         "+r"(dst_vu),  // %1
         "+r"(width)    // %2
@@ -2884,6 +3579,170 @@ void UVToVURow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
       : "cc", "memory", "q0", "q1", "q2");
 }
 
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  const uint8_t* src_u_1 = src_u + src_stride_u;
+  const uint8_t* src_v_1 = src_v + src_stride_v;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 U values
+      "vld1.8      {q1}, [%2]!                   \n"  // load 16 V values
+      "vld1.8      {q2}, [%1]!                   \n"
+      "vld1.8      {q3}, [%3]!                   \n"
+      "vpaddl.u8   q0, q0                        \n"  // half size
+      "vpaddl.u8   q1, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"
+      "vpadal.u8   q1, q3                        \n"
+      "vqrshrn.u16 d0, q0, #2                    \n"
+      "vqrshrn.u16 d1, q1, #2                    \n"
+      "subs        %5, %5, #16                   \n"  // 16 src pixels per loop
+      "vst2.8      {d0, d1}, [%4]!               \n"  // store 8 UV pixels
+      "bgt         1b                            \n"
+      : "+r"(src_u),    // %0
+        "+r"(src_u_1),  // %1
+        "+r"(src_v),    // %2
+        "+r"(src_v_1),  // %3
+        "+r"(dst_uv),   // %4
+        "+r"(width)     // %5
+      :
+      : "cc", "memory", "q0", "q1", "q2", "q3");
+}
+
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  int shift = depth - 16;  // Negative for right shift.
+  asm volatile(
+      "vdup.16     q2, %4                        \n"
+      "1:                                        \n"
+      "vld2.16     {q0, q1}, [%0]!               \n"  // load 8 UV
+      "vshl.u16    q0, q0, q2                    \n"
+      "vshl.u16    q1, q1, q2                    \n"
+      "subs        %3, %3, #8                    \n"  // 8 src pixels per loop
+      "vst1.16     {q0}, [%1]!                   \n"  // store 8 U pixels
+      "vst1.16     {q1}, [%2]!                   \n"  // store 8 V pixels
+      "bgt         1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "vdup.16     q2, %4                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // load 8 U
+      "vld1.16     {q1}, [%1]!                   \n"  // load 8 V
+      "vshl.u16    q0, q0, q2                    \n"
+      "vshl.u16    q1, q1, q2                    \n"
+      "subs        %3, %3, #8                    \n"  // 8 src pixels per loop
+      "vst2.16     {q0, q1}, [%2]!               \n"  // store 8 UV pixels
+      "bgt         1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width) {
+  asm volatile(
+      "vdup.16     q2, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vmul.u16    q0, q0, q2                    \n"
+      "vmul.u16    q1, q1, q2                    \n"
+      "vst1.16     {q0}, [%1]!                   \n"
+      "vst1.16     {q1}, [%1]!                   \n"
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  asm volatile(
+      "vdup.16     q0, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vld1.16     {q2}, [%0]!                   \n"
+      "vmovl.u16   q3, d2                        \n"
+      "vmovl.u16   q1, d3                        \n"
+      "vmovl.u16   q4, d4                        \n"
+      "vmovl.u16   q2, d5                        \n"
+      "vshl.u32    q3, q3, q0                    \n"
+      "vshl.u32    q4, q4, q0                    \n"
+      "vshl.u32    q1, q1, q0                    \n"
+      "vshl.u32    q2, q2, q0                    \n"
+      "vmovn.u32   d2, q3                        \n"
+      "vmovn.u32   d3, q1                        \n"
+      "vmovn.u32   d4, q4                        \n"
+      "vmovn.u32   d5, q2                        \n"
+      "vst1.16     {q1}, [%1]!                   \n"
+      "vst1.16     {q2}, [%1]!                   \n"
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "q0", "q1", "q2", "q3", "q4");
+}
+
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits = shr 1
+// 16384 = 10 bits = shr 2
+// 4096 = 12 bits = shr 4
+// 256 = 16 bits = shr 8
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width) {
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
+  asm volatile(
+      "vdup.16     q2, %3                        \n"
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"
+      "vld1.16     {q1}, [%0]!                   \n"
+      "vshl.u16    q0, q0, q2                    \n"  // shr = q2 is negative
+      "vshl.u16    q1, q1, q2                    \n"
+      "vqmovn.u16  d0, q0                        \n"
+      "vqmovn.u16  d1, q1                        \n"
+      "subs        %2, %2, #16                   \n"  // 16 src pixels per loop
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(shift)    // %3
+      : "cc", "memory", "q0", "q1", "q2");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__)..
 
 #ifdef __cplusplus
diff --git a/files/source/row_neon64.cc b/files/source/row_neon64.cc
index f5cbb470..0f120373 100644
--- a/files/source/row_neon64.cc
+++ b/files/source/row_neon64.cc
@@ -15,102 +15,108 @@ namespace libyuv {
 extern "C" {
 #endif
 
+// Enable LIBYUV_USE_ST2, LIBYUV_USE_ST3, LIBYUV_USE_ST4 for CPUs that prefer
+// STn over ZIP1+ST1
+// Exynos M1, M2, M3 are slow with ST2, ST3 and ST4 instructions.
+
 // This module is for GCC Neon armv8 64 bit.
 #if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
+// v0.8h: Y
+// v1.16b: 8U, 8V
+
 // Read 8 Y, 4 U and 4 V from 422
 #define READYUV422                               \
-  "ld1        {v0.8b}, [%0], #8              \n" \
-  "ld1        {v1.s}[0], [%1], #4            \n" \
-  "ld1        {v1.s}[1], [%2], #4            \n"
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ld1        {v1.s}[0], [%[src_u]], #4      \n" \
+  "ld1        {v1.s}[1], [%[src_v]], #4      \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "zip1       v1.16b, v1.16b, v1.16b         \n" \
+  "prfm       pldl1keep, [%[src_u], 128]     \n" \
+  "prfm       pldl1keep, [%[src_v], 128]     \n"
 
 // Read 8 Y, 8 U and 8 V from 444
 #define READYUV444                               \
-  "ld1        {v0.8b}, [%0], #8              \n" \
-  "ld1        {v1.d}[0], [%1], #8            \n" \
-  "ld1        {v1.d}[1], [%2], #8            \n" \
-  "uaddlp     v1.8h, v1.16b                  \n" \
-  "rshrn      v1.8b, v1.8h, #1               \n"
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ld1        {v1.d}[0], [%[src_u]], #8      \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "ld1        {v1.d}[1], [%[src_v]], #8      \n" \
+  "prfm       pldl1keep, [%[src_u], 448]     \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_v], 448]     \n"
 
 // Read 8 Y, and set 4 U and 4 V to 128
 #define READYUV400                               \
-  "ld1        {v0.8b}, [%0], #8              \n" \
-  "movi       v1.8b , #128                   \n"
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "movi       v1.16b, #128                   \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n"
 
-// Read 8 Y and 4 UV from NV12
+static const uvec8 kNV12Table = {0, 0, 2, 2, 4, 4, 6, 6,
+                                 1, 1, 3, 3, 5, 5, 7, 7};
+static const uvec8 kNV21Table = {1, 1, 3, 3, 5, 5, 7, 7,
+                                 0, 0, 2, 2, 4, 4, 6, 6};
+
+// Read 8 Y and 4 UV from NV12 or NV21
 #define READNV12                                 \
-  "ld1        {v0.8b}, [%0], #8              \n" \
-  "ld1        {v2.8b}, [%1], #8              \n" \
-  "uzp1       v1.8b, v2.8b, v2.8b            \n" \
-  "uzp2       v3.8b, v2.8b, v2.8b            \n" \
-  "ins        v1.s[1], v3.s[0]               \n"
-
-// Read 8 Y and 4 VU from NV21
-#define READNV21                                 \
-  "ld1        {v0.8b}, [%0], #8              \n" \
-  "ld1        {v2.8b}, [%1], #8              \n" \
-  "uzp1       v3.8b, v2.8b, v2.8b            \n" \
-  "uzp2       v1.8b, v2.8b, v2.8b            \n" \
-  "ins        v1.s[1], v3.s[0]               \n"
+  "ldr        d0, [%[src_y]], #8             \n" \
+  "ldr        d1, [%[src_uv]], #8            \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n" \
+  "prfm       pldl1keep, [%[src_y], 448]     \n" \
+  "tbl        v1.16b, {v1.16b}, v2.16b       \n" \
+  "prfm       pldl1keep, [%[src_uv], 448]    \n"
 
 // Read 8 YUY2
-#define READYUY2                                 \
-  "ld2        {v0.8b, v1.8b}, [%0], #16      \n" \
-  "uzp2       v3.8b, v1.8b, v1.8b            \n" \
-  "uzp1       v1.8b, v1.8b, v1.8b            \n" \
-  "ins        v1.s[1], v3.s[0]               \n"
+#define READYUY2                                     \
+  "ld2        {v0.8b, v1.8b}, [%[src_yuy2]], #16 \n" \
+  "zip1       v0.16b, v0.16b, v0.16b         \n"     \
+  "prfm       pldl1keep, [%[src_yuy2], 448]  \n"     \
+  "tbl        v1.16b, {v1.16b}, v2.16b       \n"
 
 // Read 8 UYVY
-#define READUYVY                                 \
-  "ld2        {v2.8b, v3.8b}, [%0], #16      \n" \
-  "orr        v0.8b, v3.8b, v3.8b            \n" \
-  "uzp1       v1.8b, v2.8b, v2.8b            \n" \
-  "uzp2       v3.8b, v2.8b, v2.8b            \n" \
-  "ins        v1.s[1], v3.s[0]               \n"
-
-#define YUVTORGB_SETUP                           \
-  "ld1r       {v24.8h}, [%[kUVBiasBGR]], #2  \n" \
-  "ld1r       {v25.8h}, [%[kUVBiasBGR]], #2  \n" \
-  "ld1r       {v26.8h}, [%[kUVBiasBGR]]      \n" \
-  "ld1r       {v31.4s}, [%[kYToRgb]]         \n" \
-  "ld2        {v27.8h, v28.8h}, [%[kUVToRB]] \n" \
-  "ld2        {v29.8h, v30.8h}, [%[kUVToG]]  \n"
-
-#define YUVTORGB(vR, vG, vB)                                        \
-  "uxtl       v0.8h, v0.8b                   \n" /* Extract Y    */ \
-  "shll       v2.8h, v1.8b, #8               \n" /* Replicate UV */ \
-  "ushll2     v3.4s, v0.8h, #0               \n" /* Y */            \
-  "ushll      v0.4s, v0.4h, #0               \n"                    \
-  "mul        v3.4s, v3.4s, v31.4s           \n"                    \
-  "mul        v0.4s, v0.4s, v31.4s           \n"                    \
-  "sqshrun    v0.4h, v0.4s, #16              \n"                    \
-  "sqshrun2   v0.8h, v3.4s, #16              \n" /* Y */            \
-  "uaddw      v1.8h, v2.8h, v1.8b            \n" /* Replicate UV */ \
-  "mov        v2.d[0], v1.d[1]               \n" /* Extract V */    \
-  "uxtl       v2.8h, v2.8b                   \n"                    \
-  "uxtl       v1.8h, v1.8b                   \n" /* Extract U */    \
-  "mul        v3.8h, v1.8h, v27.8h           \n"                    \
-  "mul        v5.8h, v1.8h, v29.8h           \n"                    \
-  "mul        v6.8h, v2.8h, v30.8h           \n"                    \
-  "mul        v7.8h, v2.8h, v28.8h           \n"                    \
-  "sqadd      v6.8h, v6.8h, v5.8h            \n"                    \
-  "sqadd      " #vB                                                 \
-  ".8h, v24.8h, v0.8h      \n" /* B */                              \
-  "sqadd      " #vG                                                 \
-  ".8h, v25.8h, v0.8h      \n" /* G */                              \
-  "sqadd      " #vR                                                 \
-  ".8h, v26.8h, v0.8h      \n" /* R */                              \
-  "sqadd      " #vB ".8h, " #vB                                     \
-  ".8h, v3.8h  \n" /* B */                                          \
-  "sqsub      " #vG ".8h, " #vG                                     \
-  ".8h, v6.8h  \n" /* G */                                          \
-  "sqadd      " #vR ".8h, " #vR                                     \
-  ".8h, v7.8h  \n" /* R */                                          \
-  "sqshrun    " #vB ".8b, " #vB                                     \
-  ".8h, #6     \n" /* B */                                          \
-  "sqshrun    " #vG ".8b, " #vG                                     \
-  ".8h, #6     \n"                               /* G */            \
-  "sqshrun    " #vR ".8b, " #vR ".8h, #6     \n" /* R */
+#define READUYVY                                     \
+  "ld2        {v3.8b, v4.8b}, [%[src_uyvy]], #16 \n" \
+  "zip1       v0.16b, v4.16b, v4.16b         \n"     \
+  "prfm       pldl1keep, [%[src_uyvy], 448]  \n"     \
+  "tbl        v1.16b, {v3.16b}, v2.16b       \n"
+
+// UB VR UG VG
+// YG BB BG BR
+#define YUVTORGB_SETUP                                                \
+  "ld4r       {v28.16b, v29.16b, v30.16b, v31.16b}, [%[kUVCoeff]] \n" \
+  "ld4r       {v24.8h, v25.8h, v26.8h, v27.8h}, [%[kRGBCoeffBias]] \n"
+
+// v16.8h: B
+// v17.8h: G
+// v18.8h: R
+
+// Convert from YUV to 2.14 fixed point RGB
+#define YUVTORGB                                          \
+  "umull2     v3.4s, v0.8h, v24.8h           \n"          \
+  "umull      v6.8h, v1.8b, v30.8b           \n"          \
+  "umull      v0.4s, v0.4h, v24.4h           \n"          \
+  "umlal2     v6.8h, v1.16b, v31.16b         \n" /* DG */ \
+  "uqshrn     v0.4h, v0.4s, #16              \n"          \
+  "uqshrn2    v0.8h, v3.4s, #16              \n" /* Y */  \
+  "umull      v4.8h, v1.8b, v28.8b           \n" /* DB */ \
+  "umull2     v5.8h, v1.16b, v29.16b         \n" /* DR */ \
+  "add        v17.8h, v0.8h, v26.8h          \n" /* G */  \
+  "add        v16.8h, v0.8h, v4.8h           \n" /* B */  \
+  "add        v18.8h, v0.8h, v5.8h           \n" /* R */  \
+  "uqsub      v17.8h, v17.8h, v6.8h          \n" /* G */  \
+  "uqsub      v16.8h, v16.8h, v25.8h         \n" /* B */  \
+  "uqsub      v18.8h, v18.8h, v27.8h         \n" /* R */
+
+// Convert from 2.14 fixed point RGB To 8 bit RGB
+#define RGBTORGB8                                \
+  "uqshrn     v17.8b, v17.8h, #6             \n" \
+  "uqshrn     v16.8b, v16.8h, #6             \n" \
+  "uqshrn     v18.8b, v18.8h, #6             \n"
+
+#define YUVTORGB_REGS                                                          \
+  "v0", "v1", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v24", "v25", \
+      "v26", "v27", "v28", "v29", "v30", "v31"
 
 void I444ToARGBRow_NEON(const uint8_t* src_y,
                         const uint8_t* src_u,
@@ -118,27 +124,22 @@ void I444ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV444
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV444 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
 void I422ToARGBRow_NEON(const uint8_t* src_y,
@@ -147,27 +148,48 @@ void I422ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_argb),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
+}
+
+void I444AlphaToARGBRow_NEON(const uint8_t* src_y,
+                             const uint8_t* src_u,
+                             const uint8_t* src_v,
+                             const uint8_t* src_a,
+                             uint8_t* dst_argb,
+                             const struct YuvConstants* yuvconstants,
+                             int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n"
+      "ld1         {v19.8b}, [%[src_a]], #8      \n" READYUV444
+      "prfm        pldl1keep, [%[src_a], 448]    \n" YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
 void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
@@ -177,28 +199,23 @@ void I422AlphaToARGBRow_NEON(const uint8_t* src_y,
                              uint8_t* dst_argb,
                              const struct YuvConstants* yuvconstants,
                              int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "ld1        {v23.8b}, [%3], #8             \n"
-    "subs       %w5, %w5, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%4], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(src_a),     // %3
-      "+r"(dst_argb),  // %4
-      "+r"(width)      // %5
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n"
+      "ld1         {v19.8b}, [%[src_a]], #8      \n" READYUV422
+      "prfm        pldl1keep, [%[src_a], 448]    \n" YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [src_a] "+r"(src_a),                               // %[src_a]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
 void I422ToRGBARow_NEON(const uint8_t* src_y,
@@ -207,27 +224,22 @@ void I422ToRGBARow_NEON(const uint8_t* src_y,
                         uint8_t* dst_rgba,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v20.8b, #255                   \n" /* A */
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v23, v22, v21)
-    "subs       %w4, %w4, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_rgba),  // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v15.8b, #255                  \n" /* A */
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v15.8b,v16.8b,v17.8b,v18.8b}, [%[dst_rgba]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgba] "+r"(dst_rgba),                         // %[dst_rgba]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v15");
 }
 
 void I422ToRGB24Row_NEON(const uint8_t* src_y,
@@ -236,34 +248,29 @@ void I422ToRGB24Row_NEON(const uint8_t* src_y,
                          uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_u),     // %1
-      "+r"(src_v),     // %2
-      "+r"(dst_rgb24), // %3
-      "+r"(width)      // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb24] "+r"(dst_rgb24),                       // %[dst_rgb24]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 #define ARGBTORGB565                                                        \
-  "shll       v0.8h,  v22.8b, #8             \n" /* R                    */ \
-  "shll       v21.8h, v21.8b, #8             \n" /* G                    */ \
-  "shll       v20.8h, v20.8b, #8             \n" /* B                    */ \
-  "sri        v0.8h,  v21.8h, #5             \n" /* RG                   */ \
-  "sri        v0.8h,  v20.8h, #11            \n" /* RGB                  */
+  "shll       v18.8h, v18.8b, #8             \n" /* R                    */ \
+  "shll       v17.8h, v17.8b, #8             \n" /* G                    */ \
+  "shll       v16.8h, v16.8b, #8             \n" /* B                    */ \
+  "sri        v18.8h, v17.8h, #5             \n" /* RG                   */ \
+  "sri        v18.8h, v16.8h, #11            \n" /* RGB                  */
 
 void I422ToRGB565Row_NEON(const uint8_t* src_y,
                           const uint8_t* src_u,
@@ -273,33 +280,28 @@ void I422ToRGB565Row_NEON(const uint8_t* src_y,
                           int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "1:                                        \n" READYUV422 YUVTORGB(
-          v22, v21,
-          v20) "subs       %w4, %w4, #8                   \n" ARGBTORGB565
-               "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels
-                                                               // RGB565.
-               "b.gt       1b                             \n"
-      : "+r"(src_y),       // %0
-        "+r"(src_u),       // %1
-        "+r"(src_v),       // %2
-        "+r"(dst_rgb565),  // %3
-        "+r"(width)        // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-        "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
+      "1:                                        \n" READYUV422 YUVTORGB
+      RGBTORGB8 "subs        %w[width], %w[width], #8      \n" ARGBTORGB565
+      "st1         {v18.8h}, [%[dst_rgb565]], #16 \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_rgb565] "+r"(dst_rgb565),                     // %[dst_rgb565]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS);
 }
 
 #define ARGBTOARGB1555                                                      \
-  "shll       v0.8h,  v23.8b, #8             \n" /* A                    */ \
-  "shll       v22.8h, v22.8b, #8             \n" /* R                    */ \
-  "shll       v21.8h, v21.8b, #8             \n" /* G                    */ \
-  "shll       v20.8h, v20.8b, #8             \n" /* B                    */ \
-  "sri        v0.8h,  v22.8h, #1             \n" /* AR                   */ \
-  "sri        v0.8h,  v21.8h, #6             \n" /* ARG                  */ \
-  "sri        v0.8h,  v20.8h, #11            \n" /* ARGB                 */
+  "shll       v0.8h,  v19.8b, #8             \n" /* A                    */ \
+  "shll       v18.8h, v18.8b, #8             \n" /* R                    */ \
+  "shll       v17.8h, v17.8b, #8             \n" /* G                    */ \
+  "shll       v16.8h, v16.8b, #8             \n" /* B                    */ \
+  "sri        v0.8h,  v18.8h, #1             \n" /* AR                   */ \
+  "sri        v0.8h,  v17.8h, #6             \n" /* ARG                  */ \
+  "sri        v0.8h,  v16.8h, #11            \n" /* ARGB                 */
 
 void I422ToARGB1555Row_NEON(const uint8_t* src_y,
                             const uint8_t* src_u,
@@ -309,34 +311,31 @@ void I422ToARGB1555Row_NEON(const uint8_t* src_y,
                             int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "movi       v23.8b, #255                   \n"
-      "1:                                        \n" READYUV422 YUVTORGB(
-          v22, v21,
-          v20) "subs       %w4, %w4, #8                   \n" ARGBTOARGB1555
-               "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels
-                                                               // RGB565.
-               "b.gt       1b                             \n"
-      : "+r"(src_y),         // %0
-        "+r"(src_u),         // %1
-        "+r"(src_v),         // %2
-        "+r"(dst_argb1555),  // %3
-        "+r"(width)          // %4
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-        "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
+      "movi        v19.8b, #255                  \n"
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n" ARGBTOARGB1555
+      "st1         {v0.8h}, [%[dst_argb1555]], #16 \n"  // store 8 pixels
+                                                        // RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb1555] "+r"(dst_argb1555),                 // %[dst_argb1555]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
 #define ARGBTOARGB4444                                                       \
-  /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f        */ \
-  "ushr       v20.8b, v20.8b, #4             \n" /* B                    */  \
-  "bic        v21.8b, v21.8b, v4.8b          \n" /* G                    */  \
-  "ushr       v22.8b, v22.8b, #4             \n" /* R                    */  \
-  "bic        v23.8b, v23.8b, v4.8b          \n" /* A                    */  \
-  "orr        v0.8b,  v20.8b, v21.8b         \n" /* BG                   */  \
-  "orr        v1.8b,  v22.8b, v23.8b         \n" /* RA                   */  \
+  /* Input v16.8b<=B, v17.8b<=G, v18.8b<=R, v19.8b<=A, v23.8b<=0x0f       */ \
+  "ushr       v16.8b, v16.8b, #4             \n" /* B                    */  \
+  "bic        v17.8b, v17.8b, v23.8b         \n" /* G                    */  \
+  "ushr       v18.8b, v18.8b, #4             \n" /* R                    */  \
+  "bic        v19.8b, v19.8b, v23.8b         \n" /* A                    */  \
+  "orr        v0.8b,  v16.8b, v17.8b         \n" /* BG                   */  \
+  "orr        v1.8b,  v18.8b, v19.8b         \n" /* RA                   */  \
   "zip1       v0.16b, v0.16b, v1.16b         \n" /* BGRA                 */
 
 void I422ToARGB4444Row_NEON(const uint8_t* src_y,
@@ -345,95 +344,109 @@ void I422ToARGB4444Row_NEON(const uint8_t* src_y,
                             uint8_t* dst_argb4444,
                             const struct YuvConstants* yuvconstants,
                             int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
-  "1:                                          \n"
-    READYUV422
-    YUVTORGB(v22, v21, v20)
-    "subs       %w4, %w4, #8                   \n"
-    "movi       v23.8b, #255                   \n"
-    ARGBTOARGB4444
-    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels ARGB4444.
-    "b.gt       1b                             \n"
-    : "+r"(src_y),    // %0
-      "+r"(src_u),    // %1
-      "+r"(src_v),    // %2
-      "+r"(dst_argb4444),  // %3
-      "+r"(width)     // %4
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v23.16b, #0x0f                \n"  // bits to clear with
+                                                      // vbic.
+      "1:                                        \n" READYUV422 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "movi        v19.8b, #255                  \n" ARGBTOARGB4444
+      "st1         {v0.8h}, [%[dst_argb4444]], #16 \n"  // store 8
+                                                        // pixels
+                                                        // ARGB4444.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [src_u] "+r"(src_u),                               // %[src_u]
+        [src_v] "+r"(src_v),                               // %[src_v]
+        [dst_argb4444] "+r"(dst_argb4444),                 // %[dst_argb4444]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19", "v23");
 }
 
-void I400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READYUV400
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&kYuvI601Constants.kUVToRB),
-      [kUVToG]"r"(&kYuvI601Constants.kUVToG),
-      [kUVBiasBGR]"r"(&kYuvI601Constants.kUVBiasBGR),
-      [kYToRgb]"r"(&kYuvI601Constants.kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+void I400ToARGBRow_NEON(const uint8_t* src_y,
+                        uint8_t* dst_argb,
+                        const struct YuvConstants* yuvconstants,
+                        int width) {
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "1:                                        \n" READYUV400 YUVTORGB
+          RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                               // %[src_y]
+        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
+        [width] "+r"(width)                                // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
+      : "cc", "memory", YUVTORGB_REGS, "v19");
 }
 
+#if LIBYUV_USE_ST4
 void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
   asm volatile(
-      "movi       v23.8b, #255                   \n"
+      "movi        v23.8b, #255                  \n"
       "1:                                        \n"
-      "ld1        {v20.8b}, [%0], #8             \n"
-      "orr        v21.8b, v20.8b, v20.8b         \n"
-      "orr        v22.8b, v20.8b, v20.8b         \n"
-      "subs       %w2, %w2, #8                   \n"
-      "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
-      "b.gt       1b                             \n"
+      "ld1         {v20.8b}, [%0], #8            \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v21.8b, v20.8b, v20.8b        \n"
+      "orr         v22.8b, v20.8b, v20.8b        \n"
+      "subs        %w2, %w2, #8                  \n"
+      "st4         {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"
+      "b.gt        1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
       :
       : "cc", "memory", "v20", "v21", "v22", "v23");
 }
+#else
+void J400ToARGBRow_NEON(const uint8_t* src_y, uint8_t* dst_argb, int width) {
+  asm volatile(
+      "movi        v20.8b, #255                  \n"
+      "1:                                        \n"
+      "ldr         d16, [%0], #8                 \n"
+      "subs        %w2, %w2, #8                  \n"
+      "zip1        v18.16b, v16.16b, v16.16b     \n"  // YY
+      "zip1        v19.16b, v16.16b, v20.16b     \n"  // YA
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v16.16b, v18.16b, v19.16b     \n"  // YYYA
+      "zip2        v17.16b, v18.16b, v19.16b     \n"
+      "stp         q16, q17, [%1], #32           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_y),     // %0
+        "+r"(dst_argb),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v16", "v17", "v18", "v19", "v20");
+}
+#endif  // LIBYUV_USE_ST4
 
 void NV12ToARGBRow_NEON(const uint8_t* src_y,
                         const uint8_t* src_uv,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READNV12
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
 void NV21ToARGBRow_NEON(const uint8_t* src_y,
@@ -441,26 +454,22 @@ void NV21ToARGBRow_NEON(const uint8_t* src_y,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READNV21
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_vu),    // %1
-      "+r"(dst_argb),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_vu),                              // %[src_uv]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV21Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
 void NV12ToRGB24Row_NEON(const uint8_t* src_y,
@@ -468,25 +477,21 @@ void NV12ToRGB24Row_NEON(const uint8_t* src_y,
                          uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READNV12
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    "st3        {v20.8b,v21.8b,v22.8b}, [%2], #24     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_uv),    // %1
-      "+r"(dst_rgb24),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
 }
 
 void NV21ToRGB24Row_NEON(const uint8_t* src_y,
@@ -494,25 +499,21 @@ void NV21ToRGB24Row_NEON(const uint8_t* src_y,
                          uint8_t* dst_rgb24,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-  "1:                                          \n"
-    READNV21
-    YUVTORGB(v22, v21, v20)
-    "subs       %w3, %w3, #8                   \n"
-    "st3        {v20.8b,v21.8b,v22.8b}, [%2], #24     \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_y),     // %0
-      "+r"(src_vu),    // %1
-      "+r"(dst_rgb24),  // %2
-      "+r"(width)      // %3
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st3         {v16.8b,v17.8b,v18.8b}, [%[dst_rgb24]], #24 \n"
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_vu),                              // %[src_uv]
+        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV21Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
 }
 
 void NV12ToRGB565Row_NEON(const uint8_t* src_y,
@@ -522,72 +523,63 @@ void NV12ToRGB565Row_NEON(const uint8_t* src_y,
                           int width) {
   asm volatile(
       YUVTORGB_SETUP
-      "1:                                        \n" READNV12 YUVTORGB(
-          v22, v21,
-          v20) "subs       %w3, %w3, #8                   \n" ARGBTORGB565
-               "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels
-                                                               // RGB565.
-               "b.gt       1b                             \n"
-      : "+r"(src_y),       // %0
-        "+r"(src_uv),      // %1
-        "+r"(dst_rgb565),  // %2
-        "+r"(width)        // %3
-      : [kUVToRB] "r"(&yuvconstants->kUVToRB),
-        [kUVToG] "r"(&yuvconstants->kUVToG),
-        [kUVBiasBGR] "r"(&yuvconstants->kUVBiasBGR),
-        [kYToRgb] "r"(&yuvconstants->kYToRgb)
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-        "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30");
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READNV12 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n" ARGBTORGB565
+      "st1         {v18.8h}, [%[dst_rgb565]], #16 \n"  // store 8
+                                                       // pixels
+                                                       // RGB565.
+      "b.gt        1b                            \n"
+      : [src_y] "+r"(src_y),                                // %[src_y]
+        [src_uv] "+r"(src_uv),                              // %[src_uv]
+        [dst_rgb565] "+r"(dst_rgb565),                      // %[dst_rgb565]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2");
 }
 
 void YUY2ToARGBRow_NEON(const uint8_t* src_yuy2,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READYUY2
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32      \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_yuy2),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READYUY2 YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_yuy2] "+r"(src_yuy2),                          // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
 void UYVYToARGBRow_NEON(const uint8_t* src_uyvy,
                         uint8_t* dst_argb,
                         const struct YuvConstants* yuvconstants,
                         int width) {
-  asm volatile (
-    YUVTORGB_SETUP
-    "movi       v23.8b, #255                   \n"
-  "1:                                          \n"
-    READUYVY
-    YUVTORGB(v22, v21, v20)
-    "subs       %w2, %w2, #8                   \n"
-    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32      \n"
-    "b.gt       1b                             \n"
-    : "+r"(src_uyvy),  // %0
-      "+r"(dst_argb),  // %1
-      "+r"(width)      // %2
-    : [kUVToRB]"r"(&yuvconstants->kUVToRB),
-      [kUVToG]"r"(&yuvconstants->kUVToG),
-      [kUVBiasBGR]"r"(&yuvconstants->kUVBiasBGR),
-      [kYToRgb]"r"(&yuvconstants->kYToRgb)
-    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
-      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
-  );
+  asm volatile(
+      YUVTORGB_SETUP
+      "movi        v19.8b, #255                  \n"
+      "ldr         q2, [%[kNV12Table]]           \n"
+      "1:                                        \n" READUYVY YUVTORGB RGBTORGB8
+      "subs        %w[width], %w[width], #8      \n"
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%[dst_argb]], #32 \n"
+      "b.gt        1b                            \n"
+      : [src_uyvy] "+r"(src_uyvy),                          // %[src_yuy2]
+        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
+        [width] "+r"(width)                                 // %[width]
+      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
+        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
+        [kNV12Table] "r"(&kNV12Table)
+      : "cc", "memory", YUVTORGB_REGS, "v2", "v19");
 }
 
 // Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
@@ -597,11 +589,12 @@ void SplitUVRow_NEON(const uint8_t* src_uv,
                      int width) {
   asm volatile(
       "1:                                        \n"
-      "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pairs of UV
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-      "st1        {v0.16b}, [%1], #16            \n"  // store U
-      "st1        {v1.16b}, [%2], #16            \n"  // store V
-      "b.gt       1b                             \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pairs of UV
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store U
+      "st1         {v1.16b}, [%2], #16           \n"  // store V
+      "b.gt        1b                            \n"
       : "+r"(src_uv),               // %0
         "+r"(dst_u),                // %1
         "+r"(dst_v),                // %2
@@ -611,6 +604,53 @@ void SplitUVRow_NEON(const uint8_t* src_uv,
   );
 }
 
+// Reads 16 byte Y's from tile and writes out 16 Y's.
+// MM21 Y tiles are 16x32 so src_tile_stride = 512 bytes
+// MM21 UV tiles are 8x16 so src_tile_stride = 256 bytes
+// width measured in bytes so 8 UV = 16.
+void DetileRow_NEON(const uint8_t* src,
+                    ptrdiff_t src_tile_stride,
+                    uint8_t* dst,
+                    int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], %3            \n"  // load 16 bytes
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 1792]         \n"  // 7 tiles of 256b ahead
+      "st1         {v0.16b}, [%1], #16           \n"  // store 16 bytes
+      "b.gt        1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(src_tile_stride)  // %3
+      : "cc", "memory", "v0"  // Clobber List
+  );
+}
+
+// Read 16 bytes of UV, detile, and write 8 bytes of U and 8 bytes of V.
+void DetileSplitUVRow_NEON(const uint8_t* src_uv,
+                           ptrdiff_t src_tile_stride,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld2         {v0.8b,v1.8b}, [%0], %4       \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%0, 1792]         \n"
+      "st1         {v0.8b}, [%1], #8             \n"
+      "st1         {v1.8b}, [%2], #8             \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),               // %0
+        "+r"(dst_u),                // %1
+        "+r"(dst_v),                // %2
+        "+r"(width)                 // %3
+      : "r"(src_tile_stride)        // %4
+      : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+
+#if LIBYUV_USE_ST2
 // Reads 16 U's and V's and writes out 16 pairs of UV.
 void MergeUVRow_NEON(const uint8_t* src_u,
                      const uint8_t* src_v,
@@ -618,11 +658,13 @@ void MergeUVRow_NEON(const uint8_t* src_u,
                      int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load U
-      "ld1        {v1.16b}, [%1], #16            \n"  // load V
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-      "st2        {v0.16b,v1.16b}, [%2], #32     \n"  // store 16 pairs of UV
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load U
+      "ld1         {v1.16b}, [%1], #16           \n"  // load V
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st2         {v0.16b,v1.16b}, [%2], #32    \n"  // store 16 pairs of UV
+      "b.gt        1b                            \n"
       : "+r"(src_u),                // %0
         "+r"(src_v),                // %1
         "+r"(dst_uv),               // %2
@@ -632,6 +674,86 @@ void MergeUVRow_NEON(const uint8_t* src_u,
   );
 }
 
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "dup         v2.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // load 8 U
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ld1         {v1.8h}, [%1], #16            \n"  // load 8 V
+      "ushl        v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st2         {v0.8h, v1.8h}, [%2], #32     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2");
+}
+#else
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8_t* src_u,
+                     const uint8_t* src_v,
+                     uint8_t* dst_uv,
+                     int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load U
+      "ld1         {v1.16b}, [%1], #16           \n"  // load V
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "zip1        v2.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip2        v3.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v2.16b,v3.16b}, [%2], #32    \n"  // store 16 pairs of UV
+      "b.gt        1b                            \n"
+      : "+r"(src_u),                            // %0
+        "+r"(src_v),                            // %1
+        "+r"(dst_uv),                           // %2
+        "+r"(width)                             // %3  // Output registers
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+void MergeUVRow_16_NEON(const uint16_t* src_u,
+                        const uint16_t* src_v,
+                        uint16_t* dst_uv,
+                        int depth,
+                        int width) {
+  int shift = 16 - depth;
+  asm volatile(
+      "dup         v4.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // load 8 U
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ld1         {v1.8h}, [%1], #16            \n"  // load 8 V
+      "ushl        v0.8h, v0.8h, v4.8h           \n"
+      "ushl        v1.8h, v1.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v2.8h, v0.8h, v1.8h           \n"
+      "zip2        v3.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v2.8h, v3.8h}, [%2], #32     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),   // %0
+        "+r"(src_v),   // %1
+        "+r"(dst_uv),  // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2", "v1", "v2", "v3", "v4");
+}
+#endif  // LIBYUV_USE_ST2
+
 // Reads 16 packed RGB and write to planar dst_r, dst_g, dst_b.
 void SplitRGBRow_NEON(const uint8_t* src_rgb,
                       uint8_t* dst_r,
@@ -640,12 +762,13 @@ void SplitRGBRow_NEON(const uint8_t* src_rgb,
                       int width) {
   asm volatile(
       "1:                                        \n"
-      "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 RGB
-      "subs       %w4, %w4, #16                  \n"  // 16 processed per loop
-      "st1        {v0.16b}, [%1], #16            \n"  // store R
-      "st1        {v1.16b}, [%2], #16            \n"  // store G
-      "st1        {v2.16b}, [%3], #16            \n"  // store B
-      "b.gt       1b                             \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 RGB
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store R
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%3], #16           \n"  // store B
+      "b.gt        1b                            \n"
       : "+r"(src_rgb),                    // %0
         "+r"(dst_r),                      // %1
         "+r"(dst_g),                      // %2
@@ -664,12 +787,15 @@ void MergeRGBRow_NEON(const uint8_t* src_r,
                       int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load R
-      "ld1        {v1.16b}, [%1], #16            \n"  // load G
-      "ld1        {v2.16b}, [%2], #16            \n"  // load B
-      "subs       %w4, %w4, #16                  \n"  // 16 processed per loop
-      "st3        {v0.16b,v1.16b,v2.16b}, [%3], #48 \n"  // store 16 RGB
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load R
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%2], #16           \n"  // load B
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st3         {v0.16b,v1.16b,v2.16b}, [%3], #48 \n"  // store 16 RGB
+      "b.gt        1b                            \n"
       : "+r"(src_r),                      // %0
         "+r"(src_g),                      // %1
         "+r"(src_b),                      // %2
@@ -680,14 +806,403 @@ void MergeRGBRow_NEON(const uint8_t* src_r,
   );
 }
 
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b, dst_a.
+void SplitARGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       uint8_t* dst_a,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%3], #16           \n"  // store B
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%1], #16           \n"  // store R
+      "st1         {v3.16b}, [%4], #16           \n"  // store A
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(dst_a),                            // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+#if LIBYUV_USE_ST4
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v3.16b}, [%3], #16           \n"  // load A
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "st4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%4], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(src_a),                            // %3
+        "+r"(dst_argb),                         // %4
+        "+r"(width)                             // %5
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#else
+// Reads 16 planar R's, G's, B's and A's and writes out 16 packed ARGB at a time
+void MergeARGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       const uint8_t* src_a,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v3.16b}, [%3], #16           \n"  // load A
+      "subs        %w5, %w5, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "zip1        v4.16b, v0.16b, v1.16b        \n"  // BG
+      "zip1        v5.16b, v2.16b, v3.16b        \n"  // RA
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "zip2        v6.16b, v0.16b, v1.16b        \n"  // BG
+      "zip2        v7.16b, v2.16b, v3.16b        \n"  // RA
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v0.8h, v4.8h, v5.8h           \n"  // BGRA
+      "zip2        v1.8h, v4.8h, v5.8h           \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "zip1        v2.8h, v6.8h, v7.8h           \n"
+      "zip2        v3.8h, v6.8h, v7.8h           \n"
+      "st1         {v0.16b,v1.16b,v2.16b,v3.16b}, [%4], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      :                  // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+        "v7"  // Clobber List
+  );
+}
+#endif  // LIBYUV_USE_ST4
+
+// Reads 16 packed ARGB and write to planar dst_r, dst_g, dst_b.
+void SplitXRGBRow_NEON(const uint8_t* src_rgba,
+                       uint8_t* dst_r,
+                       uint8_t* dst_g,
+                       uint8_t* dst_b,
+                       int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%3], #16           \n"  // store B
+      "st1         {v1.16b}, [%2], #16           \n"  // store G
+      "st1         {v2.16b}, [%1], #16           \n"  // store R
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),                         // %0
+        "+r"(dst_r),                            // %1
+        "+r"(dst_g),                            // %2
+        "+r"(dst_b),                            // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+// Reads 16 planar R's, G's and B's and writes out 16 packed ARGB at a time
+void MergeXRGBRow_NEON(const uint8_t* src_r,
+                       const uint8_t* src_g,
+                       const uint8_t* src_b,
+                       uint8_t* dst_argb,
+                       int width) {
+  asm volatile(
+      "movi        v3.16b, #255                  \n"  // load A(255)
+      "1:                                        \n"
+      "ld1         {v2.16b}, [%0], #16           \n"  // load R
+      "ld1         {v1.16b}, [%1], #16           \n"  // load G
+      "ld1         {v0.16b}, [%2], #16           \n"  // load B
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%3], #64 \n"  // store 16ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_r),                            // %0
+        "+r"(src_g),                            // %1
+        "+r"(src_b),                            // %2
+        "+r"(dst_argb),                         // %3
+        "+r"(width)                             // %4
+      :                                         // Input registers
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+void MergeXR30Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint8_t* dst_ar30,
+                       int depth,
+                       int width) {
+  int shift = 10 - depth;
+  asm volatile(
+      "movi        v30.16b, #255                 \n"
+      "ushr        v30.4s, v30.4s, #22           \n"  // 1023
+      "dup         v31.4s, %w5                   \n"
+      "1:                                        \n"
+      "ldr         d2, [%2], #8                  \n"  // B
+      "ldr         d1, [%1], #8                  \n"  // G
+      "ldr         d0, [%0], #8                  \n"  // R
+      "ushll       v2.4s, v2.4h, #0              \n"  // B
+      "ushll       v1.4s, v1.4h, #0              \n"  // G
+      "ushll       v0.4s, v0.4h, #0              \n"  // R
+      "ushl        v2.4s, v2.4s, v31.4s          \n"  // 000B
+      "ushl        v1.4s, v1.4s, v31.4s          \n"  // G
+      "ushl        v0.4s, v0.4s, v31.4s          \n"  // R
+      "umin        v2.4s, v2.4s, v30.4s          \n"
+      "umin        v1.4s, v1.4s, v30.4s          \n"
+      "umin        v0.4s, v0.4s, v30.4s          \n"
+      "sli         v2.4s, v1.4s, #10             \n"  // 00GB
+      "sli         v2.4s, v0.4s, #20             \n"  // 0RGB
+      "orr         v2.4s, #0xc0, lsl #24         \n"  // ARGB (AR30)
+      "subs        %w4, %w4, #4                  \n"
+      "str         q2, [%3], #16                 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "v0", "v1", "v2", "v30", "v31");
+}
+
+void MergeXR30Row_10_NEON(const uint16_t* src_r,
+                          const uint16_t* src_g,
+                          const uint16_t* src_b,
+                          uint8_t* dst_ar30,
+                          int /* depth */,
+                          int width) {
+  asm volatile(
+      "movi        v30.16b, #255                 \n"
+      "ushr        v30.4s, v30.4s, #22           \n"  // 1023
+      "1:                                        \n"
+      "ldr         d2, [%2], #8                  \n"  // B
+      "ldr         d1, [%1], #8                  \n"  // G
+      "ldr         d0, [%0], #8                  \n"  // R
+      "ushll       v2.4s, v2.4h, #0              \n"  // 000B
+      "ushll       v1.4s, v1.4h, #0              \n"  // G
+      "ushll       v0.4s, v0.4h, #0              \n"  // R
+      "umin        v2.4s, v2.4s, v30.4s          \n"
+      "umin        v1.4s, v1.4s, v30.4s          \n"
+      "umin        v0.4s, v0.4s, v30.4s          \n"
+      "sli         v2.4s, v1.4s, #10             \n"  // 00GB
+      "sli         v2.4s, v0.4s, #20             \n"  // 0RGB
+      "orr         v2.4s, #0xc0, lsl #24         \n"  // ARGB (AR30)
+      "subs        %w4, %w4, #4                  \n"
+      "str         q2, [%3], #16                 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar30),  // %3
+        "+r"(width)      // %4
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v30");
+}
+
+void MergeAR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       const uint16_t* src_a,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "dup         v30.8h, %w7                   \n"
+      "dup         v31.8h, %w6                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ldr         q3, [%3], #16                 \n"  // A
+      "umin        v2.8h, v2.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umin        v1.8h, v1.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umin        v0.8h, v0.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umin        v3.8h, v3.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "ushl        v3.8h, v3.8h, v31.8h          \n"
+      "subs        %w5, %w5, #8                  \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%4], #64 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_ar64),  // %4
+        "+r"(width)      // %5
+      : "r"(shift),      // %6
+        "r"(mask)        // %7
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeXR64Row_NEON(const uint16_t* src_r,
+                       const uint16_t* src_g,
+                       const uint16_t* src_b,
+                       uint16_t* dst_ar64,
+                       int depth,
+                       int width) {
+  int shift = 16 - depth;
+  int mask = (1 << depth) - 1;
+  asm volatile(
+
+      "movi        v3.16b, #0xff                 \n"  // A (0xffff)
+      "dup         v30.8h, %w6                   \n"
+      "dup         v31.8h, %w5                   \n"
+
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "umin        v2.8h, v2.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umin        v1.8h, v1.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umin        v0.8h, v0.8h, v30.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "subs        %w4, %w4, #8                  \n"
+      "st4         {v0.8h, v1.8h, v2.8h, v3.8h}, [%3], #64 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_ar64),  // %3
+        "+r"(width)      // %4
+      : "r"(shift),      // %5
+        "r"(mask)        // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeARGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            const uint16_t* src_a,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "dup         v31.8h, %w6                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ldr         q3, [%3], #16                 \n"  // A
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v3.8h, v3.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn       v3.8b, v3.8h                  \n"
+      "subs        %w5, %w5, #8                  \n"
+      "st4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%4], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(src_a),     // %3
+        "+r"(dst_argb),  // %4
+        "+r"(width)      // %5
+      : "r"(shift)       // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
+void MergeXRGB16To8Row_NEON(const uint16_t* src_r,
+                            const uint16_t* src_g,
+                            const uint16_t* src_b,
+                            uint8_t* dst_argb,
+                            int depth,
+                            int width) {
+  int shift = 8 - depth;
+  asm volatile(
+
+      "dup         v31.8h, %w5                   \n"
+      "movi        v3.8b, #0xff                  \n"  // A (0xff)
+      "1:                                        \n"
+      "ldr         q2, [%0], #16                 \n"  // R
+      "ldr         q1, [%1], #16                 \n"  // G
+      "ldr         q0, [%2], #16                 \n"  // B
+      "ushl        v2.8h, v2.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "ushl        v0.8h, v0.8h, v31.8h          \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "subs        %w4, %w4, #8                  \n"
+      "st4         {v0.8b, v1.8b, v2.8b, v3.8b}, [%3], #32 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_r),     // %0
+        "+r"(src_g),     // %1
+        "+r"(src_b),     // %2
+        "+r"(dst_argb),  // %3
+        "+r"(width)      // %4
+      : "r"(shift)       // %5
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31");
+}
+
 // Copy multiple of 32.
 void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       "1:                                        \n"
-      "ldp        q0, q1, [%0], #32              \n"
-      "subs       %w2, %w2, #32                  \n"  // 32 processed per loop
-      "stp        q0, q1, [%1], #32              \n"
-      "b.gt       1b                             \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #32                 \n"  // 32 processed per loop
+      "stp         q0, q1, [%1], #32             \n"
+      "b.gt        1b                            \n"
       : "+r"(src),                  // %0
         "+r"(dst),                  // %1
         "+r"(width)                 // %2  // Output registers
@@ -699,11 +1214,11 @@ void CopyRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
 // SetRow writes 'width' bytes using an 8 bit value repeated.
 void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
   asm volatile(
-      "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
+      "dup         v0.16b, %w2                   \n"  // duplicate 16 bytes
       "1:                                        \n"
-      "subs       %w1, %w1, #16                  \n"  // 16 bytes per loop
-      "st1        {v0.16b}, [%0], #16            \n"  // store
-      "b.gt       1b                             \n"
+      "subs        %w1, %w1, #16                 \n"  // 16 bytes per loop
+      "st1         {v0.16b}, [%0], #16           \n"  // store
+      "b.gt        1b                            \n"
       : "+r"(dst),   // %0
         "+r"(width)  // %1
       : "r"(v8)      // %2
@@ -712,89 +1227,157 @@ void SetRow_NEON(uint8_t* dst, uint8_t v8, int width) {
 
 void ARGBSetRow_NEON(uint8_t* dst, uint32_t v32, int width) {
   asm volatile(
-      "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+      "dup         v0.4s, %w2                    \n"  // duplicate 4 ints
       "1:                                        \n"
-      "subs       %w1, %w1, #4                   \n"  // 4 ints per loop
-      "st1        {v0.16b}, [%0], #16            \n"  // store
-      "b.gt       1b                             \n"
+      "subs        %w1, %w1, #4                  \n"  // 4 ints per loop
+      "st1         {v0.16b}, [%0], #16           \n"  // store
+      "b.gt        1b                            \n"
       : "+r"(dst),   // %0
         "+r"(width)  // %1
       : "r"(v32)     // %2
       : "cc", "memory", "v0");
 }
 
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u,
+                                     7u,  6u,  5u,  4u,  3u,  2u,  1u, 0u};
+
 void MirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
   asm volatile(
       // Start at end of source row.
-      "add        %0, %0, %w2, sxtw              \n"
-      "sub        %0, %0, #16                    \n"
-      "1:                                        \n"
-      "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
-      "subs       %w2, %w2, #16                  \n"  // 16 pixels per loop.
-      "rev64      v0.16b, v0.16b                 \n"
-      "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
-      "st1        {v0.D}[0], [%1], #8            \n"
-      "b.gt       1b                             \n"
-      : "+r"(src),           // %0
-        "+r"(dst),           // %1
-        "+r"(width)          // %2
-      : "r"((ptrdiff_t)-16)  // %3
-      : "cc", "memory", "v0");
+      "ld1         {v3.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw             \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q2, [%0, 16]                  \n"
+      "ldr         q1, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #32                 \n"  // 32 pixels per loop.
+      "tbl         v0.16b, {v2.16b}, v3.16b      \n"
+      "tbl         v1.16b, {v1.16b}, v3.16b      \n"
+      "st1         {v0.16b, v1.16b}, [%1], #32   \n"  // store 32 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src),            // %0
+        "+r"(dst),            // %1
+        "+r"(width)           // %2
+      : "r"(&kShuffleMirror)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void MirrorUVRow_NEON(const uint8_t* src_uv,
-                      uint8_t* dst_u,
-                      uint8_t* dst_v,
-                      int width) {
+// Shuffle table for reversing the UV.
+static const uvec8 kShuffleMirrorUV = {14u, 15u, 12u, 13u, 10u, 11u, 8u, 9u,
+                                       6u,  7u,  4u,  5u,  2u,  3u,  0u, 1u};
+
+void MirrorUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
   asm volatile(
       // Start at end of source row.
-      "add        %0, %0, %w3, sxtw #1           \n"
-      "sub        %0, %0, #16                    \n"
-      "1:                                        \n"
-      "ld2        {v0.8b, v1.8b}, [%0], %4       \n"  // src -= 16
-      "subs       %w3, %w3, #8                   \n"  // 8 pixels per loop.
-      "rev64      v0.8b, v0.8b                   \n"
-      "rev64      v1.8b, v1.8b                   \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // dst += 8
-      "st1        {v1.8b}, [%2], #8              \n"
-      "b.gt       1b                             \n"
-      : "+r"(src_uv),        // %0
-        "+r"(dst_u),         // %1
-        "+r"(dst_v),         // %2
-        "+r"(width)          // %3
-      : "r"((ptrdiff_t)-16)  // %4
-      : "cc", "memory", "v0", "v1");
+      "ld1         {v4.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #1          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // dst += 32
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),           // %0
+        "+r"(dst_uv),           // %1
+        "+r"(width)             // %2
+      : "r"(&kShuffleMirrorUV)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
 }
 
-void ARGBMirrorRow_NEON(const uint8_t* src, uint8_t* dst, int width) {
+void MirrorSplitUVRow_NEON(const uint8_t* src_uv,
+                           uint8_t* dst_u,
+                           uint8_t* dst_v,
+                           int width) {
   asm volatile(
       // Start at end of source row.
-      "add        %0, %0, %w2, sxtw #2           \n"
-      "sub        %0, %0, #16                    \n"
-      "1:                                        \n"
-      "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
-      "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
-      "rev64      v0.4s, v0.4s                   \n"
-      "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
-      "st1        {v0.D}[0], [%1], #8            \n"
-      "b.gt       1b                             \n"
-      : "+r"(src),           // %0
-        "+r"(dst),           // %1
-        "+r"(width)          // %2
-      : "r"((ptrdiff_t)-16)  // %3
-      : "cc", "memory", "v0");
+      "ld1         {v4.16b}, [%4]                \n"  // shuffler
+      "add         %0, %0, %w3, sxtw #1          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "uzp1        v0.16b, v2.16b, v3.16b        \n"  // U
+      "uzp2        v1.16b, v2.16b, v3.16b        \n"  // V
+      "st1         {v0.16b}, [%1], #16           \n"  // dst += 16
+      "st1         {v1.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),           // %0
+        "+r"(dst_u),            // %1
+        "+r"(dst_v),            // %2
+        "+r"(width)             // %3
+      : "r"(&kShuffleMirrorUV)  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+// Shuffle table for reversing the ARGB.
+static const uvec8 kShuffleMirrorARGB = {12u, 13u, 14u, 15u, 8u, 9u, 10u, 11u,
+                                         4u,  5u,  6u,  7u,  0u, 1u, 2u,  3u};
+
+void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
+  asm volatile(
+      // Start at end of source row.
+      "ld1         {v4.16b}, [%3]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #2          \n"
+      "sub         %0, %0, #32                   \n"
+      "1:                                        \n"
+      "ldr         q1, [%0, 16]                  \n"
+      "ldr         q0, [%0], -32                 \n"  // src -= 32
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop.
+      "tbl         v2.16b, {v1.16b}, v4.16b      \n"
+      "tbl         v3.16b, {v0.16b}, v4.16b      \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // dst += 32
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleMirrorARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+void RGB24MirrorRow_NEON(const uint8_t* src_rgb24,
+                         uint8_t* dst_rgb24,
+                         int width) {
+  asm volatile(
+      "ld1         {v3.16b}, [%4]                \n"  // shuffler
+      "add         %0, %0, %w2, sxtw #1          \n"  // Start at end of row.
+      "add         %0, %0, %w2, sxtw             \n"
+      "sub         %0, %0, #48                   \n"
+
+      "1:                                        \n"
+      "ld3         {v0.16b, v1.16b, v2.16b}, [%0], %3 \n"  // src -= 48
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "tbl         v0.16b, {v0.16b}, v3.16b      \n"
+      "tbl         v1.16b, {v1.16b}, v3.16b      \n"
+      "tbl         v2.16b, {v2.16b}, v3.16b      \n"
+      "st3         {v0.16b, v1.16b, v2.16b}, [%1], #48 \n"  // dst += 48
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb24),      // %0
+        "+r"(dst_rgb24),      // %1
+        "+r"(width)           // %2
+      : "r"((ptrdiff_t)-48),  // %3
+        "r"(&kShuffleMirror)  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
 void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
                          uint8_t* dst_argb,
                          int width) {
   asm volatile(
-      "movi       v4.8b, #255                    \n"  // Alpha
+      "movi        v4.8b, #255                   \n"  // Alpha
       "1:                                        \n"
-      "ld3        {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of RGB24.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "st4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld3         {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of
+                                                       // RGB24.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st4         {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_rgb24),  // %0
         "+r"(dst_argb),   // %1
         "+r"(width)       // %2
@@ -805,14 +1388,15 @@ void RGB24ToARGBRow_NEON(const uint8_t* src_rgb24,
 
 void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   asm volatile(
-      "movi       v5.8b, #255                    \n"  // Alpha
+      "movi        v5.8b, #255                   \n"  // Alpha
       "1:                                        \n"
-      "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
-      "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
-      "st4        {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
-      "b.gt       1b                             \n"
+      "ld3         {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v3.8b, v1.8b, v1.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v4.8b, v0.8b, v0.8b           \n"         // move r
+      "st4         {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
+      "b.gt        1b                            \n"
       : "+r"(src_raw),   // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -821,15 +1405,35 @@ void RAWToARGBRow_NEON(const uint8_t* src_raw, uint8_t* dst_argb, int width) {
   );
 }
 
+void RAWToRGBARow_NEON(const uint8_t* src_raw, uint8_t* dst_rgba, int width) {
+  asm volatile(
+      "movi        v0.8b, #255                   \n"  // Alpha
+      "1:                                        \n"
+      "ld3         {v3.8b,v4.8b,v5.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v2.8b, v4.8b, v4.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v1.8b, v5.8b, v5.8b           \n"         // move r
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store a b g r
+      "b.gt        1b                            \n"
+      : "+r"(src_raw),   // %0
+        "+r"(dst_rgba),  // %1
+        "+r"(width)      // %2
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+
 void RAWToRGB24Row_NEON(const uint8_t* src_raw, uint8_t* dst_rgb24, int width) {
   asm volatile(
       "1:                                        \n"
-      "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
-      "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
-      "st3        {v2.8b,v3.8b,v4.8b}, [%1], #24 \n"  // store b g r
-      "b.gt       1b                             \n"
+      "ld3         {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+      "subs        %w2, %w2, #8                  \n"   // 8 processed per loop.
+      "orr         v3.8b, v1.8b, v1.8b           \n"   // move g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v4.8b, v0.8b, v0.8b           \n"   // move r
+      "st3         {v2.8b,v3.8b,v4.8b}, [%1], #24 \n"  // store b g r
+      "b.gt        1b                            \n"
       : "+r"(src_raw),    // %0
         "+r"(dst_rgb24),  // %1
         "+r"(width)       // %2
@@ -855,13 +1459,13 @@ void RGB565ToARGBRow_NEON(const uint8_t* src_rgb565,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
-      "movi       v3.8b, #255                    \n"  // Alpha
+      "movi        v3.8b, #255                   \n"  // Alpha
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      RGB565TOARGB
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" RGB565TOARGB
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_rgb565),  // %0
         "+r"(dst_argb),    // %1
         "+r"(width)        // %2
@@ -911,14 +1515,14 @@ void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
                             uint8_t* dst_argb,
                             int width) {
   asm volatile(
-      "movi       v3.8b, #255                    \n"  // Alpha
+      "movi        v3.8b, #255                   \n"  // Alpha
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
       ARGB1555TOARGB
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
-                                                            // pixels
-      "b.gt       1b                             \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_argb1555),  // %0
         "+r"(dst_argb),      // %1
         "+r"(width)          // %2
@@ -927,6 +1531,8 @@ void ARGB1555ToARGBRow_NEON(const uint8_t* src_argb1555,
   );
 }
 
+// Convert v0.8h to b = v0.8b g = v1.8b r = v2.8b
+// clobbers v3
 #define ARGB4444TOARGB                                                      \
   "shrn       v1.8b,  v0.8h, #8              \n" /* v1(l) AR             */ \
   "xtn2       v1.16b, v0.8h                  \n" /* v1(h) GB             */ \
@@ -944,12 +1550,11 @@ void ARGB4444ToARGBRow_NEON(const uint8_t* src_argb4444,
                             int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      ARGB4444TOARGB
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
-                                                            // pixels
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGB4444TOARGB
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_argb4444),  // %0
         "+r"(dst_argb),      // %1
         "+r"(width)          // %2
@@ -963,28 +1568,29 @@ void ARGBToRGB24Row_NEON(const uint8_t* src_argb,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "st3        {v1.8b,v2.8b,v3.8b}, [%1], #24 \n"  // store 8 pixels of
-                                                      // RGB24.
-      "b.gt       1b                             \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st3         {v0.16b,v1.16b,v2.16b}, [%1], #48 \n"  // store 8 RGB24
+      "b.gt        1b                            \n"
       : "+r"(src_argb),   // %0
         "+r"(dst_rgb24),  // %1
         "+r"(width)       // %2
       :
-      : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
   );
 }
 
 void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "orr        v4.8b, v2.8b, v2.8b            \n"  // mov g
-      "orr        v5.8b, v1.8b, v1.8b            \n"  // mov b
-      "st3        {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
-      "b.gt       1b                             \n"
+      "ld4         {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "orr         v4.8b, v2.8b, v2.8b           \n"  // mov g
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v5.8b, v1.8b, v1.8b           \n"   // mov b
+      "st3         {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_raw),   // %1
         "+r"(width)      // %2
@@ -996,10 +1602,11 @@ void ARGBToRAWRow_NEON(const uint8_t* src_argb, uint8_t* dst_raw, int width) {
 void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
   asm volatile(
       "1:                                        \n"
-      "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of YUY2.
-      "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
-      "st1        {v0.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-      "b.gt       1b                             \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pixels of YUY2.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.16b}, [%1], #16           \n"  // store 16 pixels of Y.
+      "b.gt        1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -1011,10 +1618,11 @@ void YUY2ToYRow_NEON(const uint8_t* src_yuy2, uint8_t* dst_y, int width) {
 void UYVYToYRow_NEON(const uint8_t* src_uyvy, uint8_t* dst_y, int width) {
   asm volatile(
       "1:                                        \n"
-      "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of UYVY.
-      "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
-      "st1        {v1.16b}, [%1], #16            \n"  // store 16 pixels of Y.
-      "b.gt       1b                             \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 pixels of UYVY.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 16 pixels of Y.
+      "b.gt        1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -1029,11 +1637,12 @@ void YUY2ToUV422Row_NEON(const uint8_t* src_yuy2,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2
-      "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
-      "st1        {v1.8b}, [%1], #8              \n"  // store 8 U.
-      "st1        {v3.8b}, [%2], #8              \n"  // store 8 V.
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels = 8 UVs.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v1.8b}, [%1], #8             \n"  // store 8 U.
+      "st1         {v3.8b}, [%2], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
       : "+r"(src_yuy2),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1049,11 +1658,12 @@ void UYVYToUV422Row_NEON(const uint8_t* src_uyvy,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY
-      "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 U.
-      "st1        {v2.8b}, [%2], #8              \n"  // store 8 V.
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY
+      "subs        %w3, %w3, #16                 \n"  // 16 pixels = 8 UVs.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 U.
+      "st1         {v2.8b}, [%2], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
       : "+r"(src_uyvy),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1071,14 +1681,15 @@ void YUY2ToUVRow_NEON(const uint8_t* src_yuy2,
   const uint8_t* src_yuy2b = src_yuy2 + stride_yuy2;
   asm volatile(
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
-      "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
-      "urhadd     v1.8b, v1.8b, v5.8b            \n"        // average rows of U
-      "urhadd     v3.8b, v3.8b, v7.8b            \n"        // average rows of V
-      "st1        {v1.8b}, [%2], #8              \n"        // store 8 U.
-      "st1        {v3.8b}, [%3], #8              \n"        // store 8 V.
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels = 8 UVs.
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+      "urhadd      v1.8b, v1.8b, v5.8b           \n"  // average rows of U
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "urhadd      v3.8b, v3.8b, v7.8b           \n"  // average rows of V
+      "st1         {v1.8b}, [%2], #8             \n"  // store 8 U.
+      "st1         {v3.8b}, [%3], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
       : "+r"(src_yuy2),   // %0
         "+r"(src_yuy2b),  // %1
         "+r"(dst_u),      // %2
@@ -1098,14 +1709,15 @@ void UYVYToUVRow_NEON(const uint8_t* src_uyvy,
   const uint8_t* src_uyvyb = src_uyvy + stride_uyvy;
   asm volatile(
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
-      "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
-      "urhadd     v0.8b, v0.8b, v4.8b            \n"        // average rows of U
-      "urhadd     v2.8b, v2.8b, v6.8b            \n"        // average rows of V
-      "st1        {v0.8b}, [%2], #8              \n"        // store 8 U.
-      "st1        {v2.8b}, [%3], #8              \n"        // store 8 V.
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels = 8 UVs.
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+      "urhadd      v0.8b, v0.8b, v4.8b           \n"  // average rows of U
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "urhadd      v2.8b, v2.8b, v6.8b           \n"  // average rows of V
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 U.
+      "st1         {v2.8b}, [%3], #8             \n"  // store 8 V.
+      "b.gt        1b                            \n"
       : "+r"(src_uyvy),   // %0
         "+r"(src_uyvyb),  // %1
         "+r"(dst_u),      // %2
@@ -1123,13 +1735,14 @@ void ARGBShuffleRow_NEON(const uint8_t* src_argb,
                          const uint8_t* shuffler,
                          int width) {
   asm volatile(
-      "ld1        {v2.16b}, [%3]                 \n"  // shuffler
+      "ld1         {v2.16b}, [%3]                \n"  // shuffler
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 4 pixels.
-      "subs       %w2, %w2, #4                   \n"  // 4 processed per loop
-      "tbl        v1.16b, {v0.16b}, v2.16b       \n"  // look up 4 pixels
-      "st1        {v1.16b}, [%1], #16            \n"  // store 4.
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 4 pixels.
+      "subs        %w2, %w2, #4                  \n"  // 4 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v1.16b, {v0.16b}, v2.16b      \n"  // look up 4 pixels
+      "st1         {v1.16b}, [%1], #16           \n"  // store 4.
+      "b.gt        1b                            \n"
       : "+r"(src_argb),                   // %0
         "+r"(dst_argb),                   // %1
         "+r"(width)                       // %2
@@ -1145,13 +1758,14 @@ void I422ToYUY2Row_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       "1:                                        \n"
-      "ld2        {v0.8b, v1.8b}, [%0], #16      \n"  // load 16 Ys
-      "orr        v2.8b, v1.8b, v1.8b            \n"
-      "ld1        {v1.8b}, [%1], #8              \n"        // load 8 Us
-      "ld1        {v3.8b}, [%2], #8              \n"        // load 8 Vs
-      "subs       %w4, %w4, #16                  \n"        // 16 pixels
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
-      "b.gt       1b                             \n"
+      "ld2         {v0.8b, v1.8b}, [%0], #16     \n"  // load 16 Ys
+      "subs        %w4, %w4, #16                 \n"  // 16 pixels
+      "orr         v2.8b, v1.8b, v1.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v1.8b}, [%1], #8             \n"         // load 8 Us
+      "ld1         {v3.8b}, [%2], #8             \n"         // load 8 Vs
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+      "b.gt        1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -1168,13 +1782,14 @@ void I422ToUYVYRow_NEON(const uint8_t* src_y,
                         int width) {
   asm volatile(
       "1:                                        \n"
-      "ld2        {v1.8b,v2.8b}, [%0], #16       \n"  // load 16 Ys
-      "orr        v3.8b, v2.8b, v2.8b            \n"
-      "ld1        {v0.8b}, [%1], #8              \n"        // load 8 Us
-      "ld1        {v2.8b}, [%2], #8              \n"        // load 8 Vs
-      "subs       %w4, %w4, #16                  \n"        // 16 pixels
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
-      "b.gt       1b                             \n"
+      "ld2         {v1.8b,v2.8b}, [%0], #16      \n"  // load 16 Ys
+      "orr         v3.8b, v2.8b, v2.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v0.8b}, [%1], #8             \n"         // load 8 Us
+      "ld1         {v2.8b}, [%2], #8             \n"         // load 8 Vs
+      "subs        %w4, %w4, #16                 \n"         // 16 pixels
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+      "b.gt        1b                            \n"
       : "+r"(src_y),     // %0
         "+r"(src_u),     // %1
         "+r"(src_v),     // %2
@@ -1189,16 +1804,17 @@ void ARGBToRGB565Row_NEON(const uint8_t* src_argb,
                           int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      ARGBTORGB565
-      "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels RGB565.
-      "b.gt       1b                             \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTORGB565
+      "st1         {v18.16b}, [%1], #16          \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
       : "+r"(src_argb),    // %0
         "+r"(dst_rgb565),  // %1
         "+r"(width)        // %2
       :
-      : "cc", "memory", "v0", "v20", "v21", "v22", "v23");
+      : "cc", "memory", "v16", "v17", "v18", "v19");
 }
 
 void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
@@ -1206,20 +1822,22 @@ void ARGBToRGB565DitherRow_NEON(const uint8_t* src_argb,
                                 const uint32_t dither4,
                                 int width) {
   asm volatile(
-      "dup        v1.4s, %w2                     \n"  // dither4
+      "dup         v1.4s, %w2                    \n"  // dither4
       "1:                                        \n"
-      "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"  // load 8 pixels
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uqadd      v20.8b, v20.8b, v1.8b          \n"
-      "uqadd      v21.8b, v21.8b, v1.8b          \n"
-      "uqadd      v22.8b, v22.8b, v1.8b          \n" ARGBTORGB565
-      "st1        {v0.16b}, [%0], #16            \n"  // store 8 pixels RGB565.
-      "b.gt       1b                             \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v16.8b, v16.8b, v1.8b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v17.8b, v17.8b, v1.8b         \n"
+      "uqadd       v18.8b, v18.8b, v1.8b         \n" ARGBTORGB565
+      "st1         {v18.16b}, [%0], #16          \n"  // store 8 pixels RGB565.
+      "b.gt        1b                            \n"
       : "+r"(dst_rgb)   // %0
       : "r"(src_argb),  // %1
         "r"(dither4),   // %2
         "r"(width)      // %3
-      : "cc", "memory", "v0", "v1", "v20", "v21", "v22", "v23");
+      : "cc", "memory", "v1", "v16", "v17", "v18", "v19");
 }
 
 void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
@@ -1227,99 +1845,198 @@ void ARGBToARGB1555Row_NEON(const uint8_t* src_argb,
                             int width) {
   asm volatile(
       "1:                                        \n"
-      "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      ARGBTOARGB1555
-      "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels
-                                                      // ARGB1555.
-      "b.gt       1b                             \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTOARGB1555
+      "st1         {v0.16b}, [%1], #16           \n"  // store 8 pixels
+      "b.gt        1b                            \n"
       : "+r"(src_argb),      // %0
         "+r"(dst_argb1555),  // %1
         "+r"(width)          // %2
       :
-      : "cc", "memory", "v0", "v20", "v21", "v22", "v23");
+      : "cc", "memory", "v0", "v16", "v17", "v18", "v19");
 }
 
 void ARGBToARGB4444Row_NEON(const uint8_t* src_argb,
                             uint8_t* dst_argb4444,
                             int width) {
   asm volatile(
-      "movi       v4.16b, #0x0f                  \n"  // bits to clear with
+      "movi        v23.16b, #0x0f                \n"  // bits to clear with
                                                       // vbic.
       "1:                                        \n"
-      "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      ARGBTOARGB4444
-      "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels
-                                                      // ARGB4444.
-      "b.gt       1b                             \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8
+                                                                 // pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n" ARGBTOARGB4444
+      "st1         {v0.16b}, [%1], #16           \n"  // store 8 pixels
+      "b.gt        1b                            \n"
       : "+r"(src_argb),      // %0
         "+r"(dst_argb4444),  // %1
         "+r"(width)          // %2
       :
-      : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23");
+      : "cc", "memory", "v0", "v1", "v16", "v17", "v18", "v19", "v23");
 }
 
-void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+#if LIBYUV_USE_ST2
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
   asm volatile(
-      "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-      "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-      "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-      "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ldp         q0, q2, [%0], #32             \n"  // load 8 pixels
+      "mov         v1.16b, v0.16b                \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mov         v3.16b, v2.16b                \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st2         {v0.16b, v1.16b}, [%1], #32   \n"  // store 4 pixels
+      "st2         {v2.16b, v3.16b}, [%1], #32   \n"  // store 4 pixels
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
-        "+r"(dst_y),     // %1
+        "+r"(dst_ar64),  // %1
         "+r"(width)      // %2
       :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+      : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
-                              uint8_t* dst_a,
-                              int width) {
+static const uvec8 kShuffleARGBToABGR = {2,  1, 0, 3,  6,  5,  4,  7,
+                                         10, 9, 8, 11, 14, 13, 12, 15};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
   asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
       "1:                                        \n"
-      "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load row 16
-                                                                // pixels
-      "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-      "st1        {v3.16b}, [%1], #16            \n"  // store 16 A's.
-      "b.gt       1b                             \n"
+      "ldp         q0, q2, [%0], #32             \n"  // load 8 pixels
+      "tbl         v0.16b, {v0.16b}, v4.16b      \n"
+      "tbl         v2.16b, {v2.16b}, v4.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mov         v1.16b, v0.16b                \n"
+      "mov         v3.16b, v2.16b                \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "st2         {v0.16b, v1.16b}, [%1], #32   \n"  // store 4 pixels
+      "st2         {v2.16b, v3.16b}, [%1], #32   \n"  // store 4 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),           // %0
+        "+r"(dst_ab64),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleARGBToABGR)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+#else
+void ARGBToAR64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ar64,
+                        int width) {
+  asm volatile(
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 8 ARGB pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "zip1        v2.16b, v0.16b, v0.16b        \n"
+      "zip2        v3.16b, v0.16b, v0.16b        \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v4.16b, v1.16b, v1.16b        \n"
+      "zip2        v5.16b, v1.16b, v1.16b        \n"
+      "st1         {v2.8h, v3.8h, v4.8h, v5.8h}, [%1], #64 \n"  // 8 AR64
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
-        "+r"(dst_a),     // %1
+        "+r"(dst_ar64),  // %1
         "+r"(width)      // %2
       :
-      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
-  );
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5");
+}
+
+static const uvec8 kShuffleARGBToAB64[2] = {
+    {2, 2, 1, 1, 0, 0, 3, 3, 6, 6, 5, 5, 4, 4, 7, 7},
+    {10, 10, 9, 9, 8, 8, 11, 11, 14, 14, 13, 13, 12, 12, 15, 15}};
+
+void ARGBToAB64Row_NEON(const uint8_t* src_argb,
+                        uint16_t* dst_ab64,
+                        int width) {
+  asm volatile(
+      "ldp         q6, q7, [%3]                  \n"  // 2 shufflers
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 8 pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "tbl         v2.16b, {v0.16b}, v6.16b      \n"  // ARGB to AB64
+      "tbl         v3.16b, {v0.16b}, v7.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v4.16b, {v1.16b}, v6.16b      \n"
+      "tbl         v5.16b, {v1.16b}, v7.16b      \n"
+      "st1         {v2.8h, v3.8h, v4.8h, v5.8h}, [%1], #64 \n"  // 8 AR64
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),              // %0
+        "+r"(dst_ab64),              // %1
+        "+r"(width)                  // %2
+      : "r"(&kShuffleARGBToAB64[0])  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+#endif  // LIBYUV_USE_ST2
+
+static const uvec8 kShuffleAR64ToARGB = {1,  3,  5,  7,  9,  11, 13, 15,
+                                         17, 19, 21, 23, 25, 27, 29, 31};
+
+void AR64ToARGBRow_NEON(const uint16_t* src_ar64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 4 pixels
+      "ldp         q2, q3, [%0], #32             \n"  // load 4 pixels
+      "tbl         v0.16b, {v0.16b, v1.16b}, v4.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v2.16b, {v2.16b, v3.16b}, v4.16b \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "stp         q0, q2, [%1], #32             \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ar64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAR64ToARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
+}
+
+static const uvec8 kShuffleAB64ToARGB = {5,  3,  1,  7,  13, 11, 9,  15,
+                                         21, 19, 17, 23, 29, 27, 25, 31};
+
+void AB64ToARGBRow_NEON(const uint16_t* src_ab64,
+                        uint8_t* dst_argb,
+                        int width) {
+  asm volatile(
+      "ldr         q4, [%3]                      \n"  // shuffler
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"  // load 4 pixels
+      "ldp         q2, q3, [%0], #32             \n"  // load 4 pixels
+      "tbl         v0.16b, {v0.16b, v1.16b}, v4.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v2.16b, {v2.16b, v3.16b}, v4.16b \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "stp         q0, q2, [%1], #32             \n"  // store 8 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_ab64),           // %0
+        "+r"(dst_argb),           // %1
+        "+r"(width)               // %2
+      : "r"(&kShuffleAB64ToARGB)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
 }
 
-void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+void ARGBExtractAlphaRow_NEON(const uint8_t* src_argb,
+                              uint8_t* dst_a,
+                              int width) {
   asm volatile(
-      "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
-      "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
-      "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-      "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-      "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-      "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "st1         {v3.16b}, [%1], #16           \n"  // store 16 A's.
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
-        "+r"(dst_y),     // %1
+        "+r"(dst_a),     // %1
         "+r"(width)      // %2
       :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
+      : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
 }
 
 // 8x1 pixels.
@@ -1328,33 +2045,31 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
                          uint8_t* dst_v,
                          int width) {
   asm volatile(
-      "movi       v24.8b, #112                   \n"  // UB / VR 0.875
+      "movi        v24.8b, #112                  \n"  // UB / VR 0.875
                                                       // coefficient
-      "movi       v25.8b, #74                    \n"  // UG -0.5781 coefficient
-      "movi       v26.8b, #38                    \n"  // UR -0.2969 coefficient
-      "movi       v27.8b, #18                    \n"  // VB -0.1406 coefficient
-      "movi       v28.8b, #94                    \n"  // VG -0.7344 coefficient
-      "movi       v29.16b,#0x80                  \n"  // 128.5
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-                                                            // pixels.
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "umull      v4.8h, v0.8b, v24.8b           \n"  // B
-      "umlsl      v4.8h, v1.8b, v25.8b           \n"  // G
-      "umlsl      v4.8h, v2.8b, v26.8b           \n"  // R
-      "add        v4.8h, v4.8h, v29.8h           \n"  // +128 -> unsigned
-
-      "umull      v3.8h, v2.8b, v24.8b           \n"  // R
-      "umlsl      v3.8h, v1.8b, v28.8b           \n"  // G
-      "umlsl      v3.8h, v0.8b, v27.8b           \n"  // B
-      "add        v3.8h, v3.8h, v29.8h           \n"  // +128 -> unsigned
-
-      "uqshrn     v0.8b, v4.8h, #8               \n"  // 16 bit to 8 bit U
-      "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
-      "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
-      "b.gt       1b                             \n"
+      "movi        v25.8b, #74                   \n"  // UG -0.5781 coefficient
+      "movi        v26.8b, #38                   \n"  // UR -0.2969 coefficient
+      "movi        v27.8b, #18                   \n"  // VB -0.1406 coefficient
+      "movi        v28.8b, #94                   \n"  // VG -0.7344 coefficient
+      "movi        v29.16b,#0x80                 \n"  // 128.5
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v24.8b          \n"  // B
+      "umlsl       v4.8h, v1.8b, v25.8b          \n"  // G
+      "umlsl       v4.8h, v2.8b, v26.8b          \n"  // R
+      "prfm        pldl1keep, [%0, 448]          \n"
+
+      "umull       v3.8h, v2.8b, v24.8b          \n"  // R
+      "umlsl       v3.8h, v1.8b, v28.8b          \n"  // G
+      "umlsl       v3.8h, v0.8b, v27.8b          \n"  // B
+
+      "addhn       v0.8b, v4.8h, v29.8h          \n"  // +128 -> unsigned
+      "addhn       v1.8b, v3.8h, v29.8h          \n"  // +128 -> unsigned
+
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%2], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_u),     // %1
         "+r"(dst_v),     // %2
@@ -1381,10 +2096,8 @@ void ARGBToUV444Row_NEON(const uint8_t* src_argb,
   "mls        v4.8h, " #QG ",v24.8h          \n" /* G                    */ \
   "mls        v3.8h, " #QR ",v22.8h          \n" /* R                    */ \
   "mls        v4.8h, " #QB ",v23.8h          \n" /* B                    */ \
-  "add        v3.8h, v3.8h, v25.8h           \n" /* +128 -> unsigned     */ \
-  "add        v4.8h, v4.8h, v25.8h           \n" /* +128 -> unsigned     */ \
-  "uqshrn     v0.8b, v3.8h, #8               \n" /* 16 bit to 8 bit U    */ \
-  "uqshrn     v1.8b, v4.8h, #8               \n" /* 16 bit to 8 bit V    */
+  "addhn      v0.8b, v3.8h, v25.8h           \n" /* +128 -> unsigned     */ \
+  "addhn      v1.8b, v4.8h, v25.8h           \n" /* +128 -> unsigned     */
 // clang-format on
 
 // TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
@@ -1398,26 +2111,28 @@ void ARGBToUVRow_NEON(const uint8_t* src_argb,
   const uint8_t* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_argb_1),  // %1
     "+r"(dst_u),     // %2
@@ -1429,7 +2144,6 @@ void ARGBToUVRow_NEON(const uint8_t* src_argb,
   );
 }
 
-// TODO(fbarchard): Subsample match C code.
 void ARGBToUVJRow_NEON(const uint8_t* src_argb,
                        int src_stride_argb,
                        uint8_t* dst_u,
@@ -1437,31 +2151,33 @@ void ARGBToUVJRow_NEON(const uint8_t* src_argb,
                        int width) {
   const uint8_t* src_argb_1 = src_argb + src_stride_argb;
   asm volatile (
-    "movi       v20.8h, #63, lsl #0            \n"  // UB/VR coeff (0.500) / 2
-    "movi       v21.8h, #42, lsl #0            \n"  // UG coeff (-0.33126) / 2
-    "movi       v22.8h, #21, lsl #0            \n"  // UR coeff (-0.16874) / 2
-    "movi       v23.8h, #10, lsl #0            \n"  // VB coeff (-0.08131) / 2
-    "movi       v24.8h, #53, lsl #0            \n"  // VG coeff (-0.41869) / 2
-    "movi       v25.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
-  "1:                                          \n"
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64  \n"  // load next 16
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_argb),  // %0
     "+r"(src_argb_1),  // %1
     "+r"(dst_u),     // %2
@@ -1473,6 +2189,96 @@ void ARGBToUVJRow_NEON(const uint8_t* src_argb,
   );
 }
 
+void RGB24ToUVJRow_NEON(const uint8_t* src_rgb24,
+                        int src_stride_rgb24,
+                        uint8_t* dst_u,
+                        uint8_t* dst_v,
+                        int width) {
+  const uint8_t* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
+  asm volatile (
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_rgb24_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+
+void RAWToUVJRow_NEON(const uint8_t* src_raw,
+                      int src_stride_raw,
+                      uint8_t* dst_u,
+                      uint8_t* dst_v,
+                      int width) {
+  const uint8_t* src_raw_1 = src_raw + src_stride_raw;
+  asm volatile (
+      "movi        v20.8h, #63, lsl #0           \n"  // UB/VR coeff (0.500) / 2
+      "movi        v21.8h, #42, lsl #0           \n"  // UG coeff (-0.33126) / 2
+      "movi        v22.8h, #21, lsl #0           \n"  // UR coeff (-0.16874) / 2
+      "movi        v23.8h, #10, lsl #0           \n"  // VB coeff (-0.08131) / 2
+      "movi        v24.8h, #53, lsl #0           \n"  // VG coeff (-0.41869) / 2
+      "movi        v25.16b, #0x80                \n"  // 128.5 (0x8080 in 16-bit)
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_raw_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(width)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+
 void BGRAToUVRow_NEON(const uint8_t* src_bgra,
                       int src_stride_bgra,
                       uint8_t* dst_u,
@@ -1481,25 +2287,27 @@ void BGRAToUVRow_NEON(const uint8_t* src_bgra,
   const uint8_t* src_bgra_1 = src_bgra + src_stride_bgra;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v3.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v3.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v1.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
-    "uadalp     v0.8h, v7.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v3.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v5.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v3.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v3.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v3.8h, v2.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v1.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
+      "uadalp      v0.8h, v7.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v3.8h, v6.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v5.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v3.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_bgra),  // %0
     "+r"(src_bgra_1),  // %1
     "+r"(dst_u),     // %2
@@ -1519,25 +2327,27 @@ void ABGRToUVRow_NEON(const uint8_t* src_abgr,
   const uint8_t* src_abgr_1 = src_abgr + src_stride_abgr;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v3.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
-    "uadalp     v3.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v3.8h, #1               \n"  // 2x average
-    "urshr      v2.8h, v2.8h, #1               \n"
-    "urshr      v1.8h, v1.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v3.8h, v2.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v2.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v0.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+      "uadalp      v3.8h, v6.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v2.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v4.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v3.8h, #1              \n"  // 2x average
+      "urshr       v2.8h, v2.8h, #1              \n"
+      "urshr       v1.8h, v1.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v2.8h, v1.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_abgr),  // %0
     "+r"(src_abgr_1),  // %1
     "+r"(dst_u),     // %2
@@ -1557,25 +2367,27 @@ void RGBAToUVRow_NEON(const uint8_t* src_rgba,
   const uint8_t* src_rgba_1 = src_rgba + src_stride_rgba;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v1.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v3.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
-    "uadalp     v0.8h, v5.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v7.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v1.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v2.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v3.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+      "uadalp      v0.8h, v5.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v6.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v7.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_rgba),  // %0
     "+r"(src_rgba_1),  // %1
     "+r"(dst_u),     // %2
@@ -1595,25 +2407,27 @@ void RGB24ToUVRow_NEON(const uint8_t* src_rgb24,
   const uint8_t* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
-    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
-    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v2.8h, v2.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
+      "uadalp      v0.8h, v4.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v2.8h, v6.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v0.8h, v0.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v2.8h, v2.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v0.8h, v1.8h, v2.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_rgb24),  // %0
     "+r"(src_rgb24_1),  // %1
     "+r"(dst_u),     // %2
@@ -1633,25 +2447,27 @@ void RAWToUVRow_NEON(const uint8_t* src_raw,
   const uint8_t* src_raw_1 = src_raw + src_stride_raw;
   asm volatile (
     RGBTOUV_SETUP_REG
-  "1:                                          \n"
-    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 8 RAW pixels.
-    "uaddlp     v2.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uaddlp     v0.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
-    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
-    "uadalp     v2.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
-    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
-    "uadalp     v0.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
-
-    "urshr      v2.8h, v2.8h, #1               \n"  // 2x average
-    "urshr      v1.8h, v1.8h, #1               \n"
-    "urshr      v0.8h, v0.8h, #1               \n"
-
-    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+      "1:                                        \n"
+      "ld3         {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 RAW pixels.
+      "uaddlp      v2.8h, v2.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v0.8h, v0.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "ld3         {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
+      "uadalp      v2.8h, v6.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v5.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uadalp      v0.8h, v4.16b                 \n"  // R 16 bytes -> 8 shorts.
+
+      "urshr       v2.8h, v2.8h, #1              \n"  // 2x average
+      "urshr       v1.8h, v1.8h, #1              \n"
+      "urshr       v0.8h, v0.8h, #1              \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
     RGBTOUV(v2.8h, v1.8h, v0.8h)
-    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-    "b.gt       1b                             \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
   : "+r"(src_raw),  // %0
     "+r"(src_raw_1),  // %1
     "+r"(dst_u),     // %2
@@ -1663,7 +2479,7 @@ void RAWToUVRow_NEON(const uint8_t* src_raw,
   );
 }
 
-// 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
+// 16x2 pixels -> 8x1.  width is number of rgb pixels. e.g. 16.
 void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
                         int src_stride_rgb565,
                         uint8_t* dst_u,
@@ -1671,67 +2487,54 @@ void RGB565ToUVRow_NEON(const uint8_t* src_rgb565,
                         int width) {
   const uint8_t* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
   asm volatile(
-      "movi       v22.8h, #56, lsl #0            \n"  // UB / VR coeff (0.875) /
-                                                      // 2
-      "movi       v23.8h, #37, lsl #0            \n"  // UG coeff (-0.5781) / 2
-      "movi       v24.8h, #19, lsl #0            \n"  // UR coeff (-0.2969) / 2
-      "movi       v25.8h, #9 , lsl #0            \n"  // VB coeff (-0.1406) / 2
-      "movi       v26.8h, #47, lsl #0            \n"  // VG coeff (-0.7344) / 2
-      "movi       v27.16b, #0x80                 \n"  // 128.5 0x8080 in 16bit
+      RGBTOUV_SETUP_REG
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
       RGB565TOARGB
-      "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%0], #16            \n"  // next 8 RGB565 pixels.
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 RGB565 pixels.
       RGB565TOARGB
-      "uaddlp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
 
-      "ld1        {v0.16b}, [%1], #16            \n"  // load 8 RGB565 pixels.
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 RGB565 pixels.
       RGB565TOARGB
-      "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%1], #16            \n"  // next 8 RGB565 pixels.
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 RGB565 pixels.
       RGB565TOARGB
-      "uadalp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-      "ins        v16.D[1], v17.D[0]             \n"
-      "ins        v18.D[1], v19.D[0]             \n"
-      "ins        v20.D[1], v21.D[0]             \n"
-
-      "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-      "urshr      v5.8h, v18.8h, #1              \n"
-      "urshr      v6.8h, v20.8h, #1              \n"
-
-      "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-      "mul        v16.8h, v4.8h, v22.8h          \n"  // B
-      "mls        v16.8h, v5.8h, v23.8h          \n"  // G
-      "mls        v16.8h, v6.8h, v24.8h          \n"  // R
-      "add        v16.8h, v16.8h, v27.8h         \n"  // +128 -> unsigned
-      "mul        v17.8h, v6.8h, v22.8h          \n"  // R
-      "mls        v17.8h, v5.8h, v26.8h          \n"  // G
-      "mls        v17.8h, v4.8h, v25.8h          \n"  // B
-      "add        v17.8h, v17.8h, v27.8h         \n"  // +128 -> unsigned
-      "uqshrn     v0.8b, v16.8h, #8              \n"  // 16 bit to 8 bit U
-      "uqshrn     v1.8b, v17.8h, #8              \n"  // 16 bit to 8 bit V
-      "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-      "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-      "b.gt       1b                             \n"
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
       : "+r"(src_rgb565),    // %0
         "+r"(src_rgb565_1),  // %1
-        "+r"(dst_u),         // %2
-        "+r"(dst_v),         // %3
-        "+r"(width)          // %4
+        "+r"(dst_u),           // %2
+        "+r"(dst_v),           // %3
+        "+r"(width)            // %4
       :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
-        "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26",
-        "v27");
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v16", "v17",
+        "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27",
+        "v28");
 }
 
 // 16x2 pixels -> 8x1.  width is number of argb pixels. e.g. 16.
@@ -1744,50 +2547,43 @@ void ARGB1555ToUVRow_NEON(const uint8_t* src_argb1555,
   asm volatile(
       RGBTOUV_SETUP_REG
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
       RGB555TOARGB
-      "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB1555 pixels.
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 ARGB1555 pixels.
       RGB555TOARGB
-      "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
 
-      "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB1555 pixels.
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 ARGB1555 pixels.
       RGB555TOARGB
-      "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB1555 pixels.
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 ARGB1555 pixels.
       RGB555TOARGB
-      "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-      "ins        v16.D[1], v26.D[0]             \n"
-      "ins        v17.D[1], v27.D[0]             \n"
-      "ins        v18.D[1], v28.D[0]             \n"
-
-      "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-      "urshr      v5.8h, v17.8h, #1              \n"
-      "urshr      v6.8h, v18.8h, #1              \n"
-
-      "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-      "mul        v2.8h, v4.8h, v20.8h           \n"  // B
-      "mls        v2.8h, v5.8h, v21.8h           \n"  // G
-      "mls        v2.8h, v6.8h, v22.8h           \n"  // R
-      "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
-      "mul        v3.8h, v6.8h, v20.8h           \n"  // R
-      "mls        v3.8h, v5.8h, v24.8h           \n"  // G
-      "mls        v3.8h, v4.8h, v23.8h           \n"  // B
-      "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
-      "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
-      "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-      "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-      "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-      "b.gt       1b                             \n"
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
       : "+r"(src_argb1555),    // %0
         "+r"(src_argb1555_1),  // %1
         "+r"(dst_u),           // %2
@@ -1807,52 +2603,45 @@ void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
                           int width) {
   const uint8_t* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
   asm volatile(
-      RGBTOUV_SETUP_REG
+      RGBTOUV_SETUP_REG  // sets v20-v25
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB4444 pixels.
+      "uaddlp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%0], #16           \n"  // next 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+      "uaddlp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uaddlp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uaddlp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
 
-      "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB4444 pixels.
+      "ld1         {v0.16b}, [%1], #16           \n"  // load 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-      "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB4444 pixels.
+      "uadalp      v16.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v17.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v18.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+      "ld1         {v0.16b}, [%1], #16           \n"  // next 8 ARGB4444 pixels.
       ARGB4444TOARGB
-      "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
-      "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
-      "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
-
-      "ins        v16.D[1], v26.D[0]             \n"
-      "ins        v17.D[1], v27.D[0]             \n"
-      "ins        v18.D[1], v28.D[0]             \n"
-
-      "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
-      "urshr      v5.8h, v17.8h, #1              \n"
-      "urshr      v6.8h, v18.8h, #1              \n"
-
-      "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
-      "mul        v2.8h, v4.8h, v20.8h           \n"  // B
-      "mls        v2.8h, v5.8h, v21.8h           \n"  // G
-      "mls        v2.8h, v6.8h, v22.8h           \n"  // R
-      "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
-      "mul        v3.8h, v6.8h, v20.8h           \n"  // R
-      "mls        v3.8h, v5.8h, v24.8h           \n"  // G
-      "mls        v3.8h, v4.8h, v23.8h           \n"  // B
-      "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
-      "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
-      "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
-      "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
-      "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
-      "b.gt       1b                             \n"
+      "uadalp      v26.4h, v0.8b                 \n"  // B 8 bytes -> 4 shorts.
+      "uadalp      v27.4h, v1.8b                 \n"  // G 8 bytes -> 4 shorts.
+      "uadalp      v28.4h, v2.8b                 \n"  // R 8 bytes -> 4 shorts.
+
+      "ins         v16.D[1], v26.D[0]            \n"
+      "ins         v17.D[1], v27.D[0]            \n"
+      "ins         v18.D[1], v28.D[0]            \n"
+
+      "urshr       v0.8h, v16.8h, #1             \n"  // 2x average
+      "urshr       v1.8h, v17.8h, #1             \n"
+      "urshr       v2.8h, v18.8h, #1             \n"
+
+      "subs        %w4, %w4, #16                 \n"  // 16 processed per loop.
+      RGBTOUV(v0.8h, v1.8h, v2.8h)
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 pixels U.
+      "st1         {v1.8b}, [%3], #8             \n"  // store 8 pixels V.
+      "b.gt        1b                            \n"
       : "+r"(src_argb4444),    // %0
         "+r"(src_argb4444_1),  // %1
         "+r"(dst_u),           // %2
@@ -1868,21 +2657,22 @@ void ARGB4444ToUVRow_NEON(const uint8_t* src_argb4444,
 
 void RGB565ToYRow_NEON(const uint8_t* src_rgb565, uint8_t* dst_y, int width) {
   asm volatile(
-      "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
-      "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
-      "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
-      "movi       v27.8b, #16                    \n"  // Add 16 constant
+      "movi        v24.8b, #25                   \n"  // B * 0.1016 coefficient
+      "movi        v25.8b, #129                  \n"  // G * 0.5078 coefficient
+      "movi        v26.8b, #66                   \n"  // R * 0.2578 coefficient
+      "movi        v27.8b, #16                   \n"  // Add 16 constant
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 RGB565 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
       RGB565TOARGB
-      "umull      v3.8h, v0.8b, v24.8b           \n"  // B
-      "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
-      "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
-      "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v27.8b           \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
+      "umull       v3.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v3.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v27.8b          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
       : "+r"(src_rgb565),  // %0
         "+r"(dst_y),       // %1
         "+r"(width)        // %2
@@ -1895,21 +2685,22 @@ void ARGB1555ToYRow_NEON(const uint8_t* src_argb1555,
                          uint8_t* dst_y,
                          int width) {
   asm volatile(
-      "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
+      "movi        v4.8b, #25                    \n"  // B * 0.1016 coefficient
+      "movi        v5.8b, #129                   \n"  // G * 0.5078 coefficient
+      "movi        v6.8b, #66                    \n"  // R * 0.2578 coefficient
+      "movi        v7.8b, #16                    \n"  // Add 16 constant
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB1555 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
       ARGB1555TOARGB
-      "umull      v3.8h, v0.8b, v4.8b            \n"  // B
-      "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
-      "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
-      "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
+      "umull       v3.8h, v0.8b, v4.8b           \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v5.8b           \n"  // G
+      "umlal       v3.8h, v2.8b, v6.8b           \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v7.8b           \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
       : "+r"(src_argb1555),  // %0
         "+r"(dst_y),         // %1
         "+r"(width)          // %2
@@ -1921,21 +2712,22 @@ void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
                          uint8_t* dst_y,
                          int width) {
   asm volatile(
-      "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
-      "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
-      "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
-      "movi       v27.8b, #16                    \n"  // Add 16 constant
+      "movi        v24.8b, #25                   \n"  // B * 0.1016 coefficient
+      "movi        v25.8b, #129                  \n"  // G * 0.5078 coefficient
+      "movi        v26.8b, #66                   \n"  // R * 0.2578 coefficient
+      "movi        v27.8b, #16                   \n"  // Add 16 constant
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 8 ARGB4444 pixels.
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
       ARGB4444TOARGB
-      "umull      v3.8h, v0.8b, v24.8b           \n"  // B
-      "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
-      "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
-      "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v27.8b           \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
+      "umull       v3.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v3.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v3.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v3.8h, #8              \n"  // 16 bit to 8 bit Y
+      "uqadd       v0.8b, v0.8b, v27.8b          \n"
+      "st1         {v0.8b}, [%1], #8             \n"  // store 8 pixels Y.
+      "b.gt        1b                            \n"
       : "+r"(src_argb4444),  // %0
         "+r"(dst_y),         // %1
         "+r"(width)          // %2
@@ -1943,119 +2735,175 @@ void ARGB4444ToYRow_NEON(const uint8_t* src_argb4444,
       : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27");
 }
 
-void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+struct RgbConstants {
+  uint8_t kRGBToY[4];
+  uint16_t kAddY;
+  uint16_t pad;
+};
+
+// RGB to JPeg coefficients
+// B * 0.1140 coefficient = 29
+// G * 0.5870 coefficient = 150
+// R * 0.2990 coefficient = 77
+// Add 0.5 = 0x80
+static const struct RgbConstants kRgb24JPEGConstants = {{29, 150, 77, 0},
+                                                        128,
+                                                        0};
+
+static const struct RgbConstants kRawJPEGConstants = {{77, 150, 29, 0}, 128, 0};
+
+// RGB to BT.601 coefficients
+// B * 0.1016 coefficient = 25
+// G * 0.5078 coefficient = 129
+// R * 0.2578 coefficient = 66
+// Add 16.5 = 0x1080
+
+static const struct RgbConstants kRgb24I601Constants = {{25, 129, 66, 0},
+                                                        0x1080,
+                                                        0};
+
+static const struct RgbConstants kRawI601Constants = {{66, 129, 25, 0},
+                                                      0x1080,
+                                                      0};
+
+// ARGB expects first 3 values to contain RGB and 4th value is ignored.
+void ARGBToYMatrixRow_NEON(const uint8_t* src_argb,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
   asm volatile(
-      "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v1.8b, v4.8b           \n"  // R
-      "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
-      "umlal      v16.8h, v3.8b, v6.8b           \n"  // B
-      "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
-      : "+r"(src_bgra),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v6.16b, v0.b[0]               \n"
+      "dup         v7.16b, v0.b[1]               \n"
+      "dup         v16.16b, v0.b[2]              \n"
+      "dup         v17.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld4         {v2.16b,v3.16b,v4.16b,v5.16b}, [%0], #64 \n"  // load 16
+                                                                 // pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v6.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v6.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v7.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v7.16b         \n"
+      "umlal       v0.8h, v4.8b, v16.8b          \n"  // R
+      "umlal2      v1.8h, v4.16b, v16.16b        \n"
+      "addhn       v0.8b, v0.8h, v17.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v17.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17");
+}
+
+void ARGBToYRow_NEON(const uint8_t* src_argb, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_y, width, &kRgb24I601Constants);
+}
+
+void ARGBToYJRow_NEON(const uint8_t* src_argb, uint8_t* dst_yj, int width) {
+  ARGBToYMatrixRow_NEON(src_argb, dst_yj, width, &kRgb24JPEGConstants);
 }
 
 void ABGRToYRow_NEON(const uint8_t* src_abgr, uint8_t* dst_y, int width) {
+  ARGBToYMatrixRow_NEON(src_abgr, dst_y, width, &kRawI601Constants);
+}
+
+// RGBA expects first value to be A and ignored, then 3 values to contain RGB.
+// Same code as ARGB, except the LD4
+void RGBAToYMatrixRow_NEON(const uint8_t* src_rgba,
+                           uint8_t* dst_y,
+                           int width,
+                           const struct RgbConstants* rgbconstants) {
   asm volatile(
-      "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v0.8b, v4.8b           \n"  // R
-      "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-      "umlal      v16.8h, v2.8b, v6.8b           \n"  // B
-      "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
-      : "+r"(src_abgr),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v6.16b, v0.b[0]               \n"
+      "dup         v7.16b, v0.b[1]               \n"
+      "dup         v16.16b, v0.b[2]              \n"
+      "dup         v17.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld4         {v1.16b,v2.16b,v3.16b,v4.16b}, [%0], #64 \n"  // load 16
+                                                                 // pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v6.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v6.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v7.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v7.16b         \n"
+      "umlal       v0.8h, v4.8b, v16.8b          \n"  // R
+      "umlal2      v1.8h, v4.16b, v16.16b        \n"
+      "addhn       v0.8b, v0.8h, v17.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v17.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgba),    // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+        "v17");
 }
 
 void RGBAToYRow_NEON(const uint8_t* src_rgba, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_y, width, &kRgb24I601Constants);
+}
+
+void RGBAToYJRow_NEON(const uint8_t* src_rgba, uint8_t* dst_yj, int width) {
+  RGBAToYMatrixRow_NEON(src_rgba, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void BGRAToYRow_NEON(const uint8_t* src_bgra, uint8_t* dst_y, int width) {
+  RGBAToYMatrixRow_NEON(src_bgra, dst_y, width, &kRawI601Constants);
+}
+
+void RGBToYMatrixRow_NEON(const uint8_t* src_rgb,
+                          uint8_t* dst_y,
+                          int width,
+                          const struct RgbConstants* rgbconstants) {
   asm volatile(
-      "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v1.8b, v4.8b           \n"  // B
-      "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
-      "umlal      v16.8h, v3.8b, v6.8b           \n"  // R
-      "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
-      : "+r"(src_rgba),  // %0
-        "+r"(dst_y),     // %1
-        "+r"(width)      // %2
-      :
+      "ldr         d0, [%3]                      \n"  // load rgbconstants
+      "dup         v5.16b, v0.b[0]               \n"
+      "dup         v6.16b, v0.b[1]               \n"
+      "dup         v7.16b, v0.b[2]               \n"
+      "dup         v16.8h,  v0.h[2]              \n"
+      "1:                                        \n"
+      "ld3         {v2.16b,v3.16b,v4.16b}, [%0], #48 \n"  // load 16 pixels.
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop.
+      "umull       v0.8h, v2.8b, v5.8b           \n"  // B
+      "umull2      v1.8h, v2.16b, v5.16b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v0.8h, v3.8b, v6.8b           \n"  // G
+      "umlal2      v1.8h, v3.16b, v6.16b         \n"
+      "umlal       v0.8h, v4.8b, v7.8b           \n"  // R
+      "umlal2      v1.8h, v4.16b, v7.16b         \n"
+      "addhn       v0.8b, v0.8h, v16.8h          \n"  // 16 bit to 8 bit Y
+      "addhn       v1.8b, v1.8h, v16.8h          \n"
+      "st1         {v0.8b, v1.8b}, [%1], #16     \n"  // store 16 pixels Y.
+      "b.gt        1b                            \n"
+      : "+r"(src_rgb),     // %0
+        "+r"(dst_y),       // %1
+        "+r"(width)        // %2
+      : "r"(rgbconstants)  // %3
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
 }
 
+void RGB24ToYJRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_rgb24, dst_yj, width, &kRgb24JPEGConstants);
+}
+
+void RAWToYJRow_NEON(const uint8_t* src_raw, uint8_t* dst_yj, int width) {
+  RGBToYMatrixRow_NEON(src_raw, dst_yj, width, &kRawJPEGConstants);
+}
+
 void RGB24ToYRow_NEON(const uint8_t* src_rgb24, uint8_t* dst_y, int width) {
-  asm volatile(
-      "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v0.8b, v4.8b           \n"  // B
-      "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-      "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
-      "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
-      : "+r"(src_rgb24),  // %0
-        "+r"(dst_y),      // %1
-        "+r"(width)       // %2
-      :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+  RGBToYMatrixRow_NEON(src_rgb24, dst_y, width, &kRgb24I601Constants);
 }
 
 void RAWToYRow_NEON(const uint8_t* src_raw, uint8_t* dst_y, int width) {
-  asm volatile(
-      "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
-      "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
-      "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
-      "movi       v7.8b, #16                     \n"  // Add 16 constant
-      "1:                                        \n"
-      "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v0.8b, v4.8b           \n"  // B
-      "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
-      "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
-      "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
-      "uqadd      v0.8b, v0.8b, v7.8b            \n"
-      "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
-      "b.gt       1b                             \n"
-      : "+r"(src_raw),  // %0
-        "+r"(dst_y),    // %1
-        "+r"(width)     // %2
-      :
-      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16");
+  RGBToYMatrixRow_NEON(src_raw, dst_y, width, &kRawI601Constants);
 }
 
 // Bilinear filter 16x2 -> 16x1
@@ -2068,44 +2916,49 @@ void InterpolateRow_NEON(uint8_t* dst_ptr,
   int y0_fraction = 256 - y1_fraction;
   const uint8_t* src_ptr1 = src_ptr + src_stride;
   asm volatile(
-      "cmp        %w4, #0                        \n"
-      "b.eq       100f                           \n"
-      "cmp        %w4, #128                      \n"
-      "b.eq       50f                            \n"
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
 
-      "dup        v5.16b, %w4                    \n"
-      "dup        v4.16b, %w5                    \n"
+      "dup         v5.16b, %w4                   \n"
+      "dup         v4.16b, %w5                   \n"
       // General purpose row blend.
       "1:                                        \n"
-      "ld1        {v0.16b}, [%1], #16            \n"
-      "ld1        {v1.16b}, [%2], #16            \n"
-      "subs       %w3, %w3, #16                  \n"
-      "umull      v2.8h, v0.8b,  v4.8b           \n"
-      "umull2     v3.8h, v0.16b, v4.16b          \n"
-      "umlal      v2.8h, v1.8b,  v5.8b           \n"
-      "umlal2     v3.8h, v1.16b, v5.16b          \n"
-      "rshrn      v0.8b,  v2.8h, #8              \n"
-      "rshrn2     v0.16b, v3.8h, #8              \n"
-      "st1        {v0.16b}, [%0], #16            \n"
-      "b.gt       1b                             \n"
-      "b          99f                            \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "umull       v2.8h, v0.8b,  v4.8b          \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.8h, v0.16b, v4.16b         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.8h, v1.8b,  v5.8b          \n"
+      "umlal2      v3.8h, v1.16b, v5.16b         \n"
+      "rshrn       v0.8b,  v2.8h, #8             \n"
+      "rshrn2      v0.16b, v3.8h, #8             \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
 
       // Blend 50 / 50.
       "50:                                       \n"
-      "ld1        {v0.16b}, [%1], #16            \n"
-      "ld1        {v1.16b}, [%2], #16            \n"
-      "subs       %w3, %w3, #16                  \n"
-      "urhadd     v0.16b, v0.16b, v1.16b         \n"
-      "st1        {v0.16b}, [%0], #16            \n"
-      "b.gt       50b                            \n"
-      "b          99f                            \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       "100:                                      \n"
-      "ld1        {v0.16b}, [%1], #16            \n"
-      "subs       %w3, %w3, #16                  \n"
-      "st1        {v0.16b}, [%0], #16            \n"
-      "b.gt       100b                           \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        100b                          \n"
 
       "99:                                       \n"
       : "+r"(dst_ptr),      // %0
@@ -2118,66 +2971,215 @@ void InterpolateRow_NEON(uint8_t* dst_ptr,
       : "cc", "memory", "v0", "v1", "v3", "v4", "v5");
 }
 
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_16_NEON(uint16_t* dst_ptr,
+                            const uint16_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            int dst_width,
+                            int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+
+  asm volatile(
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+
+      "dup         v5.8h, %w4                    \n"
+      "dup         v4.8h, %w5                    \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "umull       v2.4s, v0.4h, v4.4h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.4s, v0.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.4s, v1.4h, v5.4h           \n"
+      "umlal2      v3.4s, v1.8h, v5.8h           \n"
+      "rshrn       v0.4h, v2.4s, #8              \n"
+      "rshrn2      v0.8h, v3.4s, #8              \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.8h}, [%0], #16            \n"
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction)   // %5
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5");
+}
+
+// Bilinear filter 8x2 -> 8x1
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+void InterpolateRow_16To8_NEON(uint8_t* dst_ptr,
+                               const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               int scale,
+                               int dst_width,
+                               int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
+
+  asm volatile(
+      "dup         v6.8h, %w6                    \n"
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+
+      "dup         v5.8h, %w4                    \n"
+      "dup         v4.8h, %w5                    \n"
+      // General purpose row blend.
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "umull       v2.4s, v0.4h, v4.4h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull2      v3.4s, v0.8h, v4.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal       v2.4s, v1.4h, v5.4h           \n"
+      "umlal2      v3.4s, v1.8h, v5.8h           \n"
+      "rshrn       v0.4h, v2.4s, #8              \n"
+      "rshrn2      v0.8h, v3.4s, #8              \n"
+      "ushl        v0.8h, v0.8h, v6.8h           \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%0], #8             \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
+
+      // Blend 50 / 50.
+      "50:                                       \n"
+      "ld1         {v0.8h}, [%1], #16            \n"
+      "ld1         {v1.8h}, [%2], #16            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "urhadd      v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "ushl        v0.8h, v0.8h, v6.8h           \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%0], #8             \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
+
+      // Blend 100 / 0 - Copy row unchanged.
+      "100:                                      \n"
+      "ldr         q0, [%1], #16                 \n"
+      "ushl        v0.8h, v0.8h, v2.8h           \n"  // shr = v2 is negative
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "str         d0, [%0], #8                  \n"  // store 8 pixels
+      "b.gt        100b                          \n"
+
+      "99:                                       \n"
+      : "+r"(dst_ptr),     // %0
+        "+r"(src_ptr),     // %1
+        "+r"(src_ptr1),    // %2
+        "+r"(dst_width)    // %3
+      : "r"(y1_fraction),  // %4
+        "r"(y0_fraction),  // %5
+        "r"(shift)         // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6");
+}
+
 // dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
-void ARGBBlendRow_NEON(const uint8_t* src_argb0,
+void ARGBBlendRow_NEON(const uint8_t* src_argb,
                        const uint8_t* src_argb1,
                        uint8_t* dst_argb,
                        int width) {
   asm volatile(
-      "subs       %w3, %w3, #8                   \n"
-      "b.lt       89f                            \n"
+      "subs        %w3, %w3, #8                  \n"
+      "b.lt        89f                           \n"
       // Blend 8 pixels.
       "8:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0
-                                                            // pixels
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1
-                                                            // pixels
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
-      "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
-      "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
-      "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
-      "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
-      "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
-      "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
-      "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
-      "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
-      "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
-      "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
-      "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
-      "movi       v3.8b, #255                    \n"  // a = 255
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-                                                            // pixels
-      "b.ge       8b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v16.8h, v4.8b, v3.8b          \n"  // db * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v17.8h, v5.8b, v3.8b          \n"  // dg * a
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v18.8h, v6.8b, v3.8b          \n"  // dr * a
+      "uqrshrn     v16.8b, v16.8h, #8            \n"  // db >>= 8
+      "uqrshrn     v17.8b, v17.8h, #8            \n"  // dg >>= 8
+      "uqrshrn     v18.8b, v18.8h, #8            \n"  // dr >>= 8
+      "uqsub       v4.8b, v4.8b, v16.8b          \n"  // db - (db * a / 256)
+      "uqsub       v5.8b, v5.8b, v17.8b          \n"  // dg - (dg * a / 256)
+      "uqsub       v6.8b, v6.8b, v18.8b          \n"  // dr - (dr * a / 256)
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"  // + sb
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"  // + sg
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"  // + sr
+      "movi        v3.8b, #255                   \n"  // a = 255
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+                                                             // pixels
+      "b.ge        8b                            \n"
 
       "89:                                       \n"
-      "adds       %w3, %w3, #8-1                 \n"
-      "b.lt       99f                            \n"
+      "adds        %w3, %w3, #8-1                \n"
+      "b.lt        99f                           \n"
 
       // Blend 1 pixels.
       "1:                                        \n"
-      "ld4        {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel ARGB0.
-      "ld4        {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel ARGB1.
-      "subs       %w3, %w3, #1                   \n"  // 1 processed per loop.
-      "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
-      "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
-      "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
-      "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
-      "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
-      "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
-      "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
-      "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
-      "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
-      "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
-      "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
-      "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
-      "movi       v3.8b, #255                    \n"  // a = 255
-      "st4        {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
-      "b.ge       1b                             \n"
+      "ld4         {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel
+                                                           // ARGB0.
+      "ld4         {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel
+                                                           // ARGB1.
+      "subs        %w3, %w3, #1                  \n"  // 1 processed per loop.
+      "umull       v16.8h, v4.8b, v3.8b          \n"  // db * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v17.8h, v5.8b, v3.8b          \n"  // dg * a
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v18.8h, v6.8b, v3.8b          \n"  // dr * a
+      "uqrshrn     v16.8b, v16.8h, #8            \n"  // db >>= 8
+      "uqrshrn     v17.8b, v17.8h, #8            \n"  // dg >>= 8
+      "uqrshrn     v18.8b, v18.8h, #8            \n"  // dr >>= 8
+      "uqsub       v4.8b, v4.8b, v16.8b          \n"  // db - (db * a / 256)
+      "uqsub       v5.8b, v5.8b, v17.8b          \n"  // dg - (dg * a / 256)
+      "uqsub       v6.8b, v6.8b, v18.8b          \n"  // dr - (dr * a / 256)
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"  // + sb
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"  // + sg
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"  // + sr
+      "movi        v3.8b, #255                   \n"  // a = 255
+      "st4         {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
+      "b.ge        1b                            \n"
 
       "99:                                       \n"
 
-      : "+r"(src_argb0),  // %0
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2193,17 +3195,17 @@ void ARGBAttenuateRow_NEON(const uint8_t* src_argb,
   asm volatile(
       // Attenuate 8 pixels.
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v4.8h, v0.8b, v3.8b            \n"  // b * a
-      "umull      v5.8h, v1.8b, v3.8b            \n"  // g * a
-      "umull      v6.8h, v2.8b, v3.8b            \n"  // r * a
-      "uqrshrn    v0.8b, v4.8h, #8               \n"  // b >>= 8
-      "uqrshrn    v1.8b, v5.8h, #8               \n"  // g >>= 8
-      "uqrshrn    v2.8b, v6.8h, #8               \n"  // r >>= 8
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
-                                                            // pixels
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v3.8b           \n"  // b * a
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v5.8h, v1.8b, v3.8b           \n"         // g * a
+      "umull       v6.8h, v2.8b, v3.8b           \n"         // r * a
+      "uqrshrn     v0.8b, v4.8h, #8              \n"         // b >>= 8
+      "uqrshrn     v1.8b, v5.8h, #8              \n"         // g >>= 8
+      "uqrshrn     v2.8b, v6.8h, #8              \n"         // r >>= 8
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2219,32 +3221,33 @@ void ARGBQuantizeRow_NEON(uint8_t* dst_argb,
                           int interval_offset,
                           int width) {
   asm volatile(
-      "dup        v4.8h, %w2                     \n"
-      "ushr       v4.8h, v4.8h, #1               \n"  // scale >>= 1
-      "dup        v5.8h, %w3                     \n"  // interval multiply.
-      "dup        v6.8h, %w4                     \n"  // interval add
+      "dup         v4.8h, %w2                    \n"
+      "ushr        v4.8h, v4.8h, #1              \n"  // scale >>= 1
+      "dup         v5.8h, %w3                    \n"  // interval multiply.
+      "dup         v6.8h, %w4                    \n"  // interval add
 
       // 8 pixel loop.
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0]  \n"  // load 8  ARGB.
-      "subs       %w1, %w1, #8                   \n"    // 8 processed per loop.
-      "uxtl       v0.8h, v0.8b                   \n"    // b (0 .. 255)
-      "uxtl       v1.8h, v1.8b                   \n"
-      "uxtl       v2.8h, v2.8b                   \n"
-      "sqdmulh    v0.8h, v0.8h, v4.8h            \n"  // b * scale
-      "sqdmulh    v1.8h, v1.8h, v4.8h            \n"  // g
-      "sqdmulh    v2.8h, v2.8h, v4.8h            \n"  // r
-      "mul        v0.8h, v0.8h, v5.8h            \n"  // b * interval_size
-      "mul        v1.8h, v1.8h, v5.8h            \n"  // g
-      "mul        v2.8h, v2.8h, v5.8h            \n"  // r
-      "add        v0.8h, v0.8h, v6.8h            \n"  // b + interval_offset
-      "add        v1.8h, v1.8h, v6.8h            \n"  // g
-      "add        v2.8h, v2.8h, v6.8h            \n"  // r
-      "uqxtn      v0.8b, v0.8h                   \n"
-      "uqxtn      v1.8b, v1.8h                   \n"
-      "uqxtn      v2.8b, v2.8h                   \n"
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8  ARGB.
+      "subs        %w1, %w1, #8                  \n"    // 8 processed per loop.
+      "uxtl        v0.8h, v0.8b                  \n"    // b (0 .. 255)
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v1.8h, v1.8b                  \n"
+      "uxtl        v2.8h, v2.8b                  \n"
+      "sqdmulh     v0.8h, v0.8h, v4.8h           \n"  // b * scale
+      "sqdmulh     v1.8h, v1.8h, v4.8h           \n"  // g
+      "sqdmulh     v2.8h, v2.8h, v4.8h           \n"  // r
+      "mul         v0.8h, v0.8h, v5.8h           \n"  // b * interval_size
+      "mul         v1.8h, v1.8h, v5.8h           \n"  // g
+      "mul         v2.8h, v2.8h, v5.8h           \n"  // r
+      "add         v0.8h, v0.8h, v6.8h           \n"  // b + interval_offset
+      "add         v1.8h, v1.8h, v6.8h           \n"  // g
+      "add         v2.8h, v2.8h, v6.8h           \n"  // r
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn       v1.8b, v1.8h                  \n"
+      "uqxtn       v2.8b, v2.8h                  \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(dst_argb),       // %0
         "+r"(width)           // %1
       : "r"(scale),           // %2
@@ -2261,28 +3264,29 @@ void ARGBShadeRow_NEON(const uint8_t* src_argb,
                        int width,
                        uint32_t value) {
   asm volatile(
-      "dup        v0.4s, %w3                     \n"  // duplicate scale value.
-      "zip1       v0.8b, v0.8b, v0.8b            \n"  // v0.8b aarrggbb.
-      "ushr       v0.8h, v0.8h, #1               \n"  // scale / 2.
+      "dup         v0.4s, %w3                    \n"  // duplicate scale value.
+      "zip1        v0.8b, v0.8b, v0.8b           \n"  // v0.8b aarrggbb.
+      "ushr        v0.8h, v0.8h, #1              \n"  // scale / 2.
 
       // 8 pixel loop.
       "1:                                        \n"
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "uxtl       v4.8h, v4.8b                   \n"  // b (0 .. 255)
-      "uxtl       v5.8h, v5.8b                   \n"
-      "uxtl       v6.8h, v6.8b                   \n"
-      "uxtl       v7.8h, v7.8b                   \n"
-      "sqrdmulh   v4.8h, v4.8h, v0.h[0]          \n"  // b * scale * 2
-      "sqrdmulh   v5.8h, v5.8h, v0.h[1]          \n"  // g
-      "sqrdmulh   v6.8h, v6.8h, v0.h[2]          \n"  // r
-      "sqrdmulh   v7.8h, v7.8h, v0.h[3]          \n"  // a
-      "uqxtn      v4.8b, v4.8h                   \n"
-      "uqxtn      v5.8b, v5.8h                   \n"
-      "uqxtn      v6.8b, v6.8h                   \n"
-      "uqxtn      v7.8b, v7.8h                   \n"
-      "st4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "uxtl        v4.8h, v4.8b                  \n"  // b (0 .. 255)
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v5.8h, v5.8b                  \n"
+      "uxtl        v6.8h, v6.8b                  \n"
+      "uxtl        v7.8h, v7.8b                  \n"
+      "sqrdmulh    v4.8h, v4.8h, v0.h[0]         \n"  // b * scale * 2
+      "sqrdmulh    v5.8h, v5.8h, v0.h[1]         \n"  // g
+      "sqrdmulh    v6.8h, v6.8h, v0.h[2]         \n"  // r
+      "sqrdmulh    v7.8h, v7.8h, v0.h[3]         \n"  // a
+      "uqxtn       v4.8b, v4.8h                  \n"
+      "uqxtn       v5.8b, v5.8h                  \n"
+      "uqxtn       v6.8b, v6.8h                  \n"
+      "uqxtn       v7.8b, v7.8h                  \n"
+      "st4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2292,23 +3296,24 @@ void ARGBShadeRow_NEON(const uint8_t* src_argb,
 
 // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
 // Similar to ARGBToYJ but stores ARGB.
-// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+// C code is (29 * b + 150 * g + 77 * r + 128) >> 8;
 void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
   asm volatile(
-      "movi       v24.8b, #15                    \n"  // B * 0.11400 coefficient
-      "movi       v25.8b, #75                    \n"  // G * 0.58700 coefficient
-      "movi       v26.8b, #38                    \n"  // R * 0.29900 coefficient
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "umull      v4.8h, v0.8b, v24.8b           \n"  // B
-      "umlal      v4.8h, v1.8b, v25.8b           \n"  // G
-      "umlal      v4.8h, v2.8b, v26.8b           \n"  // R
-      "sqrshrun   v0.8b, v4.8h, #7               \n"  // 15 bit to 8 bit B
-      "orr        v1.8b, v0.8b, v0.8b            \n"  // G
-      "orr        v2.8b, v0.8b, v0.8b            \n"  // R
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
-      "b.gt       1b                             \n"
+      "movi        v24.8b, #29                   \n"  // B * 0.1140 coefficient
+      "movi        v25.8b, #150                  \n"  // G * 0.5870 coefficient
+      "movi        v26.8b, #77                   \n"  // R * 0.2990 coefficient
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v24.8b          \n"  // B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v4.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v4.8h, v2.8b, v26.8b          \n"  // R
+      "uqrshrn     v0.8b, v4.8h, #8              \n"  // 16 bit to 8 bit B
+      "orr         v1.8b, v0.8b, v0.8b           \n"  // G
+      "orr         v2.8b, v0.8b, v0.8b           \n"  // R
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(width)      // %2
@@ -2323,32 +3328,33 @@ void ARGBGrayRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width) {
 
 void ARGBSepiaRow_NEON(uint8_t* dst_argb, int width) {
   asm volatile(
-      "movi       v20.8b, #17                    \n"  // BB coefficient
-      "movi       v21.8b, #68                    \n"  // BG coefficient
-      "movi       v22.8b, #35                    \n"  // BR coefficient
-      "movi       v24.8b, #22                    \n"  // GB coefficient
-      "movi       v25.8b, #88                    \n"  // GG coefficient
-      "movi       v26.8b, #45                    \n"  // GR coefficient
-      "movi       v28.8b, #24                    \n"  // BB coefficient
-      "movi       v29.8b, #98                    \n"  // BG coefficient
-      "movi       v30.8b, #50                    \n"  // BR coefficient
-      "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
-      "subs       %w1, %w1, #8                   \n"   // 8 processed per loop.
-      "umull      v4.8h, v0.8b, v20.8b           \n"   // B to Sepia B
-      "umlal      v4.8h, v1.8b, v21.8b           \n"   // G
-      "umlal      v4.8h, v2.8b, v22.8b           \n"   // R
-      "umull      v5.8h, v0.8b, v24.8b           \n"   // B to Sepia G
-      "umlal      v5.8h, v1.8b, v25.8b           \n"   // G
-      "umlal      v5.8h, v2.8b, v26.8b           \n"   // R
-      "umull      v6.8h, v0.8b, v28.8b           \n"   // B to Sepia R
-      "umlal      v6.8h, v1.8b, v29.8b           \n"   // G
-      "umlal      v6.8h, v2.8b, v30.8b           \n"   // R
-      "uqshrn     v0.8b, v4.8h, #7               \n"   // 16 bit to 8 bit B
-      "uqshrn     v1.8b, v5.8h, #7               \n"   // 16 bit to 8 bit G
-      "uqshrn     v2.8b, v6.8h, #7               \n"   // 16 bit to 8 bit R
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
-      "b.gt       1b                             \n"
+      "movi        v20.8b, #17                   \n"  // BB coefficient
+      "movi        v21.8b, #68                   \n"  // BG coefficient
+      "movi        v22.8b, #35                   \n"  // BR coefficient
+      "movi        v24.8b, #22                   \n"  // GB coefficient
+      "movi        v25.8b, #88                   \n"  // GG coefficient
+      "movi        v26.8b, #45                   \n"  // GR coefficient
+      "movi        v28.8b, #24                   \n"  // BB coefficient
+      "movi        v29.8b, #98                   \n"  // BG coefficient
+      "movi        v30.8b, #50                   \n"  // BR coefficient
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
+      "subs        %w1, %w1, #8                  \n"    // 8 processed per loop.
+      "umull       v4.8h, v0.8b, v20.8b          \n"    // B to Sepia B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umlal       v4.8h, v1.8b, v21.8b          \n"  // G
+      "umlal       v4.8h, v2.8b, v22.8b          \n"  // R
+      "umull       v5.8h, v0.8b, v24.8b          \n"  // B to Sepia G
+      "umlal       v5.8h, v1.8b, v25.8b          \n"  // G
+      "umlal       v5.8h, v2.8b, v26.8b          \n"  // R
+      "umull       v6.8h, v0.8b, v28.8b          \n"  // B to Sepia R
+      "umlal       v6.8h, v1.8b, v29.8b          \n"  // G
+      "umlal       v6.8h, v2.8b, v30.8b          \n"  // R
+      "uqshrn      v0.8b, v4.8h, #7              \n"  // 16 bit to 8 bit B
+      "uqshrn      v1.8b, v5.8h, #7              \n"  // 16 bit to 8 bit G
+      "uqshrn      v2.8b, v6.8h, #7              \n"  // 16 bit to 8 bit R
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
+      "b.gt        1b                            \n"
       : "+r"(dst_argb),  // %0
         "+r"(width)      // %1
       :
@@ -2364,51 +3370,52 @@ void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
                              const int8_t* matrix_argb,
                              int width) {
   asm volatile(
-      "ld1        {v2.16b}, [%3]                 \n"  // load 3 ARGB vectors.
-      "sxtl       v0.8h, v2.8b                   \n"  // B,G coefficients s16.
-      "sxtl2      v1.8h, v2.16b                  \n"  // R,A coefficients s16.
-
-      "1:                                        \n"
-      "ld4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 ARGB
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
-      "uxtl       v16.8h, v16.8b                 \n"  // b (0 .. 255) 16 bit
-      "uxtl       v17.8h, v17.8b                 \n"  // g
-      "uxtl       v18.8h, v18.8b                 \n"  // r
-      "uxtl       v19.8h, v19.8b                 \n"  // a
-      "mul        v22.8h, v16.8h, v0.h[0]        \n"  // B = B * Matrix B
-      "mul        v23.8h, v16.8h, v0.h[4]        \n"  // G = B * Matrix G
-      "mul        v24.8h, v16.8h, v1.h[0]        \n"  // R = B * Matrix R
-      "mul        v25.8h, v16.8h, v1.h[4]        \n"  // A = B * Matrix A
-      "mul        v4.8h, v17.8h, v0.h[1]         \n"  // B += G * Matrix B
-      "mul        v5.8h, v17.8h, v0.h[5]         \n"  // G += G * Matrix G
-      "mul        v6.8h, v17.8h, v1.h[1]         \n"  // R += G * Matrix R
-      "mul        v7.8h, v17.8h, v1.h[5]         \n"  // A += G * Matrix A
-      "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-      "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-      "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-      "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-      "mul        v4.8h, v18.8h, v0.h[2]         \n"  // B += R * Matrix B
-      "mul        v5.8h, v18.8h, v0.h[6]         \n"  // G += R * Matrix G
-      "mul        v6.8h, v18.8h, v1.h[2]         \n"  // R += R * Matrix R
-      "mul        v7.8h, v18.8h, v1.h[6]         \n"  // A += R * Matrix A
-      "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-      "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-      "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-      "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-      "mul        v4.8h, v19.8h, v0.h[3]         \n"  // B += A * Matrix B
-      "mul        v5.8h, v19.8h, v0.h[7]         \n"  // G += A * Matrix G
-      "mul        v6.8h, v19.8h, v1.h[3]         \n"  // R += A * Matrix R
-      "mul        v7.8h, v19.8h, v1.h[7]         \n"  // A += A * Matrix A
-      "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
-      "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
-      "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
-      "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
-      "sqshrun    v16.8b, v22.8h, #6             \n"  // 16 bit to 8 bit B
-      "sqshrun    v17.8b, v23.8h, #6             \n"  // 16 bit to 8 bit G
-      "sqshrun    v18.8b, v24.8h, #6             \n"  // 16 bit to 8 bit R
-      "sqshrun    v19.8b, v25.8h, #6             \n"  // 16 bit to 8 bit A
-      "st4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld1         {v2.16b}, [%3]                \n"  // load 3 ARGB vectors.
+      "sxtl        v0.8h, v2.8b                  \n"  // B,G coefficients s16.
+      "sxtl2       v1.8h, v2.16b                 \n"  // R,A coefficients s16.
+
+      "1:                                        \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 ARGB
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop.
+      "uxtl        v16.8h, v16.8b                \n"  // b (0 .. 255) 16 bit
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v17.8h, v17.8b                \n"  // g
+      "uxtl        v18.8h, v18.8b                \n"  // r
+      "uxtl        v19.8h, v19.8b                \n"  // a
+      "mul         v22.8h, v16.8h, v0.h[0]       \n"  // B = B * Matrix B
+      "mul         v23.8h, v16.8h, v0.h[4]       \n"  // G = B * Matrix G
+      "mul         v24.8h, v16.8h, v1.h[0]       \n"  // R = B * Matrix R
+      "mul         v25.8h, v16.8h, v1.h[4]       \n"  // A = B * Matrix A
+      "mul         v4.8h, v17.8h, v0.h[1]        \n"  // B += G * Matrix B
+      "mul         v5.8h, v17.8h, v0.h[5]        \n"  // G += G * Matrix G
+      "mul         v6.8h, v17.8h, v1.h[1]        \n"  // R += G * Matrix R
+      "mul         v7.8h, v17.8h, v1.h[5]        \n"  // A += G * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "mul         v4.8h, v18.8h, v0.h[2]        \n"  // B += R * Matrix B
+      "mul         v5.8h, v18.8h, v0.h[6]        \n"  // G += R * Matrix G
+      "mul         v6.8h, v18.8h, v1.h[2]        \n"  // R += R * Matrix R
+      "mul         v7.8h, v18.8h, v1.h[6]        \n"  // A += R * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "mul         v4.8h, v19.8h, v0.h[3]        \n"  // B += A * Matrix B
+      "mul         v5.8h, v19.8h, v0.h[7]        \n"  // G += A * Matrix G
+      "mul         v6.8h, v19.8h, v1.h[3]        \n"  // R += A * Matrix R
+      "mul         v7.8h, v19.8h, v1.h[7]        \n"  // A += A * Matrix A
+      "sqadd       v22.8h, v22.8h, v4.8h         \n"  // Accumulate B
+      "sqadd       v23.8h, v23.8h, v5.8h         \n"  // Accumulate G
+      "sqadd       v24.8h, v24.8h, v6.8h         \n"  // Accumulate R
+      "sqadd       v25.8h, v25.8h, v7.8h         \n"  // Accumulate A
+      "sqshrun     v16.8b, v22.8h, #6            \n"  // 16 bit to 8 bit B
+      "sqshrun     v17.8b, v23.8h, #6            \n"  // 16 bit to 8 bit G
+      "sqshrun     v18.8b, v24.8h, #6            \n"  // 16 bit to 8 bit R
+      "sqshrun     v19.8b, v25.8h, #6            \n"  // 16 bit to 8 bit A
+      "st4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_argb),   // %0
         "+r"(dst_argb),   // %1
         "+r"(width)       // %2
@@ -2419,27 +3426,29 @@ void ARGBColorMatrixRow_NEON(const uint8_t* src_argb,
 
 // TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBMultiplyRow_NEON(const uint8_t* src_argb0,
+void ARGBMultiplyRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "umull      v0.8h, v0.8b, v4.8b            \n"  // multiply B
-      "umull      v1.8h, v1.8b, v5.8b            \n"  // multiply G
-      "umull      v2.8h, v2.8b, v6.8b            \n"  // multiply R
-      "umull      v3.8h, v3.8b, v7.8b            \n"  // multiply A
-      "rshrn      v0.8b, v0.8h, #8               \n"  // 16 bit to 8 bit B
-      "rshrn      v1.8b, v1.8h, #8               \n"  // 16 bit to 8 bit G
-      "rshrn      v2.8b, v2.8h, #8               \n"  // 16 bit to 8 bit R
-      "rshrn      v3.8b, v3.8h, #8               \n"  // 16 bit to 8 bit A
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "umull       v0.8h, v0.8b, v4.8b           \n"  // multiply B
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "umull       v1.8h, v1.8b, v5.8b           \n"  // multiply G
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umull       v2.8h, v2.8b, v6.8b           \n"  // multiply R
+      "umull       v3.8h, v3.8b, v7.8b           \n"  // multiply A
+      "rshrn       v0.8b, v0.8h, #8              \n"  // 16 bit to 8 bit B
+      "rshrn       v1.8b, v1.8h, #8              \n"  // 16 bit to 8 bit G
+      "rshrn       v2.8b, v2.8h, #8              \n"  // 16 bit to 8 bit R
+      "rshrn       v3.8b, v3.8h, #8              \n"  // 16 bit to 8 bit A
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2448,23 +3457,25 @@ void ARGBMultiplyRow_NEON(const uint8_t* src_argb0,
 }
 
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-void ARGBAddRow_NEON(const uint8_t* src_argb0,
+void ARGBAddRow_NEON(const uint8_t* src_argb,
                      const uint8_t* src_argb1,
                      uint8_t* dst_argb,
                      int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uqadd      v0.8b, v0.8b, v4.8b            \n"
-      "uqadd      v1.8b, v1.8b, v5.8b            \n"
-      "uqadd      v2.8b, v2.8b, v6.8b            \n"
-      "uqadd      v3.8b, v3.8b, v7.8b            \n"
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v0.8b, v0.8b, v4.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v1.8b, v1.8b, v5.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqadd       v2.8b, v2.8b, v6.8b           \n"
+      "uqadd       v3.8b, v3.8b, v7.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2473,23 +3484,25 @@ void ARGBAddRow_NEON(const uint8_t* src_argb0,
 }
 
 // Subtract 2 rows of ARGB pixels, 8 pixels at a time.
-void ARGBSubtractRow_NEON(const uint8_t* src_argb0,
+void ARGBSubtractRow_NEON(const uint8_t* src_argb,
                           const uint8_t* src_argb1,
                           uint8_t* dst_argb,
                           int width) {
   asm volatile(
       // 8 pixel loop.
       "1:                                        \n"
-      "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
-      "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uqsub      v0.8b, v0.8b, v4.8b            \n"
-      "uqsub      v1.8b, v1.8b, v5.8b            \n"
-      "uqsub      v2.8b, v2.8b, v6.8b            \n"
-      "uqsub      v3.8b, v3.8b, v7.8b            \n"
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
-      : "+r"(src_argb0),  // %0
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqsub       v0.8b, v0.8b, v4.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqsub       v1.8b, v1.8b, v5.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqsub       v2.8b, v2.8b, v6.8b           \n"
+      "uqsub       v3.8b, v3.8b, v7.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
+      : "+r"(src_argb),   // %0
         "+r"(src_argb1),  // %1
         "+r"(dst_argb),   // %2
         "+r"(width)       // %3
@@ -2507,17 +3520,19 @@ void SobelRow_NEON(const uint8_t* src_sobelx,
                    uint8_t* dst_argb,
                    int width) {
   asm volatile(
-      "movi       v3.8b, #255                    \n"  // alpha
+      "movi        v3.8b, #255                   \n"  // alpha
       // 8 pixel loop.
       "1:                                        \n"
-      "ld1        {v0.8b}, [%0], #8              \n"  // load 8 sobelx.
-      "ld1        {v1.8b}, [%1], #8              \n"  // load 8 sobely.
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uqadd      v0.8b, v0.8b, v1.8b            \n"  // add
-      "orr        v1.8b, v0.8b, v0.8b            \n"
-      "orr        v2.8b, v0.8b, v0.8b            \n"
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld1         {v0.8b}, [%0], #8             \n"  // load 8 sobelx.
+      "ld1         {v1.8b}, [%1], #8             \n"  // load 8 sobely.
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uqadd       v0.8b, v0.8b, v1.8b           \n"  // add
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "orr         v1.8b, v0.8b, v0.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "orr         v2.8b, v0.8b, v0.8b           \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -2534,12 +3549,14 @@ void SobelToPlaneRow_NEON(const uint8_t* src_sobelx,
   asm volatile(
       // 16 pixel loop.
       "1:                                        \n"
-      "ld1        {v0.16b}, [%0], #16            \n"  // load 16 sobelx.
-      "ld1        {v1.16b}, [%1], #16            \n"  // load 16 sobely.
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
-      "uqadd      v0.16b, v0.16b, v1.16b         \n"  // add
-      "st1        {v0.16b}, [%2], #16            \n"  // store 16 pixels.
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 sobelx.
+      "ld1         {v1.16b}, [%1], #16           \n"  // load 16 sobely.
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v0.16b, v0.16b, v1.16b        \n"  // add
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st1         {v0.16b}, [%2], #16           \n"  // store 16 pixels.
+      "b.gt        1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_y),       // %2
@@ -2558,15 +3575,17 @@ void SobelXYRow_NEON(const uint8_t* src_sobelx,
                      uint8_t* dst_argb,
                      int width) {
   asm volatile(
-      "movi       v3.8b, #255                    \n"  // alpha
+      "movi        v3.8b, #255                   \n"  // alpha
       // 8 pixel loop.
       "1:                                        \n"
-      "ld1        {v2.8b}, [%0], #8              \n"  // load 8 sobelx.
-      "ld1        {v0.8b}, [%1], #8              \n"  // load 8 sobely.
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uqadd      v1.8b, v0.8b, v2.8b            \n"  // add
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
-      "b.gt       1b                             \n"
+      "ld1         {v2.8b}, [%0], #8             \n"  // load 8 sobelx.
+      "ld1         {v0.8b}, [%1], #8             \n"  // load 8 sobely.
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqadd       v1.8b, v0.8b, v2.8b           \n"  // add
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB
+      "b.gt        1b                            \n"
       : "+r"(src_sobelx),  // %0
         "+r"(src_sobely),  // %1
         "+r"(dst_argb),    // %2
@@ -2586,23 +3605,26 @@ void SobelXRow_NEON(const uint8_t* src_y0,
                     int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.8b}, [%0],%5               \n"  // top
-      "ld1        {v1.8b}, [%0],%6               \n"
-      "usubl      v0.8h, v0.8b, v1.8b            \n"
-      "ld1        {v2.8b}, [%1],%5               \n"  // center * 2
-      "ld1        {v3.8b}, [%1],%6               \n"
-      "usubl      v1.8h, v2.8b, v3.8b            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "ld1        {v2.8b}, [%2],%5               \n"  // bottom
-      "ld1        {v3.8b}, [%2],%6               \n"
-      "subs       %w4, %w4, #8                   \n"  // 8 pixels
-      "usubl      v1.8h, v2.8b, v3.8b            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "abs        v0.8h, v0.8h                   \n"
-      "uqxtn      v0.8b, v0.8h                   \n"
-      "st1        {v0.8b}, [%3], #8              \n"  // store 8 sobelx
-      "b.gt       1b                             \n"
+      "ld1         {v0.8b}, [%0],%5              \n"  // top
+      "ld1         {v1.8b}, [%0],%6              \n"
+      "usubl       v0.8h, v0.8b, v1.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ld1         {v2.8b}, [%1],%5              \n"  // center * 2
+      "ld1         {v3.8b}, [%1],%6              \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "ld1         {v2.8b}, [%2],%5              \n"  // bottom
+      "ld1         {v3.8b}, [%2],%6              \n"
+      "subs        %w4, %w4, #8                  \n"  // 8 pixels
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "abs         v0.8h, v0.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%3], #8             \n"  // store 8 sobelx
+      "b.gt        1b                            \n"
       : "+r"(src_y0),                           // %0
         "+r"(src_y1),                           // %1
         "+r"(src_y2),                           // %2
@@ -2624,23 +3646,25 @@ void SobelYRow_NEON(const uint8_t* src_y0,
                     int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.8b}, [%0],%4               \n"  // left
-      "ld1        {v1.8b}, [%1],%4               \n"
-      "usubl      v0.8h, v0.8b, v1.8b            \n"
-      "ld1        {v2.8b}, [%0],%4               \n"  // center * 2
-      "ld1        {v3.8b}, [%1],%4               \n"
-      "usubl      v1.8h, v2.8b, v3.8b            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "ld1        {v2.8b}, [%0],%5               \n"  // right
-      "ld1        {v3.8b}, [%1],%5               \n"
-      "subs       %w3, %w3, #8                   \n"  // 8 pixels
-      "usubl      v1.8h, v2.8b, v3.8b            \n"
-      "add        v0.8h, v0.8h, v1.8h            \n"
-      "abs        v0.8h, v0.8h                   \n"
-      "uqxtn      v0.8b, v0.8h                   \n"
-      "st1        {v0.8b}, [%2], #8              \n"  // store 8 sobely
-      "b.gt       1b                             \n"
+      "ld1         {v0.8b}, [%0],%4              \n"  // left
+      "ld1         {v1.8b}, [%1],%4              \n"
+      "usubl       v0.8h, v0.8b, v1.8b           \n"
+      "ld1         {v2.8b}, [%0],%4              \n"  // center * 2
+      "ld1         {v3.8b}, [%1],%4              \n"
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "ld1         {v2.8b}, [%0],%5              \n"  // right
+      "ld1         {v3.8b}, [%1],%5              \n"
+      "subs        %w3, %w3, #8                  \n"  // 8 pixels
+      "usubl       v1.8h, v2.8b, v3.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "add         v0.8h, v0.8h, v1.8h           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "abs         v0.8h, v0.8h                  \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "st1         {v0.8b}, [%2], #8             \n"  // store 8 sobely
+      "b.gt        1b                            \n"
       : "+r"(src_y0),                           // %0
         "+r"(src_y1),                           // %1
         "+r"(dst_sobely),                       // %2
@@ -2658,16 +3682,17 @@ void HalfFloat1Row_NEON(const uint16_t* src,
                         int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
-      "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
-      "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
-      "uxtl2      v3.4s, v1.8h                   \n"
-      "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
-      "scvtf      v3.4s, v3.4s                   \n"
-      "fcvtn      v1.4h, v2.4s                   \n"  // 8 half floats
-      "fcvtn2     v1.8h, v3.4s                   \n"
-      "st1        {v1.16b}, [%1], #16            \n"  // store 8 shorts
-      "b.gt       1b                             \n"
+      "ld1         {v1.16b}, [%0], #16           \n"  // load 8 shorts
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 int's
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fcvtn       v1.4h, v2.4s                  \n"  // 8 half floats
+      "fcvtn2      v1.8h, v3.4s                  \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 8 shorts
+      "b.gt        1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -2681,18 +3706,19 @@ void HalfFloatRow_NEON(const uint16_t* src,
                        int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v1.16b}, [%0], #16            \n"  // load 8 shorts
-      "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
-      "uxtl       v2.4s, v1.4h                   \n"  // 8 int's
-      "uxtl2      v3.4s, v1.8h                   \n"
-      "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
-      "scvtf      v3.4s, v3.4s                   \n"
-      "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // adjust exponent
-      "fmul       v3.4s, v3.4s, %3.s[0]          \n"
-      "uqshrn     v1.4h, v2.4s, #13              \n"  // isolate halffloat
-      "uqshrn2    v1.8h, v3.4s, #13              \n"
-      "st1        {v1.16b}, [%1], #16            \n"  // store 8 shorts
-      "b.gt       1b                             \n"
+      "ld1         {v1.16b}, [%0], #16           \n"  // load 8 shorts
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 int's
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // adjust exponent
+      "fmul        v3.4s, v3.4s, %3.s[0]         \n"
+      "uqshrn      v1.4h, v2.4s, #13             \n"  // isolate halffloat
+      "uqshrn2     v1.8h, v3.4s, #13             \n"
+      "st1         {v1.16b}, [%1], #16           \n"  // store 8 shorts
+      "b.gt        1b                            \n"
       : "+r"(src),                      // %0
         "+r"(dst),                      // %1
         "+r"(width)                     // %2
@@ -2706,17 +3732,18 @@ void ByteToFloatRow_NEON(const uint8_t* src,
                          int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v1.8b}, [%0], #8              \n"  // load 8 bytes
-      "subs       %w2, %w2, #8                   \n"  // 8 pixels per loop
-      "uxtl       v1.8h, v1.8b                   \n"  // 8 shorts
-      "uxtl       v2.4s, v1.4h                   \n"  // 8 ints
-      "uxtl2      v3.4s, v1.8h                   \n"
-      "scvtf      v2.4s, v2.4s                   \n"  // 8 floats
-      "scvtf      v3.4s, v3.4s                   \n"
-      "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // scale
-      "fmul       v3.4s, v3.4s, %3.s[0]          \n"
-      "st1        {v2.16b, v3.16b}, [%1], #32    \n"  // store 8 floats
-      "b.gt       1b                             \n"
+      "ld1         {v1.8b}, [%0], #8             \n"  // load 8 bytes
+      "subs        %w2, %w2, #8                  \n"  // 8 pixels per loop
+      "uxtl        v1.8h, v1.8b                  \n"  // 8 shorts
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uxtl        v2.4s, v1.4h                  \n"  // 8 ints
+      "uxtl2       v3.4s, v1.8h                  \n"
+      "scvtf       v2.4s, v2.4s                  \n"  // 8 floats
+      "scvtf       v3.4s, v3.4s                  \n"
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // scale
+      "fmul        v3.4s, v3.4s, %3.s[0]         \n"
+      "st1         {v2.16b, v3.16b}, [%1], #32   \n"  // store 8 floats
+      "b.gt        1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -2730,20 +3757,21 @@ float ScaleMaxSamples_NEON(const float* src,
                            int width) {
   float fmax;
   asm volatile(
-      "movi       v5.4s, #0                      \n"  // max
-      "movi       v6.4s, #0                      \n"
+      "movi        v5.4s, #0                     \n"  // max
+      "movi        v6.4s, #0                     \n"
 
       "1:                                        \n"
-      "ld1        {v1.4s, v2.4s}, [%0], #32      \n"  // load 8 samples
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "fmul       v3.4s, v1.4s, %4.s[0]          \n"  // scale
-      "fmul       v4.4s, v2.4s, %4.s[0]          \n"  // scale
-      "fmax       v5.4s, v5.4s, v1.4s            \n"  // max
-      "fmax       v6.4s, v6.4s, v2.4s            \n"
-      "st1        {v3.4s, v4.4s}, [%1], #32      \n"  // store 8 samples
-      "b.gt       1b                             \n"
-      "fmax       v5.4s, v5.4s, v6.4s            \n"  // max
-      "fmaxv      %s3, v5.4s                     \n"  // signed max acculator
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v3.4s, v1.4s, %4.s[0]         \n"  // scale
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v4.4s, v2.4s, %4.s[0]         \n"  // scale
+      "fmax        v5.4s, v5.4s, v1.4s           \n"  // max
+      "fmax        v6.4s, v6.4s, v2.4s           \n"
+      "st1         {v3.4s, v4.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      "fmax        v5.4s, v5.4s, v6.4s           \n"  // max
+      "fmaxv       %s3, v5.4s                    \n"  // signed max acculator
       : "+r"(src),                                    // %0
         "+r"(dst),                                    // %1
         "+r"(width),                                  // %2
@@ -2759,21 +3787,22 @@ float ScaleSumSamples_NEON(const float* src,
                            int width) {
   float fsum;
   asm volatile(
-      "movi       v5.4s, #0                      \n"  // max
-      "movi       v6.4s, #0                      \n"  // max
-
-      "1:                                        \n"
-      "ld1        {v1.4s, v2.4s}, [%0], #32      \n"  // load 8 samples
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "fmul       v3.4s, v1.4s, %4.s[0]          \n"  // scale
-      "fmul       v4.4s, v2.4s, %4.s[0]          \n"
-      "fmla       v5.4s, v1.4s, v1.4s            \n"  // sum of squares
-      "fmla       v6.4s, v2.4s, v2.4s            \n"
-      "st1        {v3.4s, v4.4s}, [%1], #32      \n"  // store 8 samples
-      "b.gt       1b                             \n"
-      "faddp      v5.4s, v5.4s, v6.4s            \n"
-      "faddp      v5.4s, v5.4s, v5.4s            \n"
-      "faddp      %3.4s, v5.4s, v5.4s            \n"  // sum
+      "movi        v5.4s, #0                     \n"  // max
+      "movi        v6.4s, #0                     \n"  // max
+
+      "1:                                        \n"
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v3.4s, v1.4s, %4.s[0]         \n"  // scale
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v4.4s, v2.4s, %4.s[0]         \n"
+      "fmla        v5.4s, v1.4s, v1.4s           \n"  // sum of squares
+      "fmla        v6.4s, v2.4s, v2.4s           \n"
+      "st1         {v3.4s, v4.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      "faddp       v5.4s, v5.4s, v6.4s           \n"
+      "faddp       v5.4s, v5.4s, v5.4s           \n"
+      "faddp       %3.4s, v5.4s, v5.4s           \n"  // sum
       : "+r"(src),                                    // %0
         "+r"(dst),                                    // %1
         "+r"(width),                                  // %2
@@ -2786,12 +3815,13 @@ float ScaleSumSamples_NEON(const float* src,
 void ScaleSamples_NEON(const float* src, float* dst, float scale, int width) {
   asm volatile(
       "1:                                        \n"
-      "ld1        {v1.4s, v2.4s}, [%0], #32      \n"  // load 8 samples
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "fmul       v1.4s, v1.4s, %3.s[0]          \n"  // scale
-      "fmul       v2.4s, v2.4s, %3.s[0]          \n"  // scale
-      "st1        {v1.4s, v2.4s}, [%1], #32      \n"  // store 8 samples
-      "b.gt       1b                             \n"
+      "ld1         {v1.4s, v2.4s}, [%0], #32     \n"  // load 8 samples
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "fmul        v1.4s, v1.4s, %3.s[0]         \n"  // scale
+      "fmul        v2.4s, v2.4s, %3.s[0]         \n"  // scale
+      "st1         {v1.4s, v2.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
       : "+r"(src),   // %0
         "+r"(dst),   // %1
         "+r"(width)  // %2
@@ -2808,26 +3838,31 @@ void GaussCol_NEON(const uint16_t* src0,
                    uint32_t* dst,
                    int width) {
   asm volatile(
-      "movi       v6.8h, #4                      \n"  // constant 4
-      "movi       v7.8h, #6                      \n"  // constant 6
-
-      "1:                                        \n"
-      "ld1        {v1.8h}, [%0], #16             \n"  // load 8 samples, 5 rows
-      "ld1        {v2.8h}, [%4], #16             \n"
-      "uaddl      v0.4s, v1.4h, v2.4h            \n"  // * 1
-      "uaddl2     v1.4s, v1.8h, v2.8h            \n"  // * 1
-      "ld1        {v2.8h}, [%1], #16             \n"
-      "umlal      v0.4s, v2.4h, v6.4h            \n"  // * 4
-      "umlal2     v1.4s, v2.8h, v6.8h            \n"  // * 4
-      "ld1        {v2.8h}, [%2], #16             \n"
-      "umlal      v0.4s, v2.4h, v7.4h            \n"  // * 6
-      "umlal2     v1.4s, v2.8h, v7.8h            \n"  // * 6
-      "ld1        {v2.8h}, [%3], #16             \n"
-      "umlal      v0.4s, v2.4h, v6.4h            \n"  // * 4
-      "umlal2     v1.4s, v2.8h, v6.8h            \n"  // * 4
-      "subs       %w6, %w6, #8                   \n"  // 8 processed per loop
-      "st1        {v0.4s,v1.4s}, [%5], #32       \n"  // store 8 samples
-      "b.gt       1b                             \n"
+      "movi        v6.8h, #4                     \n"  // constant 4
+      "movi        v7.8h, #6                     \n"  // constant 6
+
+      "1:                                        \n"
+      "ld1         {v1.8h}, [%0], #16            \n"  // load 8 samples, 5 rows
+      "ld1         {v2.8h}, [%4], #16            \n"
+      "uaddl       v0.4s, v1.4h, v2.4h           \n"  // * 1
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddl2      v1.4s, v1.8h, v2.8h           \n"  // * 1
+      "ld1         {v2.8h}, [%1], #16            \n"
+      "umlal       v0.4s, v2.4h, v6.4h           \n"  // * 4
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v6.8h           \n"  // * 4
+      "ld1         {v2.8h}, [%2], #16            \n"
+      "umlal       v0.4s, v2.4h, v7.4h           \n"  // * 6
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v7.8h           \n"  // * 6
+      "ld1         {v2.8h}, [%3], #16            \n"
+      "umlal       v0.4s, v2.4h, v6.4h           \n"  // * 4
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "umlal2      v1.4s, v2.8h, v6.8h           \n"  // * 4
+      "subs        %w6, %w6, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s,v1.4s}, [%5], #32      \n"  // store 8 samples
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "b.gt        1b                            \n"
       : "+r"(src0),  // %0
         "+r"(src1),  // %1
         "+r"(src2),  // %2
@@ -2845,27 +3880,28 @@ void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width) {
   const uint32_t* src2 = src + 2;
   const uint32_t* src3 = src + 3;
   asm volatile(
-      "movi       v6.4s, #4                      \n"  // constant 4
-      "movi       v7.4s, #6                      \n"  // constant 6
-
-      "1:                                        \n"
-      "ld1        {v0.4s,v1.4s,v2.4s}, [%0], %6  \n"  // load 12 source samples
-      "add        v0.4s, v0.4s, v1.4s            \n"  // * 1
-      "add        v1.4s, v1.4s, v2.4s            \n"  // * 1
-      "ld1        {v2.4s,v3.4s}, [%2], #32       \n"
-      "mla        v0.4s, v2.4s, v7.4s            \n"  // * 6
-      "mla        v1.4s, v3.4s, v7.4s            \n"  // * 6
-      "ld1        {v2.4s,v3.4s}, [%1], #32       \n"
-      "ld1        {v4.4s,v5.4s}, [%3], #32       \n"
-      "add        v2.4s, v2.4s, v4.4s            \n"  // add rows for * 4
-      "add        v3.4s, v3.4s, v5.4s            \n"
-      "mla        v0.4s, v2.4s, v6.4s            \n"  // * 4
-      "mla        v1.4s, v3.4s, v6.4s            \n"  // * 4
-      "subs       %w5, %w5, #8                   \n"  // 8 processed per loop
-      "uqrshrn    v0.4h, v0.4s, #8               \n"  // round and pack
-      "uqrshrn2   v0.8h, v1.4s, #8               \n"
-      "st1        {v0.8h}, [%4], #16             \n"  // store 8 samples
-      "b.gt       1b                             \n"
+      "movi        v6.4s, #4                     \n"  // constant 4
+      "movi        v7.4s, #6                     \n"  // constant 6
+
+      "1:                                        \n"
+      "ld1         {v0.4s,v1.4s,v2.4s}, [%0], %6 \n"  // load 12 source samples
+      "add         v0.4s, v0.4s, v1.4s           \n"  // * 1
+      "add         v1.4s, v1.4s, v2.4s           \n"  // * 1
+      "ld1         {v2.4s,v3.4s}, [%2], #32      \n"
+      "mla         v0.4s, v2.4s, v7.4s           \n"  // * 6
+      "mla         v1.4s, v3.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s,v3.4s}, [%1], #32      \n"
+      "ld1         {v4.4s,v5.4s}, [%3], #32      \n"
+      "add         v2.4s, v2.4s, v4.4s           \n"  // add rows for * 4
+      "add         v3.4s, v3.4s, v5.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mla         v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "mla         v1.4s, v3.4s, v6.4s           \n"  // * 4
+      "subs        %w5, %w5, #8                  \n"  // 8 processed per loop
+      "uqrshrn     v0.4h, v0.4s, #8              \n"  // round and pack
+      "uqrshrn2    v0.8h, v1.4s, #8              \n"
+      "st1         {v0.8h}, [%4], #16            \n"  // store 8 samples
+      "b.gt        1b                            \n"
       : "+r"(src),   // %0
         "+r"(src1),  // %1
         "+r"(src2),  // %2
@@ -2876,20 +3912,104 @@ void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width) {
       : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
 }
 
+static const vecf32 kGaussCoefficients = {4.0f, 6.0f, 1.0f / 256.0f, 0.0f};
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussCol_F32_NEON(const float* src0,
+                       const float* src1,
+                       const float* src2,
+                       const float* src3,
+                       const float* src4,
+                       float* dst,
+                       int width) {
+  asm volatile(
+      "ld2r        {v6.4s, v7.4s}, [%7]          \n"  // constants 4 and 6
+
+      "1:                                        \n"
+      "ld1         {v0.4s, v1.4s}, [%0], #32     \n"  // load 8 samples, 5 rows
+      "ld1         {v2.4s, v3.4s}, [%1], #32     \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "ld1         {v4.4s, v5.4s}, [%2], #32     \n"
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmla        v0.4s, v4.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s, v3.4s}, [%3], #32     \n"
+      "fmla        v1.4s, v5.4s, v7.4s           \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "ld1         {v4.4s, v5.4s}, [%4], #32     \n"
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "fadd        v0.4s, v0.4s, v4.4s           \n"  // * 1
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "fadd        v1.4s, v1.4s, v5.4s           \n"
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "subs        %w6, %w6, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s, v1.4s}, [%5], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src0),               // %0
+        "+r"(src1),               // %1
+        "+r"(src2),               // %2
+        "+r"(src3),               // %3
+        "+r"(src4),               // %4
+        "+r"(dst),                // %5
+        "+r"(width)               // %6
+      : "r"(&kGaussCoefficients)  // %7
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+
+// filter 5 rows with 1, 4, 6, 4, 1 coefficients to produce 1 row.
+void GaussRow_F32_NEON(const float* src, float* dst, int width) {
+  asm volatile(
+      "ld3r        {v6.4s, v7.4s, v8.4s}, [%3]   \n"  // constants 4, 6, 1/256
+
+      "1:                                        \n"
+      "ld1         {v0.4s, v1.4s, v2.4s}, [%0], %4 \n"  // load 12 samples, 5
+                                                        // rows
+      "fadd        v0.4s, v0.4s, v1.4s           \n"    // * 1
+      "ld1         {v4.4s, v5.4s}, [%0], %5      \n"
+      "fadd        v1.4s, v1.4s, v2.4s           \n"
+      "fmla        v0.4s, v4.4s, v7.4s           \n"  // * 6
+      "ld1         {v2.4s, v3.4s}, [%0], %4      \n"
+      "fmla        v1.4s, v5.4s, v7.4s           \n"
+      "ld1         {v4.4s, v5.4s}, [%0], %6      \n"
+      "fadd        v2.4s, v2.4s, v4.4s           \n"
+      "fadd        v3.4s, v3.4s, v5.4s           \n"
+      "fmla        v0.4s, v2.4s, v6.4s           \n"  // * 4
+      "fmla        v1.4s, v3.4s, v6.4s           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "fmul        v0.4s, v0.4s, v8.4s           \n"  // / 256
+      "fmul        v1.4s, v1.4s, v8.4s           \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "st1         {v0.4s, v1.4s}, [%1], #32     \n"  // store 8 samples
+      "b.gt        1b                            \n"
+      : "+r"(src),                 // %0
+        "+r"(dst),                 // %1
+        "+r"(width)                // %2
+      : "r"(&kGaussCoefficients),  // %3
+        "r"(8LL),                  // %4
+        "r"(-4LL),                 // %5
+        "r"(20LL)                  // %6
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8");
+}
+
+#if LIBYUV_USE_ST3
 // Convert biplanar NV21 to packed YUV24
 void NV21ToYUV24Row_NEON(const uint8_t* src_y,
                          const uint8_t* src_vu,
                          uint8_t* dst_yuv24,
                          int width) {
   asm volatile(
-      "1:                                          \n"
-      "ld1        {v2.16b}, [%0], #16            \n"     // load 16 Y values
-      "ld2        {v0.8b, v1.8b}, [%1], #16      \n"     // load 8 VU values
-      "zip1       v0.16b, v0.16b, v0.16b         \n"     // replicate V values
-      "zip1       v1.16b, v1.16b, v1.16b         \n"     // replicate U values
-      "subs       %w3, %w3, #16                  \n"     // 16 pixels per loop
-      "st3        {v0.16b,v1.16b,v2.16b}, [%2], #48 \n"  // store 16 YUV pixels
-      "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld1         {v2.16b}, [%0], #16           \n"  // load 16 Y values
+      "ld2         {v0.8b, v1.8b}, [%1], #16     \n"  // load 8 VU values
+      "zip1        v0.16b, v0.16b, v0.16b        \n"  // replicate V values
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "zip1        v1.16b, v1.16b, v1.16b        \n"  // replicate U values
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "subs        %w3, %w3, #16                 \n"      // 16 pixels per loop
+      "st3         {v0.16b,v1.16b,v2.16b}, [%2], #48 \n"  // store 16 YUV pixels
+      "b.gt        1b                            \n"
       : "+r"(src_y),      // %0
         "+r"(src_vu),     // %1
         "+r"(dst_yuv24),  // %2
@@ -2897,7 +4017,44 @@ void NV21ToYUV24Row_NEON(const uint8_t* src_y,
       :
       : "cc", "memory", "v0", "v1", "v2");
 }
+#else
+static const uvec8 kYUV24Shuffle[3] = {
+    {16, 17, 0, 16, 17, 1, 18, 19, 2, 18, 19, 3, 20, 21, 4, 20},
+    {21, 5, 22, 23, 6, 22, 23, 7, 24, 25, 8, 24, 25, 9, 26, 27},
+    {10, 26, 27, 11, 28, 29, 12, 28, 29, 13, 30, 31, 14, 30, 31, 15}};
+
+// Convert biplanar NV21 to packed YUV24
+// NV21 has VU in memory for chroma.
+// YUV24 is VUY in memory
+void NV21ToYUV24Row_NEON(const uint8_t* src_y,
+                         const uint8_t* src_vu,
+                         uint8_t* dst_yuv24,
+                         int width) {
+  asm volatile(
+      "ld1         {v5.16b,v6.16b,v7.16b}, [%4]  \n"  // 3 shuffler constants
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"    // load 16 Y values
+      "ld1         {v1.16b}, [%1], #16           \n"    // load 8 VU values
+      "tbl         v2.16b, {v0.16b,v1.16b}, v5.16b \n"  // weave into YUV24
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v3.16b, {v0.16b,v1.16b}, v6.16b \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "tbl         v4.16b, {v0.16b,v1.16b}, v7.16b \n"
+      "subs        %w3, %w3, #16                 \n"      // 16 pixels per loop
+      "st1         {v2.16b,v3.16b,v4.16b}, [%2], #48 \n"  // store 16 YUV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_y),            // %0
+        "+r"(src_vu),           // %1
+        "+r"(dst_yuv24),        // %2
+        "+r"(width)             // %3
+      : "r"(&kYUV24Shuffle[0])  // %4
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7");
+}
+#endif  // LIBYUV_USE_ST3
+
+// Note ST2 8b version is faster than zip+ST1
 
+// AYUV is VUYA in memory.  UV for NV12 is UV order in memory.
 void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
                       int src_stride_ayuv,
                       uint8_t* dst_uv,
@@ -2905,19 +4062,20 @@ void AYUVToUVRow_NEON(const uint8_t* src_ayuv,
   const uint8_t* src_ayuv_1 = src_ayuv + src_stride_ayuv;
   asm volatile(
 
-      "1:                                          \n"
-      "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
-                                                                // pixels.
-      "uaddlp     v0.8h, v0.16b                  \n"  // V 16 bytes -> 8 shorts.
-      "uaddlp     v1.8h, v1.16b                  \n"  // U 16 bytes -> 8 shorts.
-      "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
-      "uadalp     v0.8h, v4.16b                  \n"  // V 16 bytes -> 8 shorts.
-      "uadalp     v1.8h, v5.16b                  \n"  // U 16 bytes -> 8 shorts.
-      "uqrshrn    v3.8b, v0.8h, #2               \n"  // 2x2 average
-      "uqrshrn    v2.8b, v1.8h, #2               \n"
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
-      "st2        {v2.8b,v3.8b}, [%2], #16       \n"  // store 8 pixels UV.
-      "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ayuv
+      "uaddlp      v0.8h, v0.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v5.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqrshrn     v3.8b, v0.8h, #2              \n"  // 2x2 average
+      "uqrshrn     v2.8b, v1.8h, #2              \n"
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "st2         {v2.8b,v3.8b}, [%2], #16      \n"  // store 8 pixels UV.
+      "b.gt        1b                            \n"
       : "+r"(src_ayuv),    // %0
         "+r"(src_ayuv_1),  // %1
         "+r"(dst_uv),      // %2
@@ -2933,19 +4091,20 @@ void AYUVToVURow_NEON(const uint8_t* src_ayuv,
   const uint8_t* src_ayuv_1 = src_ayuv + src_stride_ayuv;
   asm volatile(
 
-      "1:                                          \n"
-      "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
-                                                                // pixels.
-      "uaddlp     v0.8h, v0.16b                  \n"  // V 16 bytes -> 8 shorts.
-      "uaddlp     v1.8h, v1.16b                  \n"  // U 16 bytes -> 8 shorts.
-      "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
-      "uadalp     v0.8h, v4.16b                  \n"  // V 16 bytes -> 8 shorts.
-      "uadalp     v1.8h, v5.16b                  \n"  // U 16 bytes -> 8 shorts.
-      "uqrshrn    v0.8b, v0.8h, #2               \n"  // 2x2 average
-      "uqrshrn    v1.8b, v1.8h, #2               \n"
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
-      "st2        {v0.8b,v1.8b}, [%2], #16       \n"  // store 8 pixels VU.
-      "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ayuv
+      "uaddlp      v0.8h, v0.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "ld4         {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+      "uadalp      v0.8h, v4.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v5.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uqrshrn     v0.8b, v0.8h, #2              \n"  // 2x2 average
+      "uqrshrn     v1.8b, v1.8h, #2              \n"
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop.
+      "st2         {v0.8b,v1.8b}, [%2], #16      \n"  // store 8 pixels VU.
+      "b.gt        1b                            \n"
       : "+r"(src_ayuv),    // %0
         "+r"(src_ayuv_1),  // %1
         "+r"(dst_vu),      // %2
@@ -2957,12 +4116,12 @@ void AYUVToVURow_NEON(const uint8_t* src_ayuv,
 // Copy row of AYUV Y's into Y
 void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
   asm volatile(
-      "1:                                          \n"
-      "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
-                                                                // pixels
-      "subs       %w2, %w2, #16                  \n"  // 16 pixels per loop
-      "st1        {v2.16b}, [%1], #16            \n"  // store 16 Y pixels
-      "b.gt       1b                             \n"
+      "1:                                        \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "st1         {v2.16b}, [%1], #16           \n"  // store 16 Y pixels
+      "b.gt        1b                            \n"
       : "+r"(src_ayuv),  // %0
         "+r"(dst_y),     // %1
         "+r"(width)      // %2
@@ -2970,61 +4129,170 @@ void AYUVToYRow_NEON(const uint8_t* src_ayuv, uint8_t* dst_y, int width) {
       : "cc", "memory", "v0", "v1", "v2", "v3");
 }
 
-void FloatDivToByteRow_NEON(const float* src_weights,
-                            const float* src_values,
-                            uint8_t* dst_out,
-                            uint8_t* dst_mask,
-                            int width) {
-  asm volatile(
-      "movi       v0.4s, #0                      \n"
+// Shuffle table for swapping UV bytes.
+static const uvec8 kShuffleSwapUV = {1u, 0u, 3u,  2u,  5u,  4u,  7u,  6u,
+                                     9u, 8u, 11u, 10u, 13u, 12u, 15u, 14u};
 
+// Convert UV plane of NV12 to VU of NV21.
+void SwapUVRow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+  asm volatile(
+      "ld1         {v2.16b}, [%3]                \n"  // shuffler
       "1:                                        \n"
-      "ld1        {v1.4s,v2.4s}, [%0], #32       \n"  // load 8 float weights
-      "ld1        {v3.4s,v4.4s}, [%1], #32       \n"  // load 8 float values
-      "subs       %w4, %w4, #8                   \n"  // 8 pixels per loop
-
-      "fdiv       v1.4s, v3.4s, v1.4s            \n"  // values / weights
-      "fdiv       v2.4s, v4.4s, v2.4s            \n"
-
-      "fcvtas     v1.4s, v1.4s                   \n"  // float to int
-      "fcvtas     v2.4s, v2.4s                   \n"  // float to int
-      "uqxtn      v1.4h, v1.4s                   \n"  // 8 shorts
-      "uqxtn2     v1.8h, v2.4s                   \n"
-      "uqxtn      v1.8b, v1.8h                   \n"  // 8 bytes
+      "ld1         {v0.16b}, [%0], 16            \n"  // load 16 UV values
+      "ld1         {v1.16b}, [%0], 16            \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 pixels per loop
+      "tbl         v0.16b, {v0.16b}, v2.16b      \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "tbl         v1.16b, {v1.16b}, v2.16b      \n"
+      "stp         q0, q1, [%1], 32              \n"  // store 16 VU pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),         // %0
+        "+r"(dst_vu),         // %1
+        "+r"(width)           // %2
+      : "r"(&kShuffleSwapUV)  // %3
+      : "cc", "memory", "v0", "v1", "v2");
+}
 
-      "st1        {v1.8b}, [%2], #8              \n"  // store 8 byte out
+void HalfMergeUVRow_NEON(const uint8_t* src_u,
+                         int src_stride_u,
+                         const uint8_t* src_v,
+                         int src_stride_v,
+                         uint8_t* dst_uv,
+                         int width) {
+  const uint8_t* src_u_1 = src_u + src_stride_u;
+  const uint8_t* src_v_1 = src_v + src_stride_v;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 U values
+      "ld1         {v1.16b}, [%2], #16           \n"  // load 16 V values
+      "ld1         {v2.16b}, [%1], #16           \n"
+      "ld1         {v3.16b}, [%3], #16           \n"
+      "uaddlp      v0.8h, v0.16b                 \n"  // half size
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uaddlp      v1.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v1.8h, v3.16b                 \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uqrshrn     v0.8b, v0.8h, #2              \n"
+      "uqrshrn     v1.8b, v1.8h, #2              \n"
+      "subs        %w5, %w5, #16                 \n"  // 16 src pixels per loop
+      "st2         {v0.8b, v1.8b}, [%4], #16     \n"  // store 8 UV pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_u),    // %0
+        "+r"(src_u_1),  // %1
+        "+r"(src_v),    // %2
+        "+r"(src_v_1),  // %3
+        "+r"(dst_uv),   // %4
+        "+r"(width)     // %5
+      :
+      : "cc", "memory", "v0", "v1", "v2", "v3");
+}
 
-      "fcmgt      v5.4s, v1.4s, v0.4s            \n"  // cmp weight to zero
-      "fcmgt      v6.4s, v2.4s, v0.4s            \n"
-      "uqxtn      v5.4h, v5.4s                   \n"  // 8 shorts
-      "uqxtn2     v5.8h, v6.4s                   \n"
-      "uqxtn      v5.8b, v1.8h                   \n"  // 8 bytes
+void SplitUVRow_16_NEON(const uint16_t* src_uv,
+                        uint16_t* dst_u,
+                        uint16_t* dst_v,
+                        int depth,
+                        int width) {
+  int shift = depth - 16;  // Negative for right shift.
+  asm volatile(
+      "dup         v2.8h, %w4                    \n"
+      "1:                                        \n"
+      "ld2         {v0.8h, v1.8h}, [%0], #32     \n"  // load 8 UV
+      "subs        %w3, %w3, #8                  \n"  // 8 src pixels per loop
+      "ushl        v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "st1         {v0.8h}, [%1], #16            \n"  // store 8 U pixels
+      "st1         {v1.8h}, [%2], #16            \n"  // store 8 V pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_uv),  // %0
+        "+r"(dst_u),   // %1
+        "+r"(dst_v),   // %2
+        "+r"(width)    // %3
+      : "r"(shift)     // %4
+      : "cc", "memory", "v0", "v1", "v2");
+}
 
-      "st1        {v5.8b}, [%3], #8              \n"  // store 8 byte mask
+void MultiplyRow_16_NEON(const uint16_t* src_y,
+                         uint16_t* dst_y,
+                         int scale,
+                         int width) {
+  asm volatile(
+      "dup         v2.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "mul         v0.8h, v0.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "mul         v1.8h, v1.8h, v2.8h           \n"
+      "stp         q0, q1, [%1], #32             \n"  // store 16 pixels
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "v0", "v1", "v2");
+}
 
-      "b.gt       1b                             \n"
-      : "+r"(src_weights),  // %0
-        "+r"(src_values),   // %1
-        "+r"(dst_out),      // %2
-        "+r"(dst_mask),     // %3
-        "+r"(width)         // %4
-      :
-      : "cc", "memory", "v1", "v2", "v3", "v4", "v5", "v6");
+void DivideRow_16_NEON(const uint16_t* src_y,
+                       uint16_t* dst_y,
+                       int scale,
+                       int width) {
+  asm volatile(
+      "dup         v0.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q1, q2, [%0], #32             \n"
+      "ushll       v3.4s, v1.4h, #0              \n"
+      "ushll       v4.4s, v2.4h, #0              \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "ushll2      v1.4s, v1.8h, #0              \n"
+      "ushll2      v2.4s, v2.8h, #0              \n"
+      "mul         v3.4s, v0.4s, v3.4s           \n"
+      "mul         v4.4s, v0.4s, v4.4s           \n"
+      "mul         v1.4s, v0.4s, v1.4s           \n"
+      "mul         v2.4s, v0.4s, v2.4s           \n"
+      "shrn        v3.4h, v3.4s, #16             \n"
+      "shrn        v4.4h, v4.4s, #16             \n"
+      "shrn2       v3.8h, v1.4s, #16             \n"
+      "shrn2       v4.8h, v2.4s, #16             \n"
+      "stp         q3, q3, [%1], #32             \n"  // store 16 pixels
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(scale)    // %3
+      : "cc", "memory", "v0", "v1", "v2", "v3", "v4");
 }
 
-// Convert biplanar UV channel of NV12 to NV21
-void UVToVURow_NEON(const uint8_t* src_uv, uint8_t* dst_vu, int width) {
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits = shr 1
+// 16384 = 10 bits = shr 2
+// 4096 = 12 bits = shr 4
+// 256 = 16 bits = shr 8
+void Convert16To8Row_NEON(const uint16_t* src_y,
+                          uint8_t* dst_y,
+                          int scale,
+                          int width) {
+  int shift = 15 - __builtin_clz((int32_t)scale);  // Negative shl is shr
   asm volatile(
-      "1:                                          \n"
-      "ld2        {v0.16b, v1.16b}, [%0], #32    \n"  // load 16 UV values
-      "orr        v2.16b, v0.16b, v0.16b         \n"  // move U after V
-      "subs       %w2, %w2, #16                  \n"  // 16 pixels per loop
-      "st2        {v1.16b, v2.16b}, [%1], #32    \n"  // store 16 VU pixels
-      "b.gt       1b                             \n"
-      : "+r"(src_uv),  // %0
-        "+r"(dst_vu),  // %1
-        "+r"(width)    // %2
-      :
+      "dup         v2.8h, %w3                    \n"
+      "1:                                        \n"
+      "ldp         q0, q1, [%0], #32             \n"
+      "ushl        v0.8h, v0.8h, v2.8h           \n"  // shr = v2 is negative
+      "ushl        v1.8h, v1.8h, v2.8h           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"
+      "uqxtn       v0.8b, v0.8h                  \n"
+      "uqxtn2      v0.16b, v1.8h                 \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 src pixels per loop
+      "str         q0, [%1], #16                 \n"  // store 16 pixels
+      "b.gt        1b                            \n"
+      : "+r"(src_y),  // %0
+        "+r"(dst_y),  // %1
+        "+r"(width)   // %2
+      : "r"(shift)    // %3
       : "cc", "memory", "v0", "v1", "v2");
 }
 
diff --git a/files/source/row_win.cc b/files/source/row_win.cc
index 27e3da7b..c7c1ff60 100644
--- a/files/source/row_win.cc
+++ b/files/source/row_win.cc
@@ -10,9 +10,9 @@
 
 #include "libyuv/row.h"
 
-// This module is for Visual C 32/64 bit and clangcl 32 bit
+// This module is for Visual C 32/64 bit
 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
-    (defined(_M_IX86) || (defined(_M_X64) && !defined(__clang__)))
+    !defined(__clang__) && (defined(_M_IX86) || defined(_M_X64))
 
 #if defined(_M_X64)
 #include <emmintrin.h>
@@ -27,12 +27,34 @@ extern "C" {
 // 64 bit
 #if defined(_M_X64)
 
+// Read 8 UV from 444
+#define READYUV444                                    \
+  xmm3 = _mm_loadl_epi64((__m128i*)u_buf);            \
+  xmm1 = _mm_loadl_epi64((__m128i*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);               \
+  u_buf += 8;                                         \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);            \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);               \
+  y_buf += 8;
+
+// Read 8 UV from 444, With 8 Alpha.
+#define READYUVA444                                   \
+  xmm3 = _mm_loadl_epi64((__m128i*)u_buf);            \
+  xmm1 = _mm_loadl_epi64((__m128i*)(u_buf + offset)); \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);               \
+  u_buf += 8;                                         \
+  xmm4 = _mm_loadl_epi64((__m128i*)y_buf);            \
+  xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);               \
+  y_buf += 8;                                         \
+  xmm5 = _mm_loadl_epi64((__m128i*)a_buf);            \
+  a_buf += 8;
+
 // Read 4 UV from 422, upsample to 8 UV.
 #define READYUV422                                        \
-  xmm0 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
+  xmm3 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
   xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
-  xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);                   \
-  xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);                  \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);                   \
+  xmm3 = _mm_unpacklo_epi16(xmm3, xmm3);                  \
   u_buf += 4;                                             \
   xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                \
   xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                   \
@@ -40,10 +62,10 @@ extern "C" {
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
 #define READYUVA422                                       \
-  xmm0 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
+  xmm3 = _mm_cvtsi32_si128(*(uint32_t*)u_buf);            \
   xmm1 = _mm_cvtsi32_si128(*(uint32_t*)(u_buf + offset)); \
-  xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);                   \
-  xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);                  \
+  xmm3 = _mm_unpacklo_epi8(xmm3, xmm1);                   \
+  xmm3 = _mm_unpacklo_epi16(xmm3, xmm3);                  \
   u_buf += 4;                                             \
   xmm4 = _mm_loadl_epi64((__m128i*)y_buf);                \
   xmm4 = _mm_unpacklo_epi8(xmm4, xmm4);                   \
@@ -52,24 +74,21 @@ extern "C" {
   a_buf += 8;
 
 // Convert 8 pixels: 8 UV and 8 Y.
-#define YUVTORGB(yuvconstants)                                     \
-  xmm1 = _mm_loadu_si128(&xmm0);                                   \
-  xmm2 = _mm_loadu_si128(&xmm0);                                   \
-  xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)yuvconstants->kUVToB); \
-  xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)yuvconstants->kUVToG); \
-  xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)yuvconstants->kUVToR); \
-  xmm0 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasB, xmm0);   \
-  xmm1 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasG, xmm1);   \
-  xmm2 = _mm_sub_epi16(*(__m128i*)yuvconstants->kUVBiasR, xmm2);   \
-  xmm4 = _mm_mulhi_epu16(xmm4, *(__m128i*)yuvconstants->kYToRgb);  \
-  xmm0 = _mm_adds_epi16(xmm0, xmm4);                               \
-  xmm1 = _mm_adds_epi16(xmm1, xmm4);                               \
-  xmm2 = _mm_adds_epi16(xmm2, xmm4);                               \
-  xmm0 = _mm_srai_epi16(xmm0, 6);                                  \
-  xmm1 = _mm_srai_epi16(xmm1, 6);                                  \
-  xmm2 = _mm_srai_epi16(xmm2, 6);                                  \
-  xmm0 = _mm_packus_epi16(xmm0, xmm0);                             \
-  xmm1 = _mm_packus_epi16(xmm1, xmm1);                             \
+#define YUVTORGB(yuvconstants)                                      \
+  xmm3 = _mm_sub_epi8(xmm3, _mm_set1_epi8((char)0x80));             \
+  xmm4 = _mm_mulhi_epu16(xmm4, *(__m128i*)yuvconstants->kYToRgb);   \
+  xmm4 = _mm_add_epi16(xmm4, *(__m128i*)yuvconstants->kYBiasToRgb); \
+  xmm0 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToB, xmm3);  \
+  xmm1 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToG, xmm3);  \
+  xmm2 = _mm_maddubs_epi16(*(__m128i*)yuvconstants->kUVToR, xmm3);  \
+  xmm0 = _mm_adds_epi16(xmm4, xmm0);                                \
+  xmm1 = _mm_subs_epi16(xmm4, xmm1);                                \
+  xmm2 = _mm_adds_epi16(xmm4, xmm2);                                \
+  xmm0 = _mm_srai_epi16(xmm0, 6);                                   \
+  xmm1 = _mm_srai_epi16(xmm1, 6);                                   \
+  xmm2 = _mm_srai_epi16(xmm2, 6);                                   \
+  xmm0 = _mm_packus_epi16(xmm0, xmm0);                              \
+  xmm1 = _mm_packus_epi16(xmm1, xmm1);                              \
   xmm2 = _mm_packus_epi16(xmm2, xmm2);
 
 // Store 8 ARGB values.
@@ -90,7 +109,7 @@ void I422ToARGBRow_SSSE3(const uint8_t* y_buf,
                          uint8_t* dst_argb,
                          const struct YuvConstants* yuvconstants,
                          int width) {
-  __m128i xmm0, xmm1, xmm2, xmm4;
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
   const __m128i xmm5 = _mm_set1_epi8(-1);
   const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
   while (width > 0) {
@@ -110,7 +129,7 @@ void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
                               uint8_t* dst_argb,
                               const struct YuvConstants* yuvconstants,
                               int width) {
-  __m128i xmm0, xmm1, xmm2, xmm4, xmm5;
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5;
   const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
   while (width > 0) {
     READYUVA422
@@ -121,6 +140,44 @@ void I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
 }
 #endif
 
+#if defined(HAS_I444TOARGBROW_SSSE3)
+void I444ToARGBRow_SSSE3(const uint8_t* y_buf,
+                         const uint8_t* u_buf,
+                         const uint8_t* v_buf,
+                         uint8_t* dst_argb,
+                         const struct YuvConstants* yuvconstants,
+                         int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
+  while (width > 0) {
+    READYUV444
+    YUVTORGB(yuvconstants)
+    STOREARGB
+    width -= 8;
+  }
+}
+#endif
+
+#if defined(HAS_I444ALPHATOARGBROW_SSSE3)
+void I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf,
+                              const uint8_t* u_buf,
+                              const uint8_t* v_buf,
+                              const uint8_t* a_buf,
+                              uint8_t* dst_argb,
+                              const struct YuvConstants* yuvconstants,
+                              int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3, xmm4, xmm5;
+  const ptrdiff_t offset = (uint8_t*)v_buf - (uint8_t*)u_buf;
+  while (width > 0) {
+    READYUVA444
+    YUVTORGB(yuvconstants)
+    STOREARGB
+    width -= 8;
+  }
+}
+#endif
+
 // 32 bit
 #else  // defined(_M_X64)
 #ifdef HAS_ARGBTOYROW_SSSE3
@@ -187,11 +244,11 @@ static const uvec8 kAddY16 = {16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u,
 // 7 bit fixed point 0.5.
 static const vec16 kAddYJ64 = {64, 64, 64, 64, 64, 64, 64, 64};
 
-static const uvec8 kAddUV128 = {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
-                                128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u};
-
-static const uvec16 kAddUVJ128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u,
-                                  0x8080u, 0x8080u, 0x8080u, 0x8080u};
+// 8 bit fixed point 0.5, for bias of UV.
+static const ulvec8 kBiasUV128 = {
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
 
 // Shuffle table for converting RGB24 to ARGB.
 static const uvec8 kShuffleMaskRGB24ToARGB = {
@@ -1367,7 +1424,7 @@ __declspec(naked) void RGBAToYRow_SSSE3(const uint8_t* src_argb,
   }
 }
 
-__declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
                                          uint8_t* dst_v,
@@ -1380,7 +1437,7 @@ __declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToV
     movdqa     xmm7, xmmword ptr kARGBToU
     sub        edi, edx  // stride from u to v
@@ -1439,7 +1496,7 @@ __declspec(naked) void ARGBToUVRow_SSSE3(const uint8_t* src_argb0,
   }
 }
 
-__declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb,
                                           int src_stride_argb,
                                           uint8_t* dst_u,
                                           uint8_t* dst_v,
@@ -1452,7 +1509,7 @@ __declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUVJ128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToVJ
     movdqa     xmm7, xmmword ptr kARGBToUJ
     sub        edi, edx  // stride from u to v
@@ -1513,7 +1570,7 @@ __declspec(naked) void ARGBToUVJRow_SSSE3(const uint8_t* src_argb0,
 }
 
 #ifdef HAS_ARGBTOUVROW_AVX2
-__declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb,
                                         int src_stride_argb,
                                         uint8_t* dst_u,
                                         uint8_t* dst_v,
@@ -1526,7 +1583,7 @@ __declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    vbroadcastf128 ymm5, xmmword ptr kAddUV128
+    vbroadcastf128 ymm5, xmmword ptr kBiasUV128
     vbroadcastf128 ymm6, xmmword ptr kARGBToV
     vbroadcastf128 ymm7, xmmword ptr kARGBToU
     sub        edi, edx   // stride from u to v
@@ -1581,7 +1638,7 @@ __declspec(naked) void ARGBToUVRow_AVX2(const uint8_t* src_argb0,
 #endif  // HAS_ARGBTOUVROW_AVX2
 
 #ifdef HAS_ARGBTOUVJROW_AVX2
-__declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
                                          uint8_t* dst_v,
@@ -1594,9 +1651,9 @@ __declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    vbroadcastf128 ymm5, xmmword ptr kAddUV128
-    vbroadcastf128 ymm6, xmmword ptr kARGBToV
-    vbroadcastf128 ymm7, xmmword ptr kARGBToU
+    vbroadcastf128 ymm5, xmmword ptr kBiasUV128
+    vbroadcastf128 ymm6, xmmword ptr kARGBToVJ
+    vbroadcastf128 ymm7, xmmword ptr kARGBToUJ
     sub        edi, edx   // stride from u to v
 
  convertloop:
@@ -1649,7 +1706,7 @@ __declspec(naked) void ARGBToUVJRow_AVX2(const uint8_t* src_argb0,
 }
 #endif  // HAS_ARGBTOUVJROW_AVX2
 
-__declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb,
                                             uint8_t* dst_u,
                                             uint8_t* dst_v,
                                             int width) {
@@ -1659,7 +1716,7 @@ __declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 4 + 8]  // dst_u
     mov        edi, [esp + 4 + 12]  // dst_v
     mov        ecx, [esp + 4 + 16]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kARGBToV
     movdqa     xmm7, xmmword ptr kARGBToU
     sub        edi, edx    // stride from u to v
@@ -1707,7 +1764,7 @@ __declspec(naked) void ARGBToUV444Row_SSSE3(const uint8_t* src_argb0,
   }
 }
 
-__declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
                                          uint8_t* dst_v,
@@ -1720,7 +1777,7 @@ __declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kBGRAToV
     movdqa     xmm7, xmmword ptr kBGRAToU
     sub        edi, edx  // stride from u to v
@@ -1779,7 +1836,7 @@ __declspec(naked) void BGRAToUVRow_SSSE3(const uint8_t* src_argb0,
   }
 }
 
-__declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
                                          uint8_t* dst_v,
@@ -1792,7 +1849,7 @@ __declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kABGRToV
     movdqa     xmm7, xmmword ptr kABGRToU
     sub        edi, edx  // stride from u to v
@@ -1851,7 +1908,7 @@ __declspec(naked) void ABGRToUVRow_SSSE3(const uint8_t* src_argb0,
   }
 }
 
-__declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb,
                                          int src_stride_argb,
                                          uint8_t* dst_u,
                                          uint8_t* dst_v,
@@ -1864,7 +1921,7 @@ __declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb0,
     mov        edx, [esp + 8 + 12]  // dst_u
     mov        edi, [esp + 8 + 16]  // dst_v
     mov        ecx, [esp + 8 + 20]  // width
-    movdqa     xmm5, xmmword ptr kAddUV128
+    movdqa     xmm5, xmmword ptr kBiasUV128
     movdqa     xmm6, xmmword ptr kRGBAToV
     movdqa     xmm7, xmmword ptr kRGBAToU
     sub        edi, edx  // stride from u to v
@@ -1926,137 +1983,153 @@ __declspec(naked) void RGBAToUVRow_SSSE3(const uint8_t* src_argb0,
 
 // Read 16 UV from 444
 #define READYUV444_AVX2 \
-  __asm {                                                \
-    __asm vmovdqu    xmm0, [esi] /* U */                      \
-    __asm vmovdqu    xmm1, [esi + edi] /* V */                      \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* U */                                       \
+    __asm vmovdqu    xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
     __asm vpermq     ymm1, ymm1, 0xd8                                          \
-    __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */                     \
-    __asm vmovdqu    xmm4, [eax] /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]}
 
+// Read 16 UV from 444.  With 16 Alpha.
+#define READYUVA444_AVX2 \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* U */                                       \
+    __asm vmovdqu    xmm1, [esi + edi] /* V */                                 \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpermq     ymm1, ymm1, 0xd8                                          \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
+    __asm vpermq     ymm4, ymm4, 0xd8                                          \
+    __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
+    __asm lea        eax, [eax + 16]                                           \
+    __asm vmovdqu    xmm5, [ebp] /* A */                                       \
+    __asm vpermq     ymm5, ymm5, 0xd8                                          \
+    __asm lea        ebp, [ebp + 16]}
+
 // Read 8 UV from 422, upsample to 16 UV.
 #define READYUV422_AVX2 \
-  __asm {                                                \
-    __asm vmovq      xmm0, qword ptr [esi] /* U */                      \
-    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                      \
+  __asm {                                                                      \
+    __asm vmovq      xmm3, qword ptr [esi] /* U */                             \
+    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */                     \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax] /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]}
 
 // Read 8 UV from 422, upsample to 16 UV.  With 16 Alpha.
 #define READYUVA422_AVX2 \
-  __asm {                                               \
-    __asm vmovq      xmm0, qword ptr [esi] /* U */                      \
-    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                      \
+  __asm {                                                                      \
+    __asm vmovq      xmm3, qword ptr [esi] /* U */                             \
+    __asm vmovq      xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm vpunpcklbw ymm0, ymm0, ymm1 /* UV */                     \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax] /* Y */                      \
+    __asm vpunpcklbw ymm3, ymm3, ymm1 /* UV */                                 \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]                                           \
-    __asm vmovdqu    xmm5, [ebp] /* A */                      \
+    __asm vmovdqu    xmm5, [ebp] /* A */                                       \
     __asm vpermq     ymm5, ymm5, 0xd8                                          \
     __asm lea        ebp, [ebp + 16]}
 
 // Read 8 UV from NV12, upsample to 16 UV.
 #define READNV12_AVX2 \
-  __asm {                                                  \
-    __asm vmovdqu    xmm0, [esi] /* UV */                     \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* UV */                                      \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm0, ymm0 /* UVUV (upsample) */        \
-    __asm vmovdqu    xmm4, [eax] /* Y */                      \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklwd ymm3, ymm3, ymm3 /* UVUV (upsample) */                    \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]}
 
 // Read 8 UV from NV21, upsample to 16 UV.
 #define READNV21_AVX2 \
-  __asm {                                                  \
-    __asm vmovdqu    xmm0, [esi] /* UV */                     \
+  __asm {                                                                      \
+    __asm vmovdqu    xmm3, [esi] /* UV */                                      \
     __asm lea        esi,  [esi + 16]                                          \
-    __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleNV21                      \
-    __asm vmovdqu    xmm4, [eax] /* Y */                      \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleNV21                      \
+    __asm vmovdqu    xmm4, [eax] /* Y */                                       \
     __asm vpermq     ymm4, ymm4, 0xd8                                          \
     __asm vpunpcklbw ymm4, ymm4, ymm4                                          \
     __asm lea        eax, [eax + 16]}
 
 // Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV.
 #define READYUY2_AVX2 \
-  __asm {                                                  \
-    __asm vmovdqu    ymm4, [eax] /* YUY2 */                           \
+  __asm {                                                                      \
+    __asm vmovdqu    ymm4, [eax] /* YUY2 */                                    \
     __asm vpshufb    ymm4, ymm4, ymmword ptr kShuffleYUY2Y                     \
-    __asm vmovdqu    ymm0, [eax] /* UV */                             \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleYUY2UV                    \
+    __asm vmovdqu    ymm3, [eax] /* UV */                                      \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleYUY2UV                    \
     __asm lea        eax, [eax + 32]}
 
 // Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV.
 #define READUYVY_AVX2 \
-  __asm {                                                  \
-    __asm vmovdqu    ymm4, [eax] /* UYVY */                           \
+  __asm {                                                                      \
+    __asm vmovdqu    ymm4, [eax] /* UYVY */                                    \
     __asm vpshufb    ymm4, ymm4, ymmword ptr kShuffleUYVYY                     \
-    __asm vmovdqu    ymm0, [eax] /* UV */                             \
-    __asm vpshufb    ymm0, ymm0, ymmword ptr kShuffleUYVYUV                    \
+    __asm vmovdqu    ymm3, [eax] /* UV */                                      \
+    __asm vpshufb    ymm3, ymm3, ymmword ptr kShuffleUYVYUV                    \
     __asm lea        eax, [eax + 32]}
 
 // Convert 16 pixels: 16 UV and 16 Y.
 #define YUVTORGB_AVX2(YuvConstants) \
-  __asm {                                    \
-    __asm vpmaddubsw ymm2, ymm0, ymmword ptr [YuvConstants + KUVTOR] /* R UV */\
-    __asm vpmaddubsw ymm1, ymm0, ymmword ptr [YuvConstants + KUVTOG] /* G UV */\
-    __asm vpmaddubsw ymm0, ymm0, ymmword ptr [YuvConstants + KUVTOB] /* B UV */\
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASR]               \
-    __asm vpsubw     ymm2, ymm3, ymm2                                          \
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASG]               \
-    __asm vpsubw     ymm1, ymm3, ymm1                                          \
-    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KUVBIASB]               \
-    __asm vpsubw     ymm0, ymm3, ymm0 /* Step 2: Find Y contribution to 16 R,G,B values */                       \
+  __asm {                                                                      \
+    __asm vpsubb     ymm3, ymm3, ymmword ptr kBiasUV128                        \
     __asm vpmulhuw   ymm4, ymm4, ymmword ptr [YuvConstants + KYTORGB]          \
-    __asm vpaddsw    ymm0, ymm0, ymm4 /* B += Y */                   \
-    __asm vpaddsw    ymm1, ymm1, ymm4 /* G += Y */                   \
-    __asm vpaddsw    ymm2, ymm2, ymm4 /* R += Y */                   \
+    __asm vmovdqa    ymm0, ymmword ptr [YuvConstants + KUVTOB]                 \
+    __asm vmovdqa    ymm1, ymmword ptr [YuvConstants + KUVTOG]                 \
+    __asm vmovdqa    ymm2, ymmword ptr [YuvConstants + KUVTOR]                 \
+    __asm vpmaddubsw ymm0, ymm0, ymm3 /* B UV */                               \
+    __asm vpmaddubsw ymm1, ymm1, ymm3 /* G UV */                               \
+    __asm vpmaddubsw ymm2, ymm2, ymm3 /* B UV */                               \
+    __asm vmovdqu    ymm3, ymmword ptr [YuvConstants + KYBIASTORGB]            \
+    __asm vpaddw     ymm4, ymm3, ymm4                                          \
+    __asm vpaddsw    ymm0, ymm0, ymm4                                          \
+    __asm vpsubsw    ymm1, ymm4, ymm1                                          \
+    __asm vpaddsw    ymm2, ymm2, ymm4                                          \
     __asm vpsraw     ymm0, ymm0, 6                                             \
     __asm vpsraw     ymm1, ymm1, 6                                             \
     __asm vpsraw     ymm2, ymm2, 6                                             \
-    __asm vpackuswb  ymm0, ymm0, ymm0 /* B */                        \
-    __asm vpackuswb  ymm1, ymm1, ymm1 /* G */                        \
-    __asm vpackuswb  ymm2, ymm2, ymm2 /* R */                  \
-  }
+    __asm vpackuswb  ymm0, ymm0, ymm0                                          \
+    __asm vpackuswb  ymm1, ymm1, ymm1                                          \
+    __asm vpackuswb  ymm2, ymm2, ymm2}
 
 // Store 16 ARGB values.
 #define STOREARGB_AVX2 \
-  __asm {                                                 \
-    __asm vpunpcklbw ymm0, ymm0, ymm1 /* BG */                       \
+  __asm {                                                                      \
+    __asm vpunpcklbw ymm0, ymm0, ymm1 /* BG */                                 \
     __asm vpermq     ymm0, ymm0, 0xd8                                          \
-    __asm vpunpcklbw ymm2, ymm2, ymm5 /* RA */                       \
+    __asm vpunpcklbw ymm2, ymm2, ymm5 /* RA */                                 \
     __asm vpermq     ymm2, ymm2, 0xd8                                          \
-    __asm vpunpcklwd ymm1, ymm0, ymm2 /* BGRA first 8 pixels */      \
-    __asm vpunpckhwd ymm0, ymm0, ymm2 /* BGRA next 8 pixels */       \
+    __asm vpunpcklwd ymm1, ymm0, ymm2 /* BGRA first 8 pixels */                \
+    __asm vpunpckhwd ymm0, ymm0, ymm2 /* BGRA next 8 pixels */                 \
     __asm vmovdqu    0[edx], ymm1                                              \
     __asm vmovdqu    32[edx], ymm0                                             \
     __asm lea        edx,  [edx + 64]}
 
 // Store 16 RGBA values.
 #define STORERGBA_AVX2 \
-  __asm {                                                 \
-    __asm vpunpcklbw ymm1, ymm1, ymm2 /* GR */                       \
+  __asm {                                                                      \
+    __asm vpunpcklbw ymm1, ymm1, ymm2 /* GR */                                 \
     __asm vpermq     ymm1, ymm1, 0xd8                                          \
-    __asm vpunpcklbw ymm2, ymm5, ymm0 /* AB */                       \
+    __asm vpunpcklbw ymm2, ymm5, ymm0 /* AB */                                 \
     __asm vpermq     ymm2, ymm2, 0xd8                                          \
-    __asm vpunpcklwd ymm0, ymm2, ymm1 /* ABGR first 8 pixels */      \
-    __asm vpunpckhwd ymm1, ymm2, ymm1 /* ABGR next 8 pixels */       \
+    __asm vpunpcklwd ymm0, ymm2, ymm1 /* ABGR first 8 pixels */                \
+    __asm vpunpckhwd ymm1, ymm2, ymm1 /* ABGR next 8 pixels */                 \
     __asm vmovdqu    [edx], ymm0                                               \
     __asm vmovdqu    [edx + 32], ymm1                                          \
     __asm lea        edx,  [edx + 64]}
@@ -2183,6 +2256,48 @@ __declspec(naked) void I444ToARGBRow_AVX2(
 }
 #endif  // HAS_I444TOARGBROW_AVX2
 
+#ifdef HAS_I444ALPHATOARGBROW_AVX2
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked) void I444AlphaToARGBRow_AVX2(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
+  __asm {
+  push       esi
+  push       edi
+  push       ebx
+  push       ebp
+  mov        eax, [esp + 16 + 4]  // Y
+  mov        esi, [esp + 16 + 8]  // U
+  mov        edi, [esp + 16 + 12]  // V
+  mov        ebp, [esp + 16 + 16]  // A
+  mov        edx, [esp + 16 + 20]  // argb
+  mov        ebx, [esp + 16 + 24]  // yuvconstants
+  mov        ecx, [esp + 16 + 28]  // width
+  sub        edi, esi
+  convertloop:
+  READYUVA444_AVX2
+  YUVTORGB_AVX2(ebx)
+  STOREARGB_AVX2
+
+  sub        ecx, 16
+  jg         convertloop
+
+  pop        ebp
+  pop        ebx
+  pop        edi
+  pop        esi
+  vzeroupper
+  ret
+  }
+}
+#endif  // HAS_I444AlphaTOARGBROW_AVX2
+
 #ifdef HAS_NV12TOARGBROW_AVX2
 // 16 pixels.
 // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
@@ -2361,191 +2476,202 @@ __declspec(naked) void I422ToRGBARow_AVX2(
 
 // Read 8 UV from 444.
 #define READYUV444 \
-  __asm {                                                     \
-    __asm movq       xmm0, qword ptr [esi] /* U */                             \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* U */                             \
     __asm movq       xmm1, qword ptr [esi + edi] /* V */                       \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm punpcklbw  xmm0, xmm1 /* UV */                             \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]}
 
+// Read 4 UV from 444.  With 8 Alpha.
+#define READYUVA444 \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* U */                             \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */                       \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm movq       xmm4, qword ptr [eax]                                     \
+    __asm punpcklbw  xmm4, xmm4                                                \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm movq       xmm5, qword ptr [ebp] /* A */                             \
+    __asm lea        ebp, [ebp + 8]}
+
 // Read 4 UV from 422, upsample to 8 UV.
 #define READYUV422 \
-  __asm {                                                     \
-    __asm movd       xmm0, [esi] /* U */                              \
-    __asm movd       xmm1, [esi + edi] /* V */                              \
+  __asm {                                                                      \
+    __asm movd       xmm3, [esi] /* U */                                       \
+    __asm movd       xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 4]                                           \
-    __asm punpcklbw  xmm0, xmm1 /* UV */                             \
-    __asm punpcklwd  xmm0, xmm0 /* UVUV (upsample) */                \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]}
 
 // Read 4 UV from 422, upsample to 8 UV.  With 8 Alpha.
 #define READYUVA422 \
-  __asm {                                                    \
-    __asm movd       xmm0, [esi] /* U */                              \
-    __asm movd       xmm1, [esi + edi] /* V */                              \
+  __asm {                                                                      \
+    __asm movd       xmm3, [esi] /* U */                                       \
+    __asm movd       xmm1, [esi + edi] /* V */                                 \
     __asm lea        esi,  [esi + 4]                                           \
-    __asm punpcklbw  xmm0, xmm1 /* UV */                             \
-    __asm punpcklwd  xmm0, xmm0 /* UVUV (upsample) */                \
-    __asm movq       xmm4, qword ptr [eax] /* Y */                           \
+    __asm punpcklbw  xmm3, xmm1 /* UV */                                       \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
+    __asm movq       xmm4, qword ptr [eax] /* Y */                             \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]                                            \
-    __asm movq       xmm5, qword ptr [ebp] /* A */                           \
+    __asm movq       xmm5, qword ptr [ebp] /* A */                             \
     __asm lea        ebp, [ebp + 8]}
 
 // Read 4 UV from NV12, upsample to 8 UV.
 #define READNV12 \
-  __asm {                                                       \
-    __asm movq       xmm0, qword ptr [esi] /* UV */                            \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* UV */                            \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm punpcklwd  xmm0, xmm0 /* UVUV (upsample) */                \
+    __asm punpcklwd  xmm3, xmm3 /* UVUV (upsample) */                          \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]}
 
 // Read 4 VU from NV21, upsample to 8 UV.
 #define READNV21 \
-  __asm {                                                       \
-    __asm movq       xmm0, qword ptr [esi] /* UV */                            \
+  __asm {                                                                      \
+    __asm movq       xmm3, qword ptr [esi] /* UV */                            \
     __asm lea        esi,  [esi + 8]                                           \
-    __asm pshufb     xmm0, xmmword ptr kShuffleNV21                            \
+    __asm pshufb     xmm3, xmmword ptr kShuffleNV21                            \
     __asm movq       xmm4, qword ptr [eax]                                     \
     __asm punpcklbw  xmm4, xmm4                                                \
     __asm lea        eax, [eax + 8]}
 
 // Read 4 YUY2 with 8 Y and upsample 4 UV to 8 UV.
 #define READYUY2 \
-  __asm {                                                       \
-    __asm movdqu     xmm4, [eax] /* YUY2 */                           \
+  __asm {                                                                      \
+    __asm movdqu     xmm4, [eax] /* YUY2 */                                    \
     __asm pshufb     xmm4, xmmword ptr kShuffleYUY2Y                           \
-    __asm movdqu     xmm0, [eax] /* UV */                             \
-    __asm pshufb     xmm0, xmmword ptr kShuffleYUY2UV                          \
+    __asm movdqu     xmm3, [eax] /* UV */                                      \
+    __asm pshufb     xmm3, xmmword ptr kShuffleYUY2UV                          \
     __asm lea        eax, [eax + 16]}
 
 // Read 4 UYVY with 8 Y and upsample 4 UV to 8 UV.
 #define READUYVY \
-  __asm {                                                       \
-    __asm movdqu     xmm4, [eax] /* UYVY */                           \
+  __asm {                                                                      \
+    __asm movdqu     xmm4, [eax] /* UYVY */                                    \
     __asm pshufb     xmm4, xmmword ptr kShuffleUYVYY                           \
-    __asm movdqu     xmm0, [eax] /* UV */                             \
-    __asm pshufb     xmm0, xmmword ptr kShuffleUYVYUV                          \
+    __asm movdqu     xmm3, [eax] /* UV */                                      \
+    __asm pshufb     xmm3, xmmword ptr kShuffleUYVYUV                          \
     __asm lea        eax, [eax + 16]}
 
 // Convert 8 pixels: 8 UV and 8 Y.
 #define YUVTORGB(YuvConstants) \
-  __asm {                                         \
-    __asm movdqa     xmm1, xmm0                                                \
-    __asm movdqa     xmm2, xmm0                                                \
-    __asm movdqa     xmm3, xmm0                                                \
-    __asm movdqa     xmm0, xmmword ptr [YuvConstants + KUVBIASB]               \
-    __asm pmaddubsw  xmm1, xmmword ptr [YuvConstants + KUVTOB]                 \
-    __asm psubw      xmm0, xmm1                                                \
-    __asm movdqa     xmm1, xmmword ptr [YuvConstants + KUVBIASG]               \
-    __asm pmaddubsw  xmm2, xmmword ptr [YuvConstants + KUVTOG]                 \
-    __asm psubw      xmm1, xmm2                                                \
-    __asm movdqa     xmm2, xmmword ptr [YuvConstants + KUVBIASR]               \
-    __asm pmaddubsw  xmm3, xmmword ptr [YuvConstants + KUVTOR]                 \
-    __asm psubw      xmm2, xmm3                                                \
+  __asm {                                                                      \
+    __asm psubb      xmm3, xmmword ptr kBiasUV128                              \
     __asm pmulhuw    xmm4, xmmword ptr [YuvConstants + KYTORGB]                \
-    __asm paddsw     xmm0, xmm4 /* B += Y */                         \
-    __asm paddsw     xmm1, xmm4 /* G += Y */                         \
-    __asm paddsw     xmm2, xmm4 /* R += Y */                         \
+    __asm movdqa     xmm0, xmmword ptr [YuvConstants + KUVTOB]                 \
+    __asm movdqa     xmm1, xmmword ptr [YuvConstants + KUVTOG]                 \
+    __asm movdqa     xmm2, xmmword ptr [YuvConstants + KUVTOR]                 \
+    __asm pmaddubsw  xmm0, xmm3                                                \
+    __asm pmaddubsw  xmm1, xmm3                                                \
+    __asm pmaddubsw  xmm2, xmm3                                                \
+    __asm movdqa     xmm3, xmmword ptr [YuvConstants + KYBIASTORGB]            \
+    __asm paddw      xmm4, xmm3                                                \
+    __asm paddsw     xmm0, xmm4                                                \
+    __asm paddsw     xmm2, xmm4                                                \
+    __asm psubsw     xmm4, xmm1                                                \
+    __asm movdqa     xmm1, xmm4                                                \
     __asm psraw      xmm0, 6                                                   \
     __asm psraw      xmm1, 6                                                   \
     __asm psraw      xmm2, 6                                                   \
-    __asm packuswb   xmm0, xmm0 /* B */                              \
-    __asm packuswb   xmm1, xmm1 /* G */                              \
+    __asm packuswb   xmm0, xmm0 /* B */                                        \
+    __asm packuswb   xmm1, xmm1 /* G */                                        \
     __asm packuswb   xmm2, xmm2 /* R */             \
   }
 
 // Store 8 ARGB values.
 #define STOREARGB \
-  __asm {                                                      \
-    __asm punpcklbw  xmm0, xmm1 /* BG */                             \
-    __asm punpcklbw  xmm2, xmm5 /* RA */                             \
+  __asm {                                                                      \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm5 /* RA */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2 /* BGRA first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2 /* BGRA next 4 pixels */             \
+    __asm punpcklwd  xmm0, xmm2 /* BGRA first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm0                                              \
     __asm movdqu     16[edx], xmm1                                             \
     __asm lea        edx,  [edx + 32]}
 
 // Store 8 BGRA values.
 #define STOREBGRA \
-  __asm {                                                      \
-    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */  \
-    __asm punpcklbw  xmm1, xmm0 /* GB */                             \
-    __asm punpcklbw  xmm5, xmm2 /* AR */                             \
+  __asm {                                                                      \
+    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */            \
+    __asm punpcklbw  xmm1, xmm0 /* GB */                                       \
+    __asm punpcklbw  xmm5, xmm2 /* AR */                                       \
     __asm movdqa     xmm0, xmm5                                                \
-    __asm punpcklwd  xmm5, xmm1 /* BGRA first 4 pixels */            \
-    __asm punpckhwd  xmm0, xmm1 /* BGRA next 4 pixels */             \
+    __asm punpcklwd  xmm5, xmm1 /* BGRA first 4 pixels */                      \
+    __asm punpckhwd  xmm0, xmm1 /* BGRA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm5                                              \
     __asm movdqu     16[edx], xmm0                                             \
     __asm lea        edx,  [edx + 32]}
 
 // Store 8 RGBA values.
 #define STORERGBA \
-  __asm {                                                      \
-    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */  \
-    __asm punpcklbw  xmm1, xmm2 /* GR */                             \
-    __asm punpcklbw  xmm5, xmm0 /* AB */                             \
+  __asm {                                                                      \
+    __asm pcmpeqb    xmm5, xmm5 /* generate 0xffffffff for alpha */            \
+    __asm punpcklbw  xmm1, xmm2 /* GR */                                       \
+    __asm punpcklbw  xmm5, xmm0 /* AB */                                       \
     __asm movdqa     xmm0, xmm5                                                \
-    __asm punpcklwd  xmm5, xmm1 /* RGBA first 4 pixels */            \
-    __asm punpckhwd  xmm0, xmm1 /* RGBA next 4 pixels */             \
+    __asm punpcklwd  xmm5, xmm1 /* RGBA first 4 pixels */                      \
+    __asm punpckhwd  xmm0, xmm1 /* RGBA next 4 pixels */                       \
     __asm movdqu     0[edx], xmm5                                              \
     __asm movdqu     16[edx], xmm0                                             \
     __asm lea        edx,  [edx + 32]}
 
 // Store 8 RGB24 values.
 #define STORERGB24 \
-  __asm {/* Weave into RRGB */                                                      \
-    __asm punpcklbw  xmm0, xmm1 /* BG */                             \
-    __asm punpcklbw  xmm2, xmm2 /* RR */                             \
+  __asm {/* Weave into RRGB */                                                 \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm2 /* RR */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB24 */                                                        \
-    __asm pshufb     xmm0, xmm5 /* Pack first 8 and last 4 bytes. */ \
-    __asm pshufb     xmm1, xmm6 /* Pack first 12 bytes. */           \
-    __asm palignr    xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */ \
-    __asm movq       qword ptr 0[edx], xmm0 /* First 8 bytes */               \
-    __asm movdqu     8[edx], xmm1 /* Last 16 bytes */                  \
+    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB24 */   \
+    __asm pshufb     xmm0, xmm5 /* Pack first 8 and last 4 bytes. */           \
+    __asm pshufb     xmm1, xmm6 /* Pack first 12 bytes. */                     \
+    __asm palignr    xmm1, xmm0, 12 /* last 4 bytes of xmm0 + 12 xmm1 */       \
+    __asm movq       qword ptr 0[edx], xmm0 /* First 8 bytes */                \
+    __asm movdqu     8[edx], xmm1 /* Last 16 bytes */                          \
     __asm lea        edx,  [edx + 24]}
 
 // Store 8 RGB565 values.
 #define STORERGB565 \
-  __asm {/* Weave into RRGB */                                                      \
-    __asm punpcklbw  xmm0, xmm1 /* BG */                             \
-    __asm punpcklbw  xmm2, xmm2 /* RR */                             \
+  __asm {/* Weave into RRGB */                                                 \
+    __asm punpcklbw  xmm0, xmm1 /* BG */                                       \
+    __asm punpcklbw  xmm2, xmm2 /* RR */                                       \
     __asm movdqa     xmm1, xmm0                                                \
-    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */            \
-    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB565 */                                                       \
-    __asm movdqa     xmm3, xmm0 /* B  first 4 pixels of argb */             \
-    __asm movdqa     xmm2, xmm0 /* G */                                     \
-    __asm pslld      xmm0, 8 /* R */                                     \
-    __asm psrld      xmm3, 3 /* B */                                     \
-    __asm psrld      xmm2, 5 /* G */                                     \
-    __asm psrad      xmm0, 16 /* R */                                     \
-    __asm pand       xmm3, xmm5 /* B */                                     \
-    __asm pand       xmm2, xmm6 /* G */                                     \
-    __asm pand       xmm0, xmm7 /* R */                                     \
-    __asm por        xmm3, xmm2 /* BG */                                    \
-    __asm por        xmm0, xmm3 /* BGR */                                   \
-    __asm movdqa     xmm3, xmm1 /* B  next 4 pixels of argb */              \
-    __asm movdqa     xmm2, xmm1 /* G */                                     \
-    __asm pslld      xmm1, 8 /* R */                                     \
-    __asm psrld      xmm3, 3 /* B */                                     \
-    __asm psrld      xmm2, 5 /* G */                                     \
-    __asm psrad      xmm1, 16 /* R */                                     \
-    __asm pand       xmm3, xmm5 /* B */                                     \
-    __asm pand       xmm2, xmm6 /* G */                                     \
-    __asm pand       xmm1, xmm7 /* R */                                     \
-    __asm por        xmm3, xmm2 /* BG */                                    \
-    __asm por        xmm1, xmm3 /* BGR */                                   \
+    __asm punpcklwd  xmm0, xmm2 /* BGRR first 4 pixels */                      \
+    __asm punpckhwd  xmm1, xmm2 /* BGRR next 4 pixels */ /* RRGB -> RGB565 */  \
+    __asm movdqa     xmm3, xmm0 /* B  first 4 pixels of argb */                \
+    __asm movdqa     xmm2, xmm0 /* G */                                        \
+    __asm pslld      xmm0, 8 /* R */                                           \
+    __asm psrld      xmm3, 3 /* B */                                           \
+    __asm psrld      xmm2, 5 /* G */                                           \
+    __asm psrad      xmm0, 16 /* R */                                          \
+    __asm pand       xmm3, xmm5 /* B */                                        \
+    __asm pand       xmm2, xmm6 /* G */                                        \
+    __asm pand       xmm0, xmm7 /* R */                                        \
+    __asm por        xmm3, xmm2 /* BG */                                       \
+    __asm por        xmm0, xmm3 /* BGR */                                      \
+    __asm movdqa     xmm3, xmm1 /* B  next 4 pixels of argb */                 \
+    __asm movdqa     xmm2, xmm1 /* G */                                        \
+    __asm pslld      xmm1, 8 /* R */                                           \
+    __asm psrld      xmm3, 3 /* B */                                           \
+    __asm psrld      xmm2, 5 /* G */                                           \
+    __asm psrad      xmm1, 16 /* R */                                          \
+    __asm pand       xmm3, xmm5 /* B */                                        \
+    __asm pand       xmm2, xmm6 /* G */                                        \
+    __asm pand       xmm1, xmm7 /* R */                                        \
+    __asm por        xmm3, xmm2 /* BG */                                       \
+    __asm por        xmm1, xmm3 /* BGR */                                      \
     __asm packssdw   xmm0, xmm1                                                \
-    __asm movdqu     0[edx], xmm0 /* store 8 pixels of RGB565 */              \
+    __asm movdqu     0[edx], xmm0 /* store 8 pixels of RGB565 */               \
     __asm lea        edx, [edx + 16]}
 
 // 8 pixels.
@@ -2586,6 +2712,46 @@ __declspec(naked) void I444ToARGBRow_SSSE3(
 }
 
 // 8 pixels.
+// 8 UV values, mixed with 8 Y and 8A producing 8 ARGB (32 bytes).
+__declspec(naked) void I444AlphaToARGBRow_SSSE3(
+    const uint8_t* y_buf,
+    const uint8_t* u_buf,
+    const uint8_t* v_buf,
+    const uint8_t* a_buf,
+    uint8_t* dst_argb,
+    const struct YuvConstants* yuvconstants,
+    int width) {
+  __asm {
+    push       esi
+    push       edi
+    push       ebx
+    push       ebp
+    mov        eax, [esp + 16 + 4]  // Y
+    mov        esi, [esp + 16 + 8]  // U
+    mov        edi, [esp + 16 + 12]  // V
+    mov        ebp, [esp + 16 + 16]  // A
+    mov        edx, [esp + 16 + 20]  // argb
+    mov        ebx, [esp + 16 + 24]  // yuvconstants
+    mov        ecx, [esp + 16 + 28]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUVA444
+    YUVTORGB(ebx)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        ebp
+    pop        ebx
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
 // 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
 __declspec(naked) void I422ToRGB24Row_SSSE3(
     const uint8_t* y_buf,
@@ -2898,10 +3064,12 @@ __declspec(naked) void I422ToRGBARow_SSSE3(
 }
 #endif  // HAS_I422TOARGBROW_SSSE3
 
+// I400ToARGBRow_SSE2 is disabled due to new yuvconstant parameter
 #ifdef HAS_I400TOARGBROW_SSE2
 // 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
 __declspec(naked) void I400ToARGBRow_SSE2(const uint8_t* y_buf,
                                           uint8_t* rgb_buf,
+                                          const struct YuvConstants*,
                                           int width) {
   __asm {
     mov        eax, 0x4a354a35  // 4a35 = 18997 = round(1.164 * 64 * 256)
@@ -2949,6 +3117,7 @@ __declspec(naked) void I400ToARGBRow_SSE2(const uint8_t* y_buf,
 // note: vpunpcklbw mutates and vpackuswb unmutates.
 __declspec(naked) void I400ToARGBRow_AVX2(const uint8_t* y_buf,
                                           uint8_t* rgb_buf,
+                                          const struct YuvConstants*,
                                           int width) {
   __asm {
     mov        eax, 0x4a354a35  // 4a35 = 18997 = round(1.164 * 64 * 256)
@@ -3045,15 +3214,15 @@ __declspec(naked) void MirrorRow_AVX2(const uint8_t* src,
 }
 #endif  // HAS_MIRRORROW_AVX2
 
-#ifdef HAS_MIRRORUVROW_SSSE3
+#ifdef HAS_MIRRORSPLITUVROW_SSSE3
 // Shuffle table for reversing the bytes of UV channels.
 static const uvec8 kShuffleMirrorUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u,
                                        15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u};
 
-__declspec(naked) void MirrorUVRow_SSSE3(const uint8_t* src,
-                                         uint8_t* dst_u,
-                                         uint8_t* dst_v,
-                                         int width) {
+__declspec(naked) void MirrorSplitUVRow_SSSE3(const uint8_t* src,
+                                              uint8_t* dst_u,
+                                              uint8_t* dst_v,
+                                              int width) {
   __asm {
     push      edi
     mov       eax, [esp + 4 + 4]  // src
@@ -3078,7 +3247,7 @@ __declspec(naked) void MirrorUVRow_SSSE3(const uint8_t* src,
     ret
   }
 }
-#endif  // HAS_MIRRORUVROW_SSSE3
+#endif  // HAS_MIRRORSPLITUVROW_SSSE3
 
 #ifdef HAS_ARGBMIRRORROW_SSE2
 __declspec(naked) void ARGBMirrorRow_SSE2(const uint8_t* src,
@@ -4172,13 +4341,13 @@ static const uvec8 kShuffleAlpha = {3u,  0x80, 3u,  0x80, 7u,  0x80, 7u,  0x80,
                                     11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80};
 
 // Blend 8 pixels at a time.
-__declspec(naked) void ARGBBlendRow_SSSE3(const uint8_t* src_argb0,
+__declspec(naked) void ARGBBlendRow_SSSE3(const uint8_t* src_argb,
                                           const uint8_t* src_argb1,
                                           uint8_t* dst_argb,
                                           int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
@@ -4267,7 +4436,7 @@ __declspec(naked) void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb,
                                               uint8_t* dst_argb,
                                               int width) {
   __asm {
-    mov        eax, [esp + 4]  // src_argb0
+    mov        eax, [esp + 4]  // src_argb
     mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     pcmpeqb    xmm3, xmm3  // generate mask 0xff000000
@@ -4312,7 +4481,7 @@ __declspec(naked) void ARGBAttenuateRow_AVX2(const uint8_t* src_argb,
                                              uint8_t* dst_argb,
                                              int width) {
   __asm {
-    mov        eax, [esp + 4]  // src_argb0
+    mov        eax, [esp + 4]  // src_argb
     mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
@@ -4406,7 +4575,7 @@ __declspec(naked) void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb,
                                                uint8_t* dst_argb,
                                                int width) {
   __asm {
-    mov        eax, [esp + 4]  // src_argb0
+    mov        eax, [esp + 4]  // src_argb
     mov        edx, [esp + 8]  // dst_argb
     mov        ecx, [esp + 12]  // width
     sub        edx, eax
@@ -4762,20 +4931,20 @@ __declspec(naked) void ARGBShadeRow_SSE2(const uint8_t* src_argb,
 
 #ifdef HAS_ARGBMULTIPLYROW_SSE2
 // Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
-__declspec(naked) void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBMultiplyRow_SSE2(const uint8_t* src_argb,
                                             const uint8_t* src_argb1,
                                             uint8_t* dst_argb,
                                             int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     pxor       xmm5, xmm5  // constant 0
 
  convertloop:
-    movdqu     xmm0, [eax]  // read 4 pixels from src_argb0
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
     movdqu     xmm2, [esi]  // read 4 pixels from src_argb1
     movdqu     xmm1, xmm0
     movdqu     xmm3, xmm2
@@ -4783,8 +4952,8 @@ __declspec(naked) void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
     punpckhbw  xmm1, xmm1  // next 2
     punpcklbw  xmm2, xmm5  // first 2
     punpckhbw  xmm3, xmm5  // next 2
-    pmulhuw    xmm0, xmm2  // src_argb0 * src_argb1 first 2
-    pmulhuw    xmm1, xmm3  // src_argb0 * src_argb1 next 2
+    pmulhuw    xmm0, xmm2  // src_argb * src_argb1 first 2
+    pmulhuw    xmm1, xmm3  // src_argb * src_argb1 next 2
     lea        eax, [eax + 16]
     lea        esi, [esi + 16]
     packuswb   xmm0, xmm1
@@ -4802,13 +4971,13 @@ __declspec(naked) void ARGBMultiplyRow_SSE2(const uint8_t* src_argb0,
 #ifdef HAS_ARGBADDROW_SSE2
 // Add 2 rows of ARGB pixels together, 4 pixels at a time.
 // TODO(fbarchard): Port this to posix, neon and other math functions.
-__declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb,
                                        const uint8_t* src_argb1,
                                        uint8_t* dst_argb,
                                        int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
@@ -4817,11 +4986,11 @@ __declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb0,
     jl         convertloop49
 
  convertloop4:
-    movdqu     xmm0, [eax]  // read 4 pixels from src_argb0
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
     lea        eax, [eax + 16]
     movdqu     xmm1, [esi]  // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
-    paddusb    xmm0, xmm1  // src_argb0 + src_argb1
+    paddusb    xmm0, xmm1  // src_argb + src_argb1
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4832,11 +5001,11 @@ __declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb0,
     jl         convertloop19
 
  convertloop1:
-    movd       xmm0, [eax]  // read 1 pixels from src_argb0
+    movd       xmm0, [eax]  // read 1 pixels from src_argb
     lea        eax, [eax + 4]
     movd       xmm1, [esi]  // read 1 pixels from src_argb1
     lea        esi, [esi + 4]
-    paddusb    xmm0, xmm1  // src_argb0 + src_argb1
+    paddusb    xmm0, xmm1  // src_argb + src_argb1
     movd       [edx], xmm0
     lea        edx, [edx + 4]
     sub        ecx, 1
@@ -4851,23 +5020,23 @@ __declspec(naked) void ARGBAddRow_SSE2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBSUBTRACTROW_SSE2
 // Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
-__declspec(naked) void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBSubtractRow_SSE2(const uint8_t* src_argb,
                                             const uint8_t* src_argb1,
                                             uint8_t* dst_argb,
                                             int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    movdqu     xmm0, [eax]  // read 4 pixels from src_argb0
+    movdqu     xmm0, [eax]  // read 4 pixels from src_argb
     lea        eax, [eax + 16]
     movdqu     xmm1, [esi]  // read 4 pixels from src_argb1
     lea        esi, [esi + 16]
-    psubusb    xmm0, xmm1  // src_argb0 - src_argb1
+    psubusb    xmm0, xmm1  // src_argb - src_argb1
     movdqu     [edx], xmm0
     lea        edx, [edx + 16]
     sub        ecx, 4
@@ -4881,20 +5050,20 @@ __declspec(naked) void ARGBSubtractRow_SSE2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBMULTIPLYROW_AVX2
 // Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked) void ARGBMultiplyRow_AVX2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBMultiplyRow_AVX2(const uint8_t* src_argb,
                                             const uint8_t* src_argb1,
                                             uint8_t* dst_argb,
                                             int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
     vpxor      ymm5, ymm5, ymm5  // constant 0
 
  convertloop:
-    vmovdqu    ymm1, [eax]  // read 8 pixels from src_argb0
+    vmovdqu    ymm1, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
     vmovdqu    ymm3, [esi]  // read 8 pixels from src_argb1
     lea        esi, [esi + 32]
@@ -4902,8 +5071,8 @@ __declspec(naked) void ARGBMultiplyRow_AVX2(const uint8_t* src_argb0,
     vpunpckhbw ymm1, ymm1, ymm1  // high 4
     vpunpcklbw ymm2, ymm3, ymm5  // low 4
     vpunpckhbw ymm3, ymm3, ymm5  // high 4
-    vpmulhuw   ymm0, ymm0, ymm2  // src_argb0 * src_argb1 low 4
-    vpmulhuw   ymm1, ymm1, ymm3  // src_argb0 * src_argb1 high 4
+    vpmulhuw   ymm0, ymm0, ymm2  // src_argb * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3  // src_argb * src_argb1 high 4
     vpackuswb  ymm0, ymm0, ymm1
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -4919,19 +5088,19 @@ __declspec(naked) void ARGBMultiplyRow_AVX2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBADDROW_AVX2
 // Add 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked) void ARGBAddRow_AVX2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBAddRow_AVX2(const uint8_t* src_argb,
                                        const uint8_t* src_argb1,
                                        uint8_t* dst_argb,
                                        int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb0
+    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
     vpaddusb   ymm0, ymm0, [esi]  // add 8 pixels from src_argb1
     lea        esi, [esi + 32]
@@ -4949,21 +5118,21 @@ __declspec(naked) void ARGBAddRow_AVX2(const uint8_t* src_argb0,
 
 #ifdef HAS_ARGBSUBTRACTROW_AVX2
 // Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
-__declspec(naked) void ARGBSubtractRow_AVX2(const uint8_t* src_argb0,
+__declspec(naked) void ARGBSubtractRow_AVX2(const uint8_t* src_argb,
                                             const uint8_t* src_argb1,
                                             uint8_t* dst_argb,
                                             int width) {
   __asm {
     push       esi
-    mov        eax, [esp + 4 + 4]  // src_argb0
+    mov        eax, [esp + 4 + 4]  // src_argb
     mov        esi, [esp + 4 + 8]  // src_argb1
     mov        edx, [esp + 4 + 12]  // dst_argb
     mov        ecx, [esp + 4 + 16]  // width
 
  convertloop:
-    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb0
+    vmovdqu    ymm0, [eax]  // read 8 pixels from src_argb
     lea        eax, [eax + 32]
-    vpsubusb   ymm0, ymm0, [esi]  // src_argb0 - src_argb1
+    vpsubusb   ymm0, ymm0, [esi]  // src_argb - src_argb1
     lea        esi, [esi + 32]
     vmovdqu    [edx], ymm0
     lea        edx, [edx + 32]
@@ -5450,7 +5619,7 @@ void ComputeCumulativeSumRow_SSE2(const uint8_t* row,
 
             // 1 pixel loop
   l1:
-    movd       xmm2, dword ptr [eax]  // 1 argb pixel, 4 bytes.
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel
     lea        eax, [eax + 4]
     punpcklbw  xmm2, xmm1
     punpcklwd  xmm2, xmm1
diff --git a/files/source/scale.cc b/files/source/scale.cc
index ab085496..e1335f1e 100644
--- a/files/source/scale.cc
+++ b/files/source/scale.cc
@@ -17,6 +17,7 @@
 #include "libyuv/planar_functions.h"  // For CopyPlane
 #include "libyuv/row.h"
 #include "libyuv/scale_row.h"
+#include "libyuv/scale_uv.h"  // For UVScale
 
 #ifdef __cplusplus
 namespace libyuv {
@@ -28,6 +29,7 @@ static __inline int Abs(int v) {
 }
 
 #define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
 
 // Scale plane, 1/2
 // This is an optimized version for scaling down a plane to 1/2 of
@@ -49,7 +51,7 @@ static void ScalePlaneDown2(int src_width,
           ? ScaleRowDown2_C
           : (filtering == kFilterLinear ? ScaleRowDown2Linear_C
                                         : ScaleRowDown2Box_C);
-  int row_stride = src_stride << 1;
+  int row_stride = src_stride * 2;
   (void)src_width;
   (void)src_height;
   if (!filtering) {
@@ -118,18 +120,18 @@ static void ScalePlaneDown2(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN2_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
+#if defined(HAS_SCALEROWDOWN2_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
     ScaleRowDown2 =
         filtering == kFilterNone
-            ? ScaleRowDown2_Any_MMI
-            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_MMI
-                                          : ScaleRowDown2Box_Any_MMI);
-    if (IS_ALIGNED(dst_width, 8)) {
-      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_MMI
+            ? ScaleRowDown2_Any_LSX
+            : (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_LSX
+                                          : ScaleRowDown2Box_Any_LSX);
+    if (IS_ALIGNED(dst_width, 32)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_LSX
                                                : (filtering == kFilterLinear
-                                                      ? ScaleRowDown2Linear_MMI
-                                                      : ScaleRowDown2Box_MMI);
+                                                      ? ScaleRowDown2Linear_LSX
+                                                      : ScaleRowDown2Box_LSX);
     }
   }
 #endif
@@ -161,7 +163,7 @@ static void ScalePlaneDown2_16(int src_width,
           ? ScaleRowDown2_16_C
           : (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C
                                         : ScaleRowDown2Box_16_C);
-  int row_stride = src_stride << 1;
+  int row_stride = src_stride * 2;
   (void)src_width;
   (void)src_height;
   if (!filtering) {
@@ -184,14 +186,6 @@ static void ScalePlaneDown2_16(int src_width,
                                           : ScaleRowDown2Box_16_SSE2);
   }
 #endif
-#if defined(HAS_SCALEROWDOWN2_16_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 4)) {
-    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_MMI
-                                             : (filtering == kFilterLinear
-                                                    ? ScaleRowDown2Linear_16_MMI
-                                                    : ScaleRowDown2Box_16_MMI);
-  }
-#endif
 
   if (filtering == kFilterLinear) {
     src_stride = 0;
@@ -221,7 +215,7 @@ static void ScalePlaneDown4(int src_width,
   void (*ScaleRowDown4)(const uint8_t* src_ptr, ptrdiff_t src_stride,
                         uint8_t* dst_ptr, int dst_width) =
       filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
-  int row_stride = src_stride << 2;
+  int row_stride = src_stride * 4;
   (void)src_width;
   (void)src_height;
   if (!filtering) {
@@ -264,12 +258,12 @@ static void ScalePlaneDown4(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEROWDOWN4_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
+#if defined(HAS_SCALEROWDOWN4_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
     ScaleRowDown4 =
-        filtering ? ScaleRowDown4Box_Any_MMI : ScaleRowDown4_Any_MMI;
-    if (IS_ALIGNED(dst_width, 8)) {
-      ScaleRowDown4 = filtering ? ScaleRowDown4Box_MMI : ScaleRowDown4_MMI;
+        filtering ? ScaleRowDown4Box_Any_LSX : ScaleRowDown4_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_LSX : ScaleRowDown4_LSX;
     }
   }
 #endif
@@ -297,7 +291,7 @@ static void ScalePlaneDown4_16(int src_width,
   void (*ScaleRowDown4)(const uint16_t* src_ptr, ptrdiff_t src_stride,
                         uint16_t* dst_ptr, int dst_width) =
       filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
-  int row_stride = src_stride << 2;
+  int row_stride = src_stride * 4;
   (void)src_width;
   (void)src_height;
   if (!filtering) {
@@ -316,11 +310,6 @@ static void ScalePlaneDown4_16(int src_width,
         filtering ? ScaleRowDown4Box_16_SSE2 : ScaleRowDown4_16_SSE2;
   }
 #endif
-#if defined(HAS_SCALEROWDOWN4_16_MMI)
-  if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 8)) {
-    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_MMI : ScaleRowDown4_16_MMI;
-  }
-#endif
 
   if (filtering == kFilterLinear) {
     src_stride = 0;
@@ -398,6 +387,26 @@ static void ScalePlaneDown34(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEROWDOWN34_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_LSX;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_LSX;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_LSX;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_LSX;
+    }
+    if (dst_width % 48 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_LSX;
+        ScaleRowDown34_1 = ScaleRowDown34_LSX;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_LSX;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_LSX;
+      }
+    }
+  }
+#endif
 #if defined(HAS_SCALEROWDOWN34_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     if (!filtering) {
@@ -613,6 +622,26 @@ static void ScalePlaneDown38(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEROWDOWN38_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_LSX;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_LSX;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_LSX;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_LSX;
+    }
+    if (dst_width % 12 == 0) {
+      if (!filtering) {
+        ScaleRowDown38_3 = ScaleRowDown38_LSX;
+        ScaleRowDown38_2 = ScaleRowDown38_LSX;
+      } else {
+        ScaleRowDown38_3 = ScaleRowDown38_3_Box_LSX;
+        ScaleRowDown38_2 = ScaleRowDown38_2_Box_LSX;
+      }
+    }
+  }
+#endif
 
   for (y = 0; y < dst_height - 2; y += 3) {
     ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
@@ -886,11 +915,11 @@ static void ScalePlaneBox(int src_width,
       }
     }
 #endif
-#if defined(HAS_SCALEADDROW_MMI)
-    if (TestCpuFlag(kCpuHasMMI)) {
-      ScaleAddRow = ScaleAddRow_Any_MMI;
-      if (IS_ALIGNED(src_width, 8)) {
-        ScaleAddRow = ScaleAddRow_MMI;
+#if defined(HAS_SCALEADDROW_LSX)
+    if (TestCpuFlag(kCpuHasLSX)) {
+      ScaleAddRow = ScaleAddRow_Any_LSX;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_LSX;
       }
     }
 #endif
@@ -898,7 +927,7 @@ static void ScalePlaneBox(int src_width,
     for (j = 0; j < dst_height; ++j) {
       int boxheight;
       int iy = y >> 16;
-      const uint8_t* src = src_ptr + iy * src_stride;
+      const uint8_t* src = src_ptr + iy * (int64_t)src_stride;
       y += dy;
       if (y > max_y) {
         y = max_y;
@@ -949,15 +978,10 @@ static void ScalePlaneBox_16(int src_width,
     }
 #endif
 
-#if defined(HAS_SCALEADDROW_16_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(src_width, 4)) {
-      ScaleAddRow = ScaleAddRow_16_MMI;
-    }
-#endif
     for (j = 0; j < dst_height; ++j) {
       int boxheight;
       int iy = y >> 16;
-      const uint16_t* src = src_ptr + iy * src_stride;
+      const uint16_t* src = src_ptr + iy * (int64_t)src_stride;
       y += dy;
       if (y > max_y) {
         y = max_y;
@@ -1038,11 +1062,11 @@ void ScalePlaneBilinearDown(int src_width,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(src_width, 16)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -1068,13 +1092,21 @@ void ScalePlaneBilinearDown(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEFILTERCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_LSX;
+    }
+  }
+#endif
   if (y > max_y) {
     y = max_y;
   }
 
   for (j = 0; j < dst_height; ++j) {
     int yi = y >> 16;
-    const uint8_t* src = src_ptr + yi * src_stride;
+    const uint8_t* src = src_ptr + yi * (int64_t)src_stride;
     if (filtering == kFilterLinear) {
       ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
     } else {
@@ -1123,7 +1155,7 @@ void ScalePlaneBilinearDown_16(int src_width,
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
@@ -1131,7 +1163,7 @@ void ScalePlaneBilinearDown_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
@@ -1139,7 +1171,7 @@ void ScalePlaneBilinearDown_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
     if (IS_ALIGNED(src_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
@@ -1147,7 +1179,7 @@ void ScalePlaneBilinearDown_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
+    InterpolateRow = InterpolateRow_16_Any_NEON;
     if (IS_ALIGNED(src_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
@@ -1165,7 +1197,7 @@ void ScalePlaneBilinearDown_16(int src_width,
 
   for (j = 0; j < dst_height; ++j) {
     int yi = y >> 16;
-    const uint16_t* src = src_ptr + yi * src_stride;
+    const uint16_t* src = src_ptr + yi * (int64_t)src_stride;
     if (filtering == kFilterLinear) {
       ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
     } else {
@@ -1258,6 +1290,14 @@ void ScalePlaneBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleFilterCols = ScaleFilterCols_LSX;
+    }
+  }
+#endif
   if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
     ScaleFilterCols = ScaleColsUp2_C;
 #if defined(HAS_SCALECOLS_SSE2)
@@ -1265,11 +1305,6 @@ void ScalePlaneBilinearUp(int src_width,
       ScaleFilterCols = ScaleColsUp2_SSE2;
     }
 #endif
-#if defined(HAS_SCALECOLS_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 8)) {
-      ScaleFilterCols = ScaleColsUp2_MMI;
-    }
-#endif
   }
 
   if (y > max_y) {
@@ -1277,7 +1312,7 @@ void ScalePlaneBilinearUp(int src_width,
   }
   {
     int yi = y >> 16;
-    const uint8_t* src = src_ptr + yi * src_stride;
+    const uint8_t* src = src_ptr + yi * (int64_t)src_stride;
 
     // Allocate 2 row buffers.
     const int kRowSize = (dst_width + 31) & ~31;
@@ -1292,7 +1327,9 @@ void ScalePlaneBilinearUp(int src_width,
       src += src_stride;
     }
     ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -1300,14 +1337,16 @@ void ScalePlaneBilinearUp(int src_width,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_ptr + yi * src_stride;
+          src = src_ptr + yi * (int64_t)src_stride;
         }
         if (yi != lasty) {
           ScaleFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -1323,6 +1362,327 @@ void ScalePlaneBilinearUp(int src_width,
   }
 }
 
+// Scale plane, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of I422 to I444.
+void ScalePlaneUp2_Linear(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_ptr,
+                          uint8_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint8_t* src_ptr, uint8_t* dst_ptr, int dst_width) =
+      ScaleRowUp2_Linear_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of I420 to I444.
+void ScalePlaneUp2_Bilinear(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_ptr,
+                            uint8_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale at most 14 bit plane, horizontally up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// stride is in count of uint16_t.
+// This is used to scale U and V planes of I210 to I410 and I212 to I412.
+void ScalePlaneUp2_12_Linear(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_12_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale at most 12 bit plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// stride is in count of uint16_t.
+// This is used to scale U and V planes of I010 to I410 and I012 to I412.
+void ScalePlaneUp2_12_Bilinear(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_12_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+void ScalePlaneUp2_16_Linear(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             int src_stride,
+                             int dst_stride,
+                             const uint16_t* src_ptr,
+                             uint16_t* dst_ptr) {
+  void (*ScaleRowUp)(const uint16_t* src_ptr, uint16_t* dst_ptr,
+                     int dst_width) = ScaleRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_ptr + ((src_height - 1) / 2) * (int64_t)src_stride, dst_ptr,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_ptr + (y >> 16) * (int64_t)src_stride, dst_ptr, dst_width);
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+void ScalePlaneUp2_16_Bilinear(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               int src_stride,
+                               int dst_stride,
+                               const uint16_t* src_ptr,
+                               uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_SSE2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
 void ScalePlaneBilinearUp_16(int src_width,
                              int src_height,
                              int dst_width,
@@ -1351,7 +1711,7 @@ void ScalePlaneBilinearUp_16(int src_width,
 
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
@@ -1359,7 +1719,7 @@ void ScalePlaneBilinearUp_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
@@ -1367,7 +1727,7 @@ void ScalePlaneBilinearUp_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
     if (IS_ALIGNED(dst_width, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
@@ -1375,7 +1735,7 @@ void ScalePlaneBilinearUp_16(int src_width,
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
+    InterpolateRow = InterpolateRow_16_Any_NEON;
     if (IS_ALIGNED(dst_width, 16)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
@@ -1397,19 +1757,13 @@ void ScalePlaneBilinearUp_16(int src_width,
       ScaleFilterCols = ScaleColsUp2_16_SSE2;
     }
 #endif
-#if defined(HAS_SCALECOLS_16_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 8)) {
-      ScaleFilterCols = ScaleColsUp2_16_MMI;
-    }
-#endif
   }
-
   if (y > max_y) {
     y = max_y;
   }
   {
     int yi = y >> 16;
-    const uint16_t* src = src_ptr + yi * src_stride;
+    const uint16_t* src = src_ptr + yi * (int64_t)src_stride;
 
     // Allocate 2 row buffers.
     const int kRowSize = (dst_width + 31) & ~31;
@@ -1424,7 +1778,9 @@ void ScalePlaneBilinearUp_16(int src_width,
       src += src_stride;
     }
     ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -1432,14 +1788,16 @@ void ScalePlaneBilinearUp_16(int src_width,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_ptr + yi * src_stride;
+          src = src_ptr + yi * (int64_t)src_stride;
         }
         if (yi != lasty) {
           ScaleFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -1487,15 +1845,11 @@ static void ScalePlaneSimple(int src_width,
       ScaleCols = ScaleColsUp2_SSE2;
     }
 #endif
-#if defined(HAS_SCALECOLS_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 8)) {
-      ScaleCols = ScaleColsUp2_MMI;
-    }
-#endif
   }
 
   for (i = 0; i < dst_height; ++i) {
-    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx);
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * (int64_t)src_stride, dst_width, x,
+              dx);
     dst_ptr += dst_stride;
     y += dy;
   }
@@ -1528,15 +1882,11 @@ static void ScalePlaneSimple_16(int src_width,
       ScaleCols = ScaleColsUp2_16_SSE2;
     }
 #endif
-#if defined(HAS_SCALECOLS_16_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 8)) {
-      ScaleCols = ScaleColsUp2_16_MMI;
-    }
-#endif
   }
 
   for (i = 0; i < dst_height; ++i) {
-    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx);
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * (int64_t)src_stride, dst_width, x,
+              dx);
     dst_ptr += dst_stride;
     y += dy;
   }
@@ -1544,7 +1894,6 @@ static void ScalePlaneSimple_16(int src_width,
 
 // Scale a plane.
 // This function dispatches to a specialized scaler based on scale factor.
-
 LIBYUV_API
 void ScalePlane(const uint8_t* src,
                 int src_stride,
@@ -1562,10 +1911,9 @@ void ScalePlane(const uint8_t* src,
   // Negative height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (int64_t)src_stride;
     src_stride = -src_stride;
   }
-
   // Use specialized scales to improve performance for common resolutions.
   // For example, all the 1/2 scalings will use ScalePlaneDown2()
   if (dst_width == src_width && dst_height == src_height) {
@@ -1574,10 +1922,19 @@ void ScalePlane(const uint8_t* src,
     return;
   }
   if (dst_width == src_width && filtering != kFilterBox) {
-    int dy = FixedDiv(src_height, dst_height);
+    int dy = 0;
+    int y = 0;
+    // When scaling down, use the center 2 rows to filter.
+    // When scaling up, last row of destination uses the last 2 source rows.
+    if (dst_height <= src_height) {
+      dy = FixedDiv(src_height, dst_height);
+      y = CENTERSTART(dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (src_height > 1 && dst_height > 1) {
+      dy = FixedDiv1(src_height, dst_height);
+    }
     // Arbitrary scale vertically, but unscaled horizontally.
     ScalePlaneVertical(src_height, dst_width, dst_height, src_stride,
-                       dst_stride, src, dst, 0, 0, dy, 1, filtering);
+                       dst_stride, src, dst, 0, y, dy, /*bpp=*/1, filtering);
     return;
   }
   if (dst_width <= Abs(src_width) && dst_height <= src_height) {
@@ -1614,6 +1971,17 @@ void ScalePlane(const uint8_t* src,
                   dst_stride, src, dst);
     return;
   }
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_Linear(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_Bilinear(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst);
+    return;
+  }
   if (filtering && dst_height > src_height) {
     ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
                          src_stride, dst_stride, src, dst, filtering);
@@ -1645,10 +2013,9 @@ void ScalePlane_16(const uint16_t* src,
   // Negative height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (int64_t)src_stride;
     src_stride = -src_stride;
   }
-
   // Use specialized scales to improve performance for common resolutions.
   // For example, all the 1/2 scalings will use ScalePlaneDown2()
   if (dst_width == src_width && dst_height == src_height) {
@@ -1657,10 +2024,22 @@ void ScalePlane_16(const uint16_t* src,
     return;
   }
   if (dst_width == src_width && filtering != kFilterBox) {
-    int dy = FixedDiv(src_height, dst_height);
-    // Arbitrary scale vertically, but unscaled vertically.
+    int dy = 0;
+    int y = 0;
+    // When scaling down, use the center 2 rows to filter.
+    // When scaling up, last row of destination uses the last 2 source rows.
+    if (dst_height <= src_height) {
+      dy = FixedDiv(src_height, dst_height);
+      y = CENTERSTART(dy, -32768);  // Subtract 0.5 (32768) to center filter.
+      // When scaling up, ensure the last row of destination uses the last
+      // source. Avoid divide by zero for dst_height but will do no scaling
+      // later.
+    } else if (src_height > 1 && dst_height > 1) {
+      dy = FixedDiv1(src_height, dst_height);
+    }
+    // Arbitrary scale vertically, but unscaled horizontally.
     ScalePlaneVertical_16(src_height, dst_width, dst_height, src_stride,
-                          dst_stride, src, dst, 0, 0, dy, 1, filtering);
+                          dst_stride, src, dst, 0, y, dy, /*bpp=*/1, filtering);
     return;
   }
   if (dst_width <= Abs(src_width) && dst_height <= src_height) {
@@ -1697,6 +2076,17 @@ void ScalePlane_16(const uint16_t* src,
                      dst_stride, src, dst);
     return;
   }
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_16_Linear(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_16_Bilinear(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst);
+    return;
+  }
   if (filtering && dst_height > src_height) {
     ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height,
                             src_stride, dst_stride, src, dst, filtering);
@@ -1711,6 +2101,43 @@ void ScalePlane_16(const uint16_t* src,
                       dst_stride, src, dst);
 }
 
+LIBYUV_API
+void ScalePlane_12(const uint16_t* src,
+                   int src_stride,
+                   int src_width,
+                   int src_height,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int dst_width,
+                   int dst_height,
+                   enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * (int64_t)src_stride;
+    src_stride = -src_stride;
+  }
+
+  if ((dst_width + 1) / 2 == src_width && filtering == kFilterLinear) {
+    ScalePlaneUp2_12_Linear(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst);
+    return;
+  }
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScalePlaneUp2_12_Bilinear(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst);
+    return;
+  }
+
+  ScalePlane_16(src, src_stride, src_width, src_height, dst, dst_stride,
+                dst_width, dst_height, filtering);
+}
+
 // Scale an I420 image.
 // This function in turn calls a scaling function for each plane.
 
@@ -1736,7 +2163,8 @@ int I420Scale(const uint8_t* src_y,
   int src_halfheight = SUBSAMPLE(src_height, 1, 1);
   int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
   int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
       src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
       dst_width <= 0 || dst_height <= 0) {
     return -1;
@@ -1773,7 +2201,8 @@ int I420Scale_16(const uint16_t* src_y,
   int src_halfheight = SUBSAMPLE(src_height, 1, 1);
   int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
   int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
       src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
       dst_width <= 0 || dst_height <= 0) {
     return -1;
@@ -1788,6 +2217,44 @@ int I420Scale_16(const uint16_t* src_y,
   return 0;
 }
 
+LIBYUV_API
+int I420Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_halfwidth, src_halfheight, dst_u,
+                dst_stride_u, dst_halfwidth, dst_halfheight, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_halfwidth, src_halfheight, dst_v,
+                dst_stride_v, dst_halfwidth, dst_halfheight, filtering);
+  return 0;
+}
+
 // Scale an I444 image.
 // This function in turn calls a scaling function for each plane.
 
@@ -1809,7 +2276,7 @@ int I444Scale(const uint8_t* src_y,
               int dst_width,
               int dst_height,
               enum FilterMode filtering) {
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
       src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
       dst_width <= 0 || dst_height <= 0) {
     return -1;
@@ -1842,7 +2309,7 @@ int I444Scale_16(const uint16_t* src_y,
                  int dst_width,
                  int dst_height,
                  enum FilterMode filtering) {
-  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
       src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
       dst_width <= 0 || dst_height <= 0) {
     return -1;
@@ -1857,6 +2324,185 @@ int I444Scale_16(const uint16_t* src_y,
   return 0;
 }
 
+LIBYUV_API
+int I444Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_width, src_height, dst_u, dst_stride_u,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_width, src_height, dst_v, dst_stride_v,
+                dst_width, dst_height, filtering);
+  return 0;
+}
+
+// Scale an I422 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I422Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_u,
+              int src_stride_u,
+              const uint8_t* src_v,
+              int src_stride_v,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_u,
+              int dst_stride_u,
+              uint8_t* dst_v,
+              int dst_stride_v,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+             dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+             dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I422Scale_16(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+                dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+                dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I422Scale_12(const uint16_t* src_y,
+                 int src_stride_y,
+                 const uint16_t* src_u,
+                 int src_stride_u,
+                 const uint16_t* src_v,
+                 int src_stride_v,
+                 int src_width,
+                 int src_height,
+                 uint16_t* dst_y,
+                 int dst_stride_y,
+                 uint16_t* dst_u,
+                 int dst_stride_u,
+                 uint16_t* dst_v,
+                 int dst_stride_v,
+                 int dst_width,
+                 int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+
+  if (!src_y || !src_u || !src_v || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_u || !dst_v ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_12(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+                dst_width, dst_height, filtering);
+  ScalePlane_12(src_u, src_stride_u, src_halfwidth, src_height, dst_u,
+                dst_stride_u, dst_halfwidth, dst_height, filtering);
+  ScalePlane_12(src_v, src_stride_v, src_halfwidth, src_height, dst_v,
+                dst_stride_v, dst_halfwidth, dst_height, filtering);
+  return 0;
+}
+
+// Scale an NV12 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int NV12Scale(const uint8_t* src_y,
+              int src_stride_y,
+              const uint8_t* src_uv,
+              int src_stride_uv,
+              int src_width,
+              int src_height,
+              uint8_t* dst_y,
+              int dst_stride_y,
+              uint8_t* dst_uv,
+              int dst_stride_uv,
+              int dst_width,
+              int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+
+  if (!src_y || !src_uv || src_width <= 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 || !dst_y || !dst_uv ||
+      dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height, dst_y, dst_stride_y,
+             dst_width, dst_height, filtering);
+  UVScale(src_uv, src_stride_uv, src_halfwidth, src_halfheight, dst_uv,
+          dst_stride_uv, dst_halfwidth, dst_halfheight, filtering);
+  return 0;
+}
+
 // Deprecated api
 LIBYUV_API
 int Scale(const uint8_t* src_y,
diff --git a/files/source/scale_any.cc b/files/source/scale_any.cc
index 17831372..317041f8 100644
--- a/files/source/scale_any.cc
+++ b/files/source/scale_any.cc
@@ -20,49 +20,6 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
-#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
-  void NAMEANY(uint8_t* dst_ptr, const uint8_t* src_ptr, int dst_width, int x, \
-               int dx) {                                                       \
-    int r = dst_width & MASK;                                                  \
-    int n = dst_width & ~MASK;                                                 \
-    if (n > 0) {                                                               \
-      TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                   \
-    }                                                                          \
-    TERP_C(dst_ptr + n * BPP, src_ptr, r, x + n * dx, dx);                     \
-  }
-
-#ifdef HAS_SCALEFILTERCOLS_NEON
-CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
-#endif
-#ifdef HAS_SCALEFILTERCOLS_MSA
-CANY(ScaleFilterCols_Any_MSA, ScaleFilterCols_MSA, ScaleFilterCols_C, 1, 15)
-#endif
-#ifdef HAS_SCALEARGBCOLS_NEON
-CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
-#endif
-#ifdef HAS_SCALEARGBCOLS_MSA
-CANY(ScaleARGBCols_Any_MSA, ScaleARGBCols_MSA, ScaleARGBCols_C, 4, 3)
-#endif
-#ifdef HAS_SCALEARGBCOLS_MMI
-CANY(ScaleARGBCols_Any_MMI, ScaleARGBCols_MMI, ScaleARGBCols_C, 4, 0)
-#endif
-#ifdef HAS_SCALEARGBFILTERCOLS_NEON
-CANY(ScaleARGBFilterCols_Any_NEON,
-     ScaleARGBFilterCols_NEON,
-     ScaleARGBFilterCols_C,
-     4,
-     3)
-#endif
-#ifdef HAS_SCALEARGBFILTERCOLS_MSA
-CANY(ScaleARGBFilterCols_Any_MSA,
-     ScaleARGBFilterCols_MSA,
-     ScaleARGBFilterCols_C,
-     4,
-     7)
-#endif
-#undef CANY
-
 // Fixed scale down.
 // Mask may be non-power of 2, so use MOD
 #define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
@@ -113,6 +70,22 @@ SDODD(ScaleRowDown2Box_Odd_SSSE3,
       1,
       15)
 #endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_SSSE3
+SDANY(ScaleUVRowDown2Box_Any_SSSE3,
+      ScaleUVRowDown2Box_SSSE3,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      3)
+#endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_AVX2
+SDANY(ScaleUVRowDown2Box_Any_AVX2,
+      ScaleUVRowDown2Box_AVX2,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      7)
+#endif
 #ifdef HAS_SCALEROWDOWN2_AVX2
 SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
 SDANY(ScaleRowDown2Linear_Any_AVX2,
@@ -155,6 +128,15 @@ SDODD(ScaleRowDown2Box_Odd_NEON,
       1,
       15)
 #endif
+#ifdef HAS_SCALEUVROWDOWN2BOX_NEON
+SDANY(ScaleUVRowDown2Box_Any_NEON,
+      ScaleUVRowDown2Box_NEON,
+      ScaleUVRowDown2Box_C,
+      2,
+      2,
+      7)
+#endif
+
 #ifdef HAS_SCALEROWDOWN2_MSA
 SDANY(ScaleRowDown2_Any_MSA, ScaleRowDown2_MSA, ScaleRowDown2_C, 2, 1, 31)
 SDANY(ScaleRowDown2Linear_Any_MSA,
@@ -170,26 +152,20 @@ SDANY(ScaleRowDown2Box_Any_MSA,
       1,
       31)
 #endif
-#ifdef HAS_SCALEROWDOWN2_MMI
-SDANY(ScaleRowDown2_Any_MMI, ScaleRowDown2_MMI, ScaleRowDown2_C, 2, 1, 7)
-SDANY(ScaleRowDown2Linear_Any_MMI,
-      ScaleRowDown2Linear_MMI,
+#ifdef HAS_SCALEROWDOWN2_LSX
+SDANY(ScaleRowDown2_Any_LSX, ScaleRowDown2_LSX, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_LSX,
+      ScaleRowDown2Linear_LSX,
       ScaleRowDown2Linear_C,
       2,
       1,
-      7)
-SDANY(ScaleRowDown2Box_Any_MMI,
-      ScaleRowDown2Box_MMI,
+      31)
+SDANY(ScaleRowDown2Box_Any_LSX,
+      ScaleRowDown2Box_LSX,
       ScaleRowDown2Box_C,
       2,
       1,
-      7)
-SDODD(ScaleRowDown2Box_Odd_MMI,
-      ScaleRowDown2Box_MMI,
-      ScaleRowDown2Box_Odd_C,
-      2,
-      1,
-      7)
+      31)
 #endif
 #ifdef HAS_SCALEROWDOWN4_SSSE3
 SDANY(ScaleRowDown4_Any_SSSE3, ScaleRowDown4_SSSE3, ScaleRowDown4_C, 4, 1, 7)
@@ -227,14 +203,14 @@ SDANY(ScaleRowDown4Box_Any_MSA,
       1,
       15)
 #endif
-#ifdef HAS_SCALEROWDOWN4_MMI
-SDANY(ScaleRowDown4_Any_MMI, ScaleRowDown4_MMI, ScaleRowDown4_C, 4, 1, 7)
-SDANY(ScaleRowDown4Box_Any_MMI,
-      ScaleRowDown4Box_MMI,
+#ifdef HAS_SCALEROWDOWN4_LSX
+SDANY(ScaleRowDown4_Any_LSX, ScaleRowDown4_LSX, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_LSX,
+      ScaleRowDown4Box_LSX,
       ScaleRowDown4Box_C,
       4,
       1,
-      7)
+      15)
 #endif
 #ifdef HAS_SCALEROWDOWN34_SSSE3
 SDANY(ScaleRowDown34_Any_SSSE3,
@@ -296,6 +272,26 @@ SDANY(ScaleRowDown34_1_Box_Any_MSA,
       1,
       47)
 #endif
+#ifdef HAS_SCALEROWDOWN34_LSX
+SDANY(ScaleRowDown34_Any_LSX,
+      ScaleRowDown34_LSX,
+      ScaleRowDown34_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_0_Box_Any_LSX,
+      ScaleRowDown34_0_Box_LSX,
+      ScaleRowDown34_0_Box_C,
+      4 / 3,
+      1,
+      47)
+SDANY(ScaleRowDown34_1_Box_Any_LSX,
+      ScaleRowDown34_1_Box_LSX,
+      ScaleRowDown34_1_Box_C,
+      4 / 3,
+      1,
+      47)
+#endif
 #ifdef HAS_SCALEROWDOWN38_SSSE3
 SDANY(ScaleRowDown38_Any_SSSE3,
       ScaleRowDown38_SSSE3,
@@ -356,6 +352,26 @@ SDANY(ScaleRowDown38_2_Box_Any_MSA,
       1,
       11)
 #endif
+#ifdef HAS_SCALEROWDOWN38_LSX
+SDANY(ScaleRowDown38_Any_LSX,
+      ScaleRowDown38_LSX,
+      ScaleRowDown38_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_3_Box_Any_LSX,
+      ScaleRowDown38_3_Box_LSX,
+      ScaleRowDown38_3_Box_C,
+      8 / 3,
+      1,
+      11)
+SDANY(ScaleRowDown38_2_Box_Any_LSX,
+      ScaleRowDown38_2_Box_LSX,
+      ScaleRowDown38_2_Box_C,
+      8 / 3,
+      1,
+      11)
+#endif
 
 #ifdef HAS_SCALEARGBROWDOWN2_SSE2
 SDANY(ScaleARGBRowDown2_Any_SSE2,
@@ -417,25 +433,25 @@ SDANY(ScaleARGBRowDown2Box_Any_MSA,
       4,
       3)
 #endif
-#ifdef HAS_SCALEARGBROWDOWN2_MMI
-SDANY(ScaleARGBRowDown2_Any_MMI,
-      ScaleARGBRowDown2_MMI,
+#ifdef HAS_SCALEARGBROWDOWN2_LSX
+SDANY(ScaleARGBRowDown2_Any_LSX,
+      ScaleARGBRowDown2_LSX,
       ScaleARGBRowDown2_C,
       2,
       4,
-      1)
-SDANY(ScaleARGBRowDown2Linear_Any_MMI,
-      ScaleARGBRowDown2Linear_MMI,
+      3)
+SDANY(ScaleARGBRowDown2Linear_Any_LSX,
+      ScaleARGBRowDown2Linear_LSX,
       ScaleARGBRowDown2Linear_C,
       2,
       4,
-      1)
-SDANY(ScaleARGBRowDown2Box_Any_MMI,
-      ScaleARGBRowDown2Box_MMI,
+      3)
+SDANY(ScaleARGBRowDown2Box_Any_LSX,
+      ScaleARGBRowDown2Box_LSX,
       ScaleARGBRowDown2Box_C,
       2,
       4,
-      1)
+      3)
 #endif
 #undef SDANY
 
@@ -488,17 +504,24 @@ SDAANY(ScaleARGBRowDownEvenBox_Any_MSA,
        4,
        3)
 #endif
-#ifdef HAS_SCALEARGBROWDOWNEVEN_MMI
-SDAANY(ScaleARGBRowDownEven_Any_MMI,
-       ScaleARGBRowDownEven_MMI,
+#ifdef HAS_SCALEARGBROWDOWNEVEN_LSX
+SDAANY(ScaleARGBRowDownEven_Any_LSX,
+       ScaleARGBRowDownEven_LSX,
        ScaleARGBRowDownEven_C,
        4,
-       1)
-SDAANY(ScaleARGBRowDownEvenBox_Any_MMI,
-       ScaleARGBRowDownEvenBox_MMI,
+       3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_LSX,
+       ScaleARGBRowDownEvenBox_LSX,
        ScaleARGBRowDownEvenBox_C,
        4,
-       1)
+       3)
+#endif
+#ifdef HAS_SCALEUVROWDOWNEVEN_NEON
+SDAANY(ScaleUVRowDownEven_Any_NEON,
+       ScaleUVRowDownEven_NEON,
+       ScaleUVRowDownEven_C,
+       2,
+       3)
 #endif
 
 #ifdef SASIMDONLY
@@ -533,8 +556,8 @@ SAROW(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, 1, 2, 15)
 #ifdef HAS_SCALEADDROW_MSA
 SAROW(ScaleAddRow_Any_MSA, ScaleAddRow_MSA, 1, 2, 15)
 #endif
-#ifdef HAS_SCALEADDROW_MMI
-SAROW(ScaleAddRow_Any_MMI, ScaleAddRow_MMI, 1, 2, 7)
+#ifdef HAS_SCALEADDROW_LSX
+SAROW(ScaleAddRow_Any_LSX, ScaleAddRow_LSX, 1, 2, 15)
 #endif
 #undef SAANY
 
@@ -562,13 +585,477 @@ SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15)
 #ifdef HAS_SCALEADDROW_MSA
 SAANY(ScaleAddRow_Any_MSA, ScaleAddRow_MSA, ScaleAddRow_C, 15)
 #endif
-#ifdef HAS_SCALEADDROW_MMI
-SAANY(ScaleAddRow_Any_MMI, ScaleAddRow_MMI, ScaleAddRow_C, 7)
+#ifdef HAS_SCALEADDROW_LSX
+SAANY(ScaleAddRow_Any_LSX, ScaleAddRow_LSX, ScaleAddRow_C, 15)
 #endif
 #undef SAANY
 
 #endif  // SASIMDONLY
 
+// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
+#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
+  void NAMEANY(uint8_t* dst_ptr, const uint8_t* src_ptr, int dst_width, int x, \
+               int dx) {                                                       \
+    int r = dst_width & MASK;                                                  \
+    int n = dst_width & ~MASK;                                                 \
+    if (n > 0) {                                                               \
+      TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                   \
+    }                                                                          \
+    TERP_C(dst_ptr + n * BPP, src_ptr, r, x + n * dx, dx);                     \
+  }
+
+#ifdef HAS_SCALEFILTERCOLS_NEON
+CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
+#endif
+#ifdef HAS_SCALEFILTERCOLS_MSA
+CANY(ScaleFilterCols_Any_MSA, ScaleFilterCols_MSA, ScaleFilterCols_C, 1, 15)
+#endif
+#ifdef HAS_SCALEFILTERCOLS_LSX
+CANY(ScaleFilterCols_Any_LSX, ScaleFilterCols_LSX, ScaleFilterCols_C, 1, 15)
+#endif
+#ifdef HAS_SCALEARGBCOLS_NEON
+CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
+#endif
+#ifdef HAS_SCALEARGBCOLS_MSA
+CANY(ScaleARGBCols_Any_MSA, ScaleARGBCols_MSA, ScaleARGBCols_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBCOLS_LSX
+CANY(ScaleARGBCols_Any_LSX, ScaleARGBCols_LSX, ScaleARGBCols_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_NEON
+CANY(ScaleARGBFilterCols_Any_NEON,
+     ScaleARGBFilterCols_NEON,
+     ScaleARGBFilterCols_C,
+     4,
+     3)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_MSA
+CANY(ScaleARGBFilterCols_Any_MSA,
+     ScaleARGBFilterCols_MSA,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_LSX
+CANY(ScaleARGBFilterCols_Any_LSX,
+     ScaleARGBFilterCols_LSX,
+     ScaleARGBFilterCols_C,
+     4,
+     7)
+#endif
+#undef CANY
+
+// Scale up horizontally 2 times using linear filter.
+#define SUH2LANY(NAME, SIMD, C, MASK, PTYPE)                       \
+  void NAME(const PTYPE* src_ptr, PTYPE* dst_ptr, int dst_width) { \
+    int work_width = (dst_width - 1) & ~1;                         \
+    int r = work_width & MASK;                                     \
+    int n = work_width & ~MASK;                                    \
+    dst_ptr[0] = src_ptr[0];                                       \
+    if (work_width > 0) {                                          \
+      if (n != 0) {                                                \
+        SIMD(src_ptr, dst_ptr + 1, n);                             \
+      }                                                            \
+      C(src_ptr + (n / 2), dst_ptr + n + 1, r);                    \
+    }                                                              \
+    dst_ptr[dst_width - 1] = src_ptr[(dst_width - 1) / 2];         \
+  }
+
+// Even the C versions need to be wrapped, because boundary pixels have to
+// be handled differently
+
+SUH2LANY(ScaleRowUp2_Linear_Any_C,
+         ScaleRowUp2_Linear_C,
+         ScaleRowUp2_Linear_C,
+         0,
+         uint8_t)
+
+SUH2LANY(ScaleRowUp2_Linear_16_Any_C,
+         ScaleRowUp2_Linear_16_C,
+         ScaleRowUp2_Linear_16_C,
+         0,
+         uint16_t)
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+SUH2LANY(ScaleRowUp2_Linear_Any_SSE2,
+         ScaleRowUp2_Linear_SSE2,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+SUH2LANY(ScaleRowUp2_Linear_Any_SSSE3,
+         ScaleRowUp2_Linear_SSSE3,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+SUH2LANY(ScaleRowUp2_Linear_12_Any_SSSE3,
+         ScaleRowUp2_Linear_12_SSSE3,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+SUH2LANY(ScaleRowUp2_Linear_16_Any_SSE2,
+         ScaleRowUp2_Linear_16_SSE2,
+         ScaleRowUp2_Linear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+SUH2LANY(ScaleRowUp2_Linear_Any_AVX2,
+         ScaleRowUp2_Linear_AVX2,
+         ScaleRowUp2_Linear_C,
+         31,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+SUH2LANY(ScaleRowUp2_Linear_12_Any_AVX2,
+         ScaleRowUp2_Linear_12_AVX2,
+         ScaleRowUp2_Linear_16_C,
+         31,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+SUH2LANY(ScaleRowUp2_Linear_16_Any_AVX2,
+         ScaleRowUp2_Linear_16_AVX2,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_NEON
+SUH2LANY(ScaleRowUp2_Linear_Any_NEON,
+         ScaleRowUp2_Linear_NEON,
+         ScaleRowUp2_Linear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_NEON
+SUH2LANY(ScaleRowUp2_Linear_12_Any_NEON,
+         ScaleRowUp2_Linear_12_NEON,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_NEON
+SUH2LANY(ScaleRowUp2_Linear_16_Any_NEON,
+         ScaleRowUp2_Linear_16_NEON,
+         ScaleRowUp2_Linear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#undef SUH2LANY
+
+// Scale up 2 times using bilinear filter.
+// This function produces 2 rows at a time.
+#define SU2BLANY(NAME, SIMD, C, MASK, PTYPE)                              \
+  void NAME(const PTYPE* src_ptr, ptrdiff_t src_stride, PTYPE* dst_ptr,   \
+            ptrdiff_t dst_stride, int dst_width) {                        \
+    int work_width = (dst_width - 1) & ~1;                                \
+    int r = work_width & MASK;                                            \
+    int n = work_width & ~MASK;                                           \
+    const PTYPE* sa = src_ptr;                                            \
+    const PTYPE* sb = src_ptr + src_stride;                               \
+    PTYPE* da = dst_ptr;                                                  \
+    PTYPE* db = dst_ptr + dst_stride;                                     \
+    da[0] = (3 * sa[0] + sb[0] + 2) >> 2;                                 \
+    db[0] = (sa[0] + 3 * sb[0] + 2) >> 2;                                 \
+    if (work_width > 0) {                                                 \
+      if (n != 0) {                                                       \
+        SIMD(sa, sb - sa, da + 1, db - da, n);                            \
+      }                                                                   \
+      C(sa + (n / 2), sb - sa, da + n + 1, db - da, r);                   \
+    }                                                                     \
+    da[dst_width - 1] =                                                   \
+        (3 * sa[(dst_width - 1) / 2] + sb[(dst_width - 1) / 2] + 2) >> 2; \
+    db[dst_width - 1] =                                                   \
+        (sa[(dst_width - 1) / 2] + 3 * sb[(dst_width - 1) / 2] + 2) >> 2; \
+  }
+
+SU2BLANY(ScaleRowUp2_Bilinear_Any_C,
+         ScaleRowUp2_Bilinear_C,
+         ScaleRowUp2_Bilinear_C,
+         0,
+         uint8_t)
+
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_C,
+         ScaleRowUp2_Bilinear_16_C,
+         ScaleRowUp2_Bilinear_16_C,
+         0,
+         uint16_t)
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+SU2BLANY(ScaleRowUp2_Bilinear_Any_SSE2,
+         ScaleRowUp2_Bilinear_SSE2,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_SSSE3,
+         ScaleRowUp2_Bilinear_12_SSSE3,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_SSE2,
+         ScaleRowUp2_Bilinear_16_SSE2,
+         ScaleRowUp2_Bilinear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+SU2BLANY(ScaleRowUp2_Bilinear_Any_SSSE3,
+         ScaleRowUp2_Bilinear_SSSE3,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_Any_AVX2,
+         ScaleRowUp2_Bilinear_AVX2,
+         ScaleRowUp2_Bilinear_C,
+         31,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_AVX2,
+         ScaleRowUp2_Bilinear_12_AVX2,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_AVX2,
+         ScaleRowUp2_Bilinear_16_AVX2,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_Any_NEON,
+         ScaleRowUp2_Bilinear_NEON,
+         ScaleRowUp2_Bilinear_C,
+         15,
+         uint8_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_12_Any_NEON,
+         ScaleRowUp2_Bilinear_12_NEON,
+         ScaleRowUp2_Bilinear_16_C,
+         15,
+         uint16_t)
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_NEON
+SU2BLANY(ScaleRowUp2_Bilinear_16_Any_NEON,
+         ScaleRowUp2_Bilinear_16_NEON,
+         ScaleRowUp2_Bilinear_16_C,
+         7,
+         uint16_t)
+#endif
+
+#undef SU2BLANY
+
+// Scale bi-planar plane up horizontally 2 times using linear filter.
+#define SBUH2LANY(NAME, SIMD, C, MASK, PTYPE)                         \
+  void NAME(const PTYPE* src_ptr, PTYPE* dst_ptr, int dst_width) {    \
+    int work_width = (dst_width - 1) & ~1;                            \
+    int r = work_width & MASK;                                        \
+    int n = work_width & ~MASK;                                       \
+    dst_ptr[0] = src_ptr[0];                                          \
+    dst_ptr[1] = src_ptr[1];                                          \
+    if (work_width > 0) {                                             \
+      if (n != 0) {                                                   \
+        SIMD(src_ptr, dst_ptr + 2, n);                                \
+      }                                                               \
+      C(src_ptr + n, dst_ptr + 2 * n + 2, r);                         \
+    }                                                                 \
+    dst_ptr[2 * dst_width - 2] = src_ptr[((dst_width + 1) & ~1) - 2]; \
+    dst_ptr[2 * dst_width - 1] = src_ptr[((dst_width + 1) & ~1) - 1]; \
+  }
+
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_C,
+          ScaleUVRowUp2_Linear_C,
+          ScaleUVRowUp2_Linear_C,
+          0,
+          uint8_t)
+
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_C,
+          ScaleUVRowUp2_Linear_16_C,
+          ScaleUVRowUp2_Linear_16_C,
+          0,
+          uint16_t)
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_SSSE3
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_SSSE3,
+          ScaleUVRowUp2_Linear_SSSE3,
+          ScaleUVRowUp2_Linear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_AVX2
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_AVX2,
+          ScaleUVRowUp2_Linear_AVX2,
+          ScaleUVRowUp2_Linear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_SSE41,
+          ScaleUVRowUp2_Linear_16_SSE41,
+          ScaleUVRowUp2_Linear_16_C,
+          3,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_AVX2,
+          ScaleUVRowUp2_Linear_16_AVX2,
+          ScaleUVRowUp2_Linear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_NEON
+SBUH2LANY(ScaleUVRowUp2_Linear_Any_NEON,
+          ScaleUVRowUp2_Linear_NEON,
+          ScaleUVRowUp2_Linear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_NEON
+SBUH2LANY(ScaleUVRowUp2_Linear_16_Any_NEON,
+          ScaleUVRowUp2_Linear_16_NEON,
+          ScaleUVRowUp2_Linear_16_C,
+          15,
+          uint16_t)
+#endif
+
+#undef SBUH2LANY
+
+// Scale bi-planar plane up 2 times using bilinear filter.
+// This function produces 2 rows at a time.
+#define SBU2BLANY(NAME, SIMD, C, MASK, PTYPE)                           \
+  void NAME(const PTYPE* src_ptr, ptrdiff_t src_stride, PTYPE* dst_ptr, \
+            ptrdiff_t dst_stride, int dst_width) {                      \
+    int work_width = (dst_width - 1) & ~1;                              \
+    int r = work_width & MASK;                                          \
+    int n = work_width & ~MASK;                                         \
+    const PTYPE* sa = src_ptr;                                          \
+    const PTYPE* sb = src_ptr + src_stride;                             \
+    PTYPE* da = dst_ptr;                                                \
+    PTYPE* db = dst_ptr + dst_stride;                                   \
+    da[0] = (3 * sa[0] + sb[0] + 2) >> 2;                               \
+    db[0] = (sa[0] + 3 * sb[0] + 2) >> 2;                               \
+    da[1] = (3 * sa[1] + sb[1] + 2) >> 2;                               \
+    db[1] = (sa[1] + 3 * sb[1] + 2) >> 2;                               \
+    if (work_width > 0) {                                               \
+      if (n != 0) {                                                     \
+        SIMD(sa, sb - sa, da + 2, db - da, n);                          \
+      }                                                                 \
+      C(sa + n, sb - sa, da + 2 * n + 2, db - da, r);                   \
+    }                                                                   \
+    da[2 * dst_width - 2] = (3 * sa[((dst_width + 1) & ~1) - 2] +       \
+                             sb[((dst_width + 1) & ~1) - 2] + 2) >>     \
+                            2;                                          \
+    db[2 * dst_width - 2] = (sa[((dst_width + 1) & ~1) - 2] +           \
+                             3 * sb[((dst_width + 1) & ~1) - 2] + 2) >> \
+                            2;                                          \
+    da[2 * dst_width - 1] = (3 * sa[((dst_width + 1) & ~1) - 1] +       \
+                             sb[((dst_width + 1) & ~1) - 1] + 2) >>     \
+                            2;                                          \
+    db[2 * dst_width - 1] = (sa[((dst_width + 1) & ~1) - 1] +           \
+                             3 * sb[((dst_width + 1) & ~1) - 1] + 2) >> \
+                            2;                                          \
+  }
+
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_C,
+          ScaleUVRowUp2_Bilinear_C,
+          ScaleUVRowUp2_Bilinear_C,
+          0,
+          uint8_t)
+
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_C,
+          ScaleUVRowUp2_Bilinear_16_C,
+          ScaleUVRowUp2_Bilinear_16_C,
+          0,
+          uint16_t)
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_SSSE3,
+          ScaleUVRowUp2_Bilinear_SSSE3,
+          ScaleUVRowUp2_Bilinear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_AVX2
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_AVX2,
+          ScaleUVRowUp2_Bilinear_AVX2,
+          ScaleUVRowUp2_Bilinear_C,
+          15,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_SSE41,
+          ScaleUVRowUp2_Bilinear_16_SSE41,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_AVX2,
+          ScaleUVRowUp2_Bilinear_16_AVX2,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_NEON
+SBU2BLANY(ScaleUVRowUp2_Bilinear_Any_NEON,
+          ScaleUVRowUp2_Bilinear_NEON,
+          ScaleUVRowUp2_Bilinear_C,
+          7,
+          uint8_t)
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_NEON
+SBU2BLANY(ScaleUVRowUp2_Bilinear_16_Any_NEON,
+          ScaleUVRowUp2_Bilinear_16_NEON,
+          ScaleUVRowUp2_Bilinear_16_C,
+          7,
+          uint16_t)
+#endif
+
+#undef SBU2BLANY
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/files/source/scale_argb.cc b/files/source/scale_argb.cc
index beef380a..9c3acf7f 100644
--- a/files/source/scale_argb.cc
+++ b/files/source/scale_argb.cc
@@ -58,9 +58,9 @@ static void ScaleARGBDown2(int src_width,
   assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
   // Advance to odd row, even column.
   if (filtering == kFilterBilinear) {
-    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+    src_argb += (y >> 16) * (int64_t)src_stride + (x >> 16) * 4;
   } else {
-    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+    src_argb += (y >> 16) * (int64_t)src_stride + ((x >> 16) - 1) * 4;
   }
 
 #if defined(HAS_SCALEARGBROWDOWN2_SSE2)
@@ -111,19 +111,19 @@ static void ScaleARGBDown2(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBROWDOWN2_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
+#if defined(HAS_SCALEARGBROWDOWN2_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
     ScaleARGBRowDown2 =
         filtering == kFilterNone
-            ? ScaleARGBRowDown2_Any_MMI
-            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_MMI
-                                          : ScaleARGBRowDown2Box_Any_MMI);
-    if (IS_ALIGNED(dst_width, 2)) {
+            ? ScaleARGBRowDown2_Any_LSX
+            : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_LSX
+                                          : ScaleARGBRowDown2Box_Any_LSX);
+    if (IS_ALIGNED(dst_width, 4)) {
       ScaleARGBRowDown2 =
           filtering == kFilterNone
-              ? ScaleARGBRowDown2_MMI
-              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_MMI
-                                            : ScaleARGBRowDown2Box_MMI);
+              ? ScaleARGBRowDown2_LSX
+              : (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_LSX
+                                            : ScaleARGBRowDown2Box_LSX);
     }
   }
 #endif
@@ -162,7 +162,7 @@ static void ScaleARGBDown4Box(int src_width,
                             uint8_t* dst_argb, int dst_width) =
       ScaleARGBRowDown2Box_C;
   // Advance to odd row, even column.
-  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  src_argb += (y >> 16) * (int64_t)src_stride + (x >> 16) * 4;
   (void)src_width;
   (void)src_height;
   (void)dx;
@@ -214,7 +214,7 @@ static void ScaleARGBDownEven(int src_width,
                               enum FilterMode filtering) {
   int j;
   int col_step = dx >> 16;
-  int row_stride = (dy >> 16) * src_stride;
+  int row_stride = (dy >> 16) * (int64_t)src_stride;
   void (*ScaleARGBRowDownEven)(const uint8_t* src_argb, ptrdiff_t src_stride,
                                int src_step, uint8_t* dst_argb, int dst_width) =
       filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
@@ -222,7 +222,7 @@ static void ScaleARGBDownEven(int src_width,
   (void)src_height;
   assert(IS_ALIGNED(src_width, 2));
   assert(IS_ALIGNED(src_height, 2));
-  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  src_argb += (y >> 16) * (int64_t)src_stride + (x >> 16) * 4;
 #if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2
@@ -253,13 +253,13 @@ static void ScaleARGBDownEven(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBROWDOWNEVEN_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_MMI
-                                     : ScaleARGBRowDownEven_Any_MMI;
-    if (IS_ALIGNED(dst_width, 2)) {
+#if defined(HAS_SCALEARGBROWDOWNEVEN_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_LSX
+                                     : ScaleARGBRowDownEven_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
       ScaleARGBRowDownEven =
-          filtering ? ScaleARGBRowDownEvenBox_MMI : ScaleARGBRowDownEven_MMI;
+          filtering ? ScaleARGBRowDownEvenBox_LSX : ScaleARGBRowDownEven_LSX;
     }
   }
 #endif
@@ -340,6 +340,14 @@ static void ScaleARGBBilinearDown(int src_width,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
@@ -361,6 +369,14 @@ static void ScaleARGBBilinearDown(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
   // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
   // Allocate a row of ARGB.
   {
@@ -372,7 +388,7 @@ static void ScaleARGBBilinearDown(int src_width,
     }
     for (j = 0; j < dst_height; ++j) {
       int yi = y >> 16;
-      const uint8_t* src = src_argb + yi * src_stride;
+      const uint8_t* src = src_argb + yi * (int64_t)src_stride;
       if (filtering == kFilterLinear) {
         ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
       } else {
@@ -444,11 +460,11 @@ static void ScaleARGBBilinearUp(int src_width,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(dst_width, 2)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -477,6 +493,14 @@ static void ScaleARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -498,11 +522,11 @@ static void ScaleARGBBilinearUp(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBCOLS_MMI)
-  if (!filtering && TestCpuFlag(kCpuHasMMI)) {
-    ScaleARGBFilterCols = ScaleARGBCols_Any_MMI;
-    if (IS_ALIGNED(dst_width, 1)) {
-      ScaleARGBFilterCols = ScaleARGBCols_MMI;
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (!filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_LSX;
     }
   }
 #endif
@@ -513,11 +537,6 @@ static void ScaleARGBBilinearUp(int src_width,
       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
-#if defined(HAS_SCALEARGBCOLSUP2_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBFilterCols = ScaleARGBColsUp2_MMI;
-    }
-#endif
   }
 
   if (y > max_y) {
@@ -526,7 +545,7 @@ static void ScaleARGBBilinearUp(int src_width,
 
   {
     int yi = y >> 16;
-    const uint8_t* src = src_argb + yi * src_stride;
+    const uint8_t* src = src_argb + yi * (int64_t)src_stride;
 
     // Allocate 2 rows of ARGB.
     const int kRowSize = (dst_width * 4 + 31) & ~31;
@@ -541,7 +560,9 @@ static void ScaleARGBBilinearUp(int src_width,
       src += src_stride;
     }
     ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
-    src += src_stride;
+    if (src_height > 2) {
+      src += src_stride;
+    }
 
     for (j = 0; j < dst_height; ++j) {
       yi = y >> 16;
@@ -549,14 +570,16 @@ static void ScaleARGBBilinearUp(int src_width,
         if (y > max_y) {
           y = max_y;
           yi = y >> 16;
-          src = src_argb + yi * src_stride;
+          src = src_argb + yi * (int64_t)src_stride;
         }
         if (yi != lasty) {
           ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
           rowptr += rowstride;
           rowstride = -rowstride;
           lasty = yi;
-          src += src_stride;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
         }
       }
       if (filtering == kFilterLinear) {
@@ -611,6 +634,15 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_AVX512BW)
+  if (TestCpuFlag(kCpuHasAVX512BW | kCpuHasAVX512VL) ==
+      (kCpuHasAVX512BW | kCpuHasAVX512VL)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX512BW;
+    if (IS_ALIGNED(src_width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_AVX512BW;
+    }
+  }
+#endif
 #if defined(HAS_I422TOARGBROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
     I422ToARGBRow = I422ToARGBRow_Any_NEON;
@@ -627,6 +659,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_I422TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    I422ToARGBRow = I422ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(src_width, 32)) {
+      I422ToARGBRow = I422ToARGBRow_LASX;
+    }
+  }
+#endif
 
   void (*InterpolateRow)(uint8_t * dst_argb, const uint8_t* src_argb,
                          ptrdiff_t src_stride, int dst_width,
@@ -663,6 +703,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
 
   void (*ScaleARGBFilterCols)(uint8_t * dst_argb, const uint8_t* src_argb,
                               int dst_width, int x, int dx) =
@@ -692,6 +740,14 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
+#if defined(HAS_SCALEARGBFILTERCOLS_LSX)
+  if (filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_LSX;
+    }
+  }
+#endif
 #if defined(HAS_SCALEARGBCOLS_SSE2)
   if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
     ScaleARGBFilterCols = ScaleARGBCols_SSE2;
@@ -713,11 +769,11 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBCOLS_MMI)
-  if (!filtering && TestCpuFlag(kCpuHasMMI)) {
-    ScaleARGBFilterCols = ScaleARGBCols_Any_MMI;
-    if (IS_ALIGNED(dst_width, 1)) {
-      ScaleARGBFilterCols = ScaleARGBCols_MMI;
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (!filtering && TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBCols_LSX;
     }
   }
 #endif
@@ -728,11 +784,6 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
       ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
-#if defined(HAS_SCALEARGBCOLSUP2_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBFilterCols = ScaleARGBColsUp2_MMI;
-    }
-#endif
   }
 
   const int max_y = (src_height - 1) << 16;
@@ -742,9 +793,9 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
   const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
   int yi = y >> 16;
   int uv_yi = yi >> kYShift;
-  const uint8_t* src_row_y = src_y + yi * src_stride_y;
-  const uint8_t* src_row_u = src_u + uv_yi * src_stride_u;
-  const uint8_t* src_row_v = src_v + uv_yi * src_stride_v;
+  const uint8_t* src_row_y = src_y + yi * (int64_t)src_stride_y;
+  const uint8_t* src_row_u = src_u + uv_yi * (int64_t)src_stride_u;
+  const uint8_t* src_row_v = src_v + uv_yi * (int64_t)src_stride_v;
 
   // Allocate 2 rows of ARGB.
   const int kRowSize = (dst_width * 4 + 31) & ~31;
@@ -782,9 +833,9 @@ static void ScaleYUVToARGBBilinearUp(int src_width,
         y = max_y;
         yi = y >> 16;
         uv_yi = yi >> kYShift;
-        src_row_y = src_y + yi * src_stride_y;
-        src_row_u = src_u + uv_yi * src_stride_u;
-        src_row_v = src_v + uv_yi * src_stride_v;
+        src_row_y = src_y + yi * (int64_t)src_stride_y;
+        src_row_u = src_u + uv_yi * (int64_t)src_stride_u;
+        src_row_v = src_v + uv_yi * (int64_t)src_stride_v;
       }
       if (yi != lasty) {
         // TODO(fbarchard): Convert the clipped region of row.
@@ -857,11 +908,11 @@ static void ScaleARGBSimple(int src_width,
     }
   }
 #endif
-#if defined(HAS_SCALEARGBCOLS_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    ScaleARGBCols = ScaleARGBCols_Any_MMI;
-    if (IS_ALIGNED(dst_width, 1)) {
-      ScaleARGBCols = ScaleARGBCols_MMI;
+#if defined(HAS_SCALEARGBCOLS_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ScaleARGBCols = ScaleARGBCols_Any_LSX;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBCols = ScaleARGBCols_LSX;
     }
   }
 #endif
@@ -872,16 +923,11 @@ static void ScaleARGBSimple(int src_width,
       ScaleARGBCols = ScaleARGBColsUp2_SSE2;
     }
 #endif
-#if defined(HAS_SCALEARGBCOLSUP2_MMI)
-    if (TestCpuFlag(kCpuHasMMI) && IS_ALIGNED(dst_width, 4)) {
-      ScaleARGBCols = ScaleARGBColsUp2_MMI;
-    }
-#endif
   }
 
   for (j = 0; j < dst_height; ++j) {
-    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride, dst_width, x,
-                  dx);
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * (int64_t)src_stride,
+                  dst_width, x, dx);
     dst_argb += dst_stride;
     y += dy;
   }
@@ -916,7 +962,7 @@ static void ScaleARGB(const uint8_t* src,
   // Negative src_height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;
-    src = src + (src_height - 1) * src_stride;
+    src = src + (src_height - 1) * (int64_t)src_stride;
     src_stride = -src_stride;
   }
   ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
@@ -931,7 +977,7 @@ static void ScaleARGB(const uint8_t* src,
   if (clip_y) {
     int64_t clipf = (int64_t)(clip_y)*dy;
     y += (clipf & 0xffff);
-    src += (clipf >> 16) * src_stride;
+    src += (clipf >> 16) * (int64_t)src_stride;
     dst += clip_y * dst_stride;
   }
 
@@ -965,17 +1011,17 @@ static void ScaleARGB(const uint8_t* src,
         filtering = kFilterNone;
         if (dx == 0x10000 && dy == 0x10000) {
           // Straight copy.
-          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
-                   dst, dst_stride, clip_width, clip_height);
+          ARGBCopy(src + (y >> 16) * (int64_t)src_stride + (x >> 16) * 4,
+                   src_stride, dst, dst_stride, clip_width, clip_height);
           return;
         }
       }
     }
   }
   if (dx == 0x10000 && (x & 0xffff) == 0) {
-    // Arbitrary scale vertically, but unscaled vertically.
+    // Arbitrary scale vertically, but unscaled horizontally.
     ScalePlaneVertical(src_height, clip_width, clip_height, src_stride,
-                       dst_stride, src, dst, x, y, dy, 4, filtering);
+                       dst_stride, src, dst, x, y, dy, /*bpp=*/4, filtering);
     return;
   }
   if (filtering && dy < 65536) {
diff --git a/files/source/scale_common.cc b/files/source/scale_common.cc
index 63690271..b02bdafd 100644
--- a/files/source/scale_common.cc
+++ b/files/source/scale_common.cc
@@ -400,6 +400,95 @@ void ScaleRowDown34_1_Box_16_C(const uint16_t* src_ptr,
   }
 }
 
+// Sample position: (O is src sample position, X is dst sample position)
+//
+//      v dst_ptr at here           v stop at here
+//  X O X   X O X   X O X   X O X   X O X
+//    ^ src_ptr at here
+void ScaleRowUp2_Linear_C(const uint8_t* src_ptr,
+                          uint8_t* dst_ptr,
+                          int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[2 * x + 0] = (src_ptr[x + 0] * 3 + src_ptr[x + 1] * 1 + 2) >> 2;
+    dst_ptr[2 * x + 1] = (src_ptr[x + 0] * 1 + src_ptr[x + 1] * 3 + 2) >> 2;
+  }
+}
+
+// Sample position: (O is src sample position, X is dst sample position)
+//
+//    src_ptr at here
+//  X v X   X   X   X   X   X   X   X   X
+//    O       O       O       O       O
+//  X   X   X   X   X   X   X   X   X   X
+//      ^ dst_ptr at here           ^ stop at here
+//  X   X   X   X   X   X   X   X   X   X
+//    O       O       O       O       O
+//  X   X   X   X   X   X   X   X   X   X
+void ScaleRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8_t* dst_ptr,
+                            ptrdiff_t dst_stride,
+                            int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint8_t* d = dst_ptr;
+  uint8_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[2 * x + 0] =
+        (s[x + 0] * 9 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 1 + 8) >> 4;
+    d[2 * x + 1] =
+        (s[x + 0] * 3 + s[x + 1] * 9 + t[x + 0] * 1 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 0] =
+        (s[x + 0] * 3 + s[x + 1] * 1 + t[x + 0] * 9 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 1] =
+        (s[x + 0] * 1 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 9 + 8) >> 4;
+  }
+}
+
+// Only suitable for at most 14 bit range.
+void ScaleRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                             uint16_t* dst_ptr,
+                             int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[2 * x + 0] = (src_ptr[x + 0] * 3 + src_ptr[x + 1] * 1 + 2) >> 2;
+    dst_ptr[2 * x + 1] = (src_ptr[x + 0] * 1 + src_ptr[x + 1] * 3 + 2) >> 2;
+  }
+}
+
+// Only suitable for at most 12bit range.
+void ScaleRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  uint16_t* d = dst_ptr;
+  uint16_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[2 * x + 0] =
+        (s[x + 0] * 9 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 1 + 8) >> 4;
+    d[2 * x + 1] =
+        (s[x + 0] * 3 + s[x + 1] * 9 + t[x + 0] * 1 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 0] =
+        (s[x + 0] * 3 + s[x + 1] * 1 + t[x + 0] * 9 + t[x + 1] * 3 + 8) >> 4;
+    e[2 * x + 1] =
+        (s[x + 0] * 1 + s[x + 1] * 3 + t[x + 0] * 3 + t[x + 1] * 9 + 8) >> 4;
+  }
+}
+
 // Scales a single row of pixels using point sampling.
 void ScaleCols_C(uint8_t* dst_ptr,
                  const uint8_t* src_ptr,
@@ -677,18 +766,18 @@ void ScaleRowDown38_3_Box_16_C(const uint16_t* src_ptr,
         (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
          src_ptr[stride + 1] + src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
          src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
-            (65536 / 9) >>
+            (65536u / 9u) >>
         16;
     dst_ptr[1] =
         (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
          src_ptr[stride + 4] + src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
          src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
-            (65536 / 9) >>
+            (65536u / 9u) >>
         16;
     dst_ptr[2] =
         (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7] +
          src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
-            (65536 / 6) >>
+            (65536u / 6u) >>
         16;
     src_ptr += 8;
     dst_ptr += 3;
@@ -731,15 +820,15 @@ void ScaleRowDown38_2_Box_16_C(const uint16_t* src_ptr,
   for (i = 0; i < dst_width; i += 3) {
     dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[stride + 0] +
                   src_ptr[stride + 1] + src_ptr[stride + 2]) *
-                     (65536 / 6) >>
+                     (65536u / 6u) >>
                  16;
     dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] + src_ptr[stride + 3] +
                   src_ptr[stride + 4] + src_ptr[stride + 5]) *
-                     (65536 / 6) >>
+                     (65536u / 6u) >>
                  16;
     dst_ptr[2] =
         (src_ptr[6] + src_ptr[7] + src_ptr[stride + 6] + src_ptr[stride + 7]) *
-            (65536 / 4) >>
+            (65536u / 4u) >>
         16;
     src_ptr += 8;
     dst_ptr += 3;
@@ -776,6 +865,8 @@ void ScaleAddRow_16_C(const uint16_t* src_ptr,
   }
 }
 
+// ARGB scale row functions
+
 void ScaleARGBRowDown2_C(const uint8_t* src_argb,
                          ptrdiff_t src_stride,
                          uint8_t* dst_argb,
@@ -1018,6 +1109,351 @@ void ScaleARGBFilterCols64_C(uint8_t* dst_argb,
 #undef BLENDERC
 #undef BLENDER
 
+// UV scale row functions
+// same as ARGB but 2 channels
+
+void ScaleUVRowDown2_C(const uint8_t* src_uv,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst_uv,
+                       int dst_width) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int x;
+  (void)src_stride;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleUVRowDown2Linear_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_uv,
+                             int dst_width) {
+  int x;
+  (void)src_stride;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + 1) >> 1;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + 1) >> 1;
+    src_uv += 4;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowDown2Box_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst_uv,
+                          int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + src_uv[src_stride] +
+                 src_uv[src_stride + 2] + 2) >>
+                2;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + src_uv[src_stride + 1] +
+                 src_uv[src_stride + 3] + 2) >>
+                2;
+    src_uv += 4;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowDownEven_C(const uint8_t* src_uv,
+                          ptrdiff_t src_stride,
+                          int src_stepx,
+                          uint8_t* dst_uv,
+                          int dst_width) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  (void)src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleUVRowDownEvenBox_C(const uint8_t* src_uv,
+                             ptrdiff_t src_stride,
+                             int src_stepx,
+                             uint8_t* dst_uv,
+                             int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_uv[0] = (src_uv[0] + src_uv[2] + src_uv[src_stride] +
+                 src_uv[src_stride + 2] + 2) >>
+                2;
+    dst_uv[1] = (src_uv[1] + src_uv[3] + src_uv[src_stride + 1] +
+                 src_uv[src_stride + 3] + 2) >>
+                2;
+    src_uv += src_stepx * 2;
+    dst_uv += 2;
+  }
+}
+
+void ScaleUVRowUp2_Linear_C(const uint8_t* src_ptr,
+                            uint8_t* dst_ptr,
+                            int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[4 * x + 0] =
+        (src_ptr[2 * x + 0] * 3 + src_ptr[2 * x + 2] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 1] =
+        (src_ptr[2 * x + 1] * 3 + src_ptr[2 * x + 3] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 2] =
+        (src_ptr[2 * x + 0] * 1 + src_ptr[2 * x + 2] * 3 + 2) >> 2;
+    dst_ptr[4 * x + 3] =
+        (src_ptr[2 * x + 1] * 1 + src_ptr[2 * x + 3] * 3 + 2) >> 2;
+  }
+}
+
+void ScaleUVRowUp2_Bilinear_C(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              ptrdiff_t dst_stride,
+                              int dst_width) {
+  const uint8_t* s = src_ptr;
+  const uint8_t* t = src_ptr + src_stride;
+  uint8_t* d = dst_ptr;
+  uint8_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[4 * x + 0] = (s[2 * x + 0] * 9 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 1 + 8) >>
+                   4;
+    d[4 * x + 1] = (s[2 * x + 1] * 9 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 1 + 8) >>
+                   4;
+    d[4 * x + 2] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 9 + t[2 * x + 0] * 1 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    d[4 * x + 3] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 9 + t[2 * x + 1] * 1 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 0] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 1 + t[2 * x + 0] * 9 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    e[4 * x + 1] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 1 + t[2 * x + 1] * 9 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 2] = (s[2 * x + 0] * 1 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 9 + 8) >>
+                   4;
+    e[4 * x + 3] = (s[2 * x + 1] * 1 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 9 + 8) >>
+                   4;
+  }
+}
+
+void ScaleUVRowUp2_Linear_16_C(const uint16_t* src_ptr,
+                               uint16_t* dst_ptr,
+                               int dst_width) {
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    dst_ptr[4 * x + 0] =
+        (src_ptr[2 * x + 0] * 3 + src_ptr[2 * x + 2] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 1] =
+        (src_ptr[2 * x + 1] * 3 + src_ptr[2 * x + 3] * 1 + 2) >> 2;
+    dst_ptr[4 * x + 2] =
+        (src_ptr[2 * x + 0] * 1 + src_ptr[2 * x + 2] * 3 + 2) >> 2;
+    dst_ptr[4 * x + 3] =
+        (src_ptr[2 * x + 1] * 1 + src_ptr[2 * x + 3] * 3 + 2) >> 2;
+  }
+}
+
+void ScaleUVRowUp2_Bilinear_16_C(const uint16_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint16_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint16_t* s = src_ptr;
+  const uint16_t* t = src_ptr + src_stride;
+  uint16_t* d = dst_ptr;
+  uint16_t* e = dst_ptr + dst_stride;
+  int src_width = dst_width >> 1;
+  int x;
+  assert((dst_width % 2 == 0) && (dst_width >= 0));
+  for (x = 0; x < src_width; ++x) {
+    d[4 * x + 0] = (s[2 * x + 0] * 9 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 1 + 8) >>
+                   4;
+    d[4 * x + 1] = (s[2 * x + 1] * 9 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 1 + 8) >>
+                   4;
+    d[4 * x + 2] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 9 + t[2 * x + 0] * 1 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    d[4 * x + 3] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 9 + t[2 * x + 1] * 1 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 0] = (s[2 * x + 0] * 3 + s[2 * x + 2] * 1 + t[2 * x + 0] * 9 +
+                    t[2 * x + 2] * 3 + 8) >>
+                   4;
+    e[4 * x + 1] = (s[2 * x + 1] * 3 + s[2 * x + 3] * 1 + t[2 * x + 1] * 9 +
+                    t[2 * x + 3] * 3 + 8) >>
+                   4;
+    e[4 * x + 2] = (s[2 * x + 0] * 1 + s[2 * x + 2] * 3 + t[2 * x + 0] * 3 +
+                    t[2 * x + 2] * 9 + 8) >>
+                   4;
+    e[4 * x + 3] = (s[2 * x + 1] * 1 + s[2 * x + 3] * 3 + t[2 * x + 1] * 3 +
+                    t[2 * x + 3] * 9 + 8) >>
+                   4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleUVCols_C(uint8_t* dst_uv,
+                   const uint8_t* src_uv,
+                   int dst_width,
+                   int x,
+                   int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleUVCols64_C(uint8_t* dst_uv,
+                     const uint8_t* src_uv,
+                     int dst_width,
+                     int x32,
+                     int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleUVColsUp2_C(uint8_t* dst_uv,
+                      const uint8_t* src_uv,
+                      int dst_width,
+                      int x,
+                      int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  (void)x;
+  (void)dx;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// TODO(fbarchard): Replace 0x7f ^ f with 128-f.  bug=607.
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b)*f) >> 7
+#define BLENDERC(a, b, f, s) \
+  (uint16_t)(BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleUVFilterCols_C(uint8_t* dst_uv,
+                         const uint8_t* src_uv,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleUVFilterCols64_C(uint8_t* dst_uv,
+                           const uint8_t* src_uv,
+                           int dst_width,
+                           int x32,
+                           int dx) {
+  int64_t x = (int64_t)(x32);
+  const uint16_t* src = (const uint16_t*)(src_uv);
+  uint16_t* dst = (uint16_t*)(dst_uv);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64_t xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64_t xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint16_t a = src[xi];
+    uint16_t b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
 // Scale plane vertically with bilinear interpolation.
 void ScalePlaneVertical(int src_height,
                         int dst_width,
@@ -1029,7 +1465,7 @@ void ScalePlaneVertical(int src_height,
                         int x,
                         int y,
                         int dy,
-                        int bpp,
+                        int bpp,  // bytes per pixel. 4 for ARGB.
                         enum FilterMode filtering) {
   // TODO(fbarchard): Allow higher bpp.
   int dst_width_bytes = dst_width * bpp;
@@ -1075,11 +1511,11 @@ void ScalePlaneVertical(int src_height,
     }
   }
 #endif
-#if defined(HAS_INTERPOLATEROW_MMI)
-  if (TestCpuFlag(kCpuHasMMI)) {
-    InterpolateRow = InterpolateRow_Any_MMI;
-    if (IS_ALIGNED(dst_width_bytes, 8)) {
-      InterpolateRow = InterpolateRow_MMI;
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
     }
   }
 #endif
@@ -1097,6 +1533,7 @@ void ScalePlaneVertical(int src_height,
     y += dy;
   }
 }
+
 void ScalePlaneVertical_16(int src_height,
                            int dst_width,
                            int dst_height,
@@ -1107,7 +1544,7 @@ void ScalePlaneVertical_16(int src_height,
                            int x,
                            int y,
                            int dy,
-                           int wpp,
+                           int wpp, /* words per pixel. normally 1 */
                            enum FilterMode filtering) {
   // TODO(fbarchard): Allow higher wpp.
   int dst_width_words = dst_width * wpp;
@@ -1123,32 +1560,32 @@ void ScalePlaneVertical_16(int src_height,
   src_argb += (x >> 16) * wpp;
 #if defined(HAS_INTERPOLATEROW_16_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
-    InterpolateRow = InterpolateRow_Any_16_SSE2;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_SSE2;
+    if (IS_ALIGNED(dst_width_words, 16)) {
       InterpolateRow = InterpolateRow_16_SSE2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
-    InterpolateRow = InterpolateRow_Any_16_SSSE3;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_words, 16)) {
       InterpolateRow = InterpolateRow_16_SSSE3;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
-    InterpolateRow = InterpolateRow_Any_16_AVX2;
-    if (IS_ALIGNED(dst_width_bytes, 32)) {
+    InterpolateRow = InterpolateRow_16_Any_AVX2;
+    if (IS_ALIGNED(dst_width_words, 32)) {
       InterpolateRow = InterpolateRow_16_AVX2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_16_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    InterpolateRow = InterpolateRow_Any_16_NEON;
-    if (IS_ALIGNED(dst_width_bytes, 16)) {
+    InterpolateRow = InterpolateRow_16_Any_NEON;
+    if (IS_ALIGNED(dst_width_words, 8)) {
       InterpolateRow = InterpolateRow_16_NEON;
     }
   }
@@ -1168,6 +1605,70 @@ void ScalePlaneVertical_16(int src_height,
   }
 }
 
+// Use scale to convert lsb formats to msb, depending how many bits there are:
+// 32768 = 9 bits
+// 16384 = 10 bits
+// 4096 = 12 bits
+// 256 = 16 bits
+// TODO(fbarchard): change scale to bits
+void ScalePlaneVertical_16To8(int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint16_t* src_argb,
+                              uint8_t* dst_argb,
+                              int x,
+                              int y,
+                              int dy,
+                              int wpp, /* words per pixel. normally 1 */
+                              int scale,
+                              enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  // TODO(https://crbug.com/libyuv/931): Add NEON 32 bit and AVX2 versions.
+  void (*InterpolateRow_16To8)(uint8_t * dst_argb, const uint16_t* src_argb,
+                               ptrdiff_t src_stride, int scale, int dst_width,
+                               int source_y_fraction) = InterpolateRow_16To8_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+
+#if defined(HAS_INTERPOLATEROW_16TO8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow_16To8 = InterpolateRow_16To8_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow_16To8 = InterpolateRow_16To8_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16TO8_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow_16To8 = InterpolateRow_16To8_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow_16To8 = InterpolateRow_16To8_AVX2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow_16To8(dst_argb, src_argb + yi * src_stride, src_stride,
+                         scale, dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
 // Simplify the filtering based on scale factors.
 enum FilterMode ScaleFilterReduce(int src_width,
                                   int src_height,
@@ -1181,8 +1682,8 @@ enum FilterMode ScaleFilterReduce(int src_width,
     src_height = -src_height;
   }
   if (filtering == kFilterBox) {
-    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
-    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+    // If scaling either axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width || dst_height * 2 >= src_height) {
       filtering = kFilterBilinear;
     }
   }
@@ -1217,7 +1718,7 @@ int FixedDiv_C(int num, int div) {
   return (int)(((int64_t)(num) << 16) / div);
 }
 
-// Divide num by div and return as 16.16 fixed point result.
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
 int FixedDiv1_C(int num, int div) {
   return (int)((((int64_t)(num) << 16) - 0x00010001) / (div - 1));
 }
@@ -1260,14 +1761,14 @@ void ScaleSlope(int src_width,
     if (dst_width <= Abs(src_width)) {
       *dx = FixedDiv(Abs(src_width), dst_width);
       *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_width > 1) {
+    } else if (src_width > 1 && dst_width > 1) {
       *dx = FixedDiv1(Abs(src_width), dst_width);
       *x = 0;
     }
     if (dst_height <= src_height) {
       *dy = FixedDiv(src_height, dst_height);
       *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_height > 1) {
+    } else if (src_height > 1 && dst_height > 1) {
       *dy = FixedDiv1(src_height, dst_height);
       *y = 0;
     }
@@ -1276,7 +1777,7 @@ void ScaleSlope(int src_width,
     if (dst_width <= Abs(src_width)) {
       *dx = FixedDiv(Abs(src_width), dst_width);
       *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
-    } else if (dst_width > 1) {
+    } else if (src_width > 1 && dst_width > 1) {
       *dx = FixedDiv1(Abs(src_width), dst_width);
       *x = 0;
     }
diff --git a/files/source/scale_dspr2.cc b/files/source/scale_dspr2.cc
deleted file mode 100644
index ddedcbf4..00000000
--- a/files/source/scale_dspr2.cc
+++ /dev/null
@@ -1,668 +0,0 @@
-/*
- *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
- *
- *  Use of this source code is governed by a BSD-style license
- *  that can be found in the LICENSE file in the root of the source
- *  tree. An additional intellectual property rights grant can be found
- *  in the file PATENTS. All contributing project authors may
- *  be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "libyuv/basic_types.h"
-#include "libyuv/row.h"
-
-#ifdef __cplusplus
-namespace libyuv {
-extern "C" {
-#endif
-
-// This module is for GCC MIPS DSPR2
-#if !defined(LIBYUV_DISABLE_DSPR2) && defined(__mips_dsp) && \
-    (__mips_dsp_rev >= 2) && (_MIPS_SIM == _MIPS_SIM_ABI32)
-
-void ScaleRowDown2_DSPR2(const uint8* src_ptr,
-                         ptrdiff_t src_stride,
-                         uint8* dst,
-                         int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                     \n"
-      ".set noreorder                                \n"
-
-      "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
-      "beqz           $t9, 2f                        \n"
-      " nop                                          \n"
-
-      "1:                                            \n"
-      "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
-      "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
-      "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
-      "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
-      "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
-      "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
-      "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
-      "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
-      // TODO(fbarchard): Use odd pixels instead of even.
-      "precrq.qb.ph   $t8, $t1, $t0                  \n"  // |7|5|3|1|
-      "precrq.qb.ph   $t0, $t3, $t2                  \n"  // |15|13|11|9|
-      "precrq.qb.ph   $t1, $t5, $t4                  \n"  // |23|21|19|17|
-      "precrq.qb.ph   $t2, $t7, $t6                  \n"  // |31|29|27|25|
-      "addiu          %[src_ptr], %[src_ptr], 32     \n"
-      "addiu          $t9, $t9, -1                   \n"
-      "sw             $t8, 0(%[dst])                 \n"
-      "sw             $t0, 4(%[dst])                 \n"
-      "sw             $t1, 8(%[dst])                 \n"
-      "sw             $t2, 12(%[dst])                \n"
-      "bgtz           $t9, 1b                        \n"
-      " addiu         %[dst], %[dst], 16             \n"
-
-      "2:                                            \n"
-      "andi           $t9, %[dst_width], 0xf         \n"  // residue
-      "beqz           $t9, 3f                        \n"
-      " nop                                          \n"
-
-      "21:                                           \n"
-      "lbu            $t0, 1(%[src_ptr])             \n"
-      "addiu          %[src_ptr], %[src_ptr], 2      \n"
-      "addiu          $t9, $t9, -1                   \n"
-      "sb             $t0, 0(%[dst])                 \n"
-      "bgtz           $t9, 21b                       \n"
-      " addiu         %[dst], %[dst], 1              \n"
-
-      "3:                                            \n"
-      ".set pop                                      \n"
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst)
-      : [dst_width] "r"(dst_width)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void ScaleRowDown2Box_DSPR2(const uint8* src_ptr,
-                            ptrdiff_t src_stride,
-                            uint8* dst,
-                            int dst_width) {
-  const uint8* t = src_ptr + src_stride;
-
-  __asm__ __volatile__(
-      ".set push                                    \n"
-      ".set noreorder                               \n"
-
-      "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
-      "bltz           $t9, 2f                       \n"
-      " nop                                         \n"
-
-      "1:                                           \n"
-      "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
-      "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
-      "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
-      "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
-      "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
-      "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
-      "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
-      "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
-      "addiu          $t9, $t9, -1                  \n"
-      "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
-      "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
-      "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
-      "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
-      "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
-      "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
-      "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
-      "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
-      "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
-      "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
-      "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
-      "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
-      "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
-      "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
-      "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
-      "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
-      "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
-      "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
-      "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
-      "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
-      "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
-      "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
-      "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
-      "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
-      "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
-      "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
-      "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
-      "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
-      "addiu          %[src_ptr], %[src_ptr], 16    \n"
-      "addiu          %[t], %[t], 16                \n"
-      "sb             $t0, 0(%[dst])                \n"
-      "sb             $t4, 1(%[dst])                \n"
-      "sb             $t1, 2(%[dst])                \n"
-      "sb             $t5, 3(%[dst])                \n"
-      "sb             $t2, 4(%[dst])                \n"
-      "sb             $t6, 5(%[dst])                \n"
-      "sb             $t3, 6(%[dst])                \n"
-      "sb             $t7, 7(%[dst])                \n"
-      "bgtz           $t9, 1b                       \n"
-      " addiu         %[dst], %[dst], 8             \n"
-
-      "2:                                           \n"
-      "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
-      "beqz           $t9, 3f                       \n"
-      " nop                                         \n"
-
-      "21:                                          \n"
-      "lwr            $t1, 0(%[src_ptr])            \n"
-      "lwl            $t1, 3(%[src_ptr])            \n"
-      "lwr            $t2, 0(%[t])                  \n"
-      "lwl            $t2, 3(%[t])                  \n"
-      "srl            $t8, $t1, 16                  \n"
-      "ins            $t1, $t2, 16, 16              \n"
-      "ins            $t2, $t8, 0, 16               \n"
-      "raddu.w.qb     $t1, $t1                      \n"
-      "raddu.w.qb     $t2, $t2                      \n"
-      "shra_r.w       $t1, $t1, 2                   \n"
-      "shra_r.w       $t2, $t2, 2                   \n"
-      "sb             $t1, 0(%[dst])                \n"
-      "sb             $t2, 1(%[dst])                \n"
-      "addiu          %[src_ptr], %[src_ptr], 4     \n"
-      "addiu          $t9, $t9, -2                  \n"
-      "addiu          %[t], %[t], 4                 \n"
-      "bgtz           $t9, 21b                      \n"
-      " addiu         %[dst], %[dst], 2             \n"
-
-      "3:                                           \n"
-      ".set pop                                     \n"
-
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [t] "+r"(t)
-      : [dst_width] "r"(dst_width)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void ScaleRowDown4_DSPR2(const uint8* src_ptr,
-                         ptrdiff_t src_stride,
-                         uint8* dst,
-                         int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                    \n"
-      ".set noreorder                               \n"
-
-      "srl            $t9, %[dst_width], 3          \n"
-      "beqz           $t9, 2f                       \n"
-      " nop                                         \n"
-
-      "1:                                           \n"
-      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
-      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
-      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
-      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
-      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
-      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
-      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
-      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
-      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
-      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
-      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
-      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
-      "precrq.qb.ph   $t1, $t2, $t1                 \n"  // |14|10|6|2|
-      "precrq.qb.ph   $t5, $t6, $t5                 \n"  // |30|26|22|18|
-      "addiu          %[src_ptr], %[src_ptr], 32    \n"
-      "addiu          $t9, $t9, -1                  \n"
-      "sw             $t1, 0(%[dst])                \n"
-      "sw             $t5, 4(%[dst])                \n"
-      "bgtz           $t9, 1b                       \n"
-      " addiu         %[dst], %[dst], 8             \n"
-
-      "2:                                           \n"
-      "andi           $t9, %[dst_width], 7          \n"  // residue
-      "beqz           $t9, 3f                       \n"
-      " nop                                         \n"
-
-      "21:                                          \n"
-      "lbu            $t1, 2(%[src_ptr])            \n"
-      "addiu          %[src_ptr], %[src_ptr], 4     \n"
-      "addiu          $t9, $t9, -1                  \n"
-      "sb             $t1, 0(%[dst])                \n"
-      "bgtz           $t9, 21b                      \n"
-      " addiu         %[dst], %[dst], 1             \n"
-
-      "3:                                           \n"
-      ".set pop                                     \n"
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst)
-      : [dst_width] "r"(dst_width)
-      : "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void ScaleRowDown4Box_DSPR2(const uint8* src_ptr,
-                            ptrdiff_t src_stride,
-                            uint8* dst,
-                            int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* s1 = src_ptr + stride;
-  const uint8* s2 = s1 + stride;
-  const uint8* s3 = s2 + stride;
-
-  __asm__ __volatile__(
-      ".set push                                  \n"
-      ".set noreorder                             \n"
-
-      "srl           $t9, %[dst_width], 1         \n"
-      "andi          $t8, %[dst_width], 1         \n"
-
-      "1:                                         \n"
-      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
-      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
-      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
-      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
-      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
-      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
-      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
-      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
-      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
-      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
-      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
-      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
-      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
-      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
-      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
-      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
-      "add           $t0, $t0, $t1                \n"
-      "add           $t1, $t2, $t3                \n"
-      "add           $t0, $t0, $t1                \n"
-      "add           $t4, $t4, $t5                \n"
-      "add           $t6, $t6, $t7                \n"
-      "add           $t4, $t4, $t6                \n"
-      "shra_r.w      $t0, $t0, 4                  \n"
-      "shra_r.w      $t4, $t4, 4                  \n"
-      "sb            $t0, 0(%[dst])               \n"
-      "sb            $t4, 1(%[dst])               \n"
-      "addiu         %[src_ptr], %[src_ptr], 8    \n"
-      "addiu         %[s1], %[s1], 8              \n"
-      "addiu         %[s2], %[s2], 8              \n"
-      "addiu         %[s3], %[s3], 8              \n"
-      "addiu         $t9, $t9, -1                 \n"
-      "bgtz          $t9, 1b                      \n"
-      " addiu        %[dst], %[dst], 2            \n"
-      "beqz          $t8, 2f                      \n"
-      " nop                                       \n"
-
-      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
-      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
-      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
-      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
-      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
-      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
-      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
-      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
-      "add           $t0, $t0, $t1                \n"
-      "add           $t1, $t2, $t3                \n"
-      "add           $t0, $t0, $t1                \n"
-      "shra_r.w      $t0, $t0, 4                  \n"
-      "sb            $t0, 0(%[dst])               \n"
-
-      "2:                                         \n"
-      ".set pop                                   \n"
-
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [s1] "+r"(s1), [s2] "+r"(s2),
-        [s3] "+r"(s3)
-      : [dst_width] "r"(dst_width)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void ScaleRowDown34_DSPR2(const uint8* src_ptr,
-                          ptrdiff_t src_stride,
-                          uint8* dst,
-                          int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                          \n"
-      ".set noreorder                                     \n"
-      "1:                                                 \n"
-      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
-      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
-      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
-      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
-      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
-      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
-      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
-      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
-      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
-      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
-      "addiu           %[dst_width], %[dst_width], -24    \n"
-      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
-      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
-      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
-      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
-      "addiu           %[src_ptr], %[src_ptr], 32         \n"
-      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
-      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
-      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
-      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
-      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
-      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
-      "sw              $t1, 0(%[dst])                     \n"
-      "sw              $t0, 4(%[dst])                     \n"
-      "sw              $t3, 8(%[dst])                     \n"
-      "sw              $t5, 12(%[dst])                    \n"
-      "sw              $t9, 16(%[dst])                    \n"
-      "sw              $t7, 20(%[dst])                    \n"
-      "bnez            %[dst_width], 1b                   \n"
-      " addiu          %[dst], %[dst], 24                 \n"
-      ".set pop                                           \n"
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [dst_width] "+r"(dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9");
-}
-
-void ScaleRowDown34_0_Box_DSPR2(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* d,
-                                int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-      "repl.ph           $t3, 3                          \n"  // 0x00030003
-
-      "1:                                                \n"
-      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
-      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
-      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
-      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
-      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
-      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
-      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
-      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
-      "raddu.w.qb        $t0, $t0                        \n"
-      "raddu.w.qb        $t1, $t1                        \n"
-      "shra_r.w          $t0, $t0, 1                     \n"
-      "shra_r.w          $t1, $t1, 1                     \n"
-      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
-      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
-      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
-      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
-      "addu.ph           $t2, $t2, $t4                   \n"
-      "addu.ph           $t6, $t6, $t5                   \n"
-      "sll               $t5, $t0, 1                     \n"
-      "add               $t0, $t5, $t0                   \n"
-      "shra_r.ph         $t2, $t2, 2                     \n"
-      "shra_r.ph         $t6, $t6, 2                     \n"
-      "shll.ph           $t4, $t2, 1                     \n"
-      "addq.ph           $t4, $t4, $t2                   \n"
-      "addu              $t0, $t0, $t1                   \n"
-      "addiu             %[src_ptr], %[src_ptr], 4       \n"
-      "shra_r.w          $t0, $t0, 2                     \n"
-      "addu.ph           $t6, $t6, $t4                   \n"
-      "shra_r.ph         $t6, $t6, 2                     \n"
-      "srl               $t1, $t6, 16                    \n"
-      "addiu             %[dst_width], %[dst_width], -3  \n"
-      "sb                $t1, 0(%[d])                    \n"
-      "sb                $t0, 1(%[d])                    \n"
-      "sb                $t6, 2(%[d])                    \n"
-      "bgtz              %[dst_width], 1b                \n"
-      " addiu            %[d], %[d], 3                   \n"
-      "3:                                                \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r"(src_ptr), [src_stride] "+r"(src_stride), [d] "+r"(d),
-        [dst_width] "+r"(dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
-}
-
-void ScaleRowDown34_1_Box_DSPR2(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* d,
-                                int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                           \n"
-      ".set noreorder                                      \n"
-      "repl.ph           $t2, 3                            \n"  // 0x00030003
-
-      "1:                                                  \n"
-      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
-      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
-      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
-      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
-      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
-      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
-      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
-      "raddu.w.qb        $t0, $t0                          \n"
-      "raddu.w.qb        $t1, $t1                          \n"
-      "shra_r.w          $t0, $t0, 1                       \n"
-      "shra_r.w          $t1, $t1, 1                       \n"
-      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
-      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
-      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
-      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
-      "addu.ph           $t4, $t4, $t3                     \n"
-      "addu.ph           $t6, $t6, $t5                     \n"
-      "shra_r.ph         $t6, $t6, 2                       \n"
-      "shra_r.ph         $t4, $t4, 2                       \n"
-      "addu.ph           $t6, $t6, $t4                     \n"
-      "addiu             %[src_ptr], %[src_ptr], 4         \n"
-      "shra_r.ph         $t6, $t6, 1                       \n"
-      "addu              $t0, $t0, $t1                     \n"
-      "addiu             %[dst_width], %[dst_width], -3    \n"
-      "shra_r.w          $t0, $t0, 1                       \n"
-      "srl               $t1, $t6, 16                      \n"
-      "sb                $t1, 0(%[d])                      \n"
-      "sb                $t0, 1(%[d])                      \n"
-      "sb                $t6, 2(%[d])                      \n"
-      "bgtz              %[dst_width], 1b                  \n"
-      " addiu            %[d], %[d], 3                     \n"
-      "3:                                                  \n"
-      ".set pop                                            \n"
-      : [src_ptr] "+r"(src_ptr), [src_stride] "+r"(src_stride), [d] "+r"(d),
-        [dst_width] "+r"(dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
-}
-
-void ScaleRowDown38_DSPR2(const uint8* src_ptr,
-                          ptrdiff_t src_stride,
-                          uint8* dst,
-                          int dst_width) {
-  __asm__ __volatile__(
-      ".set push                                     \n"
-      ".set noreorder                                \n"
-
-      "1:                                            \n"
-      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
-      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
-      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
-      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
-      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
-      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
-      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
-      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
-      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
-      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
-      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
-      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
-      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
-      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
-      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
-      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
-      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
-      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
-      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
-      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
-      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
-      "addiu      %[src_ptr], %[src_ptr], 32         \n"
-      "addiu      %[dst_width], %[dst_width], -12    \n"
-      "addiu      $t8,%[dst_width], -12              \n"
-      "sw         $t1, 0(%[dst])                     \n"
-      "sw         $t4, 4(%[dst])                     \n"
-      "sw         $t6, 8(%[dst])                     \n"
-      "bgez       $t8, 1b                            \n"
-      " addiu     %[dst], %[dst], 12                 \n"
-      ".set pop                                      \n"
-      : [src_ptr] "+r"(src_ptr), [dst] "+r"(dst), [dst_width] "+r"(dst_width)
-      :
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8");
-}
-
-void ScaleRowDown38_2_Box_DSPR2(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr,
-                                int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* t = src_ptr + stride;
-  const int c = 0x2AAA;
-
-  __asm__ __volatile__(
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-
-      "1:                                                \n"
-      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
-      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
-      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
-      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
-      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
-      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
-      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
-      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
-      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
-      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
-      "srl             $t4, $t4, 2                       \n"  // t4 / 4
-      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
-      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
-      "addu            $t6, $t5, $t6                     \n"
-      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
-      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
-      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
-      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
-      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
-      "addu            $t0, $t0, $t2                     \n"
-      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
-      "addiu           %[src_ptr], %[src_ptr], 8         \n"
-      "addiu           %[t], %[t], 8                     \n"
-      "addiu           %[dst_width], %[dst_width], -3    \n"
-      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
-      "srl             $t6, $t6, 16                      \n"
-      "srl             $t0, $t0, 16                      \n"
-      "sb              $t4, -1(%[dst_ptr])               \n"
-      "sb              $t6, -2(%[dst_ptr])               \n"
-      "bgtz            %[dst_width], 1b                  \n"
-      " sb             $t0, -3(%[dst_ptr])               \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r"(src_ptr), [dst_ptr] "+r"(dst_ptr), [t] "+r"(t),
-        [dst_width] "+r"(dst_width)
-      : [c] "r"(c)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6");
-}
-
-void ScaleRowDown38_3_Box_DSPR2(const uint8* src_ptr,
-                                ptrdiff_t src_stride,
-                                uint8* dst_ptr,
-                                int dst_width) {
-  intptr_t stride = src_stride;
-  const uint8* s1 = src_ptr + stride;
-  stride += stride;
-  const uint8* s2 = src_ptr + stride;
-  const int c1 = 0x1C71;
-  const int c2 = 0x2AAA;
-
-  __asm__ __volatile__(
-      ".set push                                         \n"
-      ".set noreorder                                    \n"
-
-      "1:                                                \n"
-      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
-      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
-      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
-      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
-      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
-      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
-      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
-      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
-      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
-      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
-      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
-      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
-      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
-      "addu            $t7, $t7, $t8                     \n"
-      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
-      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
-      "addu            $t6, $t6, $t8                     \n"
-      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
-      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
-      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
-      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
-      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
-      "addu            $t7, $t7, $t8                     \n"
-      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
-      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
-      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
-      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
-      "raddu.w.qb      $t0, $t0                          \n"
-      "raddu.w.qb      $t2, $t2                          \n"
-      "raddu.w.qb      $t4, $t4                          \n"
-      "addu            $t0, $t0, $t2                     \n"
-      "addu            $t0, $t0, $t4                     \n"
-      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
-      "addiu           %[src_ptr], %[src_ptr], 8         \n"
-      "addiu           %[s1], %[s1], 8                   \n"
-      "addiu           %[s2], %[s2], 8                   \n"
-      "addiu           %[dst_width], %[dst_width], -3    \n"
-      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
-      "srl             $t6, $t6, 16                      \n"
-      "srl             $t7, $t7, 16                      \n"
-      "srl             $t0, $t0, 16                      \n"
-      "sb              $t6, -1(%[dst_ptr])               \n"
-      "sb              $t7, -2(%[dst_ptr])               \n"
-      "bgtz            %[dst_width], 1b                  \n"
-      " sb             $t0, -3(%[dst_ptr])               \n"
-      ".set pop                                          \n"
-      : [src_ptr] "+r"(src_ptr), [dst_ptr] "+r"(dst_ptr), [s1] "+r"(s1),
-        [s2] "+r"(s2), [dst_width] "+r"(dst_width)
-      : [c1] "r"(c1), [c2] "r"(c2)
-      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8");
-}
-
-void ScaleAddRow_DSPR2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
-  int x;
-  for (x = 0; x < ((src_width - 1)); x += 8) {
-    uint32 tmp_t1, tmp_t2, tmp_t3, tmp_t4;
-    uint32 tmp_t5, tmp_t6, tmp_t7, tmp_t8;
-    __asm__ __volatile__(
-        ".set push                                                \n"
-        ".set noreorder                                           \n"
-        "lw                %[tmp_t5],   0(%[src_ptr])             \n"
-        "lw                %[tmp_t6],   4(%[src_ptr])             \n"
-        "lw                %[tmp_t1],   0(%[dst_ptr])             \n"
-        "lw                %[tmp_t2],   4(%[dst_ptr])             \n"
-        "lw                %[tmp_t3],   8(%[dst_ptr])             \n"
-        "lw                %[tmp_t4],   12(%[dst_ptr])            \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t5]                 \n"
-        "preceu.ph.qbl     %[tmp_t8],   %[tmp_t5]                 \n"
-        "addu.ph           %[tmp_t1],   %[tmp_t1],     %[tmp_t7]  \n"
-        "addu.ph           %[tmp_t2],   %[tmp_t2],     %[tmp_t8]  \n"
-        "preceu.ph.qbr     %[tmp_t7],   %[tmp_t6]                 \n"
-        "preceu.ph.qbl     %[tmp_t8],   %[tmp_t6]                 \n"
-        "addu.ph           %[tmp_t3],   %[tmp_t3],     %[tmp_t7]  \n"
-        "addu.ph           %[tmp_t4],   %[tmp_t4],     %[tmp_t8]  \n"
-        "sw                %[tmp_t1],   0(%[dst_ptr])             \n"
-        "sw                %[tmp_t2],   4(%[dst_ptr])             \n"
-        "sw                %[tmp_t3],   8(%[dst_ptr])             \n"
-        "sw                %[tmp_t4],   12(%[dst_ptr])            \n"
-        ".set pop                                                 \n"
-        :
-        [tmp_t1] "=&r"(tmp_t1), [tmp_t2] "=&r"(tmp_t2), [tmp_t3] "=&r"(tmp_t3),
-        [tmp_t4] "=&r"(tmp_t4), [tmp_t5] "=&r"(tmp_t5), [tmp_t6] "=&r"(tmp_t6),
-        [tmp_t7] "=&r"(tmp_t7), [tmp_t8] "=&r"(tmp_t8), [src_ptr] "+r"(src_ptr)
-        : [dst_ptr] "r"(dst_ptr));
-    src_ptr += 8;
-    dst_ptr += 8;
-  }
-
-  if ((src_width)&7) {
-    for (x = 0; x < ((src_width - 1) & 7); x += 1) {
-      dst_ptr[0] += src_ptr[0];
-      src_ptr += 1;
-      dst_ptr += 1;
-    }
-  }
-}
-
-#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
-
-#ifdef __cplusplus
-}  // extern "C"
-}  // namespace libyuv
-#endif
diff --git a/files/source/scale_gcc.cc b/files/source/scale_gcc.cc
index 90a49f30..edaf2e29 100644
--- a/files/source/scale_gcc.cc
+++ b/files/source/scale_gcc.cc
@@ -17,8 +17,7 @@ extern "C" {
 #endif
 
 // This module is for GCC x86 and x64.
-#if !defined(LIBYUV_DISABLE_X86) && \
-    (defined(__x86_64__) || (defined(__i386__) && !defined(_MSC_VER)))
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
 
 // Offsets for source bytes 0 to 9
 static const uvec8 kShuf0 = {0,   1,   3,   4,   5,   7,   8,   9,
@@ -102,16 +101,16 @@ void ScaleRowDown2_SSSE3(const uint8_t* src_ptr,
       // 16 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "psrlw     $0x8,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "psrlw       $0x8,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -125,25 +124,25 @@ void ScaleRowDown2Linear_SSSE3(const uint8_t* src_ptr,
                                int dst_width) {
   (void)src_stride;
   asm volatile(
-      "pcmpeqb    %%xmm4,%%xmm4                  \n"
-      "psrlw      $0xf,%%xmm4                    \n"
-      "packuswb   %%xmm4,%%xmm4                  \n"
-      "pxor       %%xmm5,%%xmm5                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pmaddubsw  %%xmm4,%%xmm0                  \n"
-      "pmaddubsw  %%xmm4,%%xmm1                  \n"
-      "pavgw      %%xmm5,%%xmm0                  \n"
-      "pavgw      %%xmm5,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                  \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -156,33 +155,33 @@ void ScaleRowDown2Box_SSSE3(const uint8_t* src_ptr,
                             uint8_t* dst_ptr,
                             int dst_width) {
   asm volatile(
-      "pcmpeqb    %%xmm4,%%xmm4                  \n"
-      "psrlw      $0xf,%%xmm4                    \n"
-      "packuswb   %%xmm4,%%xmm4                  \n"
-      "pxor       %%xmm5,%%xmm5                  \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x00(%0,%3,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%3,1),%%xmm3            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pmaddubsw  %%xmm4,%%xmm0                  \n"
-      "pmaddubsw  %%xmm4,%%xmm1                  \n"
-      "pmaddubsw  %%xmm4,%%xmm2                  \n"
-      "pmaddubsw  %%xmm4,%%xmm3                  \n"
-      "paddw      %%xmm2,%%xmm0                  \n"
-      "paddw      %%xmm3,%%xmm1                  \n"
-      "psrlw      $0x1,%%xmm0                    \n"
-      "psrlw      $0x1,%%xmm1                    \n"
-      "pavgw      %%xmm5,%%xmm0                  \n"
-      "pavgw      %%xmm5,%%xmm1                  \n"
-      "packuswb   %%xmm1,%%xmm0                  \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "psrlw       $0x1,%%xmm0                   \n"
+      "psrlw       $0x1,%%xmm1                   \n"
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pavgw       %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),               // %0
         "+r"(dst_ptr),               // %1
         "+r"(dst_width)              // %2
@@ -196,27 +195,25 @@ void ScaleRowDown2_AVX2(const uint8_t* src_ptr,
                         uint8_t* dst_ptr,
                         int dst_width) {
   (void)src_stride;
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x8,%%ymm1,%%ymm1            \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
-      : "+r"(src_ptr),   // %0
-        "+r"(dst_ptr),   // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1");
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 void ScaleRowDown2Linear_AVX2(const uint8_t* src_ptr,
@@ -225,26 +222,26 @@ void ScaleRowDown2Linear_AVX2(const uint8_t* src_ptr,
                               int dst_width) {
   (void)src_stride;
   asm volatile(
-      "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-      "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
@@ -258,34 +255,34 @@ void ScaleRowDown2Box_AVX2(const uint8_t* src_ptr,
                            uint8_t* dst_ptr,
                            int dst_width) {
   asm volatile(
-      "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-      "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x00(%0,%3,1),%%ymm2           \n"
-      "vmovdqu    0x20(%0,%3,1),%%ymm3           \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-      "vpsrlw     $0x1,%%ymm0,%%ymm0             \n"
-      "vpsrlw     $0x1,%%ymm1,%%ymm1             \n"
-      "vpavgw     %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpavgw     %%ymm5,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "lea        0x20(%1),%1                    \n"
-      "sub        $0x20,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,1),%%ymm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"
+      "vpsrlw      $0x1,%%ymm1,%%ymm1            \n"
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpavgw      %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_ptr),               // %0
         "+r"(dst_ptr),               // %1
@@ -301,24 +298,24 @@ void ScaleRowDown4_SSSE3(const uint8_t* src_ptr,
                          int dst_width) {
   (void)src_stride;
   asm volatile(
-      "pcmpeqb   %%xmm5,%%xmm5                   \n"
-      "psrld     $0x18,%%xmm5                    \n"
-      "pslld     $0x10,%%xmm5                    \n"
+      "pcmpeqb     %%xmm5,%%xmm5                 \n"
+      "psrld       $0x18,%%xmm5                  \n"
+      "pslld       $0x10,%%xmm5                  \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pand      %%xmm5,%%xmm0                   \n"
-      "pand      %%xmm5,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm0                   \n"
-      "psrlw     $0x8,%%xmm0                     \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pand        %%xmm5,%%xmm0                 \n"
+      "pand        %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x8,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -332,46 +329,46 @@ void ScaleRowDown4Box_SSSE3(const uint8_t* src_ptr,
                             int dst_width) {
   intptr_t stridex3;
   asm volatile(
-      "pcmpeqb    %%xmm4,%%xmm4                  \n"
-      "psrlw      $0xf,%%xmm4                    \n"
-      "movdqa     %%xmm4,%%xmm5                  \n"
-      "packuswb   %%xmm4,%%xmm4                  \n"
-      "psllw      $0x3,%%xmm5                    \n"
-      "lea       0x00(%4,%4,2),%3                \n"
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"
+      "psrlw       $0xf,%%xmm4                   \n"
+      "movdqa      %%xmm4,%%xmm5                 \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "psllw       $0x3,%%xmm5                   \n"
+      "lea         0x00(%4,%4,2),%3              \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x00(%0,%4,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%4,1),%%xmm3            \n"
-      "pmaddubsw  %%xmm4,%%xmm0                  \n"
-      "pmaddubsw  %%xmm4,%%xmm1                  \n"
-      "pmaddubsw  %%xmm4,%%xmm2                  \n"
-      "pmaddubsw  %%xmm4,%%xmm3                  \n"
-      "paddw      %%xmm2,%%xmm0                  \n"
-      "paddw      %%xmm3,%%xmm1                  \n"
-      "movdqu    0x00(%0,%4,2),%%xmm2            \n"
-      "movdqu    0x10(%0,%4,2),%%xmm3            \n"
-      "pmaddubsw  %%xmm4,%%xmm2                  \n"
-      "pmaddubsw  %%xmm4,%%xmm3                  \n"
-      "paddw      %%xmm2,%%xmm0                  \n"
-      "paddw      %%xmm3,%%xmm1                  \n"
-      "movdqu    0x00(%0,%3,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%3,1),%%xmm3            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pmaddubsw  %%xmm4,%%xmm2                  \n"
-      "pmaddubsw  %%xmm4,%%xmm3                  \n"
-      "paddw      %%xmm2,%%xmm0                  \n"
-      "paddw      %%xmm3,%%xmm1                  \n"
-      "phaddw     %%xmm1,%%xmm0                  \n"
-      "paddw      %%xmm5,%%xmm0                  \n"
-      "psrlw      $0x4,%%xmm0                    \n"
-      "packuswb   %%xmm0,%%xmm0                  \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "lea       0x8(%1),%1                      \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%4,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,1),%%xmm3          \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"
+      "pmaddubsw   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "movdqu      0x00(%0,%4,2),%%xmm2          \n"
+      "movdqu      0x10(%0,%4,2),%%xmm3          \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm3                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"
+      "phaddw      %%xmm1,%%xmm0                 \n"
+      "paddw       %%xmm5,%%xmm0                 \n"
+      "psrlw       $0x4,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),               // %0
         "+r"(dst_ptr),               // %1
         "+r"(dst_width),             // %2
@@ -387,26 +384,26 @@ void ScaleRowDown4_AVX2(const uint8_t* src_ptr,
                         int dst_width) {
   (void)src_stride;
   asm volatile(
-      "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
-      "vpsrld     $0x18,%%ymm5,%%ymm5            \n"
-      "vpslld     $0x10,%%ymm5,%%ymm5            \n"
+      "vpcmpeqb    %%ymm5,%%ymm5,%%ymm5          \n"
+      "vpsrld      $0x18,%%ymm5,%%ymm5           \n"
+      "vpslld      $0x10,%%ymm5,%%ymm5           \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpand      %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpand      %%ymm5,%%ymm1,%%ymm1           \n"
-      "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
-      "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%xmm0,(%1)                    \n"
-      "lea        0x10(%1),%1                    \n"
-      "sub        $0x10,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpand       %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpand       %%ymm5,%%ymm1,%%ymm1          \n"
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpsrlw      $0x8,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
@@ -420,46 +417,46 @@ void ScaleRowDown4Box_AVX2(const uint8_t* src_ptr,
                            uint8_t* dst_ptr,
                            int dst_width) {
   asm volatile(
-      "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
-      "vpsrlw     $0xf,%%ymm4,%%ymm4             \n"
-      "vpsllw     $0x3,%%ymm4,%%ymm5             \n"
-      "vpackuswb  %%ymm4,%%ymm4,%%ymm4           \n"
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpsllw      $0x3,%%ymm4,%%ymm5            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm0                    \n"
-      "vmovdqu    0x20(%0),%%ymm1                \n"
-      "vmovdqu    0x00(%0,%3,1),%%ymm2           \n"
-      "vmovdqu    0x20(%0,%3,1),%%ymm3           \n"
-      "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
-      "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-      "vmovdqu    0x00(%0,%3,2),%%ymm2           \n"
-      "vmovdqu    0x20(%0,%3,2),%%ymm3           \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-      "vmovdqu    0x00(%0,%4,1),%%ymm2           \n"
-      "vmovdqu    0x20(%0,%4,1),%%ymm3           \n"
-      "lea        0x40(%0),%0                    \n"
-      "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
-      "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
-      "vpaddw     %%ymm2,%%ymm0,%%ymm0           \n"
-      "vpaddw     %%ymm3,%%ymm1,%%ymm1           \n"
-      "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"
-      "vpsrlw     $0x4,%%ymm0,%%ymm0             \n"
-      "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
-      "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
-      "vmovdqu    %%xmm0,(%1)                    \n"
-      "lea        0x10(%1),%1                    \n"
-      "sub        $0x10,%2                       \n"
-      "jg         1b                             \n"
+      "vmovdqu     (%0),%%ymm0                   \n"
+      "vmovdqu     0x20(%0),%%ymm1               \n"
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,1),%%ymm3          \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm4,%%ymm1,%%ymm1          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x00(%0,%3,2),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%3,2),%%ymm3          \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vmovdqu     0x00(%0,%4,1),%%ymm2          \n"
+      "vmovdqu     0x20(%0,%4,1),%%ymm3          \n"
+      "lea         0x40(%0),%0                   \n"
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm3,%%ymm3          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"
+      "vpaddw      %%ymm3,%%ymm1,%%ymm1          \n"
+      "vphaddw     %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vpaddw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpsrlw      $0x4,%%ymm0,%%ymm0            \n"
+      "vpackuswb   %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_ptr),                   // %0
         "+r"(dst_ptr),                   // %1
@@ -476,37 +473,35 @@ void ScaleRowDown34_SSSE3(const uint8_t* src_ptr,
                           int dst_width) {
   (void)src_stride;
   asm volatile(
-      "movdqa    %0,%%xmm3                       \n"
-      "movdqa    %1,%%xmm4                       \n"
-      "movdqa    %2,%%xmm5                       \n"
+      "movdqa      %0,%%xmm3                     \n"
+      "movdqa      %1,%%xmm4                     \n"
+      "movdqa      %2,%%xmm5                     \n"
       :
       : "m"(kShuf0),  // %0
         "m"(kShuf1),  // %1
         "m"(kShuf2)   // %2
   );
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm2                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "movdqa    %%xmm2,%%xmm1                   \n"
-      "palignr   $0x8,%%xmm0,%%xmm1              \n"
-      "pshufb    %%xmm3,%%xmm0                   \n"
-      "pshufb    %%xmm4,%%xmm1                   \n"
-      "pshufb    %%xmm5,%%xmm2                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movq      %%xmm1,0x8(%1)                  \n"
-      "movq      %%xmm2,0x10(%1)                 \n"
-      "lea       0x18(%1),%1                     \n"
-      "sub       $0x18,%2                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_ptr),   // %0
-        "+r"(dst_ptr),   // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm2               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "palignr     $0x8,%%xmm0,%%xmm1            \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"
+      "pshufb      %%xmm4,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movq        %%xmm1,0x8(%1)                \n"
+      "movq        %%xmm2,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
 
 void ScaleRowDown34_1_Box_SSSE3(const uint8_t* src_ptr,
@@ -514,65 +509,63 @@ void ScaleRowDown34_1_Box_SSSE3(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int dst_width) {
   asm volatile(
-      "movdqa    %0,%%xmm2                       \n"  // kShuf01
-      "movdqa    %1,%%xmm3                       \n"  // kShuf11
-      "movdqa    %2,%%xmm4                       \n"  // kShuf21
+      "movdqa      %0,%%xmm2                     \n"  // kShuf01
+      "movdqa      %1,%%xmm3                     \n"  // kShuf11
+      "movdqa      %2,%%xmm4                     \n"  // kShuf21
       :
       : "m"(kShuf01),  // %0
         "m"(kShuf11),  // %1
         "m"(kShuf21)   // %2
   );
   asm volatile(
-      "movdqa    %0,%%xmm5                       \n"  // kMadd01
-      "movdqa    %1,%%xmm0                       \n"  // kMadd11
-      "movdqa    %2,%%xmm1                       \n"  // kRound34
+      "movdqa      %0,%%xmm5                     \n"  // kMadd01
+      "movdqa      %1,%%xmm0                     \n"  // kMadd11
+      "movdqa      %2,%%xmm1                     \n"  // kRound34
       :
       : "m"(kMadd01),  // %0
         "m"(kMadd11),  // %1
         "m"(kRound34)  // %2
   );
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm6                     \n"
-      "movdqu    0x00(%0,%3,1),%%xmm7            \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm5,%%xmm6                   \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,(%1)                     \n"
-      "movdqu    0x8(%0),%%xmm6                  \n"
-      "movdqu    0x8(%0,%3,1),%%xmm7             \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm3,%%xmm6                   \n"
-      "pmaddubsw %%xmm0,%%xmm6                   \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,0x8(%1)                  \n"
-      "movdqu    0x10(%0),%%xmm6                 \n"
-      "movdqu    0x10(%0,%3,1),%%xmm7            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm4,%%xmm6                   \n"
-      "pmaddubsw %4,%%xmm6                       \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,0x10(%1)                 \n"
-      "lea       0x18(%1),%1                     \n"
-      "sub       $0x18,%2                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_ptr),                // %0
-        "+r"(dst_ptr),                // %1
-        "+r"(dst_width)               // %2
-      : "r"((intptr_t)(src_stride)),  // %3
-        "m"(kMadd21)                  // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
-        "xmm7");
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm7          \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,(%1)                   \n"
+      "movdqu      0x8(%0),%%xmm6                \n"
+      "movdqu      0x8(%0,%3,1),%%xmm7           \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x8(%1)                \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "movdqu      0x10(%0,%3,1),%%xmm7          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %4,%%xmm6                     \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),                // %0
+                 "+r"(dst_ptr),                // %1
+                 "+r"(dst_width)               // %2
+               : "r"((intptr_t)(src_stride)),  // %3
+                 "m"(kMadd21)                  // %4
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
 }
 
 void ScaleRowDown34_0_Box_SSSE3(const uint8_t* src_ptr,
@@ -580,69 +573,67 @@ void ScaleRowDown34_0_Box_SSSE3(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int dst_width) {
   asm volatile(
-      "movdqa    %0,%%xmm2                       \n"  // kShuf01
-      "movdqa    %1,%%xmm3                       \n"  // kShuf11
-      "movdqa    %2,%%xmm4                       \n"  // kShuf21
+      "movdqa      %0,%%xmm2                     \n"  // kShuf01
+      "movdqa      %1,%%xmm3                     \n"  // kShuf11
+      "movdqa      %2,%%xmm4                     \n"  // kShuf21
       :
       : "m"(kShuf01),  // %0
         "m"(kShuf11),  // %1
         "m"(kShuf21)   // %2
   );
   asm volatile(
-      "movdqa    %0,%%xmm5                       \n"  // kMadd01
-      "movdqa    %1,%%xmm0                       \n"  // kMadd11
-      "movdqa    %2,%%xmm1                       \n"  // kRound34
+      "movdqa      %0,%%xmm5                     \n"  // kMadd01
+      "movdqa      %1,%%xmm0                     \n"  // kMadd11
+      "movdqa      %2,%%xmm1                     \n"  // kRound34
       :
       : "m"(kMadd01),  // %0
         "m"(kMadd11),  // %1
         "m"(kRound34)  // %2
   );
 
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm6                     \n"
-      "movdqu    0x00(%0,%3,1),%%xmm7            \n"
-      "pavgb     %%xmm6,%%xmm7                   \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm2,%%xmm6                   \n"
-      "pmaddubsw %%xmm5,%%xmm6                   \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,(%1)                     \n"
-      "movdqu    0x8(%0),%%xmm6                  \n"
-      "movdqu    0x8(%0,%3,1),%%xmm7             \n"
-      "pavgb     %%xmm6,%%xmm7                   \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm3,%%xmm6                   \n"
-      "pmaddubsw %%xmm0,%%xmm6                   \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,0x8(%1)                  \n"
-      "movdqu    0x10(%0),%%xmm6                 \n"
-      "movdqu    0x10(%0,%3,1),%%xmm7            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pavgb     %%xmm6,%%xmm7                   \n"
-      "pavgb     %%xmm7,%%xmm6                   \n"
-      "pshufb    %%xmm4,%%xmm6                   \n"
-      "pmaddubsw %4,%%xmm6                       \n"
-      "paddsw    %%xmm1,%%xmm6                   \n"
-      "psrlw     $0x2,%%xmm6                     \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movq      %%xmm6,0x10(%1)                 \n"
-      "lea       0x18(%1),%1                     \n"
-      "sub       $0x18,%2                        \n"
-      "jg        1b                              \n"
-      : "+r"(src_ptr),                // %0
-        "+r"(dst_ptr),                // %1
-        "+r"(dst_width)               // %2
-      : "r"((intptr_t)(src_stride)),  // %3
-        "m"(kMadd21)                  // %4
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
-        "xmm7");
+      "movdqu      (%0),%%xmm6                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm7          \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "pmaddubsw   %%xmm5,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,(%1)                   \n"
+      "movdqu      0x8(%0),%%xmm6                \n"
+      "movdqu      0x8(%0,%3,1),%%xmm7           \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "pmaddubsw   %%xmm0,%%xmm6                 \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x8(%1)                \n"
+      "movdqu      0x10(%0),%%xmm6               \n"
+      "movdqu      0x10(%0,%3,1),%%xmm7          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm6,%%xmm7                 \n"
+      "pavgb       %%xmm7,%%xmm6                 \n"
+      "pshufb      %%xmm4,%%xmm6                 \n"
+      "pmaddubsw   %4,%%xmm6                     \n"
+      "paddsw      %%xmm1,%%xmm6                 \n"
+      "psrlw       $0x2,%%xmm6                   \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movq        %%xmm6,0x10(%1)               \n"
+      "lea         0x18(%1),%1                   \n"
+      "sub         $0x18,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),                // %0
+                 "+r"(dst_ptr),                // %1
+                 "+r"(dst_width)               // %2
+               : "r"((intptr_t)(src_stride)),  // %3
+                 "m"(kMadd21)                  // %4
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
 }
 
 void ScaleRowDown38_SSSE3(const uint8_t* src_ptr,
@@ -651,23 +642,23 @@ void ScaleRowDown38_SSSE3(const uint8_t* src_ptr,
                           int dst_width) {
   (void)src_stride;
   asm volatile(
-      "movdqa    %3,%%xmm4                       \n"
-      "movdqa    %4,%%xmm5                       \n"
+      "movdqa      %3,%%xmm4                     \n"
+      "movdqa      %4,%%xmm5                     \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "pshufb    %%xmm5,%%xmm1                   \n"
-      "paddusb   %%xmm1,%%xmm0                   \n"
-      "movq      %%xmm0,(%1)                     \n"
-      "movhlps   %%xmm0,%%xmm1                   \n"
-      "movd      %%xmm1,0x8(%1)                  \n"
-      "lea       0xc(%1),%1                      \n"
-      "sub       $0xc,%2                         \n"
-      "jg        1b                              \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "paddusb     %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%1)                   \n"
+      "movhlps     %%xmm0,%%xmm1                 \n"
+      "movd        %%xmm1,0x8(%1)                \n"
+      "lea         0xc(%1),%1                    \n"
+      "sub         $0xc,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -681,44 +672,43 @@ void ScaleRowDown38_2_Box_SSSE3(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int dst_width) {
   asm volatile(
-      "movdqa    %0,%%xmm2                       \n"
-      "movdqa    %1,%%xmm3                       \n"
-      "movdqa    %2,%%xmm4                       \n"
-      "movdqa    %3,%%xmm5                       \n"
+      "movdqa      %0,%%xmm2                     \n"
+      "movdqa      %1,%%xmm3                     \n"
+      "movdqa      %2,%%xmm4                     \n"
+      "movdqa      %3,%%xmm5                     \n"
       :
       : "m"(kShufAb0),  // %0
         "m"(kShufAb1),  // %1
         "m"(kShufAb2),  // %2
         "m"(kScaleAb2)  // %3
   );
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%3,1),%%xmm1            \n"
-      "lea       0x10(%0),%0                     \n"
-      "pavgb     %%xmm1,%%xmm0                   \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "pshufb    %%xmm2,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm6                   \n"
-      "pshufb    %%xmm3,%%xmm6                   \n"
-      "paddusw   %%xmm6,%%xmm1                   \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "paddusw   %%xmm0,%%xmm1                   \n"
-      "pmulhuw   %%xmm5,%%xmm1                   \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movd      %%xmm1,(%1)                     \n"
-      "psrlq     $0x10,%%xmm1                    \n"
-      "movd      %%xmm1,0x2(%1)                  \n"
-      "lea       0x6(%1),%1                      \n"
-      "sub       $0x6,%2                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_ptr),               // %0
-        "+r"(dst_ptr),               // %1
-        "+r"(dst_width)              // %2
-      : "r"((intptr_t)(src_stride))  // %3
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm1          \n"
+      "lea         0x10(%0),%0                   \n"
+      "pavgb       %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm2,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "pshufb      %%xmm3,%%xmm6                 \n"
+      "paddusw     %%xmm6,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "paddusw     %%xmm0,%%xmm1                 \n"
+      "pmulhuw     %%xmm5,%%xmm1                 \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movd        %%xmm1,(%1)                   \n"
+      "psrlq       $0x10,%%xmm1                  \n"
+      "movd        %%xmm1,0x2(%1)                \n"
+      "lea         0x6(%1),%1                    \n"
+      "sub         $0x6,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),               // %0
+                 "+r"(dst_ptr),               // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6");
 }
 
 void ScaleRowDown38_3_Box_SSSE3(const uint8_t* src_ptr,
@@ -726,126 +716,1105 @@ void ScaleRowDown38_3_Box_SSSE3(const uint8_t* src_ptr,
                                 uint8_t* dst_ptr,
                                 int dst_width) {
   asm volatile(
-      "movdqa    %0,%%xmm2                       \n"
-      "movdqa    %1,%%xmm3                       \n"
-      "movdqa    %2,%%xmm4                       \n"
-      "pxor      %%xmm5,%%xmm5                   \n"
+      "movdqa      %0,%%xmm2                     \n"
+      "movdqa      %1,%%xmm3                     \n"
+      "movdqa      %2,%%xmm4                     \n"
+      "pxor        %%xmm5,%%xmm5                 \n"
       :
       : "m"(kShufAc),    // %0
         "m"(kShufAc3),   // %1
         "m"(kScaleAc33)  // %2
   );
+  asm volatile(LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x00(%0,%3,1),%%xmm6          \n"
+      "movhlps     %%xmm0,%%xmm1                 \n"
+      "movhlps     %%xmm6,%%xmm7                 \n"
+      "punpcklbw   %%xmm5,%%xmm0                 \n"
+      "punpcklbw   %%xmm5,%%xmm1                 \n"
+      "punpcklbw   %%xmm5,%%xmm6                 \n"
+      "punpcklbw   %%xmm5,%%xmm7                 \n"
+      "paddusw     %%xmm6,%%xmm0                 \n"
+      "paddusw     %%xmm7,%%xmm1                 \n"
+      "movdqu      0x00(%0,%3,2),%%xmm6          \n"
+      "lea         0x10(%0),%0                   \n"
+      "movhlps     %%xmm6,%%xmm7                 \n"
+      "punpcklbw   %%xmm5,%%xmm6                 \n"
+      "punpcklbw   %%xmm5,%%xmm7                 \n"
+      "paddusw     %%xmm6,%%xmm0                 \n"
+      "paddusw     %%xmm7,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm6                 \n"
+      "psrldq      $0x2,%%xmm0                   \n"
+      "paddusw     %%xmm0,%%xmm6                 \n"
+      "psrldq      $0x2,%%xmm0                   \n"
+      "paddusw     %%xmm0,%%xmm6                 \n"
+      "pshufb      %%xmm2,%%xmm6                 \n"
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "psrldq      $0x2,%%xmm1                   \n"
+      "paddusw     %%xmm1,%%xmm7                 \n"
+      "psrldq      $0x2,%%xmm1                   \n"
+      "paddusw     %%xmm1,%%xmm7                 \n"
+      "pshufb      %%xmm3,%%xmm7                 \n"
+      "paddusw     %%xmm7,%%xmm6                 \n"
+      "pmulhuw     %%xmm4,%%xmm6                 \n"
+      "packuswb    %%xmm6,%%xmm6                 \n"
+      "movd        %%xmm6,(%1)                   \n"
+      "psrlq       $0x10,%%xmm6                  \n"
+      "movd        %%xmm6,0x2(%1)                \n"
+      "lea         0x6(%1),%1                    \n"
+      "sub         $0x6,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),               // %0
+                 "+r"(dst_ptr),               // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5",
+                 "xmm6", "xmm7");
+}
+
+static const uvec8 kLinearShuffleFar = {2,  3,  0, 1, 6,  7,  4,  5,
+                                        10, 11, 8, 9, 14, 15, 12, 13};
+
+static const uvec8 kLinearMadd31 = {3, 1, 1, 3, 3, 1, 1, 3,
+                                    3, 1, 1, 3, 3, 1, 1, 3};
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSE2
+void ScaleRowUp2_Linear_SSE2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
   asm volatile(
+      "pxor        %%xmm0,%%xmm0                 \n"  // 0
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $1,%%xmm6                     \n"  // all 2
 
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x00(%0,%3,1),%%xmm6            \n"
-      "movhlps   %%xmm0,%%xmm1                   \n"
-      "movhlps   %%xmm6,%%xmm7                   \n"
-      "punpcklbw %%xmm5,%%xmm0                   \n"
-      "punpcklbw %%xmm5,%%xmm1                   \n"
-      "punpcklbw %%xmm5,%%xmm6                   \n"
-      "punpcklbw %%xmm5,%%xmm7                   \n"
-      "paddusw   %%xmm6,%%xmm0                   \n"
-      "paddusw   %%xmm7,%%xmm1                   \n"
-      "movdqu    0x00(%0,%3,2),%%xmm6            \n"
-      "lea       0x10(%0),%0                     \n"
-      "movhlps   %%xmm6,%%xmm7                   \n"
-      "punpcklbw %%xmm5,%%xmm6                   \n"
-      "punpcklbw %%xmm5,%%xmm7                   \n"
-      "paddusw   %%xmm6,%%xmm0                   \n"
-      "paddusw   %%xmm7,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm6                   \n"
-      "psrldq    $0x2,%%xmm0                     \n"
-      "paddusw   %%xmm0,%%xmm6                   \n"
-      "psrldq    $0x2,%%xmm0                     \n"
-      "paddusw   %%xmm0,%%xmm6                   \n"
-      "pshufb    %%xmm2,%%xmm6                   \n"
-      "movdqa    %%xmm1,%%xmm7                   \n"
-      "psrldq    $0x2,%%xmm1                     \n"
-      "paddusw   %%xmm1,%%xmm7                   \n"
-      "psrldq    $0x2,%%xmm1                     \n"
-      "paddusw   %%xmm1,%%xmm7                   \n"
-      "pshufb    %%xmm3,%%xmm7                   \n"
-      "paddusw   %%xmm7,%%xmm6                   \n"
-      "pmulhuw   %%xmm4,%%xmm6                   \n"
-      "packuswb  %%xmm6,%%xmm6                   \n"
-      "movd      %%xmm6,(%1)                     \n"
-      "psrlq     $0x10,%%xmm6                    \n"
-      "movd      %%xmm6,0x2(%1)                  \n"
-      "lea       0x6(%1),%1                      \n"
-      "sub       $0x6,%2                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_ptr),               // %0
-        "+r"(dst_ptr),               // %1
-        "+r"(dst_width)              // %2
-      : "r"((intptr_t)(src_stride))  // %3
+      "movq        (%0),%%xmm1                   \n"  // 01234567
+      "movq        1(%0),%%xmm2                  \n"  // 12345678
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm1,%%xmm1                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm4                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 11223344 (16)
+      "paddw       %%xmm5,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "paddw       %%xmm6,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 01122334 (16)
+      "paddw       %%xmm5,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm5                 \n"  // 3*near+far+2 (lo)
+      "psrlw       $2,%%xmm5                     \n"  // 3/4*near+1/4*far (lo)
+
+      "punpckhbw   %%xmm0,%%xmm1                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm2,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm6,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm3                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSE2
+void ScaleRowUp2_Bilinear_SSE2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  asm volatile(
+      LABELALIGN
+      "1:                                        \n"
+      "pxor        %%xmm0,%%xmm0                 \n"  // 0
+      // above line
+      "movq        (%0),%%xmm1                   \n"  // 01234567
+      "movq        1(%0),%%xmm2                  \n"  // 12345678
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm1,%%xmm1                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "punpcklbw   %%xmm0,%%xmm4                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 11223344 (16)
+      "paddw       %%xmm5,%%xmm4                 \n"  // near+far
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 01122334 (16)
+      "paddw       %%xmm5,%%xmm5                 \n"  // 2*near
+      "paddw       %%xmm5,%%xmm4                 \n"  // 3*near+far (1, lo)
+
+      "punpckhbw   %%xmm0,%%xmm1                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm2,%%xmm1                 \n"
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      // below line
+      "movq        (%0,%3),%%xmm6                \n"  // 01234567
+      "movq        1(%0,%3),%%xmm2               \n"  // 12345678
+      "movdqa      %%xmm6,%%xmm3                 \n"
+      "punpcklbw   %%xmm2,%%xmm3                 \n"  // 0112233445566778
+      "punpcklbw   %%xmm6,%%xmm6                 \n"  // 0011223344556677
+      "punpcklbw   %%xmm2,%%xmm2                 \n"  // 1122334455667788
+
+      "movdqa      %%xmm6,%%xmm5                 \n"
+      "punpcklbw   %%xmm0,%%xmm5                 \n"  // 00112233 (16)
+      "movdqa      %%xmm2,%%xmm7                 \n"
+      "punpcklbw   %%xmm0,%%xmm7                 \n"  // 11223344 (16)
+      "paddw       %%xmm7,%%xmm5                 \n"  // near+far
+      "movdqa      %%xmm3,%%xmm7                 \n"
+      "punpcklbw   %%xmm0,%%xmm7                 \n"  // 01122334 (16)
+      "paddw       %%xmm7,%%xmm7                 \n"  // 2*near
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far (2, lo)
+
+      "punpckhbw   %%xmm0,%%xmm6                 \n"  // 44556677 (16)
+      "punpckhbw   %%xmm0,%%xmm2                 \n"  // 55667788 (16)
+      "paddw       %%xmm6,%%xmm2                 \n"  // near+far
+      "punpckhbw   %%xmm0,%%xmm3                 \n"  // 45566778 (16)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far (2, hi)
+
+      // xmm4 xmm1
+      // xmm5 xmm2
+      "pcmpeqw     %%xmm0,%%xmm0                 \n"
+      "psrlw       $15,%%xmm0                    \n"
+      "psllw       $3,%%xmm0                     \n"  // all 8
+
+      "movdqa      %%xmm4,%%xmm3                 \n"
+      "movdqa      %%xmm5,%%xmm6                 \n"
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm0,%%xmm6                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm4,%%xmm3                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm6,%%xmm3                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm3                     \n"  // ^ div by 16
+
+      "movdqa      %%xmm1,%%xmm7                 \n"
+      "movdqa      %%xmm2,%%xmm6                 \n"
+      "paddw       %%xmm7,%%xmm7                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm0,%%xmm6                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm1,%%xmm7                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm6,%%xmm7                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm7                     \n"  // ^ div by 16
+
+      "packuswb    %%xmm7,%%xmm3                 \n"
+      "movdqu      %%xmm3,(%1)                   \n"  // save above line
+
+      "movdqa      %%xmm5,%%xmm3                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm3,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm4,%%xmm5                 \n"  // 9 3 3 1 + 8 (lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16
+
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "paddw       %%xmm0,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm3                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (hi)
+      "psrlw       $4,%%xmm2                     \n"  // ^ div by 16
+
+      "packuswb    %%xmm2,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // save below line
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
       : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
         "xmm7");
 }
+#endif
 
-// Reads 16xN bytes and produces 16 shorts at a time.
-void ScaleAddRow_SSE2(const uint8_t* src_ptr,
-                      uint16_t* dst_ptr,
-                      int src_width) {
+#ifdef HAS_SCALEROWUP2_LINEAR_12_SSSE3
+void ScaleRowUp2_Linear_12_SSSE3(const uint16_t* src_ptr,
+                                 uint16_t* dst_ptr,
+                                 int dst_width) {
+  asm volatile(
+      "movdqa      %3,%%xmm5                     \n"
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // 01234567 (16)
+      "movdqu      2(%0),%%xmm1                  \n"  // 12345678 (16)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm2                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm1,%%xmm0                 \n"  // 01122334 (16)
+
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm5,%%xmm3                 \n"  // 54657687 (far)
+      "pshufb      %%xmm5,%%xmm1                 \n"  // 10213243 (far)
+
+      "paddw       %%xmm4,%%xmm1                 \n"  // far+2
+      "paddw       %%xmm4,%%xmm3                 \n"  // far+2
+      "paddw       %%xmm0,%%xmm1                 \n"  // near+far+2
+      "paddw       %%xmm2,%%xmm3                 \n"  // near+far+2
+      "paddw       %%xmm0,%%xmm0                 \n"  // 2*near
+      "paddw       %%xmm2,%%xmm2                 \n"  // 2*near
+      "paddw       %%xmm1,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far+2 (hi)
+
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm2,16(%1)                 \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),          // %0
+        "+r"(dst_ptr),          // %1
+        "+r"(dst_width)         // %2
+      : "m"(kLinearShuffleFar)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_SSSE3
+void ScaleRowUp2_Bilinear_12_SSSE3(const uint16_t* src_ptr,
+                                   ptrdiff_t src_stride,
+                                   uint16_t* dst_ptr,
+                                   ptrdiff_t dst_stride,
+                                   int dst_width) {
   asm volatile(
+      "pcmpeqw     %%xmm7,%%xmm7                 \n"
+      "psrlw       $15,%%xmm7                    \n"
+      "psllw       $3,%%xmm7                     \n"  // all 8
+      "movdqa      %5,%%xmm6                     \n"
 
-      "pxor      %%xmm5,%%xmm5                   \n"
+      LABELALIGN
+      "1:                                        \n"
+      // above line
+      "movdqu      (%0),%%xmm0                   \n"  // 01234567 (16)
+      "movdqu      2(%0),%%xmm1                  \n"  // 12345678 (16)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhwd   %%xmm1,%%xmm2                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm1,%%xmm0                 \n"  // 01122334 (16)
+      "movdqa      %%xmm2,%%xmm3                 \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "pshufb      %%xmm6,%%xmm3                 \n"  // 54657687 (far)
+      "pshufb      %%xmm6,%%xmm1                 \n"  // 10213243 (far)
+      "paddw       %%xmm0,%%xmm1                 \n"  // near+far
+      "paddw       %%xmm2,%%xmm3                 \n"  // near+far
+      "paddw       %%xmm0,%%xmm0                 \n"  // 2*near
+      "paddw       %%xmm2,%%xmm2                 \n"  // 2*near
+      "paddw       %%xmm1,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddw       %%xmm3,%%xmm2                 \n"  // 3*near+far (1, hi)
+
+      // below line
+      "movdqu      (%0,%3,2),%%xmm1              \n"  // 01234567 (16)
+      "movdqu      2(%0,%3,2),%%xmm4             \n"  // 12345678 (16)
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhwd   %%xmm4,%%xmm3                 \n"  // 45566778 (16)
+      "punpcklwd   %%xmm4,%%xmm1                 \n"  // 01122334 (16)
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "pshufb      %%xmm6,%%xmm5                 \n"  // 54657687 (far)
+      "pshufb      %%xmm6,%%xmm4                 \n"  // 10213243 (far)
+      "paddw       %%xmm1,%%xmm4                 \n"  // near+far
+      "paddw       %%xmm3,%%xmm5                 \n"  // near+far
+      "paddw       %%xmm1,%%xmm1                 \n"  // 2*near
+      "paddw       %%xmm3,%%xmm3                 \n"  // 2*near
+      "paddw       %%xmm4,%%xmm1                 \n"  // 3*near+far (2, lo)
+      "paddw       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16
+      "movdqu      %%xmm4,(%1)                   \n"
+
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm7,%%xmm5                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm4                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16
+      "movdqu      %%xmm4,0x10(%1)               \n"
+
+      "movdqa      %%xmm1,%%xmm4                 \n"
+      "paddw       %%xmm7,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm4,%%xmm1                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16
+      "movdqu      %%xmm1,(%1,%4,2)              \n"
+
+      "movdqa      %%xmm3,%%xmm4                 \n"
+      "paddw       %%xmm7,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm4,%%xmm4                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm4,%%xmm3                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm3                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm3                     \n"  // ^ div by 16
+      "movdqu      %%xmm3,0x10(%1,%4,2)          \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearShuffleFar)        // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_SSE2
+void ScaleRowUp2_Linear_16_SSE2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqd     %%xmm4,%%xmm4                 \n"
+      "psrld       $31,%%xmm4                    \n"
+      "pslld       $1,%%xmm4                     \n"  // all 2
 
-      // 16 pixel loop.
       LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm3                     \n"
-      "lea       0x10(%0),%0                     \n"  // src_ptr += 16
-      "movdqu    (%1),%%xmm0                     \n"
-      "movdqu    0x10(%1),%%xmm1                 \n"
-      "movdqa    %%xmm3,%%xmm2                   \n"
-      "punpcklbw %%xmm5,%%xmm2                   \n"
-      "punpckhbw %%xmm5,%%xmm3                   \n"
-      "paddusw   %%xmm2,%%xmm0                   \n"
-      "paddusw   %%xmm3,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "movdqu    %%xmm1,0x10(%1)                 \n"
-      "lea       0x20(%1),%1                     \n"
-      "sub       $0x10,%2                        \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"  // 0123 (16b)
+      "movq        2(%0),%%xmm1                  \n"  // 1234 (16b)
+
+      "punpcklwd   %%xmm5,%%xmm0                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm5,%%xmm1                 \n"  // 1234 (32b)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+
+      "pshufd      $0b10110001,%%xmm2,%%xmm2     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm3,%%xmm3     \n"  // 2143 (odd, far)
+
+      "paddd       %%xmm4,%%xmm2                 \n"  // far+2 (lo)
+      "paddd       %%xmm4,%%xmm3                 \n"  // far+2 (hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far+2 (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far+2 (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+
+      "psrld       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrld       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+      "packssdw    %%xmm1,%%xmm0                 \n"
+      "pshufd      $0b11011000,%%xmm0,%%xmm0     \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 pixel to 8 pixel
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
-        "+r"(src_width)  // %2
+        "+r"(dst_width)  // %2
       :
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
 }
+#endif
 
-#ifdef HAS_SCALEADDROW_AVX2
-// Reads 32 bytes and accumulates to 32 shorts at a time.
-void ScaleAddRow_AVX2(const uint8_t* src_ptr,
-                      uint16_t* dst_ptr,
-                      int src_width) {
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_SSE2
+void ScaleRowUp2_Bilinear_16_SSE2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "pcmpeqd     %%xmm6,%%xmm6                 \n"
+      "psrld       $31,%%xmm6                    \n"
+      "pslld       $3,%%xmm6                     \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0011 (near) (32b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1122 (near) (32b, 1u1v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (far) (1, lo)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (far) (1, hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (1, lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (1, hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (1, lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0),%%xmm0                   \n"  // 0123 (16b)
+      "movq        2(%0),%%xmm1                  \n"  // 1234 (16b)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1234 (32b)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b10110001,%%xmm2,%%xmm2     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm3,%%xmm3     \n"  // 2143 (odd, far)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0,%3,2),%%xmm2              \n"
+      "movq        2(%0,%3,2),%%xmm3             \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"  // 0123 (32b)
+      "punpcklwd   %%xmm7,%%xmm3                 \n"  // 1234 (32b)
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufd      $0b10110001,%%xmm4,%%xmm4     \n"  // 1032 (even, far)
+      "pshufd      $0b10110001,%%xmm5,%%xmm5     \n"  // 2143 (odd, far)
+      "paddd       %%xmm2,%%xmm4                 \n"  // near+far (lo)
+      "paddd       %%xmm3,%%xmm5                 \n"  // near+far (hi)
+      "paddd       %%xmm2,%%xmm2                 \n"  // 2*near (lo)
+      "paddd       %%xmm3,%%xmm3                 \n"  // 2*near (hi)
+      "paddd       %%xmm4,%%xmm2                 \n"  // 3*near+far (2, lo)
+      "paddd       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddd       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddd       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddd       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrld       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddd       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddd       %%xmm2,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddd       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrld       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddd       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (2, hi)
+      "paddd       %%xmm1,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrld       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"  // 6*near+2*far (2, hi)
+      "paddd       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddd       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddd       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrld       $4,%%xmm2                     \n"  // ^ div by 16 (2, hi)
+
+      "packssdw    %%xmm0,%%xmm4                 \n"
+      "pshufd      $0b11011000,%%xmm4,%%xmm4     \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packssdw    %%xmm2,%%xmm5                 \n"
+      "pshufd      $0b11011000,%%xmm5,%%xmm5     \n"
+      "movdqu      %%xmm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 pixel to 8 pixel
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_SSSE3
+void ScaleRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+      "movdqa      %3,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 01234567
+      "movq        1(%0),%%xmm1                  \n"  // 12345678
+      "punpcklwd   %%xmm0,%%xmm0                 \n"  // 0101232345456767
+      "punpcklwd   %%xmm1,%%xmm1                 \n"  // 1212343456567878
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 4545565667677878
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 0101121223233434
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"  // 3*near+far (hi)
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"  // 3*near+far (lo)
+      "paddw       %%xmm4,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm4,%%xmm2                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far (hi)
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(dst_ptr),      // %1
+        "+r"(dst_width)     // %2
+      : "m"(kLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_SSSE3
+void ScaleRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8_t* dst_ptr,
+                                ptrdiff_t dst_stride,
+                                int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $3,%%xmm6                     \n"  // all 8
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 01234567
+      "movq        1(%0),%%xmm1                  \n"  // 12345678
+      "punpcklwd   %%xmm0,%%xmm0                 \n"  // 0101232345456767
+      "punpcklwd   %%xmm1,%%xmm1                 \n"  // 1212343456567878
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm1,%%xmm2                 \n"  // 4545565667677878
+      "punpckldq   %%xmm1,%%xmm0                 \n"  // 0101121223233434
+      "pmaddubsw   %%xmm7,%%xmm2                 \n"  // 3*near+far (1, hi)
+      "pmaddubsw   %%xmm7,%%xmm0                 \n"  // 3*near+far (1, lo)
+
+      "movq        (%0,%3),%%xmm1                \n"
+      "movq        1(%0,%3),%%xmm4               \n"
+      "punpcklwd   %%xmm1,%%xmm1                 \n"
+      "punpcklwd   %%xmm4,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhdq   %%xmm4,%%xmm3                 \n"
+      "punpckldq   %%xmm4,%%xmm1                 \n"
+      "pmaddubsw   %%xmm7,%%xmm3                 \n"  // 3*near+far (2, hi)
+      "pmaddubsw   %%xmm7,%%xmm1                 \n"  // 3*near+far (2, lo)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm1,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm1,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm1,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm1                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16 (2, hi)
+
+      "packuswb    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearMadd31)            // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_AVX2
+void ScaleRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+      "vbroadcastf128 %3,%%ymm3                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0),%%xmm1                  \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklwd  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpunpcklwd  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm3,%%ymm2,%%ymm1          \n"  // 3*near+far (hi)
+      "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // 3*near+far+2 (hi)
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrlw      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(dst_ptr),      // %1
+        "+r"(dst_width)     // %2
+      : "m"(kLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_AVX2
+void ScaleRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $15,%%ymm6,%%ymm6             \n"
+      "vpsllw      $3,%%ymm6,%%ymm6              \n"  // all 8
+      "vbroadcastf128 %5,%%ymm7                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0),%%xmm1                  \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklwd  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpunpcklwd  %%ymm1,%%ymm1,%%ymm1          \n"
+      "vpunpckhdq  %%ymm1,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm1          \n"  // 3*near+far (1, hi)
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm0          \n"  // 3*near+far (1, lo)
+
+      "vmovdqu     (%0,%3),%%xmm2                \n"  // 0123456789ABCDEF
+      "vmovdqu     1(%0,%3),%%xmm3               \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm2,%%ymm2     \n"
+      "vpermq      $0b11011000,%%ymm3,%%ymm3     \n"
+      "vpunpcklwd  %%ymm2,%%ymm2,%%ymm2          \n"
+      "vpunpcklwd  %%ymm3,%%ymm3,%%ymm3          \n"
+      "vpunpckhdq  %%ymm3,%%ymm2,%%ymm4          \n"
+      "vpunpckldq  %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm7,%%ymm4,%%ymm3          \n"  // 3*near+far (2, hi)
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm2          \n"  // 3*near+far (2, lo)
+
+      // ymm0 ymm1
+      // ymm2 ymm3
+
+      "vpaddw      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddw      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddw      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrlw      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddw      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddw      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddw      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrlw      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddw      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddw      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddw      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddw      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrlw      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackuswb   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackuswb   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4)                \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearMadd31)            // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_12_AVX2
+void ScaleRowUp2_Linear_12_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "vbroadcastf128 %3,%%ymm5                  \n"
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // 0123456789ABCDEF (16b)
+      "vmovdqu     2(%0),%%ymm1                  \n"  // 123456789ABCDEF0 (16b)
+
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 012389AB4567CDEF
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 12349ABC5678DEF0
+
+      "vpunpckhwd  %%ymm1,%%ymm0,%%ymm2          \n"  // 899AABBCCDDEEFF0 (near)
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm2,%%ymm3          \n"  // 98A9BACBDCEDFE0F (far)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // far+2
+      "vpaddw      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far+2
+      "vpaddw      %%ymm2,%%ymm3,%%ymm3          \n"  // near+far+2
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 3*near+far+2
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 3*near+far+2
+
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far
+      "vpsrlw      $2,%%ymm2,%%ymm2              \n"  // 3/4*near+1/4*far
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm2,32(%1)                 \n"
+
+      "lea         0x20(%0),%0                   \n"
+      "lea         0x40(%1),%1                   \n"  // 16 sample to 32 sample
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),          // %0
+        "+r"(dst_ptr),          // %1
+        "+r"(dst_width)         // %2
+      : "m"(kLinearShuffleFar)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_12_AVX2
+void ScaleRowUp2_Bilinear_12_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
   asm volatile(
+      "vbroadcastf128 %5,%%ymm5                  \n"
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $3,%%ymm4,%%ymm4              \n"  // all 8
 
-      "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b)
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 0123000045670000
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 1234000056780000
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm2          \n"  // 3*near+far (1)
+
+      "vmovdqu     (%0,%3,2),%%xmm0              \n"  // 01234567 (16b)
+      "vmovdqu     2(%0,%3,2),%%xmm1             \n"  // 12345678 (16b)
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"  // 0123000045670000
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"  // 1234000056780000
+      "vpunpcklwd  %%ymm1,%%ymm0,%%ymm0          \n"  // 0112233445566778 (near)
+      "vpshufb     %%ymm5,%%ymm0,%%ymm1          \n"  // 1021324354657687 (far)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm1          \n"  // near+far
+      "vpaddw      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near
+      "vpaddw      %%ymm0,%%ymm1,%%ymm3          \n"  // 3*near+far (2)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm0          \n"  // 6*near+2*far (1)
+      "vpaddw      %%ymm4,%%ymm3,%%ymm1          \n"  // 3*near+far+8 (2)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9*near+3*far (1)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9 3 3 1 + 8 (1)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16
+      "vmovdqu     %%ymm0,(%1)                   \n"  // store above
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm0          \n"  // 6*near+2*far (2)
+      "vpaddw      %%ymm4,%%ymm2,%%ymm1          \n"  // 3*near+far+8 (1)
+      "vpaddw      %%ymm0,%%ymm3,%%ymm0          \n"  // 9*near+3*far (2)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9 3 3 1 + 8 (2)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16
+      "vmovdqu     %%ymm0,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 sample to 16 sample
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kLinearShuffleFar)        // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_LINEAR_16_AVX2
+void ScaleRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $31,%%ymm4,%%ymm4             \n"
+      "vpslld      $1,%%ymm4,%%ymm4              \n"  // all 2
 
       LABELALIGN
       "1:                                        \n"
-      "vmovdqu    (%0),%%ymm3                    \n"
-      "lea        0x20(%0),%0                    \n"  // src_ptr += 32
-      "vpermq     $0xd8,%%ymm3,%%ymm3            \n"
-      "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
-      "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
-      "vpaddusw   (%1),%%ymm2,%%ymm0             \n"
-      "vpaddusw   0x20(%1),%%ymm3,%%ymm1         \n"
-      "vmovdqu    %%ymm0,(%1)                    \n"
-      "vmovdqu    %%ymm1,0x20(%1)                \n"
-      "lea       0x40(%1),%1                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b, 1u1v)
+
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+
+      "vpshufd     $0b10110001,%%ymm0,%%ymm2     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm1,%%ymm3     \n"  // 21436587 (hi, far)
+
+      "vpaddd      %%ymm4,%%ymm2,%%ymm2          \n"  // far+2 (lo)
+      "vpaddd      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far+2 (hi)
+
+      "vpsrld      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrld      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackusdw   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpshufd     $0b11011000,%%ymm0,%%ymm0     \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 pixel to 16 pixel
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
       "vzeroupper                                \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
-        "+r"(src_width)  // %2
+        "+r"(dst_width)  // %2
       :
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEROWUP2_BILINEAR_16_AVX2
+void ScaleRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrld      $31,%%ymm6,%%ymm6             \n"
+      "vpslld      $3,%%ymm6,%%ymm6              \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0),%%xmm1                  \n"  // 12345678 (16b, 1u1v)
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b10110001,%%ymm0,%%ymm2     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm1,%%ymm3     \n"  // 21436587 (hi, far)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far (1, lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far (1, hi)
+
+      "vmovdqu     (%0,%3,2),%%xmm2              \n"  // 01234567 (16b, 1u1v)
+      "vmovdqu     2(%0,%3,2),%%xmm3             \n"  // 12345678 (16b, 1u1v)
+      "vpmovzxwd   %%xmm2,%%ymm2                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm3,%%ymm3                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b10110001,%%ymm2,%%ymm4     \n"  // 10325476 (lo, far)
+      "vpshufd     $0b10110001,%%ymm3,%%ymm5     \n"  // 21436587 (hi, far)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm4          \n"  // near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm5          \n"  // near+far (hi)
+      "vpaddd      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near (lo)
+      "vpaddd      %%ymm3,%%ymm3,%%ymm3          \n"  // 2*near (hi)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm2          \n"  // 3*near+far (2, lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm3          \n"  // 3*near+far (2, hi)
+
+      "vpaddd      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddd      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddd      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddd      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrld      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddd      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddd      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddd      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddd      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrld      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddd      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddd      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddd      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrld      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddd      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddd      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddd      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddd      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrld      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackusdw   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vpshufd     $0b11011000,%%ymm4,%%ymm4     \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackusdw   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vpshufd     $0b11011000,%%ymm5,%%ymm5     \n"
+      "vmovdqu     %%ymm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 pixel to 16 pixel
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+void ScaleAddRow_SSE2(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+      asm volatile("pxor        %%xmm5,%%xmm5                 \n"
+
+               // 16 pixel loop.
+               LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm3                   \n"
+      "lea         0x10(%0),%0                   \n"  // src_ptr += 16
+      "movdqu      (%1),%%xmm0                   \n"
+      "movdqu      0x10(%1),%%xmm1               \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "punpcklbw   %%xmm5,%%xmm2                 \n"
+      "punpckhbw   %%xmm5,%%xmm3                 \n"
+      "paddusw     %%xmm2,%%xmm0                 \n"
+      "paddusw     %%xmm3,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "movdqu      %%xmm1,0x10(%1)               \n"
+      "lea         0x20(%1),%1                   \n"
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(src_width)  // %2
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
+}
+
+#ifdef HAS_SCALEADDROW_AVX2
+// Reads 32 bytes and accumulates to 32 shorts at a time.
+void ScaleAddRow_AVX2(const uint8_t* src_ptr,
+                      uint16_t* dst_ptr,
+                      int src_width) {
+      asm volatile("vpxor       %%ymm5,%%ymm5,%%ymm5          \n"
+
+               LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm3                   \n"
+      "lea         0x20(%0),%0                   \n"  // src_ptr += 32
+      "vpermq      $0xd8,%%ymm3,%%ymm3           \n"
+      "vpunpcklbw  %%ymm5,%%ymm3,%%ymm2          \n"
+      "vpunpckhbw  %%ymm5,%%ymm3,%%ymm3          \n"
+      "vpaddusw    (%1),%%ymm2,%%ymm0            \n"
+      "vpaddusw    0x20(%1),%%ymm3,%%ymm1        \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+      "vmovdqu     %%ymm1,0x20(%1)               \n"
+      "lea         0x40(%1),%1                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+               : "+r"(src_ptr),   // %0
+                 "+r"(dst_ptr),   // %1
+                 "+r"(src_width)  // %2
+               :
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5");
 }
 #endif  // HAS_SCALEADDROW_AVX2
 
@@ -866,69 +1835,69 @@ void ScaleFilterCols_SSSE3(uint8_t* dst_ptr,
                            int dx) {
   intptr_t x0, x1, temp_pixel;
   asm volatile(
-      "movd      %6,%%xmm2                       \n"
-      "movd      %7,%%xmm3                       \n"
-      "movl      $0x04040000,%k2                 \n"
-      "movd      %k2,%%xmm5                      \n"
-      "pcmpeqb   %%xmm6,%%xmm6                   \n"
-      "psrlw     $0x9,%%xmm6                     \n"  // 0x007f007f
-      "pcmpeqb   %%xmm7,%%xmm7                   \n"
-      "psrlw     $15,%%xmm7                      \n"  // 0x00010001
-
-      "pextrw    $0x1,%%xmm2,%k3                 \n"
-      "subl      $0x2,%5                         \n"
-      "jl        29f                             \n"
-      "movdqa    %%xmm2,%%xmm0                   \n"
-      "paddd     %%xmm3,%%xmm0                   \n"
-      "punpckldq %%xmm0,%%xmm2                   \n"
-      "punpckldq %%xmm3,%%xmm3                   \n"
-      "paddd     %%xmm3,%%xmm3                   \n"
-      "pextrw    $0x3,%%xmm2,%k4                 \n"
+      "movd        %6,%%xmm2                     \n"
+      "movd        %7,%%xmm3                     \n"
+      "movl        $0x04040000,%k2               \n"
+      "movd        %k2,%%xmm5                    \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x9,%%xmm6                   \n"  // 0x007f007f
+      "pcmpeqb     %%xmm7,%%xmm7                 \n"
+      "psrlw       $15,%%xmm7                    \n"  // 0x00010001
+
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "subl        $0x2,%5                       \n"
+      "jl          29f                           \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "punpckldq   %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm3,%%xmm3                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
 
       LABELALIGN
       "2:                                        \n"
-      "movdqa    %%xmm2,%%xmm1                   \n"
-      "paddd     %%xmm3,%%xmm2                   \n"
-      "movzwl    0x00(%1,%3,1),%k2               \n"
-      "movd      %k2,%%xmm0                      \n"
-      "psrlw     $0x9,%%xmm1                     \n"
-      "movzwl    0x00(%1,%4,1),%k2               \n"
-      "movd      %k2,%%xmm4                      \n"
-      "pshufb    %%xmm5,%%xmm1                   \n"
-      "punpcklwd %%xmm4,%%xmm0                   \n"
-      "psubb     %8,%%xmm0                       \n"  // make pixels signed.
-      "pxor      %%xmm6,%%xmm1                   \n"  // 128 - f = (f ^ 127 ) +
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "movzwl      0x00(%1,%3,1),%k2             \n"
+      "movd        %k2,%%xmm0                    \n"
+      "psrlw       $0x9,%%xmm1                   \n"
+      "movzwl      0x00(%1,%4,1),%k2             \n"
+      "movd        %k2,%%xmm4                    \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "punpcklwd   %%xmm4,%%xmm0                 \n"
+      "psubb       %8,%%xmm0                     \n"  // make pixels signed.
+      "pxor        %%xmm6,%%xmm1                 \n"  // 128 - f = (f ^ 127 ) +
                                                       // 1
-      "paddusb   %%xmm7,%%xmm1                   \n"
-      "pmaddubsw %%xmm0,%%xmm1                   \n"
-      "pextrw    $0x1,%%xmm2,%k3                 \n"
-      "pextrw    $0x3,%%xmm2,%k4                 \n"
-      "paddw     %9,%%xmm1                       \n"  // make pixels unsigned.
-      "psrlw     $0x7,%%xmm1                     \n"
-      "packuswb  %%xmm1,%%xmm1                   \n"
-      "movd      %%xmm1,%k2                      \n"
-      "mov       %w2,(%0)                        \n"
-      "lea       0x2(%0),%0                      \n"
-      "subl      $0x2,%5                         \n"
-      "jge       2b                              \n"
+      "paddusb     %%xmm7,%%xmm1                 \n"
+      "pmaddubsw   %%xmm0,%%xmm1                 \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+      "paddw       %9,%%xmm1                     \n"  // make pixels unsigned.
+      "psrlw       $0x7,%%xmm1                   \n"
+      "packuswb    %%xmm1,%%xmm1                 \n"
+      "movd        %%xmm1,%k2                    \n"
+      "mov         %w2,(%0)                      \n"
+      "lea         0x2(%0),%0                    \n"
+      "subl        $0x2,%5                       \n"
+      "jge         2b                            \n"
 
       LABELALIGN
       "29:                                       \n"
-      "addl      $0x1,%5                         \n"
-      "jl        99f                             \n"
-      "movzwl    0x00(%1,%3,1),%k2               \n"
-      "movd      %k2,%%xmm0                      \n"
-      "psrlw     $0x9,%%xmm2                     \n"
-      "pshufb    %%xmm5,%%xmm2                   \n"
-      "psubb     %8,%%xmm0                       \n"  // make pixels signed.
-      "pxor      %%xmm6,%%xmm2                   \n"
-      "paddusb   %%xmm7,%%xmm2                   \n"
-      "pmaddubsw %%xmm0,%%xmm2                   \n"
-      "paddw     %9,%%xmm2                       \n"  // make pixels unsigned.
-      "psrlw     $0x7,%%xmm2                     \n"
-      "packuswb  %%xmm2,%%xmm2                   \n"
-      "movd      %%xmm2,%k2                      \n"
-      "mov       %b2,(%0)                        \n"
+      "addl        $0x1,%5                       \n"
+      "jl          99f                           \n"
+      "movzwl      0x00(%1,%3,1),%k2             \n"
+      "movd        %k2,%%xmm0                    \n"
+      "psrlw       $0x9,%%xmm2                   \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "psubb       %8,%%xmm0                     \n"  // make pixels signed.
+      "pxor        %%xmm6,%%xmm2                 \n"
+      "paddusb     %%xmm7,%%xmm2                 \n"
+      "pmaddubsw   %%xmm0,%%xmm2                 \n"
+      "paddw       %9,%%xmm2                     \n"  // make pixels unsigned.
+      "psrlw       $0x7,%%xmm2                   \n"
+      "packuswb    %%xmm2,%%xmm2                 \n"
+      "movd        %%xmm2,%k2                    \n"
+      "mov         %b2,(%0)                      \n"
       "99:                                       \n"
       : "+r"(dst_ptr),      // %0
         "+r"(src_ptr),      // %1
@@ -962,26 +1931,24 @@ void ScaleColsUp2_SSE2(uint8_t* dst_ptr,
                        int dx) {
   (void)x;
   (void)dx;
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%1),%%xmm0                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpcklbw %%xmm0,%%xmm0                   \n"
-      "punpckhbw %%xmm1,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%0)                     \n"
-      "movdqu    %%xmm1,0x10(%0)                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "sub       $0x20,%2                        \n"
-      "jg        1b                              \n"
-
-      : "+r"(dst_ptr),   // %0
-        "+r"(src_ptr),   // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1");
+      "movdqu      (%1),%%xmm0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpcklbw   %%xmm0,%%xmm0                 \n"
+      "punpckhbw   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x20,%2                      \n"
+      "jg          1b                            \n"
+
+               : "+r"(dst_ptr),   // %0
+                 "+r"(src_ptr),   // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
@@ -989,23 +1956,21 @@ void ScaleARGBRowDown2_SSE2(const uint8_t* src_argb,
                             uint8_t* dst_argb,
                             int dst_width) {
   (void)src_stride;
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "shufps    $0xdd,%%xmm1,%%xmm0             \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb),  // %0
-        "+r"(dst_argb),  // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1");
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "shufps      $0xdd,%%xmm1,%%xmm0           \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),  // %0
+                 "+r"(dst_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 void ScaleARGBRowDown2Linear_SSE2(const uint8_t* src_argb,
@@ -1013,56 +1978,52 @@ void ScaleARGBRowDown2Linear_SSE2(const uint8_t* src_argb,
                                   uint8_t* dst_argb,
                                   int dst_width) {
   (void)src_stride;
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb),  // %0
-        "+r"(dst_argb),  // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1");
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "lea         0x20(%0),%0                   \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),  // %0
+                 "+r"(dst_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 void ScaleARGBRowDown2Box_SSE2(const uint8_t* src_argb,
                                ptrdiff_t src_stride,
                                uint8_t* dst_argb,
                                int dst_width) {
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%0),%%xmm0                     \n"
-      "movdqu    0x10(%0),%%xmm1                 \n"
-      "movdqu    0x00(%0,%3,1),%%xmm2            \n"
-      "movdqu    0x10(%0,%3,1),%%xmm3            \n"
-      "lea       0x20(%0),%0                     \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "pavgb     %%xmm3,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%1)                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "sub       $0x4,%2                         \n"
-      "jg        1b                              \n"
-      : "+r"(src_argb),              // %0
-        "+r"(dst_argb),              // %1
-        "+r"(dst_width)              // %2
-      : "r"((intptr_t)(src_stride))  // %3
-      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
+      "movdqu      (%0),%%xmm0                   \n"
+      "movdqu      0x10(%0),%%xmm1               \n"
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"
+      "movdqu      0x10(%0,%3,1),%%xmm3          \n"
+      "lea         0x20(%0),%0                   \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+               : "+r"(src_argb),              // %0
+                 "+r"(dst_argb),              // %1
+                 "+r"(dst_width)              // %2
+               : "r"((intptr_t)(src_stride))  // %3
+               : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3");
 }
 
 // Reads 4 pixels at a time.
@@ -1076,23 +2037,23 @@ void ScaleARGBRowDownEven_SSE2(const uint8_t* src_argb,
   intptr_t src_stepx_x12;
   (void)src_stride;
   asm volatile(
-      "lea       0x00(,%1,4),%1                  \n"
-      "lea       0x00(%1,%1,2),%4                \n"
+      "lea         0x00(,%1,4),%1                \n"
+      "lea         0x00(%1,%1,2),%4              \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movd      (%0),%%xmm0                     \n"
-      "movd      0x00(%0,%1,1),%%xmm1            \n"
-      "punpckldq %%xmm1,%%xmm0                   \n"
-      "movd      0x00(%0,%1,2),%%xmm2            \n"
-      "movd      0x00(%0,%4,1),%%xmm3            \n"
-      "lea       0x00(%0,%1,4),%0                \n"
-      "punpckldq %%xmm3,%%xmm2                   \n"
-      "punpcklqdq %%xmm2,%%xmm0                  \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jg        1b                              \n"
+      "movd        (%0),%%xmm0                   \n"
+      "movd        0x00(%0,%1,1),%%xmm1          \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movd        0x00(%0,%1,2),%%xmm2          \n"
+      "movd        0x00(%0,%4,1),%%xmm3          \n"
+      "lea         0x00(%0,%1,4),%0              \n"
+      "punpckldq   %%xmm3,%%xmm2                 \n"
+      "punpcklqdq  %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),       // %0
         "+r"(src_stepx_x4),   // %1
         "+r"(dst_argb),       // %2
@@ -1113,32 +2074,32 @@ void ScaleARGBRowDownEvenBox_SSE2(const uint8_t* src_argb,
   intptr_t src_stepx_x12;
   intptr_t row1 = (intptr_t)(src_stride);
   asm volatile(
-      "lea       0x00(,%1,4),%1                  \n"
-      "lea       0x00(%1,%1,2),%4                \n"
-      "lea       0x00(%0,%5,1),%5                \n"
+      "lea         0x00(,%1,4),%1                \n"
+      "lea         0x00(%1,%1,2),%4              \n"
+      "lea         0x00(%0,%5,1),%5              \n"
 
       LABELALIGN
       "1:                                        \n"
-      "movq      (%0),%%xmm0                     \n"
-      "movhps    0x00(%0,%1,1),%%xmm0            \n"
-      "movq      0x00(%0,%1,2),%%xmm1            \n"
-      "movhps    0x00(%0,%4,1),%%xmm1            \n"
-      "lea       0x00(%0,%1,4),%0                \n"
-      "movq      (%5),%%xmm2                     \n"
-      "movhps    0x00(%5,%1,1),%%xmm2            \n"
-      "movq      0x00(%5,%1,2),%%xmm3            \n"
-      "movhps    0x00(%5,%4,1),%%xmm3            \n"
-      "lea       0x00(%5,%1,4),%5                \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "pavgb     %%xmm3,%%xmm1                   \n"
-      "movdqa    %%xmm0,%%xmm2                   \n"
-      "shufps    $0x88,%%xmm1,%%xmm0             \n"
-      "shufps    $0xdd,%%xmm1,%%xmm2             \n"
-      "pavgb     %%xmm2,%%xmm0                   \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%3                         \n"
-      "jg        1b                              \n"
+      "movq        (%0),%%xmm0                   \n"
+      "movhps      0x00(%0,%1,1),%%xmm0          \n"
+      "movq        0x00(%0,%1,2),%%xmm1          \n"
+      "movhps      0x00(%0,%4,1),%%xmm1          \n"
+      "lea         0x00(%0,%1,4),%0              \n"
+      "movq        (%5),%%xmm2                   \n"
+      "movhps      0x00(%5,%1,1),%%xmm2          \n"
+      "movq        0x00(%5,%1,2),%%xmm3          \n"
+      "movhps      0x00(%5,%4,1),%%xmm3          \n"
+      "lea         0x00(%5,%1,4),%5              \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "pavgb       %%xmm3,%%xmm1                 \n"
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "shufps      $0x88,%%xmm1,%%xmm0           \n"
+      "shufps      $0xdd,%%xmm1,%%xmm2           \n"
+      "pavgb       %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%3                       \n"
+      "jg          1b                            \n"
       : "+r"(src_argb),        // %0
         "+r"(src_stepx_x4),    // %1
         "+r"(dst_argb),        // %2
@@ -1156,56 +2117,56 @@ void ScaleARGBCols_SSE2(uint8_t* dst_argb,
                         int dx) {
   intptr_t x0, x1;
   asm volatile(
-      "movd      %5,%%xmm2                       \n"
-      "movd      %6,%%xmm3                       \n"
-      "pshufd    $0x0,%%xmm2,%%xmm2              \n"
-      "pshufd    $0x11,%%xmm3,%%xmm0             \n"
-      "paddd     %%xmm0,%%xmm2                   \n"
-      "paddd     %%xmm3,%%xmm3                   \n"
-      "pshufd    $0x5,%%xmm3,%%xmm0              \n"
-      "paddd     %%xmm0,%%xmm2                   \n"
-      "paddd     %%xmm3,%%xmm3                   \n"
-      "pshufd    $0x0,%%xmm3,%%xmm3              \n"
-      "pextrw    $0x1,%%xmm2,%k0                 \n"
-      "pextrw    $0x3,%%xmm2,%k1                 \n"
-      "cmp       $0x0,%4                         \n"
-      "jl        99f                             \n"
-      "sub       $0x4,%4                         \n"
-      "jl        49f                             \n"
+      "movd        %5,%%xmm2                     \n"
+      "movd        %6,%%xmm3                     \n"
+      "pshufd      $0x0,%%xmm2,%%xmm2            \n"
+      "pshufd      $0x11,%%xmm3,%%xmm0           \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pshufd      $0x5,%%xmm3,%%xmm0            \n"
+      "paddd       %%xmm0,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pshufd      $0x0,%%xmm3,%%xmm3            \n"
+      "pextrw      $0x1,%%xmm2,%k0               \n"
+      "pextrw      $0x3,%%xmm2,%k1               \n"
+      "cmp         $0x0,%4                       \n"
+      "jl          99f                           \n"
+      "sub         $0x4,%4                       \n"
+      "jl          49f                           \n"
 
       LABELALIGN
       "40:                                       \n"
-      "movd      0x00(%3,%0,4),%%xmm0            \n"
-      "movd      0x00(%3,%1,4),%%xmm1            \n"
-      "pextrw    $0x5,%%xmm2,%k0                 \n"
-      "pextrw    $0x7,%%xmm2,%k1                 \n"
-      "paddd     %%xmm3,%%xmm2                   \n"
-      "punpckldq %%xmm1,%%xmm0                   \n"
-      "movd      0x00(%3,%0,4),%%xmm1            \n"
-      "movd      0x00(%3,%1,4),%%xmm4            \n"
-      "pextrw    $0x1,%%xmm2,%k0                 \n"
-      "pextrw    $0x3,%%xmm2,%k1                 \n"
-      "punpckldq %%xmm4,%%xmm1                   \n"
-      "punpcklqdq %%xmm1,%%xmm0                  \n"
-      "movdqu    %%xmm0,(%2)                     \n"
-      "lea       0x10(%2),%2                     \n"
-      "sub       $0x4,%4                         \n"
-      "jge       40b                             \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        0x00(%3,%1,4),%%xmm1          \n"
+      "pextrw      $0x5,%%xmm2,%k0               \n"
+      "pextrw      $0x7,%%xmm2,%k1               \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movd        0x00(%3,%0,4),%%xmm1          \n"
+      "movd        0x00(%3,%1,4),%%xmm4          \n"
+      "pextrw      $0x1,%%xmm2,%k0               \n"
+      "pextrw      $0x3,%%xmm2,%k1               \n"
+      "punpckldq   %%xmm4,%%xmm1                 \n"
+      "punpcklqdq  %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%2)                   \n"
+      "lea         0x10(%2),%2                   \n"
+      "sub         $0x4,%4                       \n"
+      "jge         40b                           \n"
 
       "49:                                       \n"
-      "test      $0x2,%4                         \n"
-      "je        29f                             \n"
-      "movd      0x00(%3,%0,4),%%xmm0            \n"
-      "movd      0x00(%3,%1,4),%%xmm1            \n"
-      "pextrw    $0x5,%%xmm2,%k0                 \n"
-      "punpckldq %%xmm1,%%xmm0                   \n"
-      "movq      %%xmm0,(%2)                     \n"
-      "lea       0x8(%2),%2                      \n"
+      "test        $0x2,%4                       \n"
+      "je          29f                           \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        0x00(%3,%1,4),%%xmm1          \n"
+      "pextrw      $0x5,%%xmm2,%k0               \n"
+      "punpckldq   %%xmm1,%%xmm0                 \n"
+      "movq        %%xmm0,(%2)                   \n"
+      "lea         0x8(%2),%2                    \n"
       "29:                                       \n"
-      "test      $0x1,%4                         \n"
-      "je        99f                             \n"
-      "movd      0x00(%3,%0,4),%%xmm0            \n"
-      "movd      %%xmm0,(%2)                     \n"
+      "test        $0x1,%4                       \n"
+      "je          99f                           \n"
+      "movd        0x00(%3,%0,4),%%xmm0          \n"
+      "movd        %%xmm0,(%2)                   \n"
       "99:                                       \n"
       : "=&a"(x0),       // %0
         "=&d"(x1),       // %1
@@ -1226,26 +2187,24 @@ void ScaleARGBColsUp2_SSE2(uint8_t* dst_argb,
                            int dx) {
   (void)x;
   (void)dx;
-  asm volatile(
-
-      LABELALIGN
+  asm volatile(LABELALIGN
       "1:                                        \n"
-      "movdqu    (%1),%%xmm0                     \n"
-      "lea       0x10(%1),%1                     \n"
-      "movdqa    %%xmm0,%%xmm1                   \n"
-      "punpckldq %%xmm0,%%xmm0                   \n"
-      "punpckhdq %%xmm1,%%xmm1                   \n"
-      "movdqu    %%xmm0,(%0)                     \n"
-      "movdqu    %%xmm1,0x10(%0)                 \n"
-      "lea       0x20(%0),%0                     \n"
-      "sub       $0x8,%2                         \n"
-      "jg        1b                              \n"
-
-      : "+r"(dst_argb),  // %0
-        "+r"(src_argb),  // %1
-        "+r"(dst_width)  // %2
-        ::"memory",
-        "cc", "xmm0", "xmm1");
+      "movdqu      (%1),%%xmm0                   \n"
+      "lea         0x10(%1),%1                   \n"
+      "movdqa      %%xmm0,%%xmm1                 \n"
+      "punpckldq   %%xmm0,%%xmm0                 \n"
+      "punpckhdq   %%xmm1,%%xmm1                 \n"
+      "movdqu      %%xmm0,(%0)                   \n"
+      "movdqu      %%xmm1,0x10(%0)               \n"
+      "lea         0x20(%0),%0                   \n"
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+
+               : "+r"(dst_argb),  // %0
+                 "+r"(src_argb),  // %1
+                 "+r"(dst_width)  // %2
+                 ::"memory",
+                 "cc", "xmm0", "xmm1");
 }
 
 // Shuffle table for arranging 2 pixels into pairs for pmaddubsw
@@ -1267,63 +2226,64 @@ void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
                                int dx) {
   intptr_t x0, x1;
   asm volatile(
-      "movdqa    %0,%%xmm4                       \n"
-      "movdqa    %1,%%xmm5                       \n"
+      "movdqa      %0,%%xmm4                     \n"
+      "movdqa      %1,%%xmm5                     \n"
       :
       : "m"(kShuffleColARGB),   // %0
         "m"(kShuffleFractions)  // %1
   );
 
   asm volatile(
-      "movd      %5,%%xmm2                       \n"
-      "movd      %6,%%xmm3                       \n"
-      "pcmpeqb   %%xmm6,%%xmm6                   \n"
-      "psrlw     $0x9,%%xmm6                     \n"
-      "pextrw    $0x1,%%xmm2,%k3                 \n"
-      "sub       $0x2,%2                         \n"
-      "jl        29f                             \n"
-      "movdqa    %%xmm2,%%xmm0                   \n"
-      "paddd     %%xmm3,%%xmm0                   \n"
-      "punpckldq %%xmm0,%%xmm2                   \n"
-      "punpckldq %%xmm3,%%xmm3                   \n"
-      "paddd     %%xmm3,%%xmm3                   \n"
-      "pextrw    $0x3,%%xmm2,%k4                 \n"
+      "movd        %5,%%xmm2                     \n"
+      "movd        %6,%%xmm3                     \n"
+      "pcmpeqb     %%xmm6,%%xmm6                 \n"
+      "psrlw       $0x9,%%xmm6                   \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "sub         $0x2,%2                       \n"
+      "jl          29f                           \n"
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "paddd       %%xmm3,%%xmm0                 \n"
+      "punpckldq   %%xmm0,%%xmm2                 \n"
+      "punpckldq   %%xmm3,%%xmm3                 \n"
+      "paddd       %%xmm3,%%xmm3                 \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
 
       LABELALIGN
       "2:                                        \n"
-      "movdqa    %%xmm2,%%xmm1                   \n"
-      "paddd     %%xmm3,%%xmm2                   \n"
-      "movq      0x00(%1,%3,4),%%xmm0            \n"
-      "psrlw     $0x9,%%xmm1                     \n"
-      "movhps    0x00(%1,%4,4),%%xmm0            \n"
-      "pshufb    %%xmm5,%%xmm1                   \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "pxor      %%xmm6,%%xmm1                   \n"
-      "pmaddubsw %%xmm1,%%xmm0                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "pextrw    $0x1,%%xmm2,%k3                 \n"
-      "pextrw    $0x3,%%xmm2,%k4                 \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movq      %%xmm0,(%0)                     \n"
-      "lea       0x8(%0),%0                      \n"
-      "sub       $0x2,%2                         \n"
-      "jge       2b                              \n"
+      "movdqa      %%xmm2,%%xmm1                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"
+      "movq        0x00(%1,%3,4),%%xmm0          \n"
+      "psrlw       $0x9,%%xmm1                   \n"
+      "movhps      0x00(%1,%4,4),%%xmm0          \n"
+      "pshufb      %%xmm5,%%xmm1                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pxor        %%xmm6,%%xmm1                 \n"
+      "pmaddubsw   %%xmm1,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "pextrw      $0x1,%%xmm2,%k3               \n"
+      "pextrw      $0x3,%%xmm2,%k4               \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movq        %%xmm0,(%0)                   \n"
+      "lea         0x8(%0),%0                    \n"
+      "sub         $0x2,%2                       \n"
+      "jge         2b                            \n"
 
       LABELALIGN
       "29:                                       \n"
-      "add       $0x1,%2                         \n"
-      "jl        99f                             \n"
-      "psrlw     $0x9,%%xmm2                     \n"
-      "movq      0x00(%1,%3,4),%%xmm0            \n"
-      "pshufb    %%xmm5,%%xmm2                   \n"
-      "pshufb    %%xmm4,%%xmm0                   \n"
-      "pxor      %%xmm6,%%xmm2                   \n"
-      "pmaddubsw %%xmm2,%%xmm0                   \n"
-      "psrlw     $0x7,%%xmm0                     \n"
-      "packuswb  %%xmm0,%%xmm0                   \n"
-      "movd      %%xmm0,(%0)                     \n"
-
-      LABELALIGN "99:                            \n"  // clang-format error.
+      "add         $0x1,%2                       \n"
+      "jl          99f                           \n"
+      "psrlw       $0x9,%%xmm2                   \n"
+      "movq        0x00(%1,%3,4),%%xmm0          \n"
+      "pshufb      %%xmm5,%%xmm2                 \n"
+      "pshufb      %%xmm4,%%xmm0                 \n"
+      "pxor        %%xmm6,%%xmm2                 \n"
+      "pmaddubsw   %%xmm2,%%xmm0                 \n"
+      "psrlw       $0x7,%%xmm0                   \n"
+      "packuswb    %%xmm0,%%xmm0                 \n"
+      "movd        %%xmm0,(%0)                   \n"
+
+      LABELALIGN
+      "99:                                       \n"  // clang-format error.
 
       : "+r"(dst_argb),    // %0
         "+r"(src_argb),    // %1
@@ -1339,10 +2299,10 @@ void ScaleARGBFilterCols_SSSE3(uint8_t* dst_argb,
 int FixedDiv_X86(int num, int div) {
   asm volatile(
       "cdq                                       \n"
-      "shld      $0x10,%%eax,%%edx               \n"
-      "shl       $0x10,%%eax                     \n"
-      "idiv      %1                              \n"
-      "mov       %0, %%eax                       \n"
+      "shld        $0x10,%%eax,%%edx             \n"
+      "shl         $0x10,%%eax                   \n"
+      "idiv        %1                            \n"
+      "mov         %0, %%eax                     \n"
       : "+a"(num)  // %0
       : "c"(div)   // %1
       : "memory", "cc", "edx");
@@ -1353,19 +2313,637 @@ int FixedDiv_X86(int num, int div) {
 int FixedDiv1_X86(int num, int div) {
   asm volatile(
       "cdq                                       \n"
-      "shld      $0x10,%%eax,%%edx               \n"
-      "shl       $0x10,%%eax                     \n"
-      "sub       $0x10001,%%eax                  \n"
-      "sbb       $0x0,%%edx                      \n"
-      "sub       $0x1,%1                         \n"
-      "idiv      %1                              \n"
-      "mov       %0, %%eax                       \n"
+      "shld        $0x10,%%eax,%%edx             \n"
+      "shl         $0x10,%%eax                   \n"
+      "sub         $0x10001,%%eax                \n"
+      "sbb         $0x0,%%edx                    \n"
+      "sub         $0x1,%1                       \n"
+      "idiv        %1                            \n"
+      "mov         %0, %%eax                     \n"
       : "+a"(num)  // %0
       : "c"(div)   // %1
       : "memory", "cc", "edx");
   return num;
 }
 
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3) || \
+    defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+
+// Shuffle table for splitting UV into upper and lower part of register.
+static const uvec8 kShuffleSplitUV = {0u, 2u, 4u, 6u, 8u, 10u, 12u, 14u,
+                                      1u, 3u, 5u, 7u, 9u, 11u, 13u, 15u};
+static const uvec8 kShuffleMergeUV = {0u,   8u,   2u,   10u,  4u,   12u,
+                                      6u,   14u,  0x80, 0x80, 0x80, 0x80,
+                                      0x80, 0x80, 0x80, 0x80};
+#endif
+
+#ifdef HAS_SCALEUVROWDOWN2BOX_SSSE3
+
+void ScaleUVRowDown2Box_SSSE3(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  asm volatile(
+      "pcmpeqb     %%xmm4,%%xmm4                 \n"  // 01010101
+      "psrlw       $0xf,%%xmm4                   \n"
+      "packuswb    %%xmm4,%%xmm4                 \n"
+      "pxor        %%xmm5, %%xmm5                \n"  // zero
+      "movdqa      %4,%%xmm1                     \n"  // split shuffler
+      "movdqa      %5,%%xmm3                     \n"  // merge shuffler
+
+      LABELALIGN
+      "1:                                        \n"
+      "movdqu      (%0),%%xmm0                   \n"  // 8 UV row 0
+      "movdqu      0x00(%0,%3,1),%%xmm2          \n"  // 8 UV row 1
+      "lea         0x10(%0),%0                   \n"
+      "pshufb      %%xmm1,%%xmm0                 \n"  // uuuuvvvv
+      "pshufb      %%xmm1,%%xmm2                 \n"
+      "pmaddubsw   %%xmm4,%%xmm0                 \n"  // horizontal add
+      "pmaddubsw   %%xmm4,%%xmm2                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // vertical add
+      "psrlw       $0x1,%%xmm0                   \n"  // round
+      "pavgw       %%xmm5,%%xmm0                 \n"
+      "pshufb      %%xmm3,%%xmm0                 \n"  // merge uv
+      "movq        %%xmm0,(%1)                   \n"
+      "lea         0x8(%1),%1                    \n"  // 4 UV
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "m"(kShuffleSplitUV),         // %4
+        "m"(kShuffleMergeUV)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_SCALEUVROWDOWN2BOX_SSSE3
+
+#ifdef HAS_SCALEUVROWDOWN2BOX_AVX2
+void ScaleUVRowDown2Box_AVX2(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  asm volatile(
+      "vpcmpeqb    %%ymm4,%%ymm4,%%ymm4          \n"  // 01010101
+      "vpsrlw      $0xf,%%ymm4,%%ymm4            \n"
+      "vpackuswb   %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpxor       %%ymm5,%%ymm5,%%ymm5          \n"  // zero
+      "vbroadcastf128 %4,%%ymm1                  \n"  // split shuffler
+      "vbroadcastf128 %5,%%ymm3                  \n"  // merge shuffler
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%ymm0                   \n"  // 16 UV row 0
+      "vmovdqu     0x00(%0,%3,1),%%ymm2          \n"  // 16 UV row 1
+      "lea         0x20(%0),%0                   \n"
+      "vpshufb     %%ymm1,%%ymm0,%%ymm0          \n"  // uuuuvvvv
+      "vpshufb     %%ymm1,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm4,%%ymm0,%%ymm0          \n"  // horizontal add
+      "vpmaddubsw  %%ymm4,%%ymm2,%%ymm2          \n"
+      "vpaddw      %%ymm2,%%ymm0,%%ymm0          \n"  // vertical add
+      "vpsrlw      $0x1,%%ymm0,%%ymm0            \n"  // round
+      "vpavgw      %%ymm5,%%ymm0,%%ymm0          \n"
+      "vpshufb     %%ymm3,%%ymm0,%%ymm0          \n"  // merge uv
+      "vpermq      $0xd8,%%ymm0,%%ymm0           \n"  // combine qwords
+      "vmovdqu     %%xmm0,(%1)                   \n"
+      "lea         0x10(%1),%1                   \n"  // 8 UV
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "m"(kShuffleSplitUV),         // %4
+        "m"(kShuffleMergeUV)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif  // HAS_SCALEUVROWDOWN2BOX_AVX2
+
+static const uvec8 kUVLinearMadd31 = {3, 1, 3, 1, 1, 3, 1, 3,
+                                      3, 1, 3, 1, 1, 3, 1, 3};
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_SSSE3
+void ScaleUVRowUp2_Linear_SSSE3(const uint8_t* src_ptr,
+                                uint8_t* dst_ptr,
+                                int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm4,%%xmm4                 \n"
+      "psrlw       $15,%%xmm4                    \n"
+      "psllw       $1,%%xmm4                     \n"  // all 2
+      "movdqa      %3,%%xmm3                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 00112233 (1u1v)
+      "movq        2(%0),%%xmm1                  \n"  // 11223344 (1u1v)
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // 0101121223233434 (2u2v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm0,%%xmm2                 \n"  // 2323232334343434 (2u2v)
+      "punpckldq   %%xmm0,%%xmm0                 \n"  // 0101010112121212 (2u2v)
+      "pmaddubsw   %%xmm3,%%xmm2                 \n"  // 3*near+far (1u1v16, hi)
+      "pmaddubsw   %%xmm3,%%xmm0                 \n"  // 3*near+far (1u1v16, lo)
+      "paddw       %%xmm4,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddw       %%xmm4,%%xmm2                 \n"  // 3*near+far+2 (hi)
+      "psrlw       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrlw       $2,%%xmm2                     \n"  // 3/4*near+1/4*far (hi)
+      "packuswb    %%xmm2,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),        // %0
+        "+r"(dst_ptr),        // %1
+        "+r"(dst_width)       // %2
+      : "m"(kUVLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_SSSE3
+void ScaleUVRowUp2_Bilinear_SSSE3(const uint8_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint8_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  asm volatile(
+      "pcmpeqw     %%xmm6,%%xmm6                 \n"
+      "psrlw       $15,%%xmm6                    \n"
+      "psllw       $3,%%xmm6                     \n"  // all 8
+      "movdqa      %5,%%xmm7                     \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 00112233 (1u1v)
+      "movq        2(%0),%%xmm1                  \n"  // 11223344 (1u1v)
+      "punpcklbw   %%xmm1,%%xmm0                 \n"  // 0101121223233434 (2u2v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "punpckhdq   %%xmm0,%%xmm2                 \n"  // 2323232334343434 (2u2v)
+      "punpckldq   %%xmm0,%%xmm0                 \n"  // 0101010112121212 (2u2v)
+      "pmaddubsw   %%xmm7,%%xmm2                 \n"  // 3*near+far (1u1v16, hi)
+      "pmaddubsw   %%xmm7,%%xmm0                 \n"  // 3*near+far (1u1v16, lo)
+
+      "movq        (%0,%3),%%xmm1                \n"
+      "movq        2(%0,%3),%%xmm4               \n"
+      "punpcklbw   %%xmm4,%%xmm1                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "punpckhdq   %%xmm1,%%xmm3                 \n"
+      "punpckldq   %%xmm1,%%xmm1                 \n"
+      "pmaddubsw   %%xmm7,%%xmm3                 \n"  // 3*near+far (2, hi)
+      "pmaddubsw   %%xmm7,%%xmm1                 \n"  // 3*near+far (2, lo)
+
+      // xmm0 xmm2
+      // xmm1 xmm3
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddw       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddw       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddw       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrlw       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm1,%%xmm5                 \n"
+      "paddw       %%xmm1,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddw       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddw       %%xmm1,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddw       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrlw       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm2,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm2,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddw       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (2, hi)
+      "paddw       %%xmm2,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddw       %%xmm1,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrlw       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm1                 \n"
+      "paddw       %%xmm3,%%xmm1                 \n"  // 6*near+2*far (2, hi)
+      "paddw       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (1, hi)
+      "paddw       %%xmm3,%%xmm1                 \n"  // 9*near+3*far (2, hi)
+      "paddw       %%xmm2,%%xmm1                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrlw       $4,%%xmm1                     \n"  // ^ div by 16 (2, hi)
+
+      "packuswb    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packuswb    %%xmm1,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4)                \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kUVLinearMadd31)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_AVX2
+
+void ScaleUVRowUp2_Linear_AVX2(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrlw      $15,%%ymm4,%%ymm4             \n"
+      "vpsllw      $1,%%ymm4,%%ymm4              \n"  // all 2
+      "vbroadcastf128 %3,%%ymm3                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"
+      "vmovdqu     2(%0),%%xmm1                  \n"
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpunpckhdq  %%ymm0,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm3,%%ymm2,%%ymm1          \n"  // 3*near+far (hi)
+      "vpmaddubsw  %%ymm3,%%ymm0,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddw      %%ymm4,%%ymm1,%%ymm1          \n"  // 3*near+far+2 (hi)
+      "vpsrlw      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrlw      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackuswb   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 uv to 16 uv
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),        // %0
+        "+r"(dst_ptr),        // %1
+        "+r"(dst_width)       // %2
+      : "m"(kUVLinearMadd31)  // %3
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_AVX2
+void ScaleUVRowUp2_Bilinear_AVX2(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  asm volatile(
+      "vpcmpeqw    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrlw      $15,%%ymm6,%%ymm6             \n"
+      "vpsllw      $3,%%ymm6,%%ymm6              \n"  // all 8
+      "vbroadcastf128 %5,%%ymm7                  \n"
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"
+      "vmovdqu     2(%0),%%xmm1                  \n"
+      "vpermq      $0b11011000,%%ymm0,%%ymm0     \n"
+      "vpermq      $0b11011000,%%ymm1,%%ymm1     \n"
+      "vpunpcklbw  %%ymm1,%%ymm0,%%ymm0          \n"
+      "vpunpckhdq  %%ymm0,%%ymm0,%%ymm2          \n"
+      "vpunpckldq  %%ymm0,%%ymm0,%%ymm0          \n"
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm1          \n"  // 3*near+far (1, hi)
+      "vpmaddubsw  %%ymm7,%%ymm0,%%ymm0          \n"  // 3*near+far (1, lo)
+
+      "vmovdqu     (%0,%3),%%xmm2                \n"  // 0123456789ABCDEF
+      "vmovdqu     2(%0,%3),%%xmm3               \n"  // 123456789ABCDEF0
+      "vpermq      $0b11011000,%%ymm2,%%ymm2     \n"
+      "vpermq      $0b11011000,%%ymm3,%%ymm3     \n"
+      "vpunpcklbw  %%ymm3,%%ymm2,%%ymm2          \n"
+      "vpunpckhdq  %%ymm2,%%ymm2,%%ymm4          \n"
+      "vpunpckldq  %%ymm2,%%ymm2,%%ymm2          \n"
+      "vpmaddubsw  %%ymm7,%%ymm4,%%ymm3          \n"  // 3*near+far (2, hi)
+      "vpmaddubsw  %%ymm7,%%ymm2,%%ymm2          \n"  // 3*near+far (2, lo)
+
+      // ymm0 ymm1
+      // ymm2 ymm3
+
+      "vpaddw      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddw      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddw      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddw      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrlw      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddw      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddw      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddw      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddw      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrlw      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddw      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddw      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddw      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddw      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrlw      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddw      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddw      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddw      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddw      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrlw      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackuswb   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackuswb   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4)                \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 8 uv to 16 uv
+      "sub         $0x10,%2                      \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride)),  // %4
+        "m"(kUVLinearMadd31)          // %5
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_SSE41
+void ScaleUVRowUp2_Linear_16_SSE41(const uint16_t* src_ptr,
+                                   uint16_t* dst_ptr,
+                                   int dst_width) {
+  asm volatile(
+      "pxor        %%xmm5,%%xmm5                 \n"
+      "pcmpeqd     %%xmm4,%%xmm4                 \n"
+      "psrld       $31,%%xmm4                    \n"
+      "pslld       $1,%%xmm4                     \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+
+      "punpcklwd   %%xmm5,%%xmm0                 \n"  // 0011 (32b, 1u1v)
+      "punpcklwd   %%xmm5,%%xmm1                 \n"  // 1122 (32b, 1u1v)
+
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (lo, far)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (hi, far)
+
+      "paddd       %%xmm4,%%xmm2                 \n"  // far+2 (lo)
+      "paddd       %%xmm4,%%xmm3                 \n"  // far+2 (hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far+2 (lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far+2 (hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far+2 (lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far+2 (hi)
+
+      "psrld       $2,%%xmm0                     \n"  // 3/4*near+1/4*far (lo)
+      "psrld       $2,%%xmm1                     \n"  // 3/4*near+1/4*far (hi)
+      "packusdw    %%xmm1,%%xmm0                 \n"
+      "movdqu      %%xmm0,(%1)                   \n"
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 2 uv to 4 uv
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_SSE41
+void ScaleUVRowUp2_Bilinear_16_SSE41(const uint16_t* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint16_t* dst_ptr,
+                                     ptrdiff_t dst_stride,
+                                     int dst_width) {
+  asm volatile(
+      "pxor        %%xmm7,%%xmm7                 \n"
+      "pcmpeqd     %%xmm6,%%xmm6                 \n"
+      "psrld       $31,%%xmm6                    \n"
+      "pslld       $3,%%xmm6                     \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+      "movq        (%0),%%xmm0                   \n"  // 0011 (16b, 1u1v)
+      "movq        4(%0),%%xmm1                  \n"  // 1122 (16b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm0                 \n"  // 0011 (near) (32b, 1u1v)
+      "punpcklwd   %%xmm7,%%xmm1                 \n"  // 1122 (near) (32b, 1u1v)
+      "movdqa      %%xmm0,%%xmm2                 \n"
+      "movdqa      %%xmm1,%%xmm3                 \n"
+      "pshufd      $0b01001110,%%xmm2,%%xmm2     \n"  // 1100 (far) (1, lo)
+      "pshufd      $0b01001110,%%xmm3,%%xmm3     \n"  // 2211 (far) (1, hi)
+      "paddd       %%xmm0,%%xmm2                 \n"  // near+far (1, lo)
+      "paddd       %%xmm1,%%xmm3                 \n"  // near+far (1, hi)
+      "paddd       %%xmm0,%%xmm0                 \n"  // 2*near (1, lo)
+      "paddd       %%xmm1,%%xmm1                 \n"  // 2*near (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 3*near+far (1, lo)
+      "paddd       %%xmm3,%%xmm1                 \n"  // 3*near+far (1, hi)
+
+      "movq        (%0,%3,2),%%xmm2              \n"
+      "movq        4(%0,%3,2),%%xmm3             \n"
+      "punpcklwd   %%xmm7,%%xmm2                 \n"
+      "punpcklwd   %%xmm7,%%xmm3                 \n"
+      "movdqa      %%xmm2,%%xmm4                 \n"
+      "movdqa      %%xmm3,%%xmm5                 \n"
+      "pshufd      $0b01001110,%%xmm4,%%xmm4     \n"  // 1100 (far) (2, lo)
+      "pshufd      $0b01001110,%%xmm5,%%xmm5     \n"  // 2211 (far) (2, hi)
+      "paddd       %%xmm2,%%xmm4                 \n"  // near+far (2, lo)
+      "paddd       %%xmm3,%%xmm5                 \n"  // near+far (2, hi)
+      "paddd       %%xmm2,%%xmm2                 \n"  // 2*near (2, lo)
+      "paddd       %%xmm3,%%xmm3                 \n"  // 2*near (2, hi)
+      "paddd       %%xmm4,%%xmm2                 \n"  // 3*near+far (2, lo)
+      "paddd       %%xmm5,%%xmm3                 \n"  // 3*near+far (2, hi)
+
+      "movdqa      %%xmm0,%%xmm4                 \n"
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm0,%%xmm4                 \n"  // 6*near+2*far (1, lo)
+      "paddd       %%xmm6,%%xmm5                 \n"  // 3*near+far+8 (2, lo)
+      "paddd       %%xmm0,%%xmm4                 \n"  // 9*near+3*far (1, lo)
+      "paddd       %%xmm5,%%xmm4                 \n"  // 9 3 3 1 + 8 (1, lo)
+      "psrld       $4,%%xmm4                     \n"  // ^ div by 16 (1, lo)
+
+      "movdqa      %%xmm2,%%xmm5                 \n"
+      "paddd       %%xmm2,%%xmm5                 \n"  // 6*near+2*far (2, lo)
+      "paddd       %%xmm6,%%xmm0                 \n"  // 3*near+far+8 (1, lo)
+      "paddd       %%xmm2,%%xmm5                 \n"  // 9*near+3*far (2, lo)
+      "paddd       %%xmm0,%%xmm5                 \n"  // 9 3 3 1 + 8 (2, lo)
+      "psrld       $4,%%xmm5                     \n"  // ^ div by 16 (2, lo)
+
+      "movdqa      %%xmm1,%%xmm0                 \n"
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm1,%%xmm0                 \n"  // 6*near+2*far (1, hi)
+      "paddd       %%xmm6,%%xmm2                 \n"  // 3*near+far+8 (2, hi)
+      "paddd       %%xmm1,%%xmm0                 \n"  // 9*near+3*far (1, hi)
+      "paddd       %%xmm2,%%xmm0                 \n"  // 9 3 3 1 + 8 (1, hi)
+      "psrld       $4,%%xmm0                     \n"  // ^ div by 16 (1, hi)
+
+      "movdqa      %%xmm3,%%xmm2                 \n"
+      "paddd       %%xmm3,%%xmm2                 \n"  // 6*near+2*far (2, hi)
+      "paddd       %%xmm6,%%xmm1                 \n"  // 3*near+far+8 (1, hi)
+      "paddd       %%xmm3,%%xmm2                 \n"  // 9*near+3*far (2, hi)
+      "paddd       %%xmm1,%%xmm2                 \n"  // 9 3 3 1 + 8 (2, hi)
+      "psrld       $4,%%xmm2                     \n"  // ^ div by 16 (2, hi)
+
+      "packusdw    %%xmm0,%%xmm4                 \n"
+      "movdqu      %%xmm4,(%1)                   \n"  // store above
+      "packusdw    %%xmm2,%%xmm5                 \n"
+      "movdqu      %%xmm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x8(%0),%0                    \n"
+      "lea         0x10(%1),%1                   \n"  // 2 uv to 4 uv
+      "sub         $0x4,%2                       \n"
+      "jg          1b                            \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6",
+        "xmm7");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_LINEAR_16_AVX2
+void ScaleUVRowUp2_Linear_16_AVX2(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm4,%%ymm4,%%ymm4          \n"
+      "vpsrld      $31,%%ymm4,%%ymm4             \n"
+      "vpslld      $1,%%ymm4,%%ymm4              \n"  // all 2
+
+      LABELALIGN
+      "1:                                        \n"
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0),%%xmm1                  \n"  // 11223344 (16b, 1u1v)
+
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+
+      "vpshufd     $0b01001110,%%ymm0,%%ymm2     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm1,%%ymm3     \n"  // 22114433 (hi, far)
+
+      "vpaddd      %%ymm4,%%ymm2,%%ymm2          \n"  // far+2 (lo)
+      "vpaddd      %%ymm4,%%ymm3,%%ymm3          \n"  // far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far+2 (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far+2 (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far+2 (hi)
+
+      "vpsrld      $2,%%ymm0,%%ymm0              \n"  // 3/4*near+1/4*far (lo)
+      "vpsrld      $2,%%ymm1,%%ymm1              \n"  // 3/4*near+1/4*far (hi)
+      "vpackusdw   %%ymm1,%%ymm0,%%ymm0          \n"
+      "vmovdqu     %%ymm0,(%1)                   \n"
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(dst_ptr),   // %1
+        "+r"(dst_width)  // %2
+      :
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4");
+}
+#endif
+
+#ifdef HAS_SCALEUVROWUP2_BILINEAR_16_AVX2
+void ScaleUVRowUp2_Bilinear_16_AVX2(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  asm volatile(
+      "vpcmpeqd    %%ymm6,%%ymm6,%%ymm6          \n"
+      "vpsrld      $31,%%ymm6,%%ymm6             \n"
+      "vpslld      $3,%%ymm6,%%ymm6              \n"  // all 8
+
+      LABELALIGN
+      "1:                                        \n"
+
+      "vmovdqu     (%0),%%xmm0                   \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0),%%xmm1                  \n"  // 11223344 (16b, 1u1v)
+      "vpmovzxwd   %%xmm0,%%ymm0                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm1,%%ymm1                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b01001110,%%ymm0,%%ymm2     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm1,%%ymm3     \n"  // 22114433 (hi, far)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm2          \n"  // near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm3          \n"  // near+far (hi)
+      "vpaddd      %%ymm0,%%ymm0,%%ymm0          \n"  // 2*near (lo)
+      "vpaddd      %%ymm1,%%ymm1,%%ymm1          \n"  // 2*near (hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 3*near+far (lo)
+      "vpaddd      %%ymm1,%%ymm3,%%ymm1          \n"  // 3*near+far (hi)
+
+      "vmovdqu     (%0,%3,2),%%xmm2              \n"  // 00112233 (16b, 1u1v)
+      "vmovdqu     4(%0,%3,2),%%xmm3             \n"  // 11223344 (16b, 1u1v)
+      "vpmovzxwd   %%xmm2,%%ymm2                 \n"  // 01234567 (32b, 1u1v)
+      "vpmovzxwd   %%xmm3,%%ymm3                 \n"  // 12345678 (32b, 1u1v)
+      "vpshufd     $0b01001110,%%ymm2,%%ymm4     \n"  // 11003322 (lo, far)
+      "vpshufd     $0b01001110,%%ymm3,%%ymm5     \n"  // 22114433 (hi, far)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm4          \n"  // near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm5          \n"  // near+far (hi)
+      "vpaddd      %%ymm2,%%ymm2,%%ymm2          \n"  // 2*near (lo)
+      "vpaddd      %%ymm3,%%ymm3,%%ymm3          \n"  // 2*near (hi)
+      "vpaddd      %%ymm2,%%ymm4,%%ymm2          \n"  // 3*near+far (lo)
+      "vpaddd      %%ymm3,%%ymm5,%%ymm3          \n"  // 3*near+far (hi)
+
+      "vpaddd      %%ymm0,%%ymm0,%%ymm4          \n"  // 6*near+2*far (1, lo)
+      "vpaddd      %%ymm6,%%ymm2,%%ymm5          \n"  // 3*near+far+8 (2, lo)
+      "vpaddd      %%ymm4,%%ymm0,%%ymm4          \n"  // 9*near+3*far (1, lo)
+      "vpaddd      %%ymm4,%%ymm5,%%ymm4          \n"  // 9 3 3 1 + 8 (1, lo)
+      "vpsrld      $4,%%ymm4,%%ymm4              \n"  // ^ div by 16 (1, lo)
+
+      "vpaddd      %%ymm2,%%ymm2,%%ymm5          \n"  // 6*near+2*far (2, lo)
+      "vpaddd      %%ymm6,%%ymm0,%%ymm0          \n"  // 3*near+far+8 (1, lo)
+      "vpaddd      %%ymm5,%%ymm2,%%ymm5          \n"  // 9*near+3*far (2, lo)
+      "vpaddd      %%ymm5,%%ymm0,%%ymm5          \n"  // 9 3 3 1 + 8 (2, lo)
+      "vpsrld      $4,%%ymm5,%%ymm5              \n"  // ^ div by 16 (2, lo)
+
+      "vpaddd      %%ymm1,%%ymm1,%%ymm0          \n"  // 6*near+2*far (1, hi)
+      "vpaddd      %%ymm6,%%ymm3,%%ymm2          \n"  // 3*near+far+8 (2, hi)
+      "vpaddd      %%ymm0,%%ymm1,%%ymm0          \n"  // 9*near+3*far (1, hi)
+      "vpaddd      %%ymm0,%%ymm2,%%ymm0          \n"  // 9 3 3 1 + 8 (1, hi)
+      "vpsrld      $4,%%ymm0,%%ymm0              \n"  // ^ div by 16 (1, hi)
+
+      "vpaddd      %%ymm3,%%ymm3,%%ymm2          \n"  // 6*near+2*far (2, hi)
+      "vpaddd      %%ymm6,%%ymm1,%%ymm1          \n"  // 3*near+far+8 (1, hi)
+      "vpaddd      %%ymm2,%%ymm3,%%ymm2          \n"  // 9*near+3*far (2, hi)
+      "vpaddd      %%ymm2,%%ymm1,%%ymm2          \n"  // 9 3 3 1 + 8 (2, hi)
+      "vpsrld      $4,%%ymm2,%%ymm2              \n"  // ^ div by 16 (2, hi)
+
+      "vpackusdw   %%ymm0,%%ymm4,%%ymm4          \n"
+      "vmovdqu     %%ymm4,(%1)                   \n"  // store above
+      "vpackusdw   %%ymm2,%%ymm5,%%ymm5          \n"
+      "vmovdqu     %%ymm5,(%1,%4,2)              \n"  // store below
+
+      "lea         0x10(%0),%0                   \n"
+      "lea         0x20(%1),%1                   \n"  // 4 uv to 8 uv
+      "sub         $0x8,%2                       \n"
+      "jg          1b                            \n"
+      "vzeroupper                                \n"
+      : "+r"(src_ptr),                // %0
+        "+r"(dst_ptr),                // %1
+        "+r"(dst_width)               // %2
+      : "r"((intptr_t)(src_stride)),  // %3
+        "r"((intptr_t)(dst_stride))   // %4
+      : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6");
+}
+#endif
+
 #endif  // defined(__x86_64__) || defined(__i386__)
 
 #ifdef __cplusplus
diff --git a/files/source/scale_lsx.cc b/files/source/scale_lsx.cc
new file mode 100644
index 00000000..bfe5e9fb
--- /dev/null
+++ b/files/source/scale_lsx.cc
@@ -0,0 +1,739 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Copyright (c) 2022 Loongson Technology Corporation Limited
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+
+#include "libyuv/scale_row.h"
+
+#if !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
+#include "libyuv/loongson_intrinsics.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define LOAD_DATA(_src, _in, _out)                                       \
+  {                                                                      \
+    int _tmp1, _tmp2, _tmp3, _tmp4;                                      \
+    DUP4_ARG2(__lsx_vpickve2gr_w, _in, 0, _in, 1, _in, 2, _in, 3, _tmp1, \
+              _tmp2, _tmp3, _tmp4);                                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp1], 0);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp2], 1);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp3], 2);                      \
+    _out = __lsx_vinsgr2vr_w(_out, _src[_tmp4], 3);                      \
+  }
+
+void ScaleARGBRowDown2_LSX(const uint8_t* src_argb,
+                           ptrdiff_t src_stride,
+                           uint8_t* dst_argb,
+                           int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  (void)src_stride;
+  __m128i src0, src1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    dst0 = __lsx_vpickod_w(src1, src0);
+    __lsx_vst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Linear_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  (void)src_stride;
+  __m128i src0, src1, tmp0, tmp1, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_argb, 0, src_argb, 16, src0, src1);
+    tmp0 = __lsx_vpickev_w(src1, src0);
+    tmp1 = __lsx_vpickod_w(src1, src0);
+    dst0 = __lsx_vavgr_bu(tmp1, tmp0);
+    __lsx_vst(dst0, dst_argb, 0);
+    src_argb += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDown2Box_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  const uint8_t* s = src_argb;
+  const uint8_t* t = src_argb + src_stride;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, tmp2, tmp3, dst0;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i shuff = {0x0703060205010400, 0x0F0B0E0A0D090C08};
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, s, 0, s, 16, src0, src1);
+    DUP2_ARG2(__lsx_vld, t, 0, t, 16, src2, src3);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff, src1, src1, shuff, src2, src2,
+              shuff, src3, src3, shuff, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vhaddw_hu_bu, tmp0, tmp0, tmp1, tmp1, tmp2, tmp2, tmp3,
+              tmp3, reg0, reg1, reg2, reg3);
+    DUP2_ARG2(__lsx_vsadd_hu, reg0, reg2, reg1, reg3, reg0, reg1);
+    dst0 = __lsx_vsrarni_b_h(reg1, reg0, 2);
+    __lsx_vst(dst0, dst_argb, 0);
+    s += 32;
+    t += 32;
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEven_LSX(const uint8_t* src_argb,
+                              ptrdiff_t src_stride,
+                              int32_t src_stepx,
+                              uint8_t* dst_argb,
+                              int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  int32_t stepx = src_stepx << 2;
+  (void)src_stride;
+  __m128i dst0, dst1, dst2, dst3;
+
+  for (x = 0; x < len; x++) {
+    dst0 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst1 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst2 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    dst3 = __lsx_vldrepl_w(src_argb, 0);
+    src_argb += stepx;
+    __lsx_vstelm_w(dst0, dst_argb, 0, 0);
+    __lsx_vstelm_w(dst1, dst_argb, 4, 0);
+    __lsx_vstelm_w(dst2, dst_argb, 8, 0);
+    __lsx_vstelm_w(dst3, dst_argb, 12, 0);
+    dst_argb += 16;
+  }
+}
+
+void ScaleARGBRowDownEvenBox_LSX(const uint8_t* src_argb,
+                                 ptrdiff_t src_stride,
+                                 int src_stepx,
+                                 uint8_t* dst_argb,
+                                 int dst_width) {
+  int x;
+  int len = dst_width / 4;
+  int32_t stepx = src_stepx * 4;
+  const uint8_t* next_argb = src_argb + src_stride;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, dst0;
+
+  for (x = 0; x < len; x++) {
+    tmp0 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp1 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp2 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp3 = __lsx_vldrepl_d(src_argb, 0);
+    src_argb += stepx;
+    tmp4 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp5 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp6 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    tmp7 = __lsx_vldrepl_d(next_argb, 0);
+    next_argb += stepx;
+    DUP4_ARG2(__lsx_vilvl_d, tmp1, tmp0, tmp3, tmp2, tmp5, tmp4, tmp7, tmp6,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpackev_w, tmp1, tmp0, tmp3, tmp2, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_w, tmp1, tmp0, tmp3, tmp2, tmp4, tmp5);
+    DUP2_ARG2(__lsx_vadd_h, reg0, tmp4, reg1, tmp5, reg0, reg1);
+    dst0 = __lsx_vsrarni_b_h(reg1, reg0, 2);
+    dst0 = __lsx_vshuf4i_b(dst0, 0xD8);
+    __lsx_vst(dst0, dst_argb, 0);
+    dst_argb += 16;
+  }
+}
+
+void ScaleRowDown2_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  __m128i src0, src1, src2, src3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Linear_LSX(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, dst0, dst1;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vpickod_b, src1, src0, src3, src2, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vavgr_bu, tmp0, tmp1, tmp2, tmp3, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown2Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  int len = dst_width / 32;
+  const uint8_t* src_nex = src_ptr + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i dst0, dst1;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG3(__lsx_vsrarni_b_h, tmp1, tmp0, 2, tmp3, tmp2, 2, dst0, dst1);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    src_ptr += 64;
+    src_nex += 64;
+    dst += 32;
+  }
+}
+
+void ScaleRowDown4_LSX(const uint8_t* src_ptr,
+                       ptrdiff_t src_stride,
+                       uint8_t* dst,
+                       int dst_width) {
+  int x;
+  int len = dst_width / 16;
+  __m128i src0, src1, src2, src3, tmp0, tmp1, dst0;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vpickev_b, src1, src0, src3, src2, tmp0, tmp1);
+    dst0 = __lsx_vpickod_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst, 0);
+    src_ptr += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown4Box_LSX(const uint8_t* src_ptr,
+                          ptrdiff_t src_stride,
+                          uint8_t* dst,
+                          int dst_width) {
+  int x;
+  int len = dst_width / 16;
+  const uint8_t* ptr1 = src_ptr + src_stride;
+  const uint8_t* ptr2 = ptr1 + src_stride;
+  const uint8_t* ptr3 = ptr2 + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, dst0;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, ptr1, 0, ptr1, 16, ptr1, 32, ptr1, 48, src4, src5,
+              src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vld, ptr2, 0, ptr2, 16, ptr2, 32, ptr2, 48, src0, src1,
+              src2, src3);
+    DUP4_ARG2(__lsx_vld, ptr3, 0, ptr3, 16, ptr3, 32, ptr3, 48, src4, src5,
+              src6, src7);
+    DUP4_ARG2(__lsx_vaddwev_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp0, tmp2, tmp4, tmp6);
+    DUP4_ARG2(__lsx_vaddwod_h_bu, src0, src4, src1, src5, src2, src6, src3,
+              src7, tmp1, tmp3, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_h, reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+              reg0, reg1, reg2, reg3);
+    DUP4_ARG2(__lsx_vhaddw_wu_hu, reg0, reg0, reg1, reg1, reg2, reg2, reg3,
+              reg3, reg0, reg1, reg2, reg3);
+    DUP2_ARG3(__lsx_vsrarni_h_w, reg1, reg0, 4, reg3, reg2, 4, tmp0, tmp1);
+    dst0 = __lsx_vpickev_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst, 0);
+    src_ptr += 64;
+    ptr1 += 64;
+    ptr2 += 64;
+    ptr3 += 64;
+    dst += 16;
+  }
+}
+
+void ScaleRowDown38_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x, len;
+  __m128i src0, src1, tmp0;
+  __m128i shuff = {0x13100E0B08060300, 0x000000001E1B1816};
+
+  assert(dst_width % 3 == 0);
+  len = dst_width / 12;
+  (void)src_stride;
+
+  for (x = 0; x < len; x++) {
+    DUP2_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src0, src1);
+    tmp0 = __lsx_vshuf_b(src1, src0, shuff);
+    __lsx_vstelm_d(tmp0, dst, 0, 0);
+    __lsx_vstelm_w(tmp0, dst, 8, 2);
+    src_ptr += 32;
+    dst += 12;
+  }
+}
+
+void ScaleRowDown38_2_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, len;
+  const uint8_t* src_nex = src_ptr + src_stride;
+  __m128i src0, src1, src2, src3, dst0;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3;
+  __m128i shuff = {0x0A08160604120200, 0x000000001E0E0C1A};
+  __m128i const_0x2AAA = __lsx_vreplgr2vr_h(0x2AAA);
+  __m128i const_0x4000 = __lsx_vreplgr2vr_w(0x4000);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  len = dst_width / 12;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_nex, 0, src_nex, 16, src0,
+              src1, src2, src3);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpickev_h, tmp2, tmp0, tmp3, tmp1, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_h, tmp1, tmp0, tmp3, tmp2, reg2, reg3);
+    tmp4 = __lsx_vpickev_w(reg3, reg2);
+    tmp5 = __lsx_vadd_h(reg0, reg1);
+    tmp6 = __lsx_vadd_h(tmp5, tmp4);
+    tmp7 = __lsx_vmuh_h(tmp6, const_0x2AAA);
+    tmp0 = __lsx_vpickod_w(reg3, reg2);
+    tmp1 = __lsx_vhaddw_wu_hu(tmp0, tmp0);
+    tmp2 = __lsx_vmul_w(tmp1, const_0x4000);
+    dst0 = __lsx_vshuf_b(tmp2, tmp7, shuff);
+    __lsx_vstelm_d(dst0, dst_ptr, 0, 0);
+    __lsx_vstelm_w(dst0, dst_ptr, 8, 2);
+    src_ptr += 32;
+    src_nex += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleRowDown38_3_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* dst_ptr,
+                              int dst_width) {
+  int x, len;
+  const uint8_t* ptr1 = src_ptr + src_stride;
+  const uint8_t* ptr2 = ptr1 + src_stride;
+  __m128i src0, src1, src2, src3, src4, src5;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, dst0;
+  __m128i zero = __lsx_vldi(0);
+  __m128i shuff = {0x0A08160604120200, 0x000000001E0E0C1A};
+  __m128i const_0x1C71 = __lsx_vreplgr2vr_h(0x1C71);
+  __m128i const_0x2AAA = __lsx_vreplgr2vr_w(0x2AAA);
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  len = dst_width / 12;
+
+  for (x = 0; x < len; x++) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, ptr1, 0, ptr1, 16, src0, src1,
+              src2, src3);
+    DUP2_ARG2(__lsx_vld, ptr2, 0, ptr2, 16, src4, src5);
+    DUP2_ARG2(__lsx_vaddwev_h_bu, src0, src2, src1, src3, tmp0, tmp2);
+    DUP2_ARG2(__lsx_vaddwod_h_bu, src0, src2, src1, src3, tmp1, tmp3);
+    DUP2_ARG2(__lsx_vpackev_b, zero, src4, zero, src5, tmp4, tmp6);
+    DUP2_ARG2(__lsx_vpackod_b, zero, src4, zero, src5, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vadd_h, tmp0, tmp4, tmp1, tmp5, tmp2, tmp6, tmp3, tmp7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG2(__lsx_vpickev_h, tmp2, tmp0, tmp3, tmp1, reg0, reg1);
+    DUP2_ARG2(__lsx_vpackod_h, tmp1, tmp0, tmp3, tmp2, reg2, reg3);
+    tmp4 = __lsx_vpickev_w(reg3, reg2);
+    tmp5 = __lsx_vadd_h(reg0, reg1);
+    tmp6 = __lsx_vadd_h(tmp5, tmp4);
+    tmp7 = __lsx_vmuh_h(tmp6, const_0x1C71);
+    tmp0 = __lsx_vpickod_w(reg3, reg2);
+    tmp1 = __lsx_vhaddw_wu_hu(tmp0, tmp0);
+    tmp2 = __lsx_vmul_w(tmp1, const_0x2AAA);
+    dst0 = __lsx_vshuf_b(tmp2, tmp7, shuff);
+    __lsx_vstelm_d(dst0, dst_ptr, 0, 0);
+    __lsx_vstelm_w(dst0, dst_ptr, 8, 2);
+    src_ptr += 32;
+    ptr1 += 32;
+    ptr2 += 32;
+    dst_ptr += 12;
+  }
+}
+
+void ScaleAddRow_LSX(const uint8_t* src_ptr, uint16_t* dst_ptr, int src_width) {
+  int x;
+  int len = src_width / 16;
+  __m128i src0, tmp0, tmp1, dst0, dst1;
+  __m128i zero = __lsx_vldi(0);
+
+  assert(src_width > 0);
+
+  for (x = 0; x < len; x++) {
+    src0 = __lsx_vld(src_ptr, 0);
+    DUP2_ARG2(__lsx_vld, dst_ptr, 0, dst_ptr, 16, dst0, dst1);
+    tmp0 = __lsx_vilvl_b(zero, src0);
+    tmp1 = __lsx_vilvh_b(zero, src0);
+    DUP2_ARG2(__lsx_vadd_h, dst0, tmp0, dst1, tmp1, dst0, dst1);
+    __lsx_vst(dst0, dst_ptr, 0);
+    __lsx_vst(dst1, dst_ptr, 16);
+    src_ptr += 16;
+    dst_ptr += 16;
+  }
+}
+
+void ScaleFilterCols_LSX(uint8_t* dst_ptr,
+                         const uint8_t* src_ptr,
+                         int dst_width,
+                         int x,
+                         int dx) {
+  int j;
+  int len = dst_width / 16;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec0, vec1, dst0;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const1 = __lsx_vreplgr2vr_w(0xFFFF);
+  __m128i const2 = __lsx_vreplgr2vr_w(0x40);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+
+  vec0 = __lsx_vmul_w(vec_dx, const_tmp);
+  vec1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, vec0);
+
+  for (j = 0; j < len; j++) {
+    tmp0 = __lsx_vsrai_w(vec_x, 16);
+    tmp4 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp1 = __lsx_vsrai_w(vec_x, 16);
+    tmp5 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp2 = __lsx_vsrai_w(vec_x, 16);
+    tmp6 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp3 = __lsx_vsrai_w(vec_x, 16);
+    tmp7 = __lsx_vand_v(vec_x, const1);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    DUP4_ARG2(__lsx_vsrai_w, tmp4, 9, tmp5, 9, tmp6, 9, tmp7, 9, tmp4, tmp5,
+              tmp6, tmp7);
+    LOAD_DATA(src_ptr, tmp0, reg0);
+    LOAD_DATA(src_ptr, tmp1, reg1);
+    LOAD_DATA(src_ptr, tmp2, reg2);
+    LOAD_DATA(src_ptr, tmp3, reg3);
+    DUP4_ARG2(__lsx_vaddi_wu, tmp0, 1, tmp1, 1, tmp2, 1, tmp3, 1, tmp0, tmp1,
+              tmp2, tmp3);
+    LOAD_DATA(src_ptr, tmp0, reg4);
+    LOAD_DATA(src_ptr, tmp1, reg5);
+    LOAD_DATA(src_ptr, tmp2, reg6);
+    LOAD_DATA(src_ptr, tmp3, reg7);
+    DUP4_ARG2(__lsx_vsub_w, reg4, reg0, reg5, reg1, reg6, reg2, reg7, reg3,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vmul_w, reg4, tmp4, reg5, tmp5, reg6, tmp6, reg7, tmp7,
+              reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_w, reg4, const2, reg5, const2, reg6, const2, reg7,
+              const2, reg4, reg5, reg6, reg7);
+    DUP4_ARG2(__lsx_vsrai_w, reg4, 7, reg5, 7, reg6, 7, reg7, 7, reg4, reg5,
+              reg6, reg7);
+    DUP4_ARG2(__lsx_vadd_w, reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7,
+              reg0, reg1, reg2, reg3);
+    DUP2_ARG2(__lsx_vpickev_h, reg1, reg0, reg3, reg2, tmp0, tmp1);
+    dst0 = __lsx_vpickev_b(tmp1, tmp0);
+    __lsx_vst(dst0, dst_ptr, 0);
+    dst_ptr += 16;
+  }
+}
+
+void ScaleARGBCols_LSX(uint8_t* dst_argb,
+                       const uint8_t* src_argb,
+                       int dst_width,
+                       int x,
+                       int dx) {
+  const uint32_t* src = (const uint32_t*)src_argb;
+  uint32_t* dst = (uint32_t*)dst_argb;
+  int j;
+  int len = dst_width / 4;
+  __m128i tmp0, tmp1, tmp2, dst0;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+
+  tmp0 = __lsx_vmul_w(vec_dx, const_tmp);
+  tmp1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, tmp0);
+
+  for (j = 0; j < len; j++) {
+    tmp2 = __lsx_vsrai_w(vec_x, 16);
+    vec_x = __lsx_vadd_w(vec_x, tmp1);
+    LOAD_DATA(src, tmp2, dst0);
+    __lsx_vst(dst0, dst, 0);
+    dst += 4;
+  }
+}
+
+void ScaleARGBFilterCols_LSX(uint8_t* dst_argb,
+                             const uint8_t* src_argb,
+                             int dst_width,
+                             int x,
+                             int dx) {
+  const uint32_t* src = (const uint32_t*)src_argb;
+  int j;
+  int len = dst_width / 8;
+  __m128i src0, src1, src2, src3;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  __m128i reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  __m128i vec0, vec1, dst0, dst1;
+  __m128i vec_x = __lsx_vreplgr2vr_w(x);
+  __m128i vec_dx = __lsx_vreplgr2vr_w(dx);
+  __m128i const_tmp = {0x0000000100000000, 0x0000000300000002};
+  __m128i const_7f = __lsx_vldi(0x7F);
+
+  vec0 = __lsx_vmul_w(vec_dx, const_tmp);
+  vec1 = __lsx_vslli_w(vec_dx, 2);
+  vec_x = __lsx_vadd_w(vec_x, vec0);
+
+  for (j = 0; j < len; j++) {
+    tmp0 = __lsx_vsrai_w(vec_x, 16);
+    reg0 = __lsx_vsrai_w(vec_x, 9);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    tmp1 = __lsx_vsrai_w(vec_x, 16);
+    reg1 = __lsx_vsrai_w(vec_x, 9);
+    vec_x = __lsx_vadd_w(vec_x, vec1);
+    DUP2_ARG2(__lsx_vand_v, reg0, const_7f, reg1, const_7f, reg0, reg1);
+    DUP2_ARG2(__lsx_vshuf4i_b, reg0, 0, reg1, 0, reg0, reg1);
+    DUP2_ARG2(__lsx_vxor_v, reg0, const_7f, reg1, const_7f, reg2, reg3);
+    DUP2_ARG2(__lsx_vilvl_b, reg0, reg2, reg1, reg3, reg4, reg6);
+    DUP2_ARG2(__lsx_vilvh_b, reg0, reg2, reg1, reg3, reg5, reg7);
+    LOAD_DATA(src, tmp0, src0);
+    LOAD_DATA(src, tmp1, src1);
+    DUP2_ARG2(__lsx_vaddi_wu, tmp0, 1, tmp1, 1, tmp0, tmp1);
+    LOAD_DATA(src, tmp0, src2);
+    LOAD_DATA(src, tmp1, src3);
+    DUP2_ARG2(__lsx_vilvl_b, src2, src0, src3, src1, tmp4, tmp6);
+    DUP2_ARG2(__lsx_vilvh_b, src2, src0, src3, src1, tmp5, tmp7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, reg4, tmp5, reg5, tmp6, reg6, tmp7, reg7,
+              tmp0, tmp1, tmp2, tmp3);
+    DUP2_ARG3(__lsx_vsrani_b_h, tmp1, tmp0, 7, tmp3, tmp2, 7, dst0, dst1);
+    __lsx_vst(dst0, dst_argb, 0);
+    __lsx_vst(dst1, dst_argb, 16);
+    dst_argb += 32;
+  }
+}
+
+void ScaleRowDown34_LSX(const uint8_t* src_ptr,
+                        ptrdiff_t src_stride,
+                        uint8_t* dst,
+                        int dst_width) {
+  int x;
+  (void)src_stride;
+  __m128i src0, src1, src2, src3;
+  __m128i dst0, dst1, dst2;
+  __m128i shuff0 = {0x0908070504030100, 0x141311100F0D0C0B};
+  __m128i shuff1 = {0x0F0D0C0B09080705, 0x1918171514131110};
+  __m128i shuff2 = {0x141311100F0D0C0B, 0x1F1D1C1B19181715};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP2_ARG3(__lsx_vshuf_b, src1, src0, shuff0, src2, src1, shuff1, dst0,
+              dst1);
+    dst2 = __lsx_vshuf_b(src3, src2, shuff2);
+    __lsx_vst(dst0, dst, 0);
+    __lsx_vst(dst1, dst, 16);
+    __lsx_vst(dst2, dst, 32);
+    src_ptr += 64;
+    dst += 48;
+  }
+}
+
+void ScaleRowDown34_0_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* src_nex = src_ptr + src_stride;
+  int x;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
+  __m128i tmp10, tmp11, dst0, dst1, dst2;
+  __m128i const0 = {0x0103030101010103, 0x0101010303010101};
+  __m128i const1 = {0x0301010101030301, 0x0103030101010103};
+  __m128i const2 = {0x0101010303010101, 0x0301010101030301};
+  __m128i shuff0 = {0x0504030202010100, 0x0A09090807060605};
+  __m128i shuff1 = {0x0F0E0E0D0D0C0B0A, 0x1514131212111110};
+  __m128i shuff2 = {0x0A09090807060605, 0x0F0E0E0D0D0C0B0A};
+  __m128i shift0 = {0x0002000200010002, 0x0001000200020001};
+  __m128i shift1 = {0x0002000100020002, 0x0002000200010002};
+  __m128i shift2 = {0x0001000200020001, 0x0002000100020002};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff0, src1, src0, shuff1, src1, src1,
+              shuff2, src2, src2, shuff0, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG3(__lsx_vshuf_b, src3, src2, shuff1, src3, src3, shuff2, src4, src4,
+              shuff0, src5, src4, shuff1, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vshuf_b, src5, src5, shuff2, src6, src6, shuff0, src7, src6,
+              shuff1, src7, src7, shuff2, tmp8, tmp9, tmp10, tmp11);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp0, const0, tmp1, const1, tmp2, const2, tmp3,
+              const0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, const1, tmp5, const2, tmp6, const0, tmp7,
+              const1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp8, const2, tmp9, const0, tmp10, const1, tmp11,
+              const2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrar_h, src0, shift0, src1, shift1, src2, shift2, src3,
+              shift0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vsrar_h, src4, shift1, src5, shift2, src6, shift0, src7,
+              shift1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vsrar_h, tmp0, shift2, tmp1, shift0, tmp2, shift1, tmp3,
+              shift2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vslli_h, src0, 1, src1, 1, src2, 1, src3, 1, tmp5, tmp6,
+              tmp7, tmp8);
+    DUP2_ARG2(__lsx_vslli_h, src4, 1, src5, 1, tmp9, tmp10);
+    DUP4_ARG2(__lsx_vadd_h, src0, tmp5, src1, tmp6, src2, tmp7, src3, tmp8,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp9, src5, tmp10, src4, src5);
+    DUP4_ARG2(__lsx_vadd_h, src0, src6, src1, src7, src2, tmp0, src3, tmp1,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp2, src5, tmp3, src4, src5);
+    DUP2_ARG3(__lsx_vsrarni_b_h, src1, src0, 2, src3, src2, 2, dst0, dst1);
+    dst2 = __lsx_vsrarni_b_h(src5, src4, 2);
+    __lsx_vst(dst0, d, 0);
+    __lsx_vst(dst1, d, 16);
+    __lsx_vst(dst2, d, 32);
+    src_ptr += 64;
+    src_nex += 64;
+    d += 48;
+  }
+}
+
+void ScaleRowDown34_1_Box_LSX(const uint8_t* src_ptr,
+                              ptrdiff_t src_stride,
+                              uint8_t* d,
+                              int dst_width) {
+  const uint8_t* src_nex = src_ptr + src_stride;
+  int x;
+  __m128i src0, src1, src2, src3, src4, src5, src6, src7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
+  __m128i tmp10, tmp11, dst0, dst1, dst2;
+  __m128i const0 = {0x0103030101010103, 0x0101010303010101};
+  __m128i const1 = {0x0301010101030301, 0x0103030101010103};
+  __m128i const2 = {0x0101010303010101, 0x0301010101030301};
+  __m128i shuff0 = {0x0504030202010100, 0x0A09090807060605};
+  __m128i shuff1 = {0x0F0E0E0D0D0C0B0A, 0x1514131212111110};
+  __m128i shuff2 = {0x0A09090807060605, 0x0F0E0E0D0D0C0B0A};
+  __m128i shift0 = {0x0002000200010002, 0x0001000200020001};
+  __m128i shift1 = {0x0002000100020002, 0x0002000200010002};
+  __m128i shift2 = {0x0001000200020001, 0x0002000100020002};
+
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+
+  for (x = 0; x < dst_width; x += 48) {
+    DUP4_ARG2(__lsx_vld, src_ptr, 0, src_ptr, 16, src_ptr, 32, src_ptr, 48,
+              src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vld, src_nex, 0, src_nex, 16, src_nex, 32, src_nex, 48,
+              src4, src5, src6, src7);
+    DUP4_ARG3(__lsx_vshuf_b, src0, src0, shuff0, src1, src0, shuff1, src1, src1,
+              shuff2, src2, src2, shuff0, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG3(__lsx_vshuf_b, src3, src2, shuff1, src3, src3, shuff2, src4, src4,
+              shuff0, src5, src4, shuff1, tmp4, tmp5, tmp6, tmp7);
+    DUP4_ARG3(__lsx_vshuf_b, src5, src5, shuff2, src6, src6, shuff0, src7, src6,
+              shuff1, src7, src7, shuff2, tmp8, tmp9, tmp10, tmp11);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp0, const0, tmp1, const1, tmp2, const2, tmp3,
+              const0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp4, const1, tmp5, const2, tmp6, const0, tmp7,
+              const1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vdp2_h_bu, tmp8, const2, tmp9, const0, tmp10, const1, tmp11,
+              const2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vsrar_h, src0, shift0, src1, shift1, src2, shift2, src3,
+              shift0, src0, src1, src2, src3);
+    DUP4_ARG2(__lsx_vsrar_h, src4, shift1, src5, shift2, src6, shift0, src7,
+              shift1, src4, src5, src6, src7);
+    DUP4_ARG2(__lsx_vsrar_h, tmp0, shift2, tmp1, shift0, tmp2, shift1, tmp3,
+              shift2, tmp0, tmp1, tmp2, tmp3);
+    DUP4_ARG2(__lsx_vadd_h, src0, src6, src1, src7, src2, tmp0, src3, tmp1,
+              src0, src1, src2, src3);
+    DUP2_ARG2(__lsx_vadd_h, src4, tmp2, src5, tmp3, src4, src5);
+    DUP2_ARG3(__lsx_vsrarni_b_h, src1, src0, 1, src3, src2, 1, dst0, dst1);
+    dst2 = __lsx_vsrarni_b_h(src5, src4, 1);
+    __lsx_vst(dst0, d, 0);
+    __lsx_vst(dst1, d, 16);
+    __lsx_vst(dst2, d, 32);
+    src_ptr += 64;
+    src_nex += 64;
+    d += 48;
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // !defined(LIBYUV_DISABLE_LSX) && defined(__loongarch_sx)
diff --git a/files/source/scale_mmi.cc b/files/source/scale_mmi.cc
index 990463c2..1226ef3e 100644
--- a/files/source/scale_mmi.cc
+++ b/files/source/scale_mmi.cc
@@ -1103,6 +1103,61 @@ void ScaleRowUp2_16_MMI(const uint16_t* src_ptr,
       : "memory");
 }
 
+void ScaleRowDown34_MMI(const uint8_t* src_ptr,
+                      ptrdiff_t src_stride,
+                      uint8_t* dst,
+                      int dst_width) {
+  (void)src_stride;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  uint64_t src[2];
+  uint64_t tmp[2];
+  __asm__ volatile (
+    "1:                                                           \n\t"
+    "gsldlc1    %[src0],         0x07(%[src_ptr])                 \n\t"
+    "gsldrc1    %[src0],         0x00(%[src_ptr])                 \n\t"
+    "gsldlc1    %[src1],         0x0f(%[src_ptr])                 \n\t"
+    "gsldrc1    %[src1],         0x08(%[src_ptr])                 \n\t"
+    "and        %[tmp1],         %[src0],        %[mask1]         \n\t"
+    "psrlw      %[tmp0],         %[src0],        %[rmov]          \n\t"
+    "psllw      %[tmp0],         %[tmp0],        %[lmov1]         \n\t"
+    "or         %[src0],         %[tmp0],        %[tmp1]          \n\t"
+    "punpckhwd  %[tmp0],         %[src0],        %[src0]          \n\t"
+    "psllw      %[tmp1],         %[tmp0],        %[rmov]          \n\t"
+    "or         %[src0],         %[src0],        %[tmp1]          \n\t"
+    "psrlw      %[tmp0],         %[tmp0],        %[rmov8]         \n\t"
+    "pextrh     %[tmp0],         %[tmp0],        %[zero]          \n\t"
+    "pinsrh_2   %[src0],         %[src0],        %[tmp0]          \n\t"
+    "pextrh     %[tmp0],         %[src1],        %[zero]          \n\t"
+    "pinsrh_3   %[src0],         %[src0],        %[tmp0]          \n\t"
+
+    "punpckhwd  %[tmp0],         %[src1],        %[src1]          \n\t"
+    "pextrh     %[tmp1],         %[tmp0],        %[zero]          \n\t"
+    "psrlw      %[src1],         %[src1],        %[rmov]          \n\t"
+    "psllw      %[tmp1],         %[tmp1],        %[rmov8]         \n\t"
+    "or         %[src1],         %[src1],        %[tmp1]          \n\t"
+    "and        %[tmp0],         %[tmp0],        %[mask2]         \n\t"
+    "or         %[src1],         %[src1],        %[tmp0]          \n\t"
+
+    "gssdlc1    %[src0],         0x07(%[dst_ptr])                 \n\t"
+    "gssdrc1    %[src0],         0x00(%[dst_ptr])                 \n\t"
+    "gsswlc1    %[src1],         0x0b(%[dst_ptr])                 \n\t"
+    "gsswrc1    %[src1],         0x08(%[dst_ptr])                 \n\t"
+
+    "daddiu     %[src_ptr],      %[src_ptr],     0x10             \n\t"
+    "daddi      %[width],        %[width],      -0x0c             \n\t"
+    "daddiu     %[dst_ptr],      %[dst_ptr],     0x0c             \n\t"
+    "bnez       %[width],        1b                               \n\t"
+
+    : [src0]"=&f"(src[0]),              [src1]"=&f"(src[1]),
+      [tmp0]"=&f"(tmp[0]),              [tmp1]"=&f"(tmp[1])
+    : [src_ptr]"r"(src_ptr),            [dst_ptr]"r"(dst),
+      [lmov]"f"(0xc),                   [rmov]"f"(0x18),
+      [mask1]"f"(0xffff0000ffff),       [rmov8]"f"(0x8),
+      [zero]"f"(0x0),                   [mask2]"f"(0xff000000),
+      [width]"r"(dst_width),            [lmov1]"f"(0x10)
+    : "memory"
+  );
+}
 // clang-format on
 
 #endif  // !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
diff --git a/files/source/scale_neon.cc b/files/source/scale_neon.cc
index 366b155b..6a0d6e1b 100644
--- a/files/source/scale_neon.cc
+++ b/files/source/scale_neon.cc
@@ -31,10 +31,10 @@ void ScaleRowDown2_NEON(const uint8_t* src_ptr,
   asm volatile(
       "1:                                        \n"
       // load even pixels into q0, odd into q1
-      "vld2.8     {q0, q1}, [%0]!                \n"
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop
-      "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
-      "bgt        1b                             \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vst1.8      {q1}, [%1]!                   \n"  // store odd pixels
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -51,11 +51,11 @@ void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "vld2.8     {q0, q1}, [%0]!                \n"  // load 32 pixels
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop
-      "vrhadd.u8  q0, q0, q1                     \n"  // rounding half add
-      "vst1.8     {q0}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "vld2.8      {q0, q1}, [%0]!               \n"  // load 32 pixels
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "vrhadd.u8   q0, q0, q1                    \n"  // rounding half add
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -71,21 +71,21 @@ void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
                            int dst_width) {
   asm volatile(
       // change the stride to row 2 pointer
-      "add        %1, %0                         \n"
-      "1:                                        \n"
-      "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
-      "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
-      "subs       %3, %3, #16                    \n"  // 16 processed per loop
-      "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
-      "vpaddl.u8  q1, q1                         \n"
-      "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent +
+      "add         %1, %0                        \n"
+      "1:                                        \n"
+      "vld1.8      {q0, q1}, [%0]!               \n"  // load row 1 and post inc
+      "vld1.8      {q2, q3}, [%1]!               \n"  // load row 2 and post inc
+      "subs        %3, %3, #16                   \n"  // 16 processed per loop
+      "vpaddl.u8   q0, q0                        \n"  // row 1 add adjacent
+      "vpaddl.u8   q1, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"  // row 2 add adjacent +
                                                       // row1
-      "vpadal.u8  q1, q3                         \n"
-      "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and
+      "vpadal.u8   q1, q3                        \n"
+      "vrshrn.u16  d0, q0, #2                    \n"  // downshift, round and
                                                       // pack
-      "vrshrn.u16 d1, q1, #2                     \n"
-      "vst1.8     {q0}, [%2]!                    \n"
-      "bgt        1b                             \n"
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vst1.8      {q0}, [%2]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -102,10 +102,10 @@ void ScaleRowDown4_NEON(const uint8_t* src_ptr,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // src line 0
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop
-      "vst1.8     {d2}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vst1.8      {d2}, [%1]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -122,20 +122,20 @@ void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
   const uint8_t* src_ptr3 = src_ptr + src_stride * 3;
   asm volatile(
       "1:                                        \n"
-      "vld1.8     {q0}, [%0]!                    \n"  // load up 16x4
-      "vld1.8     {q1}, [%3]!                    \n"
-      "vld1.8     {q2}, [%4]!                    \n"
-      "vld1.8     {q3}, [%5]!                    \n"
-      "subs       %2, %2, #4                     \n"
-      "vpaddl.u8  q0, q0                         \n"
-      "vpadal.u8  q0, q1                         \n"
-      "vpadal.u8  q0, q2                         \n"
-      "vpadal.u8  q0, q3                         \n"
-      "vpaddl.u16 q0, q0                         \n"
-      "vrshrn.u32 d0, q0, #4                     \n"  // divide by 16 w/rounding
-      "vmovn.u16  d0, q0                         \n"
-      "vst1.32    {d0[0]}, [%1]!                 \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q0}, [%0]!                   \n"  // load up 16x4
+      "vld1.8      {q1}, [%3]!                   \n"
+      "vld1.8      {q2}, [%4]!                   \n"
+      "vld1.8      {q3}, [%5]!                   \n"
+      "subs        %2, %2, #4                    \n"
+      "vpaddl.u8   q0, q0                        \n"
+      "vpadal.u8   q0, q1                        \n"
+      "vpadal.u8   q0, q2                        \n"
+      "vpadal.u8   q0, q3                        \n"
+      "vpaddl.u16  q0, q0                        \n"
+      "vrshrn.u32  d0, q0, #4                    \n"  // divide by 16 w/rounding
+      "vmovn.u16   d0, q0                        \n"
+      "vst1.32     {d0[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),    // %0
         "+r"(dst_ptr),    // %1
         "+r"(dst_width),  // %2
@@ -156,11 +156,11 @@ void ScaleRowDown34_NEON(const uint8_t* src_ptr,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // src line 0
-      "subs       %2, %2, #24                    \n"
-      "vmov       d2, d3                         \n"  // order d0, d1, d2
-      "vst3.8     {d0, d1, d2}, [%1]!            \n"
-      "bgt        1b                             \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "subs        %2, %2, #24                   \n"
+      "vmov        d2, d3                        \n"  // order d0, d1, d2
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -173,49 +173,49 @@ void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int dst_width) {
   asm volatile(
-      "vmov.u8    d24, #3                        \n"
-      "add        %3, %0                         \n"
+      "vmov.u8     d24, #3                       \n"
+      "add         %3, %0                        \n"
       "1:                                        \n"
-      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"  // src line 0
-      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"  // src line 1
-      "subs         %2, %2, #24                  \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"  // src line 1
+      "subs        %2, %2, #24                   \n"
 
       // filter src line 0 with src line 1
       // expand chars to shorts to allow for room
       // when adding lines together
-      "vmovl.u8     q8, d4                       \n"
-      "vmovl.u8     q9, d5                       \n"
-      "vmovl.u8     q10, d6                      \n"
-      "vmovl.u8     q11, d7                      \n"
+      "vmovl.u8    q8, d4                        \n"
+      "vmovl.u8    q9, d5                        \n"
+      "vmovl.u8    q10, d6                       \n"
+      "vmovl.u8    q11, d7                       \n"
 
       // 3 * line_0 + line_1
-      "vmlal.u8     q8, d0, d24                  \n"
-      "vmlal.u8     q9, d1, d24                  \n"
-      "vmlal.u8     q10, d2, d24                 \n"
-      "vmlal.u8     q11, d3, d24                 \n"
+      "vmlal.u8    q8, d0, d24                   \n"
+      "vmlal.u8    q9, d1, d24                   \n"
+      "vmlal.u8    q10, d2, d24                  \n"
+      "vmlal.u8    q11, d3, d24                  \n"
 
-      // (3 * line_0 + line_1) >> 2
-      "vqrshrn.u16  d0, q8, #2                   \n"
-      "vqrshrn.u16  d1, q9, #2                   \n"
-      "vqrshrn.u16  d2, q10, #2                  \n"
-      "vqrshrn.u16  d3, q11, #2                  \n"
+      // (3 * line_0 + line_1 + 2) >> 2
+      "vqrshrn.u16 d0, q8, #2                    \n"
+      "vqrshrn.u16 d1, q9, #2                    \n"
+      "vqrshrn.u16 d2, q10, #2                   \n"
+      "vqrshrn.u16 d3, q11, #2                   \n"
 
-      // a0 = (src[0] * 3 + s[1] * 1) >> 2
-      "vmovl.u8     q8, d1                       \n"
-      "vmlal.u8     q8, d0, d24                  \n"
-      "vqrshrn.u16  d0, q8, #2                   \n"
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "vmovl.u8    q8, d1                        \n"
+      "vmlal.u8    q8, d0, d24                   \n"
+      "vqrshrn.u16 d0, q8, #2                    \n"
 
-      // a1 = (src[1] * 1 + s[2] * 1) >> 1
-      "vrhadd.u8    d1, d1, d2                   \n"
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "vrhadd.u8   d1, d1, d2                    \n"
 
-      // a2 = (src[2] * 1 + s[3] * 3) >> 2
-      "vmovl.u8     q8, d2                       \n"
-      "vmlal.u8     q8, d3, d24                  \n"
-      "vqrshrn.u16  d2, q8, #2                   \n"
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "vmovl.u8    q8, d2                        \n"
+      "vmlal.u8    q8, d3, d24                   \n"
+      "vqrshrn.u16 d2, q8, #2                    \n"
 
-      "vst3.8       {d0, d1, d2}, [%1]!          \n"
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
 
-      "bgt          1b                           \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),    // %0
         "+r"(dst_ptr),    // %1
         "+r"(dst_width),  // %2
@@ -230,31 +230,31 @@ void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int dst_width) {
   asm volatile(
-      "vmov.u8    d24, #3                        \n"
-      "add        %3, %0                         \n"
+      "vmov.u8     d24, #3                       \n"
+      "add         %3, %0                        \n"
       "1:                                        \n"
-      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"  // src line 0
-      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"  // src line 1
-      "subs         %2, %2, #24                  \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"  // src line 0
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"  // src line 1
+      "subs        %2, %2, #24                   \n"
       // average src line 0 with src line 1
-      "vrhadd.u8    q0, q0, q2                   \n"
-      "vrhadd.u8    q1, q1, q3                   \n"
+      "vrhadd.u8   q0, q0, q2                    \n"
+      "vrhadd.u8   q1, q1, q3                    \n"
 
-      // a0 = (src[0] * 3 + s[1] * 1) >> 2
-      "vmovl.u8     q3, d1                       \n"
-      "vmlal.u8     q3, d0, d24                  \n"
-      "vqrshrn.u16  d0, q3, #2                   \n"
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "vmovl.u8    q3, d1                        \n"
+      "vmlal.u8    q3, d0, d24                   \n"
+      "vqrshrn.u16 d0, q3, #2                    \n"
 
-      // a1 = (src[1] * 1 + s[2] * 1) >> 1
-      "vrhadd.u8    d1, d1, d2                   \n"
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "vrhadd.u8   d1, d1, d2                    \n"
 
-      // a2 = (src[2] * 1 + s[3] * 3) >> 2
-      "vmovl.u8     q3, d2                       \n"
-      "vmlal.u8     q3, d3, d24                  \n"
-      "vqrshrn.u16  d2, q3, #2                   \n"
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "vmovl.u8    q3, d2                        \n"
+      "vmlal.u8    q3, d3, d24                   \n"
+      "vqrshrn.u16 d2, q3, #2                    \n"
 
-      "vst3.8       {d0, d1, d2}, [%1]!          \n"
-      "bgt          1b                           \n"
+      "vst3.8      {d0, d1, d2}, [%1]!           \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),    // %0
         "+r"(dst_ptr),    // %1
         "+r"(dst_width),  // %2
@@ -282,15 +282,15 @@ void ScaleRowDown38_NEON(const uint8_t* src_ptr,
                          int dst_width) {
   (void)src_stride;
   asm volatile(
-      "vld1.8     {q3}, [%3]                     \n"
-      "1:                                        \n"
-      "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
-      "subs       %2, %2, #12                    \n"
-      "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
-      "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
-      "vst1.8     {d4}, [%1]!                    \n"
-      "vst1.32    {d5[0]}, [%1]!                 \n"
-      "bgt        1b                             \n"
+      "vld1.8      {q3}, [%3]                    \n"
+      "1:                                        \n"
+      "vld1.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "subs        %2, %2, #12                   \n"
+      "vtbl.u8     d4, {d0, d1, d2, d3}, d6      \n"
+      "vtbl.u8     d5, {d0, d1, d2, d3}, d7      \n"
+      "vst1.8      {d4}, [%1]!                   \n"
+      "vst1.32     {d5[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -306,57 +306,57 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
   const uint8_t* src_ptr1 = src_ptr + src_stride * 2;
 
   asm volatile(
-      "vld1.16    {q13}, [%5]                    \n"
-      "vld1.8     {q14}, [%6]                    \n"
-      "vld1.8     {q15}, [%7]                    \n"
-      "add        %3, %0                         \n"
+      "vld1.16     {q13}, [%5]                   \n"
+      "vld1.8      {q14}, [%6]                   \n"
+      "vld1.8      {q15}, [%7]                   \n"
+      "add         %3, %0                        \n"
       "1:                                        \n"
 
       // d0 = 00 40 01 41 02 42 03 43
       // d1 = 10 50 11 51 12 52 13 53
       // d2 = 20 60 21 61 22 62 23 63
       // d3 = 30 70 31 71 32 72 33 73
-      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
-      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
-      "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
-      "subs         %2, %2, #12                  \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"
+      "vld4.8      {d16, d17, d18, d19}, [%4]!   \n"
+      "subs        %2, %2, #12                   \n"
 
       // Shuffle the input data around to get align the data
       //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
       // d0 = 00 10 01 11 02 12 03 13
       // d1 = 40 50 41 51 42 52 43 53
-      "vtrn.u8      d0, d1                       \n"
-      "vtrn.u8      d4, d5                       \n"
-      "vtrn.u8      d16, d17                     \n"
+      "vtrn.u8     d0, d1                        \n"
+      "vtrn.u8     d4, d5                        \n"
+      "vtrn.u8     d16, d17                      \n"
 
       // d2 = 20 30 21 31 22 32 23 33
       // d3 = 60 70 61 71 62 72 63 73
-      "vtrn.u8      d2, d3                       \n"
-      "vtrn.u8      d6, d7                       \n"
-      "vtrn.u8      d18, d19                     \n"
+      "vtrn.u8     d2, d3                        \n"
+      "vtrn.u8     d6, d7                        \n"
+      "vtrn.u8     d18, d19                      \n"
 
       // d0 = 00+10 01+11 02+12 03+13
       // d2 = 40+50 41+51 42+52 43+53
-      "vpaddl.u8    q0, q0                       \n"
-      "vpaddl.u8    q2, q2                       \n"
-      "vpaddl.u8    q8, q8                       \n"
+      "vpaddl.u8   q0, q0                        \n"
+      "vpaddl.u8   q2, q2                        \n"
+      "vpaddl.u8   q8, q8                        \n"
 
       // d3 = 60+70 61+71 62+72 63+73
-      "vpaddl.u8    d3, d3                       \n"
-      "vpaddl.u8    d7, d7                       \n"
-      "vpaddl.u8    d19, d19                     \n"
+      "vpaddl.u8   d3, d3                        \n"
+      "vpaddl.u8   d7, d7                        \n"
+      "vpaddl.u8   d19, d19                      \n"
 
       // combine source lines
-      "vadd.u16     q0, q2                       \n"
-      "vadd.u16     q0, q8                       \n"
-      "vadd.u16     d4, d3, d7                   \n"
-      "vadd.u16     d4, d19                      \n"
+      "vadd.u16    q0, q2                        \n"
+      "vadd.u16    q0, q8                        \n"
+      "vadd.u16    d4, d3, d7                    \n"
+      "vadd.u16    d4, d19                       \n"
 
       // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
       //             + s[6 + st * 1] + s[7 + st * 1]
       //             + s[6 + st * 2] + s[7 + st * 2]) / 6
       "vqrdmulh.s16 q2, q2, q13                  \n"
-      "vmovn.u16    d4, q2                       \n"
+      "vmovn.u16   d4, q2                        \n"
 
       // Shuffle 2,3 reg around so that 2 can be added to the
       //  0,1 reg and 3 can be added to the 4,5 reg. This
@@ -364,24 +364,24 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
       //  registers are already expanded. Then do transposes
       //  to get aligned.
       // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-      "vmovl.u8     q1, d2                       \n"
-      "vmovl.u8     q3, d6                       \n"
-      "vmovl.u8     q9, d18                      \n"
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q3, d6                        \n"
+      "vmovl.u8    q9, d18                       \n"
 
       // combine source lines
-      "vadd.u16     q1, q3                       \n"
-      "vadd.u16     q1, q9                       \n"
+      "vadd.u16    q1, q3                        \n"
+      "vadd.u16    q1, q9                        \n"
 
       // d4 = xx 20 xx 30 xx 22 xx 32
       // d5 = xx 21 xx 31 xx 23 xx 33
-      "vtrn.u32     d2, d3                       \n"
+      "vtrn.u32    d2, d3                        \n"
 
       // d4 = xx 20 xx 21 xx 22 xx 23
       // d5 = xx 30 xx 31 xx 32 xx 33
-      "vtrn.u16     d2, d3                       \n"
+      "vtrn.u16    d2, d3                        \n"
 
       // 0+1+2, 3+4+5
-      "vadd.u16     q0, q1                       \n"
+      "vadd.u16    q0, q1                        \n"
 
       // Need to divide, but can't downshift as the the value
       //  isn't a power of 2. So multiply by 65536 / n
@@ -390,14 +390,14 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
 
       // Align for table lookup, vtbl requires registers to
       //  be adjacent
-      "vmov.u8      d2, d4                       \n"
+      "vmov.u8     d2, d4                        \n"
 
-      "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-      "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+      "vtbl.u8     d3, {d0, d1, d2}, d28         \n"
+      "vtbl.u8     d4, {d0, d1, d2}, d29         \n"
 
-      "vst1.8       {d3}, [%1]!                  \n"
-      "vst1.32      {d4[0]}, [%1]!               \n"
-      "bgt          1b                           \n"
+      "vst1.8      {d3}, [%1]!                   \n"
+      "vst1.32     {d4[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),       // %0
         "+r"(dst_ptr),       // %1
         "+r"(dst_width),     // %2
@@ -416,46 +416,46 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int dst_width) {
   asm volatile(
-      "vld1.16    {q13}, [%4]                    \n"
-      "vld1.8     {q14}, [%5]                    \n"
-      "add        %3, %0                         \n"
+      "vld1.16     {q13}, [%4]                   \n"
+      "vld1.8      {q14}, [%5]                   \n"
+      "add         %3, %0                        \n"
       "1:                                        \n"
 
       // d0 = 00 40 01 41 02 42 03 43
       // d1 = 10 50 11 51 12 52 13 53
       // d2 = 20 60 21 61 22 62 23 63
       // d3 = 30 70 31 71 32 72 33 73
-      "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
-      "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
-      "subs         %2, %2, #12                  \n"
+      "vld4.8      {d0, d1, d2, d3}, [%0]!       \n"
+      "vld4.8      {d4, d5, d6, d7}, [%3]!       \n"
+      "subs        %2, %2, #12                   \n"
 
       // Shuffle the input data around to get align the data
       //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
       // d0 = 00 10 01 11 02 12 03 13
       // d1 = 40 50 41 51 42 52 43 53
-      "vtrn.u8      d0, d1                       \n"
-      "vtrn.u8      d4, d5                       \n"
+      "vtrn.u8     d0, d1                        \n"
+      "vtrn.u8     d4, d5                        \n"
 
       // d2 = 20 30 21 31 22 32 23 33
       // d3 = 60 70 61 71 62 72 63 73
-      "vtrn.u8      d2, d3                       \n"
-      "vtrn.u8      d6, d7                       \n"
+      "vtrn.u8     d2, d3                        \n"
+      "vtrn.u8     d6, d7                        \n"
 
       // d0 = 00+10 01+11 02+12 03+13
       // d2 = 40+50 41+51 42+52 43+53
-      "vpaddl.u8    q0, q0                       \n"
-      "vpaddl.u8    q2, q2                       \n"
+      "vpaddl.u8   q0, q0                        \n"
+      "vpaddl.u8   q2, q2                        \n"
 
       // d3 = 60+70 61+71 62+72 63+73
-      "vpaddl.u8    d3, d3                       \n"
-      "vpaddl.u8    d7, d7                       \n"
+      "vpaddl.u8   d3, d3                        \n"
+      "vpaddl.u8   d7, d7                        \n"
 
       // combine source lines
-      "vadd.u16     q0, q2                       \n"
-      "vadd.u16     d4, d3, d7                   \n"
+      "vadd.u16    q0, q2                        \n"
+      "vadd.u16    d4, d3, d7                    \n"
 
       // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-      "vqrshrn.u16  d4, q2, #2                   \n"
+      "vqrshrn.u16 d4, q2, #2                    \n"
 
       // Shuffle 2,3 reg around so that 2 can be added to the
       //  0,1 reg and 3 can be added to the 4,5 reg. This
@@ -463,22 +463,22 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
       //  registers are already expanded. Then do transposes
       //  to get aligned.
       // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-      "vmovl.u8     q1, d2                       \n"
-      "vmovl.u8     q3, d6                       \n"
+      "vmovl.u8    q1, d2                        \n"
+      "vmovl.u8    q3, d6                        \n"
 
       // combine source lines
-      "vadd.u16     q1, q3                       \n"
+      "vadd.u16    q1, q3                        \n"
 
       // d4 = xx 20 xx 30 xx 22 xx 32
       // d5 = xx 21 xx 31 xx 23 xx 33
-      "vtrn.u32     d2, d3                       \n"
+      "vtrn.u32    d2, d3                        \n"
 
       // d4 = xx 20 xx 21 xx 22 xx 23
       // d5 = xx 30 xx 31 xx 32 xx 33
-      "vtrn.u16     d2, d3                       \n"
+      "vtrn.u16    d2, d3                        \n"
 
       // 0+1+2, 3+4+5
-      "vadd.u16     q0, q1                       \n"
+      "vadd.u16    q0, q1                        \n"
 
       // Need to divide, but can't downshift as the the value
       //  isn't a power of 2. So multiply by 65536 / n
@@ -487,14 +487,14 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
 
       // Align for table lookup, vtbl requires registers to
       //  be adjacent
-      "vmov.u8      d2, d4                       \n"
+      "vmov.u8     d2, d4                        \n"
 
-      "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
-      "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+      "vtbl.u8     d3, {d0, d1, d2}, d28         \n"
+      "vtbl.u8     d4, {d0, d1, d2}, d29         \n"
 
-      "vst1.8       {d3}, [%1]!                  \n"
-      "vst1.32      {d4[0]}, [%1]!               \n"
-      "bgt          1b                           \n"
+      "vst1.8      {d3}, [%1]!                   \n"
+      "vst1.32     {d4[0]}, [%1]!                \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),       // %0
         "+r"(dst_ptr),       // %1
         "+r"(dst_width),     // %2
@@ -504,6 +504,484 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
       : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc");
 }
 
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u8     d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 01234567
+      "vld1.8      {d5}, [%3]!                   \n"  // 12345678
+
+      "vmovl.u8    q0, d4                        \n"  // 01234567 (16b)
+      "vmovl.u8    q1, d5                        \n"  // 12345678 (16b)
+      "vmlal.u8    q0, d5, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u8    q1, d4, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u16  d1, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u16  d0, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.8      {d0, d1}, [%1]!               \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 1;
+  const uint8_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+      "vmov.u8     d28, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 01234567
+      "vld1.8      {d5}, [%5]!                   \n"  // 12345678
+
+      "vmovl.u8    q0, d4                        \n"  // 01234567 (16b)
+      "vmovl.u8    q1, d5                        \n"  // 12345678 (16b)
+      "vmlal.u8    q0, d5, d28                   \n"  // 3*near+far (1, odd)
+      "vmlal.u8    q1, d4, d28                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d8}, [%1]!                   \n"
+      "vld1.8      {d9}, [%6]!                   \n"
+
+      "vmovl.u8    q2, d8                        \n"
+      "vmovl.u8    q3, d9                        \n"
+      "vmlal.u8    q2, d9, d28                   \n"  // 3*near+far (2, odd)
+      "vmlal.u8    q3, d8, d28                   \n"  // 3*near+far (2, even)
+
+      // e  o
+      // q1 q0
+      // q3 q2
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      // e  o
+      // q5 q4
+      // q1 q0
+
+      "vrshrn.u16  d2, q1, #4                    \n"  // 2, even
+      "vrshrn.u16  d3, q0, #4                    \n"  // 2, odd
+      "vrshrn.u16  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u16  d1, q4, #4                    \n"  // 1, odd
+
+      "vst2.8      {d0, d1}, [%2]!               \n"  // store
+      "vst2.8      {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "d28",
+        "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q1}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q0}, [%3]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q2, q0                        \n"
+      "vmla.u16    q0, q1, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q1, q2, q15                   \n"  // 3*near+far (even)
+
+      "vrshr.u16   q0, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshr.u16   q1, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.16     {d0, d1, d2, d3}, [%1]!       \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q1}, [%5]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q2, q0                        \n"
+      "vmla.u16    q0, q1, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q1, q2, q15                   \n"  // 3*near+far (even)
+
+      "vld1.16     {q2}, [%1]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q3}, [%6]!                   \n"  // 12345678 (16b)
+
+      "vmovq       q4, q2                        \n"
+      "vmla.u16    q2, q3, q15                   \n"  // 3*near+far (odd)
+      "vmla.u16    q3, q4, q15                   \n"  // 3*near+far (even)
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      "vrshr.u16   q2, q1, #4                    \n"  // 2, even
+      "vrshr.u16   q3, q0, #4                    \n"  // 2, odd
+      "vrshr.u16   q0, q5, #4                    \n"  // 1, even
+      "vrshr.u16   q1, q4, #4                    \n"  // 1, odd
+
+      "vst2.16     {d0, d1, d2, d3}, [%2]!       \n"  // store
+      "vst2.16     {d4, d5, d6, d7}, [%3]!       \n"  // store
+      "subs        %4, %4, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5",
+        "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "vmov.u16    d31, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 01234567 (16b)
+      "vld1.16     {q1}, [%3]!                   \n"  // 12345678 (16b)
+
+      "vmovl.u16   q2, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q3, d1                        \n"  // 4567 (32b)
+      "vmovl.u16   q4, d2                        \n"  // 1234 (32b)
+      "vmovl.u16   q5, d3                        \n"  // 5678 (32b)
+
+      "vmlal.u16   q2, d2, d31                   \n"
+      "vmlal.u16   q3, d3, d31                   \n"
+      "vmlal.u16   q4, d0, d31                   \n"
+      "vmlal.u16   q5, d1, d31                   \n"
+
+      "vrshrn.u32  d0, q4, #2                    \n"
+      "vrshrn.u32  d1, q5, #2                    \n"
+      "vrshrn.u32  d2, q2, #2                    \n"
+      "vrshrn.u32  d3, q3, #2                    \n"
+
+      "vst2.16     {q0, q1}, [%1]!               \n"  // store
+      "subs        %2, %2, #16                   \n"  // 8 sample -> 16 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q15"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "vmov.u16    d31, #3                       \n"
+      "vmov.u32    q14, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {d0}, [%0]!                   \n"  // 0123 (16b)
+      "vld1.16     {d1}, [%5]!                   \n"  // 1234 (16b)
+      "vmovl.u16   q2, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q3, d1                        \n"  // 1234 (32b)
+      "vmlal.u16   q2, d1, d31                   \n"
+      "vmlal.u16   q3, d0, d31                   \n"
+
+      "vld1.16     {d0}, [%1]!                   \n"  // 0123 (16b)
+      "vld1.16     {d1}, [%6]!                   \n"  // 1234 (16b)
+      "vmovl.u16   q4, d0                        \n"  // 0123 (32b)
+      "vmovl.u16   q5, d1                        \n"  // 1234 (32b)
+      "vmlal.u16   q4, d1, d31                   \n"
+      "vmlal.u16   q5, d0, d31                   \n"
+
+      "vmovq       q0, q4                        \n"
+      "vmovq       q1, q5                        \n"
+      "vmla.u32    q4, q2, q14                   \n"
+      "vmla.u32    q5, q3, q14                   \n"
+      "vmla.u32    q2, q0, q14                   \n"
+      "vmla.u32    q3, q1, q14                   \n"
+
+      "vrshrn.u32  d1, q4, #4                    \n"
+      "vrshrn.u32  d0, q5, #4                    \n"
+      "vrshrn.u32  d3, q2, #4                    \n"
+      "vrshrn.u32  d2, q3, #4                    \n"
+
+      "vst2.16     {d0, d1}, [%2]!               \n"  // store
+      "vst2.16     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #8                    \n"  // 4 sample -> 8 sample
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q14",
+        "d31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  const uint8_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "vmov.u8     d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d5}, [%3]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q0, d4                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q1, d5                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q0, d5, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u8    q1, d4, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u16  d1, q0, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u16  d0, q1, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.16     {d0, d1}, [%1]!               \n"  // store
+      "subs        %2, %2, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "d30"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 2;
+  const uint8_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "vmov.u16    q15, #3                       \n"
+      "vmov.u8     d28, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d4}, [%0]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d5}, [%5]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q0, d4                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q1, d5                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q0, d5, d28                   \n"  // 3*near+far (1, odd)
+      "vmlal.u8    q1, d4, d28                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d8}, [%1]!                   \n"  // 00112233 (1u1v)
+      "vld1.8      {d9}, [%6]!                   \n"  // 11223344 (1u1v)
+
+      "vmovl.u8    q2, d8                        \n"  // 00112233 (1u1v, 16b)
+      "vmovl.u8    q3, d9                        \n"  // 11223344 (1u1v, 16b)
+      "vmlal.u8    q2, d9, d28                   \n"  // 3*near+far (2, odd)
+      "vmlal.u8    q3, d8, d28                   \n"  // 3*near+far (2, even)
+
+      // e  o
+      // q1 q0
+      // q3 q2
+
+      "vmovq       q4, q2                        \n"
+      "vmovq       q5, q3                        \n"
+      "vmla.u16    q4, q0, q15                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u16    q5, q1, q15                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u16    q0, q2, q15                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u16    q1, q3, q15                   \n"  // 9 3 3 1 (2, even)
+
+      // e  o
+      // q5 q4
+      // q1 q0
+
+      "vrshrn.u16  d2, q1, #4                    \n"  // 2, even
+      "vrshrn.u16  d3, q0, #4                    \n"  // 2, odd
+      "vrshrn.u16  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u16  d1, q4, #4                    \n"  // 1, odd
+
+      "vst2.16     {d0, d1}, [%2]!               \n"  // store
+      "vst2.16     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "d28",
+        "q15"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  const uint16_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "vmov.u16    d30, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.16     {q0}, [%0]!                   \n"  // 00112233 (1u1v, 16)
+      "vld1.16     {q1}, [%3]!                   \n"  // 11223344 (1u1v, 16)
+
+      "vmovl.u16   q2, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q3, d2                        \n"  // 1122 (1u1v, 32b)
+      "vmovl.u16   q4, d1                        \n"  // 2233 (1u1v, 32b)
+      "vmovl.u16   q5, d3                        \n"  // 3344 (1u1v, 32b)
+      "vmlal.u16   q2, d2, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u16   q3, d0, d30                   \n"  // 3*near+far (even)
+      "vmlal.u16   q4, d3, d30                   \n"  // 3*near+far (odd)
+      "vmlal.u16   q5, d1, d30                   \n"  // 3*near+far (even)
+
+      "vrshrn.u32  d1, q2, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u32  d0, q3, #2                    \n"  // 3/4*near+1/4*far (even)
+      "vrshrn.u32  d3, q4, #2                    \n"  // 3/4*near+1/4*far (odd)
+      "vrshrn.u32  d2, q5, #2                    \n"  // 3/4*near+1/4*far (even)
+
+      "vst2.32     {d0, d1}, [%1]!               \n"  // store
+      "vst2.32     {d2, d3}, [%1]!               \n"  // store
+      "subs        %2, %2, #8                    \n"  // 4 uv -> 8 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(dst_ptr),    // %1
+        "+r"(dst_width),  // %2
+        "+r"(src_temp)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5",
+        "d30"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 2;
+  const uint16_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "vmov.u16    d30, #3                       \n"
+      "vmov.u32    q14, #3                       \n"
+
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0]!                   \n"  // 0011 (1u1v)
+      "vld1.8      {d1}, [%5]!                   \n"  // 1122 (1u1v)
+      "vmovl.u16   q2, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q3, d1                        \n"  // 1122 (1u1v, 32b)
+      "vmlal.u16   q2, d1, d30                   \n"  // 3*near+far (1, odd)
+      "vmlal.u16   q3, d0, d30                   \n"  // 3*near+far (1, even)
+
+      "vld1.8      {d0}, [%1]!                   \n"  // 0011 (1u1v)
+      "vld1.8      {d1}, [%6]!                   \n"  // 1122 (1u1v)
+      "vmovl.u16   q4, d0                        \n"  // 0011 (1u1v, 32b)
+      "vmovl.u16   q5, d1                        \n"  // 1122 (1u1v, 32b)
+      "vmlal.u16   q4, d1, d30                   \n"  // 3*near+far (2, odd)
+      "vmlal.u16   q5, d0, d30                   \n"  // 3*near+far (2, even)
+
+      "vmovq       q0, q4                        \n"
+      "vmovq       q1, q5                        \n"
+      "vmla.u32    q4, q2, q14                   \n"  // 9 3 3 1 (1, odd)
+      "vmla.u32    q5, q3, q14                   \n"  // 9 3 3 1 (1, even)
+      "vmla.u32    q2, q0, q14                   \n"  // 9 3 3 1 (2, odd)
+      "vmla.u32    q3, q1, q14                   \n"  // 9 3 3 1 (2, even)
+
+      "vrshrn.u32  d1, q4, #4                    \n"  // 1, odd
+      "vrshrn.u32  d0, q5, #4                    \n"  // 1, even
+      "vrshrn.u32  d3, q2, #4                    \n"  // 2, odd
+      "vrshrn.u32  d2, q3, #4                    \n"  // 2, even
+
+      "vst2.32     {d0, d1}, [%2]!               \n"  // store
+      "vst2.32     {d2, d3}, [%3]!               \n"  // store
+      "subs        %4, %4, #4                    \n"  // 2 uv -> 4 uv
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(dst_ptr),    // %2
+        "+r"(dst_ptr1),   // %3
+        "+r"(dst_width),  // %4
+        "+r"(src_temp),   // %5
+        "+r"(src_temp1)   // %6
+      :
+      : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q14",
+        "d30"  // Clobber List
+  );
+}
+
 // Add a row of bytes to a row of shorts.  Used for box filter.
 // Reads 16 bytes and accumulates to 16 shorts at a time.
 void ScaleAddRow_NEON(const uint8_t* src_ptr,
@@ -511,13 +989,13 @@ void ScaleAddRow_NEON(const uint8_t* src_ptr,
                       int src_width) {
   asm volatile(
       "1:                                        \n"
-      "vld1.16    {q1, q2}, [%1]                 \n"  // load accumulator
-      "vld1.8     {q0}, [%0]!                    \n"  // load 16 bytes
-      "vaddw.u8   q2, q2, d1                     \n"  // add
-      "vaddw.u8   q1, q1, d0                     \n"
-      "vst1.16    {q1, q2}, [%1]!                \n"  // store accumulator
-      "subs       %2, %2, #16                    \n"  // 16 processed per loop
-      "bgt        1b                             \n"
+      "vld1.16     {q1, q2}, [%1]                \n"  // load accumulator
+      "vld1.8      {q0}, [%0]!                   \n"  // load 16 bytes
+      "vaddw.u8    q2, q2, d1                    \n"  // add
+      "vaddw.u8    q1, q1, d0                    \n"
+      "vst1.16     {q1, q2}, [%1]!               \n"  // store accumulator
+      "subs        %2, %2, #16                   \n"  // 16 processed per loop
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(src_width)  // %2
@@ -547,17 +1025,17 @@ void ScaleFilterCols_NEON(uint8_t* dst_ptr,
   int* tmp = dx_offset;
   const uint8_t* src_tmp = src_ptr;
   asm volatile (
-    "vdup.32    q0, %3                         \n"  // x
-    "vdup.32    q1, %4                         \n"  // dx
-    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
-    "vshl.i32   q3, q1, #2                     \n"  // 4 * dx
-    "vmul.s32   q1, q1, q2                     \n"
+      "vdup.32     q0, %3                        \n"  // x
+      "vdup.32     q1, %4                        \n"  // dx
+      "vld1.32     {q2}, [%5]                    \n"  // 0 1 2 3
+      "vshl.i32    q3, q1, #2                    \n"  // 4 * dx
+      "vmul.s32    q1, q1, q2                    \n"
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "vadd.s32   q1, q1, q0                     \n"
+      "vadd.s32    q1, q1, q0                    \n"
     // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
-    "vadd.s32   q2, q1, q3                     \n"
-    "vshl.i32   q0, q3, #1                     \n"  // 8 * dx
-  "1:                                          \n"
+      "vadd.s32    q2, q1, q3                    \n"
+      "vshl.i32    q0, q3, #1                    \n"  // 8 * dx
+      "1:                                        \n"
     LOAD2_DATA8_LANE(0)
     LOAD2_DATA8_LANE(1)
     LOAD2_DATA8_LANE(2)
@@ -566,27 +1044,27 @@ void ScaleFilterCols_NEON(uint8_t* dst_ptr,
     LOAD2_DATA8_LANE(5)
     LOAD2_DATA8_LANE(6)
     LOAD2_DATA8_LANE(7)
-    "vmov       q10, q1                        \n"
-    "vmov       q11, q2                        \n"
-    "vuzp.16    q10, q11                       \n"
-    "vmovl.u8   q8, d6                         \n"
-    "vmovl.u8   q9, d7                         \n"
-    "vsubl.s16  q11, d18, d16                  \n"
-    "vsubl.s16  q12, d19, d17                  \n"
-    "vmovl.u16  q13, d20                       \n"
-    "vmovl.u16  q10, d21                       \n"
-    "vmul.s32   q11, q11, q13                  \n"
-    "vmul.s32   q12, q12, q10                  \n"
-    "vrshrn.s32  d18, q11, #16                 \n"
-    "vrshrn.s32  d19, q12, #16                 \n"
-    "vadd.s16   q8, q8, q9                     \n"
-    "vmovn.s16  d6, q8                         \n"
-
-    "vst1.8     {d6}, [%0]!                    \n"  // store pixels
-    "vadd.s32   q1, q1, q0                     \n"
-    "vadd.s32   q2, q2, q0                     \n"
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
-    "bgt        1b                             \n"
+      "vmov        q10, q1                       \n"
+      "vmov        q11, q2                       \n"
+      "vuzp.16     q10, q11                      \n"
+      "vmovl.u8    q8, d6                        \n"
+      "vmovl.u8    q9, d7                        \n"
+      "vsubl.s16   q11, d18, d16                 \n"
+      "vsubl.s16   q12, d19, d17                 \n"
+      "vmovl.u16   q13, d20                      \n"
+      "vmovl.u16   q10, d21                      \n"
+      "vmul.s32    q11, q11, q13                 \n"
+      "vmul.s32    q12, q12, q10                 \n"
+      "vrshrn.s32  d18, q11, #16                 \n"
+      "vrshrn.s32  d19, q12, #16                 \n"
+      "vadd.s16    q8, q8, q9                    \n"
+      "vmovn.s16   d6, q8                        \n"
+
+      "vst1.8      {d6}, [%0]!                   \n"  // store pixels
+      "vadd.s32    q1, q1, q0                    \n"
+      "vadd.s32    q2, q2, q0                    \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "bgt         1b                            \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
     "+r"(dst_width),        // %2
@@ -609,75 +1087,75 @@ void ScaleFilterRows_NEON(uint8_t* dst_ptr,
                           int dst_width,
                           int source_y_fraction) {
   asm volatile(
-      "cmp          %4, #0                       \n"
-      "beq          100f                         \n"
-      "add          %2, %1                       \n"
-      "cmp          %4, #64                      \n"
-      "beq          75f                          \n"
-      "cmp          %4, #128                     \n"
-      "beq          50f                          \n"
-      "cmp          %4, #192                     \n"
-      "beq          25f                          \n"
-
-      "vdup.8       d5, %4                       \n"
-      "rsb          %4, #256                     \n"
-      "vdup.8       d4, %4                       \n"
+      "cmp         %4, #0                        \n"
+      "beq         100f                          \n"
+      "add         %2, %1                        \n"
+      "cmp         %4, #64                       \n"
+      "beq         75f                           \n"
+      "cmp         %4, #128                      \n"
+      "beq         50f                           \n"
+      "cmp         %4, #192                      \n"
+      "beq         25f                           \n"
+
+      "vdup.8      d5, %4                        \n"
+      "rsb         %4, #256                      \n"
+      "vdup.8      d4, %4                        \n"
       // General purpose row blend.
       "1:                                        \n"
-      "vld1.8       {q0}, [%1]!                  \n"
-      "vld1.8       {q1}, [%2]!                  \n"
-      "subs         %3, %3, #16                  \n"
-      "vmull.u8     q13, d0, d4                  \n"
-      "vmull.u8     q14, d1, d4                  \n"
-      "vmlal.u8     q13, d2, d5                  \n"
-      "vmlal.u8     q14, d3, d5                  \n"
-      "vrshrn.u16   d0, q13, #8                  \n"
-      "vrshrn.u16   d1, q14, #8                  \n"
-      "vst1.8       {q0}, [%0]!                  \n"
-      "bgt          1b                           \n"
-      "b            99f                          \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vmull.u8    q13, d0, d4                   \n"
+      "vmull.u8    q14, d1, d4                   \n"
+      "vmlal.u8    q13, d2, d5                   \n"
+      "vmlal.u8    q14, d3, d5                   \n"
+      "vrshrn.u16  d0, q13, #8                   \n"
+      "vrshrn.u16  d1, q14, #8                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         1b                            \n"
+      "b           99f                           \n"
 
       // Blend 25 / 75.
       "25:                                       \n"
-      "vld1.8       {q0}, [%1]!                  \n"
-      "vld1.8       {q1}, [%2]!                  \n"
-      "subs         %3, %3, #16                  \n"
-      "vrhadd.u8    q0, q1                       \n"
-      "vrhadd.u8    q0, q1                       \n"
-      "vst1.8       {q0}, [%0]!                  \n"
-      "bgt          25b                          \n"
-      "b            99f                          \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         25b                           \n"
+      "b           99f                           \n"
 
       // Blend 50 / 50.
       "50:                                       \n"
-      "vld1.8       {q0}, [%1]!                  \n"
-      "vld1.8       {q1}, [%2]!                  \n"
-      "subs         %3, %3, #16                  \n"
-      "vrhadd.u8    q0, q1                       \n"
-      "vst1.8       {q0}, [%0]!                  \n"
-      "bgt          50b                          \n"
-      "b            99f                          \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "vld1.8      {q1}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         50b                           \n"
+      "b           99f                           \n"
 
       // Blend 75 / 25.
       "75:                                       \n"
-      "vld1.8       {q1}, [%1]!                  \n"
-      "vld1.8       {q0}, [%2]!                  \n"
-      "subs         %3, %3, #16                  \n"
-      "vrhadd.u8    q0, q1                       \n"
-      "vrhadd.u8    q0, q1                       \n"
-      "vst1.8       {q0}, [%0]!                  \n"
-      "bgt          75b                          \n"
-      "b            99f                          \n"
+      "vld1.8      {q1}, [%1]!                   \n"
+      "vld1.8      {q0}, [%2]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vrhadd.u8   q0, q1                        \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         75b                           \n"
+      "b           99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       "100:                                      \n"
-      "vld1.8       {q0}, [%1]!                  \n"
-      "subs         %3, %3, #16                  \n"
-      "vst1.8       {q0}, [%0]!                  \n"
-      "bgt          100b                         \n"
+      "vld1.8      {q0}, [%1]!                   \n"
+      "subs        %3, %3, #16                   \n"
+      "vst1.8      {q0}, [%0]!                   \n"
+      "bgt         100b                          \n"
 
       "99:                                       \n"
-      "vst1.8       {d1[7]}, [%0]                \n"
+      "vst1.8      {d1[7]}, [%0]                 \n"
       : "+r"(dst_ptr),           // %0
         "+r"(src_ptr),           // %1
         "+r"(src_stride),        // %2
@@ -694,12 +1172,12 @@ void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "vld4.32    {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.32    {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop
-      "vmov       q2, q1                         \n"  // load next 8 ARGB
-      "vst2.32    {q2, q3}, [%1]!                \n"  // store odd pixels
-      "bgt        1b                             \n"
+      "vld4.32     {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.32     {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vmov        q2, q1                        \n"  // load next 8 ARGB
+      "vst2.32     {q2, q3}, [%1]!               \n"  // store odd pixels
+      "bgt         1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -722,13 +1200,13 @@ void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "vld4.32    {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.32    {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop
-      "vrhadd.u8  q0, q0, q1                     \n"  // rounding half add
-      "vrhadd.u8  q1, q2, q3                     \n"  // rounding half add
-      "vst2.32    {q0, q1}, [%1]!                \n"
-      "bgt       1b                              \n"
+      "vld4.32     {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.32     {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "vrhadd.u8   q0, q0, q1                    \n"  // rounding half add
+      "vrhadd.u8   q1, q2, q3                    \n"  // rounding half add
+      "vst2.32     {q0, q1}, [%1]!               \n"
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(dst_width)  // %2
@@ -743,27 +1221,27 @@ void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
                                int dst_width) {
   asm volatile(
       // change the stride to row 2 pointer
-      "add        %1, %1, %0                     \n"
-      "1:                                        \n"
-      "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
-      "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB
-      "subs       %3, %3, #8                     \n"  // 8 processed per loop.
-      "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
-      "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
-      "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
-      "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
-      "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB
-      "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB
-      "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
-      "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
-      "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
-      "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
-      "vrshrn.u16 d0, q0, #2                     \n"  // round and pack to bytes
-      "vrshrn.u16 d1, q1, #2                     \n"
-      "vrshrn.u16 d2, q2, #2                     \n"
-      "vrshrn.u16 d3, q3, #2                     \n"
-      "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
-      "bgt        1b                             \n"
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld4.8      {d0, d2, d4, d6}, [%0]!       \n"  // load 8 ARGB pixels.
+      "vld4.8      {d1, d3, d5, d7}, [%0]!       \n"  // load next 8 ARGB
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vpaddl.u8   q0, q0                        \n"  // B 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // G 16 bytes -> 8 shorts.
+      "vpaddl.u8   q2, q2                        \n"  // R 16 bytes -> 8 shorts.
+      "vpaddl.u8   q3, q3                        \n"  // A 16 bytes -> 8 shorts.
+      "vld4.8      {d16, d18, d20, d22}, [%1]!   \n"  // load 8 more ARGB
+      "vld4.8      {d17, d19, d21, d23}, [%1]!   \n"  // load last 8 ARGB
+      "vpadal.u8   q0, q8                        \n"  // B 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q9                        \n"  // G 16 bytes -> 8 shorts.
+      "vpadal.u8   q2, q10                       \n"  // R 16 bytes -> 8 shorts.
+      "vpadal.u8   q3, q11                       \n"  // A 16 bytes -> 8 shorts.
+      "vrshrn.u16  d0, q0, #2                    \n"  // round and pack to bytes
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vrshrn.u16  d2, q2, #2                    \n"
+      "vrshrn.u16  d3, q3, #2                    \n"
+      "vst4.8      {d0, d1, d2, d3}, [%2]!       \n"
+      "bgt         1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -781,15 +1259,15 @@ void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
                                int dst_width) {
   (void)src_stride;
   asm volatile(
-      "mov        r12, %3, lsl #2                \n"
-      "1:                                        \n"
-      "vld1.32    {d0[0]}, [%0], r12             \n"
-      "vld1.32    {d0[1]}, [%0], r12             \n"
-      "vld1.32    {d1[0]}, [%0], r12             \n"
-      "vld1.32    {d1[1]}, [%0], r12             \n"
-      "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
-      "vst1.8     {q0}, [%1]!                    \n"
-      "bgt        1b                             \n"
+      "mov         r12, %3, lsl #2               \n"
+      "1:                                        \n"
+      "vld1.32     {d0[0]}, [%0], r12            \n"
+      "vld1.32     {d0[1]}, [%0], r12            \n"
+      "vld1.32     {d1[0]}, [%0], r12            \n"
+      "vld1.32     {d1[1]}, [%0], r12            \n"
+      "subs        %2, %2, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {q0}, [%1]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(dst_width)  // %2
@@ -805,30 +1283,30 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
                                   uint8_t* dst_argb,
                                   int dst_width) {
   asm volatile(
-      "mov        r12, %4, lsl #2                \n"
-      "add        %1, %1, %0                     \n"
-      "1:                                        \n"
-      "vld1.8     {d0}, [%0], r12                \n"  // 4 2x2 blocks -> 2x1
-      "vld1.8     {d1}, [%1], r12                \n"
-      "vld1.8     {d2}, [%0], r12                \n"
-      "vld1.8     {d3}, [%1], r12                \n"
-      "vld1.8     {d4}, [%0], r12                \n"
-      "vld1.8     {d5}, [%1], r12                \n"
-      "vld1.8     {d6}, [%0], r12                \n"
-      "vld1.8     {d7}, [%1], r12                \n"
-      "vaddl.u8   q0, d0, d1                     \n"
-      "vaddl.u8   q1, d2, d3                     \n"
-      "vaddl.u8   q2, d4, d5                     \n"
-      "vaddl.u8   q3, d6, d7                     \n"
-      "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
-      "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
-      "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
-      "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
-      "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
-      "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
-      "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
-      "vst1.8     {q0}, [%2]!                    \n"
-      "bgt        1b                             \n"
+      "mov         r12, %4, lsl #2               \n"
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld1.8      {d0}, [%0], r12               \n"  // 4 2x2 blocks -> 2x1
+      "vld1.8      {d1}, [%1], r12               \n"
+      "vld1.8      {d2}, [%0], r12               \n"
+      "vld1.8      {d3}, [%1], r12               \n"
+      "vld1.8      {d4}, [%0], r12               \n"
+      "vld1.8      {d5}, [%1], r12               \n"
+      "vld1.8      {d6}, [%0], r12               \n"
+      "vld1.8      {d7}, [%1], r12               \n"
+      "vaddl.u8    q0, d0, d1                    \n"
+      "vaddl.u8    q1, d2, d3                    \n"
+      "vaddl.u8    q2, d4, d5                    \n"
+      "vaddl.u8    q3, d6, d7                    \n"
+      "vswp.8      d1, d2                        \n"  // ab_cd -> ac_bd
+      "vswp.8      d5, d6                        \n"  // ef_gh -> eg_fh
+      "vadd.u16    q0, q0, q1                    \n"  // (a+b)_(c+d)
+      "vadd.u16    q2, q2, q3                    \n"  // (e+f)_(g+h)
+      "vrshrn.u16  d0, q0, #2                    \n"  // first 2 pixels.
+      "vrshrn.u16  d1, q2, #2                    \n"  // next 2 pixels.
+      "subs        %3, %3, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {q0}, [%2]!                   \n"
+      "bgt         1b                            \n"
       : "+r"(src_argb),    // %0
         "+r"(src_stride),  // %1
         "+r"(dst_argb),    // %2
@@ -865,8 +1343,8 @@ void ScaleARGBCols_NEON(uint8_t* dst_argb,
       LOAD1_DATA32_LANE(d3, 1)
       // clang-format on
       "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
-      "subs       %2, %2, #8                     \n"  // 8 processed per loop
-      "bgt        1b                             \n"
+      "subs        %2, %2, #8                    \n"  // 8 processed per loop
+      "bgt         1b                            \n"
       : "+r"(dst_argb),   // %0
         "+r"(src_argb),   // %1
         "+r"(dst_width),  // %2
@@ -897,16 +1375,16 @@ void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
   int* tmp = dx_offset;
   const uint8_t* src_tmp = src_argb;
   asm volatile (
-    "vdup.32    q0, %3                         \n"  // x
-    "vdup.32    q1, %4                         \n"  // dx
-    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
-    "vshl.i32   q9, q1, #2                     \n"  // 4 * dx
-    "vmul.s32   q1, q1, q2                     \n"
-    "vmov.i8    q3, #0x7f                      \n"  // 0x7F
-    "vmov.i16   q15, #0x7f                     \n"  // 0x7F
+      "vdup.32     q0, %3                        \n"  // x
+      "vdup.32     q1, %4                        \n"  // dx
+      "vld1.32     {q2}, [%5]                    \n"  // 0 1 2 3
+      "vshl.i32    q9, q1, #2                    \n"  // 4 * dx
+      "vmul.s32    q1, q1, q2                    \n"
+      "vmov.i8     q3, #0x7f                     \n"  // 0x7F
+      "vmov.i16    q15, #0x7f                    \n"  // 0x7F
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "vadd.s32   q8, q1, q0                     \n"
-  "1:                                          \n"
+      "vadd.s32    q8, q1, q0                    \n"
+      "1:                                        \n"
     // d0, d1: a
     // d2, d3: b
     LOAD2_DATA32_LANE(d0, d2, 0)
@@ -950,6 +1428,64 @@ void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
 
 #undef LOAD2_DATA32_LANE
 
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "vld2.8      {d0, d2}, [%0]!               \n"  // load 8 UV pixels.
+      "vld2.8      {d1, d3}, [%0]!               \n"  // load next 8 UV
+      "subs        %3, %3, #8                    \n"  // 8 processed per loop.
+      "vpaddl.u8   q0, q0                        \n"  // U 16 bytes -> 8 shorts.
+      "vpaddl.u8   q1, q1                        \n"  // V 16 bytes -> 8 shorts.
+      "vld2.8      {d16, d18}, [%1]!             \n"  // load 8 more UV
+      "vld2.8      {d17, d19}, [%1]!             \n"  // load last 8 UV
+      "vpadal.u8   q0, q8                        \n"  // U 16 bytes -> 8 shorts.
+      "vpadal.u8   q1, q9                        \n"  // V 16 bytes -> 8 shorts.
+      "vrshrn.u16  d0, q0, #2                    \n"  // round and pack to bytes
+      "vrshrn.u16  d1, q1, #2                    \n"
+      "vst2.8      {d0, d1}, [%2]!               \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "q0", "q1", "q8", "q9");
+}
+
+// Reads 4 pixels at a time.
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,  // pixel step
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
+  const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
+  const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "vld1.16     {d0[0]}, [%0], %6             \n"
+      "vld1.16     {d0[1]}, [%1], %6             \n"
+      "vld1.16     {d0[2]}, [%2], %6             \n"
+      "vld1.16     {d0[3]}, [%3], %6             \n"
+      "subs        %5, %5, #4                    \n"  // 4 pixels per loop.
+      "vst1.8      {d0}, [%4]!                   \n"
+      "bgt         1b                            \n"
+      : "+r"(src_ptr),      // %0
+        "+r"(src1_ptr),     // %1
+        "+r"(src2_ptr),     // %2
+        "+r"(src3_ptr),     // %3
+        "+r"(dst_ptr),      // %4
+        "+r"(dst_width)     // %5
+      : "r"(src_stepx * 8)  // %6
+      : "memory", "cc", "d0");
+}
+
 #endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/files/source/scale_neon64.cc b/files/source/scale_neon64.cc
index 0a7b80ce..9f9636e6 100644
--- a/files/source/scale_neon64.cc
+++ b/files/source/scale_neon64.cc
@@ -29,10 +29,11 @@ void ScaleRowDown2_NEON(const uint8_t* src_ptr,
   asm volatile(
       "1:                                        \n"
       // load even pixels into v0, odd into v1
-      "ld2        {v0.16b,v1.16b}, [%0], #32     \n"
-      "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-      "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
-      "b.gt       1b                             \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v1.16b}, [%1], #16           \n"  // store odd pixels
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -50,11 +51,12 @@ void ScaleRowDown2Linear_NEON(const uint8_t* src_ptr,
   asm volatile(
       "1:                                        \n"
       // load even pixels into v0, odd into v1
-      "ld2        {v0.16b,v1.16b}, [%0], #32     \n"
-      "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
-      "urhadd     v0.16b, v0.16b, v1.16b         \n"  // rounding half add
-      "st1        {v0.16b}, [%1], #16            \n"
-      "b.gt       1b                             \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"  // rounding half add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%1], #16           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -70,19 +72,21 @@ void ScaleRowDown2Box_NEON(const uint8_t* src_ptr,
                            int dst_width) {
   asm volatile(
       // change the stride to row 2 pointer
-      "add        %1, %1, %0                     \n"
+      "add         %1, %1, %0                    \n"
       "1:                                        \n"
-      "ld1        {v0.16b, v1.16b}, [%0], #32    \n"  // load row 1 and post inc
-      "ld1        {v2.16b, v3.16b}, [%1], #32    \n"  // load row 2 and post inc
-      "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
-      "uaddlp     v0.8h, v0.16b                  \n"  // row 1 add adjacent
-      "uaddlp     v1.8h, v1.16b                  \n"
-      "uadalp     v0.8h, v2.16b                  \n"  // += row 2 add adjacent
-      "uadalp     v1.8h, v3.16b                  \n"
-      "rshrn      v0.8b, v0.8h, #2               \n"  // round and pack
-      "rshrn2     v0.16b, v1.8h, #2              \n"
-      "st1        {v0.16b}, [%2], #16            \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b, v1.16b}, [%0], #32   \n"  // load row 1 and post inc
+      "ld1         {v2.16b, v3.16b}, [%1], #32   \n"  // load row 2 and post inc
+      "subs        %w3, %w3, #16                 \n"  // 16 processed per loop
+      "uaddlp      v0.8h, v0.16b                 \n"  // row 1 add adjacent
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddlp      v1.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"  // += row 2 add adjacent
+      "uadalp      v1.8h, v3.16b                 \n"
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "rshrn2      v0.16b, v1.8h, #2             \n"
+      "st1         {v0.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -99,10 +103,11 @@ void ScaleRowDown4_NEON(const uint8_t* src_ptr,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "ld4     {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32  \n"  // src line 0
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "st1     {v2.8b}, [%1], #8                 \n"
-      "b.gt       1b                             \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v2.8b}, [%1], #8             \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -119,19 +124,23 @@ void ScaleRowDown4Box_NEON(const uint8_t* src_ptr,
   const uint8_t* src_ptr3 = src_ptr + src_stride * 3;
   asm volatile(
       "1:                                        \n"
-      "ld1     {v0.16b}, [%0], #16               \n"  // load up 16x4
-      "ld1     {v1.16b}, [%2], #16               \n"
-      "ld1     {v2.16b}, [%3], #16               \n"
-      "ld1     {v3.16b}, [%4], #16               \n"
-      "subs    %w5, %w5, #4                      \n"
-      "uaddlp  v0.8h, v0.16b                     \n"
-      "uadalp  v0.8h, v1.16b                     \n"
-      "uadalp  v0.8h, v2.16b                     \n"
-      "uadalp  v0.8h, v3.16b                     \n"
-      "addp    v0.8h, v0.8h, v0.8h               \n"
-      "rshrn   v0.8b, v0.8h, #4                  \n"  // divide by 16 w/rounding
-      "st1    {v0.s}[0], [%1], #4                \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.16b}, [%0], #16           \n"  // load up 16x4
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "ld1         {v2.16b}, [%3], #16           \n"
+      "ld1         {v3.16b}, [%4], #16           \n"
+      "subs        %w5, %w5, #4                  \n"
+      "uaddlp      v0.8h, v0.16b                 \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v0.8h, v1.16b                 \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "uadalp      v0.8h, v2.16b                 \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+      "uadalp      v0.8h, v3.16b                 \n"
+      "prfm        pldl1keep, [%4, 448]          \n"
+      "addp        v0.8h, v0.8h, v0.8h           \n"
+      "rshrn       v0.8b, v0.8h, #4              \n"  // divide by 16 w/rounding
+      "st1         {v0.s}[0], [%1], #4           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(src_ptr1),  // %2
@@ -151,12 +160,13 @@ void ScaleRowDown34_NEON(const uint8_t* src_ptr,
                          int dst_width) {
   (void)src_stride;
   asm volatile(
-      "1:                                                \n"
-      "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32    \n"  // src line 0
-      "subs      %w2, %w2, #24                           \n"
-      "orr       v2.16b, v3.16b, v3.16b                  \n"  // order v0,v1,v2
-      "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24          \n"
-      "b.gt      1b                                      \n"
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "subs        %w2, %w2, #24                 \n"
+      "orr         v2.16b, v3.16b, v3.16b        \n"  // order v0,v1,v2
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -169,49 +179,51 @@ void ScaleRowDown34_0_Box_NEON(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int dst_width) {
   asm volatile(
-      "movi      v20.8b, #3                              \n"
-      "add       %3, %3, %0                              \n"
-      "1:                                                \n"
-      "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32    \n"  // src line 0
-      "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32    \n"  // src line 1
-      "subs         %w2, %w2, #24                        \n"
+      "movi        v20.8b, #3                    \n"
+      "add         %3, %3, %0                    \n"
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"  // src line 1
+      "subs        %w2, %w2, #24                 \n"
 
       // filter src line 0 with src line 1
       // expand chars to shorts to allow for room
       // when adding lines together
-      "ushll     v16.8h, v4.8b, #0                       \n"
-      "ushll     v17.8h, v5.8b, #0                       \n"
-      "ushll     v18.8h, v6.8b, #0                       \n"
-      "ushll     v19.8h, v7.8b, #0                       \n"
+      "ushll       v16.8h, v4.8b, #0             \n"
+      "ushll       v17.8h, v5.8b, #0             \n"
+      "ushll       v18.8h, v6.8b, #0             \n"
+      "ushll       v19.8h, v7.8b, #0             \n"
 
       // 3 * line_0 + line_1
-      "umlal     v16.8h, v0.8b, v20.8b                   \n"
-      "umlal     v17.8h, v1.8b, v20.8b                   \n"
-      "umlal     v18.8h, v2.8b, v20.8b                   \n"
-      "umlal     v19.8h, v3.8b, v20.8b                   \n"
-
-      // (3 * line_0 + line_1) >> 2
-      "uqrshrn   v0.8b, v16.8h, #2                       \n"
-      "uqrshrn   v1.8b, v17.8h, #2                       \n"
-      "uqrshrn   v2.8b, v18.8h, #2                       \n"
-      "uqrshrn   v3.8b, v19.8h, #2                       \n"
-
-      // a0 = (src[0] * 3 + s[1] * 1) >> 2
-      "ushll     v16.8h, v1.8b, #0                       \n"
-      "umlal     v16.8h, v0.8b, v20.8b                   \n"
-      "uqrshrn   v0.8b, v16.8h, #2                       \n"
-
-      // a1 = (src[1] * 1 + s[2] * 1) >> 1
-      "urhadd    v1.8b, v1.8b, v2.8b                     \n"
-
-      // a2 = (src[2] * 1 + s[3] * 3) >> 2
-      "ushll     v16.8h, v2.8b, #0                       \n"
-      "umlal     v16.8h, v3.8b, v20.8b                   \n"
-      "uqrshrn   v2.8b, v16.8h, #2                       \n"
-
-      "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24          \n"
-
-      "b.gt      1b                                      \n"
+      "umlal       v16.8h, v0.8b, v20.8b         \n"
+      "umlal       v17.8h, v1.8b, v20.8b         \n"
+      "umlal       v18.8h, v2.8b, v20.8b         \n"
+      "umlal       v19.8h, v3.8b, v20.8b         \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // (3 * line_0 + line_1 + 2) >> 2
+      "uqrshrn     v0.8b, v16.8h, #2             \n"
+      "uqrshrn     v1.8b, v17.8h, #2             \n"
+      "uqrshrn     v2.8b, v18.8h, #2             \n"
+      "uqrshrn     v3.8b, v19.8h, #2             \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "ushll       v16.8h, v1.8b, #0             \n"
+      "umlal       v16.8h, v0.8b, v20.8b         \n"
+      "uqrshrn     v0.8b, v16.8h, #2             \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "urhadd      v1.8b, v1.8b, v2.8b           \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "ushll       v16.8h, v2.8b, #0             \n"
+      "umlal       v16.8h, v3.8b, v20.8b         \n"
+      "uqrshrn     v2.8b, v16.8h, #2             \n"
+
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),    // %0
         "+r"(dst_ptr),    // %1
         "+r"(dst_width),  // %2
@@ -226,33 +238,35 @@ void ScaleRowDown34_1_Box_NEON(const uint8_t* src_ptr,
                                uint8_t* dst_ptr,
                                int dst_width) {
   asm volatile(
-      "movi      v20.8b, #3                              \n"
-      "add       %3, %3, %0                              \n"
-      "1:                                                \n"
-      "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32    \n"  // src line 0
-      "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32    \n"  // src line 1
-      "subs         %w2, %w2, #24                        \n"
+      "movi        v20.8b, #3                    \n"
+      "add         %3, %3, %0                    \n"
+      "1:                                        \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // src line 0
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32 \n"  // src line 1
+      "subs        %w2, %w2, #24                 \n"
       // average src line 0 with src line 1
-      "urhadd    v0.8b, v0.8b, v4.8b                     \n"
-      "urhadd    v1.8b, v1.8b, v5.8b                     \n"
-      "urhadd    v2.8b, v2.8b, v6.8b                     \n"
-      "urhadd    v3.8b, v3.8b, v7.8b                     \n"
-
-      // a0 = (src[0] * 3 + s[1] * 1) >> 2
-      "ushll     v4.8h, v1.8b, #0                        \n"
-      "umlal     v4.8h, v0.8b, v20.8b                    \n"
-      "uqrshrn   v0.8b, v4.8h, #2                        \n"
-
-      // a1 = (src[1] * 1 + s[2] * 1) >> 1
-      "urhadd    v1.8b, v1.8b, v2.8b                     \n"
-
-      // a2 = (src[2] * 1 + s[3] * 3) >> 2
-      "ushll     v4.8h, v2.8b, #0                        \n"
-      "umlal     v4.8h, v3.8b, v20.8b                    \n"
-      "uqrshrn   v2.8b, v4.8h, #2                        \n"
-
-      "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24          \n"
-      "b.gt      1b                                      \n"
+      "urhadd      v0.8b, v0.8b, v4.8b           \n"
+      "urhadd      v1.8b, v1.8b, v5.8b           \n"
+      "urhadd      v2.8b, v2.8b, v6.8b           \n"
+      "urhadd      v3.8b, v3.8b, v7.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      // a0 = (src[0] * 3 + s[1] * 1 + 2) >> 2
+      "ushll       v4.8h, v1.8b, #0              \n"
+      "umlal       v4.8h, v0.8b, v20.8b          \n"
+      "uqrshrn     v0.8b, v4.8h, #2              \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
+
+      // a1 = (src[1] * 1 + s[2] * 1 + 1) >> 1
+      "urhadd      v1.8b, v1.8b, v2.8b           \n"
+
+      // a2 = (src[2] * 1 + s[3] * 3 + 2) >> 2
+      "ushll       v4.8h, v2.8b, #0              \n"
+      "umlal       v4.8h, v3.8b, v20.8b          \n"
+      "uqrshrn     v2.8b, v4.8h, #2              \n"
+
+      "st3         {v0.8b,v1.8b,v2.8b}, [%1], #24 \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),    // %0
         "+r"(dst_ptr),    // %1
         "+r"(dst_width),  // %2
@@ -279,14 +293,15 @@ void ScaleRowDown38_NEON(const uint8_t* src_ptr,
                          int dst_width) {
   (void)src_stride;
   asm volatile(
-      "ld1       {v3.16b}, [%3]                          \n"
-      "1:                                                \n"
-      "ld1       {v0.16b,v1.16b}, [%0], #32              \n"
-      "subs      %w2, %w2, #12                           \n"
-      "tbl       v2.16b, {v0.16b,v1.16b}, v3.16b         \n"
-      "st1       {v2.8b}, [%1], #8                       \n"
-      "st1       {v2.s}[2], [%1], #4                     \n"
-      "b.gt      1b                                      \n"
+      "ld1         {v3.16b}, [%3]                \n"
+      "1:                                        \n"
+      "ld1         {v0.16b,v1.16b}, [%0], #32    \n"
+      "subs        %w2, %w2, #12                 \n"
+      "tbl         v2.16b, {v0.16b,v1.16b}, v3.16b \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v2.8b}, [%1], #8             \n"
+      "st1         {v2.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(dst_width)  // %2
@@ -303,68 +318,68 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
   ptrdiff_t tmp_src_stride = src_stride;
 
   asm volatile(
-      "ld1       {v29.8h}, [%5]                          \n"
-      "ld1       {v30.16b}, [%6]                         \n"
-      "ld1       {v31.8h}, [%7]                          \n"
-      "add       %2, %2, %0                              \n"
-      "1:                                                \n"
+      "ld1         {v29.8h}, [%5]                \n"
+      "ld1         {v30.16b}, [%6]               \n"
+      "ld1         {v31.8h}, [%7]                \n"
+      "add         %2, %2, %0                    \n"
+      "1:                                        \n"
 
       // 00 40 01 41 02 42 03 43
       // 10 50 11 51 12 52 13 53
       // 20 60 21 61 22 62 23 63
       // 30 70 31 71 32 72 33 73
-      "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32    \n"
-      "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32    \n"
-      "ld4       {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32  \n"
-      "subs      %w4, %w4, #12                           \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
+      "ld4         {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32 \n"
+      "subs        %w4, %w4, #12                 \n"
 
       // Shuffle the input data around to get align the data
       //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
       // 00 10 01 11 02 12 03 13
       // 40 50 41 51 42 52 43 53
-      "trn1      v20.8b, v0.8b, v1.8b                    \n"
-      "trn2      v21.8b, v0.8b, v1.8b                    \n"
-      "trn1      v22.8b, v4.8b, v5.8b                    \n"
-      "trn2      v23.8b, v4.8b, v5.8b                    \n"
-      "trn1      v24.8b, v16.8b, v17.8b                  \n"
-      "trn2      v25.8b, v16.8b, v17.8b                  \n"
+      "trn1        v20.8b, v0.8b, v1.8b          \n"
+      "trn2        v21.8b, v0.8b, v1.8b          \n"
+      "trn1        v22.8b, v4.8b, v5.8b          \n"
+      "trn2        v23.8b, v4.8b, v5.8b          \n"
+      "trn1        v24.8b, v16.8b, v17.8b        \n"
+      "trn2        v25.8b, v16.8b, v17.8b        \n"
 
       // 20 30 21 31 22 32 23 33
       // 60 70 61 71 62 72 63 73
-      "trn1      v0.8b, v2.8b, v3.8b                     \n"
-      "trn2      v1.8b, v2.8b, v3.8b                     \n"
-      "trn1      v4.8b, v6.8b, v7.8b                     \n"
-      "trn2      v5.8b, v6.8b, v7.8b                     \n"
-      "trn1      v16.8b, v18.8b, v19.8b                  \n"
-      "trn2      v17.8b, v18.8b, v19.8b                  \n"
+      "trn1        v0.8b, v2.8b, v3.8b           \n"
+      "trn2        v1.8b, v2.8b, v3.8b           \n"
+      "trn1        v4.8b, v6.8b, v7.8b           \n"
+      "trn2        v5.8b, v6.8b, v7.8b           \n"
+      "trn1        v16.8b, v18.8b, v19.8b        \n"
+      "trn2        v17.8b, v18.8b, v19.8b        \n"
 
       // 00+10 01+11 02+12 03+13
       // 40+50 41+51 42+52 43+53
-      "uaddlp    v20.4h, v20.8b                          \n"
-      "uaddlp    v21.4h, v21.8b                          \n"
-      "uaddlp    v22.4h, v22.8b                          \n"
-      "uaddlp    v23.4h, v23.8b                          \n"
-      "uaddlp    v24.4h, v24.8b                          \n"
-      "uaddlp    v25.4h, v25.8b                          \n"
+      "uaddlp      v20.4h, v20.8b                \n"
+      "uaddlp      v21.4h, v21.8b                \n"
+      "uaddlp      v22.4h, v22.8b                \n"
+      "uaddlp      v23.4h, v23.8b                \n"
+      "uaddlp      v24.4h, v24.8b                \n"
+      "uaddlp      v25.4h, v25.8b                \n"
 
       // 60+70 61+71 62+72 63+73
-      "uaddlp    v1.4h, v1.8b                            \n"
-      "uaddlp    v5.4h, v5.8b                            \n"
-      "uaddlp    v17.4h, v17.8b                          \n"
+      "uaddlp      v1.4h, v1.8b                  \n"
+      "uaddlp      v5.4h, v5.8b                  \n"
+      "uaddlp      v17.4h, v17.8b                \n"
 
       // combine source lines
-      "add       v20.4h, v20.4h, v22.4h                  \n"
-      "add       v21.4h, v21.4h, v23.4h                  \n"
-      "add       v20.4h, v20.4h, v24.4h                  \n"
-      "add       v21.4h, v21.4h, v25.4h                  \n"
-      "add       v2.4h, v1.4h, v5.4h                     \n"
-      "add       v2.4h, v2.4h, v17.4h                    \n"
+      "add         v20.4h, v20.4h, v22.4h        \n"
+      "add         v21.4h, v21.4h, v23.4h        \n"
+      "add         v20.4h, v20.4h, v24.4h        \n"
+      "add         v21.4h, v21.4h, v25.4h        \n"
+      "add         v2.4h, v1.4h, v5.4h           \n"
+      "add         v2.4h, v2.4h, v17.4h          \n"
 
       // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
       //             + s[6 + st * 1] + s[7 + st * 1]
       //             + s[6 + st * 2] + s[7 + st * 2]) / 6
-      "sqrdmulh  v2.8h, v2.8h, v29.8h                    \n"
-      "xtn       v2.8b,  v2.8h                           \n"
+      "sqrdmulh    v2.8h, v2.8h, v29.8h          \n"
+      "xtn         v2.8b,  v2.8h                 \n"
 
       // Shuffle 2,3 reg around so that 2 can be added to the
       //  0,1 reg and 3 can be added to the 4,5 reg. This
@@ -372,35 +387,38 @@ void OMITFP ScaleRowDown38_3_Box_NEON(const uint8_t* src_ptr,
       //  registers are already expanded. Then do transposes
       //  to get aligned.
       // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
-      "ushll     v16.8h, v16.8b, #0                      \n"
-      "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+      "ushll       v16.8h, v16.8b, #0            \n"
+      "uaddl       v0.8h, v0.8b, v4.8b           \n"
 
       // combine source lines
-      "add       v0.8h, v0.8h, v16.8h                    \n"
+      "add         v0.8h, v0.8h, v16.8h          \n"
 
       // xx 20 xx 21 xx 22 xx 23
       // xx 30 xx 31 xx 32 xx 33
-      "trn1      v1.8h, v0.8h, v0.8h                     \n"
-      "trn2      v4.8h, v0.8h, v0.8h                     \n"
-      "xtn       v0.4h, v1.4s                            \n"
-      "xtn       v4.4h, v4.4s                            \n"
+      "trn1        v1.8h, v0.8h, v0.8h           \n"
+      "trn2        v4.8h, v0.8h, v0.8h           \n"
+      "xtn         v0.4h, v1.4s                  \n"
+      "xtn         v4.4h, v4.4s                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
 
       // 0+1+2, 3+4+5
-      "add       v20.8h, v20.8h, v0.8h                   \n"
-      "add       v21.8h, v21.8h, v4.8h                   \n"
+      "add         v20.8h, v20.8h, v0.8h         \n"
+      "add         v21.8h, v21.8h, v4.8h         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
 
       // Need to divide, but can't downshift as the the value
       //  isn't a power of 2. So multiply by 65536 / n
       //  and take the upper 16 bits.
-      "sqrdmulh  v0.8h, v20.8h, v31.8h                   \n"
-      "sqrdmulh  v1.8h, v21.8h, v31.8h                   \n"
+      "sqrdmulh    v0.8h, v20.8h, v31.8h         \n"
+      "sqrdmulh    v1.8h, v21.8h, v31.8h         \n"
+      "prfm        pldl1keep, [%3, 448]          \n"
 
       // Align for table lookup, vtbl requires registers to be adjacent
-      "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
+      "tbl         v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
 
-      "st1       {v3.8b}, [%1], #8                       \n"
-      "st1       {v3.s}[2], [%1], #4                     \n"
-      "b.gt      1b                                      \n"
+      "st1         {v3.8b}, [%1], #8             \n"
+      "st1         {v3.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),         // %0
         "+r"(dst_ptr),         // %1
         "+r"(tmp_src_stride),  // %2
@@ -422,53 +440,53 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
   // TODO(fbarchard): use src_stride directly for clang 3.5+.
   ptrdiff_t tmp_src_stride = src_stride;
   asm volatile(
-      "ld1       {v30.8h}, [%4]                          \n"
-      "ld1       {v31.16b}, [%5]                         \n"
-      "add       %2, %2, %0                              \n"
-      "1:                                                \n"
+      "ld1         {v30.8h}, [%4]                \n"
+      "ld1         {v31.16b}, [%5]               \n"
+      "add         %2, %2, %0                    \n"
+      "1:                                        \n"
 
       // 00 40 01 41 02 42 03 43
       // 10 50 11 51 12 52 13 53
       // 20 60 21 61 22 62 23 63
       // 30 70 31 71 32 72 33 73
-      "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32    \n"
-      "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32    \n"
-      "subs      %w3, %w3, #12                           \n"
+      "ld4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"
+      "ld4         {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32 \n"
+      "subs        %w3, %w3, #12                 \n"
 
       // Shuffle the input data around to get align the data
       //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
       // 00 10 01 11 02 12 03 13
       // 40 50 41 51 42 52 43 53
-      "trn1      v16.8b, v0.8b, v1.8b                    \n"
-      "trn2      v17.8b, v0.8b, v1.8b                    \n"
-      "trn1      v18.8b, v4.8b, v5.8b                    \n"
-      "trn2      v19.8b, v4.8b, v5.8b                    \n"
+      "trn1        v16.8b, v0.8b, v1.8b          \n"
+      "trn2        v17.8b, v0.8b, v1.8b          \n"
+      "trn1        v18.8b, v4.8b, v5.8b          \n"
+      "trn2        v19.8b, v4.8b, v5.8b          \n"
 
       // 20 30 21 31 22 32 23 33
       // 60 70 61 71 62 72 63 73
-      "trn1      v0.8b, v2.8b, v3.8b                     \n"
-      "trn2      v1.8b, v2.8b, v3.8b                     \n"
-      "trn1      v4.8b, v6.8b, v7.8b                     \n"
-      "trn2      v5.8b, v6.8b, v7.8b                     \n"
+      "trn1        v0.8b, v2.8b, v3.8b           \n"
+      "trn2        v1.8b, v2.8b, v3.8b           \n"
+      "trn1        v4.8b, v6.8b, v7.8b           \n"
+      "trn2        v5.8b, v6.8b, v7.8b           \n"
 
       // 00+10 01+11 02+12 03+13
       // 40+50 41+51 42+52 43+53
-      "uaddlp    v16.4h, v16.8b                          \n"
-      "uaddlp    v17.4h, v17.8b                          \n"
-      "uaddlp    v18.4h, v18.8b                          \n"
-      "uaddlp    v19.4h, v19.8b                          \n"
+      "uaddlp      v16.4h, v16.8b                \n"
+      "uaddlp      v17.4h, v17.8b                \n"
+      "uaddlp      v18.4h, v18.8b                \n"
+      "uaddlp      v19.4h, v19.8b                \n"
 
       // 60+70 61+71 62+72 63+73
-      "uaddlp    v1.4h, v1.8b                            \n"
-      "uaddlp    v5.4h, v5.8b                            \n"
+      "uaddlp      v1.4h, v1.8b                  \n"
+      "uaddlp      v5.4h, v5.8b                  \n"
 
       // combine source lines
-      "add       v16.4h, v16.4h, v18.4h                  \n"
-      "add       v17.4h, v17.4h, v19.4h                  \n"
-      "add       v2.4h, v1.4h, v5.4h                     \n"
+      "add         v16.4h, v16.4h, v18.4h        \n"
+      "add         v17.4h, v17.4h, v19.4h        \n"
+      "add         v2.4h, v1.4h, v5.4h           \n"
 
       // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
-      "uqrshrn   v2.8b, v2.8h, #2                        \n"
+      "uqrshrn     v2.8b, v2.8h, #2              \n"
 
       // Shuffle 2,3 reg around so that 2 can be added to the
       //  0,1 reg and 3 can be added to the 4,5 reg. This
@@ -478,33 +496,35 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
       // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
 
       // combine source lines
-      "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+      "uaddl       v0.8h, v0.8b, v4.8b           \n"
 
       // xx 20 xx 21 xx 22 xx 23
       // xx 30 xx 31 xx 32 xx 33
-      "trn1      v1.8h, v0.8h, v0.8h                     \n"
-      "trn2      v4.8h, v0.8h, v0.8h                     \n"
-      "xtn       v0.4h, v1.4s                            \n"
-      "xtn       v4.4h, v4.4s                            \n"
+      "trn1        v1.8h, v0.8h, v0.8h           \n"
+      "trn2        v4.8h, v0.8h, v0.8h           \n"
+      "xtn         v0.4h, v1.4s                  \n"
+      "xtn         v4.4h, v4.4s                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
 
       // 0+1+2, 3+4+5
-      "add       v16.8h, v16.8h, v0.8h                   \n"
-      "add       v17.8h, v17.8h, v4.8h                   \n"
+      "add         v16.8h, v16.8h, v0.8h         \n"
+      "add         v17.8h, v17.8h, v4.8h         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
 
       // Need to divide, but can't downshift as the the value
       //  isn't a power of 2. So multiply by 65536 / n
       //  and take the upper 16 bits.
-      "sqrdmulh  v0.8h, v16.8h, v30.8h                   \n"
-      "sqrdmulh  v1.8h, v17.8h, v30.8h                   \n"
+      "sqrdmulh    v0.8h, v16.8h, v30.8h         \n"
+      "sqrdmulh    v1.8h, v17.8h, v30.8h         \n"
 
       // Align for table lookup, vtbl requires registers to
       //  be adjacent
 
-      "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
+      "tbl         v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
 
-      "st1       {v3.8b}, [%1], #8                       \n"
-      "st1       {v3.s}[2], [%1], #4                     \n"
-      "b.gt      1b                                      \n"
+      "st1         {v3.8b}, [%1], #8             \n"
+      "st1         {v3.s}[2], [%1], #4           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),         // %0
         "+r"(dst_ptr),         // %1
         "+r"(tmp_src_stride),  // %2
@@ -515,6 +535,488 @@ void ScaleRowDown38_2_Box_NEON(const uint8_t* src_ptr,
         "v19", "v30", "v31", "memory", "cc");
 }
 
+void ScaleRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 01234567
+      "ldr         d1, [%1], #8                  \n"  // 12345678
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 12345678 (16b)
+
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (even)
+
+      "rshrn       v2.8b, v2.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.8b, v3.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.8b, v2.8b}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8_t* dst_ptr,
+                               ptrdiff_t dst_stride,
+                               int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 1;
+  const uint8_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+      "movi        v30.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 01234567
+      "ldr         d1, [%2], #8                  \n"  // 12345678
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 12345678 (16b)
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (1, odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v4.8h, v0.8b, #0              \n"  // 01234567 (16b)
+      "ushll       v5.8h, v1.8b, #0              \n"  // 12345678 (16b)
+      "umlal       v4.8h, v1.8b, v31.8b          \n"  // 3*near+far (2, odd)
+      "umlal       v5.8h, v0.8b, v31.8b          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v30.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v30.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v30.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v30.8h          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v2.8b, v2.8h, #4              \n"  // 2, odd
+      "rshrn       v1.8b, v3.8h, #4              \n"  // 2, even
+      "rshrn       v4.8b, v4.8h, #4              \n"  // 1, odd
+      "rshrn       v3.8b, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v1.8b, v2.8b}, [%5], #16     \n"  // store 1
+      "st2         {v3.8b, v4.8b}, [%4], #16     \n"  // store 2
+      "subs        %w6, %w6, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_12_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v2.16b, v0.16b                \n"
+      "mla         v0.8h, v1.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v1.8h, v2.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "urshr       v2.8h, v0.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "urshr       v1.8h, v1.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.8h, v2.8h}, [%2], #32     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_12_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v2.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v3.8h}, [%2], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v0.16b, v2.16b                \n"
+      "mla         v2.8h, v3.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v3.8h, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "ld1         {v4.8h}, [%1], #16            \n"  // 01234567 (16b)
+      "ld1         {v5.8h}, [%3], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "mov         v0.16b, v4.16b                \n"
+      "mla         v4.8h, v5.8h, v31.8h          \n"  // 3*near+far (odd)
+      "mla         v5.8h, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v31.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v31.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v31.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v31.8h          \n"  // 9 3 3 1 (2, even)
+
+      "urshr       v2.8h, v2.8h, #4              \n"  // 2, odd
+      "urshr       v1.8h, v3.8h, #4              \n"  // 2, even
+      "urshr       v4.8h, v4.8h, #4              \n"  // 1, odd
+      "urshr       v3.8h, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v3.8h, v4.8h}, [%4], #32     \n"  // store 1
+      "st2         {v1.8h, v2.8h}, [%5], #32     \n"  // store 2
+
+      "subs        %w6, %w6, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                uint16_t* dst_ptr,
+                                int dst_width) {
+  const uint16_t* src_temp = src_ptr + 1;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll2      v3.4s, v0.8h, #0              \n"  // 4567 (32b)
+      "ushll       v4.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "ushll2      v5.4s, v1.8h, #0              \n"  // 5678 (32b)
+
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal2      v3.4s, v1.8h, v31.8h          \n"  // 3*near+far (2, odd)
+      "umlal       v4.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+      "umlal2      v5.4s, v0.8h, v31.8h          \n"  // 3*near+far (2, even)
+
+      "rshrn       v0.4h, v4.4s, #2              \n"  // 3/4*near+1/4*far
+      "rshrn2      v0.8h, v5.4s, #2              \n"  // 3/4*near+1/4*far (even)
+      "rshrn       v1.4h, v2.4s, #2              \n"  // 3/4*near+1/4*far
+      "rshrn2      v1.8h, v3.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+
+      "st2         {v0.8h, v1.8h}, [%2], #32     \n"  // store
+      "subs        %w3, %w3, #16                 \n"  // 8 sample -> 16 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v31"  // Clobber List
+  );
+}
+
+void ScaleRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                  ptrdiff_t src_stride,
+                                  uint16_t* dst_ptr,
+                                  ptrdiff_t dst_stride,
+                                  int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 1;
+  const uint16_t* src_temp1 = src_ptr1 + 1;
+
+  asm volatile(
+      "movi        v31.4h, #3                    \n"
+      "movi        v30.4s, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 0123 (16b)
+      "ldr         d1, [%2], #8                  \n"  // 1234 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"  // 0123 (16b)
+      "ldr         d1, [%3], #8                  \n"  // 1234 (16b)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v4.4s, v0.4h, #0              \n"  // 0123 (32b)
+      "ushll       v5.4s, v1.4h, #0              \n"  // 1234 (32b)
+      "umlal       v4.4s, v1.4h, v31.4h          \n"  // 3*near+far (2, odd)
+      "umlal       v5.4s, v0.4h, v31.4h          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.4s, v2.4s, v30.4s          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.4s, v3.4s, v30.4s          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.4s, v0.4s, v30.4s          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.4s, v1.4s, v30.4s          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v1.4h, v4.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v0.4h, v5.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v5.4h, v2.4s, #4              \n"  // 3/4*near+1/4*far
+      "rshrn       v4.4h, v3.4s, #4              \n"  // 3/4*near+1/4*far
+
+      "st2         {v0.4h, v1.4h}, [%4], #16     \n"  // store 1
+      "st2         {v4.4h, v5.4h}, [%5], #16     \n"  // store 2
+
+      "subs        %w6, %w6, #8                  \n"  // 4 sample -> 8 sample
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_NEON(const uint8_t* src_ptr,
+                               uint8_t* dst_ptr,
+                               int dst_width) {
+  const uint8_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"  // 00112233 (1u1v)
+      "ldr         d1, [%1], #8                  \n"  // 11223344 (1u1v)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"  // 00112233 (1u1v, 16b)
+      "ushll       v3.8h, v1.8b, #0              \n"  // 11223344 (1u1v, 16b)
+
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (even)
+
+      "rshrn       v2.8b, v2.8h, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.8b, v3.8h, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.4h, v2.4h}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_NEON(const uint8_t* src_ptr,
+                                 ptrdiff_t src_stride,
+                                 uint8_t* dst_ptr,
+                                 ptrdiff_t dst_stride,
+                                 int dst_width) {
+  const uint8_t* src_ptr1 = src_ptr + src_stride;
+  uint8_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint8_t* src_temp = src_ptr + 2;
+  const uint8_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "movi        v31.8b, #3                    \n"
+      "movi        v30.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"
+      "ldr         d1, [%2], #8                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.8h, v0.8b, #0              \n"
+      "ushll       v3.8h, v1.8b, #0              \n"
+      "umlal       v2.8h, v1.8b, v31.8b          \n"  // 3*near+far (1, odd)
+      "umlal       v3.8h, v0.8b, v31.8b          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v4.8h, v0.8b, #0              \n"
+      "ushll       v5.8h, v1.8b, #0              \n"
+      "umlal       v4.8h, v1.8b, v31.8b          \n"  // 3*near+far (2, odd)
+      "umlal       v5.8h, v0.8b, v31.8b          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.8h, v2.8h, v30.8h          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.8h, v3.8h, v30.8h          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.8h, v0.8h, v30.8h          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.8h, v1.8h, v30.8h          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v2.8b, v2.8h, #4              \n"  // 2, odd
+      "rshrn       v1.8b, v3.8h, #4              \n"  // 2, even
+      "rshrn       v4.8b, v4.8h, #4              \n"  // 1, odd
+      "rshrn       v3.8b, v5.8h, #4              \n"  // 1, even
+
+      "st2         {v1.4h, v2.4h}, [%5], #16     \n"  // store 2
+      "st2         {v3.4h, v4.4h}, [%4], #16     \n"  // store 1
+      "subs        %w6, %w6, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Linear_16_NEON(const uint16_t* src_ptr,
+                                  uint16_t* dst_ptr,
+                                  int dst_width) {
+  const uint16_t* src_temp = src_ptr + 2;
+  asm volatile(
+      "movi        v31.8h, #3                    \n"
+
+      "1:                                        \n"
+      "ld1         {v0.8h}, [%0], #16            \n"  // 01234567 (16b)
+      "ld1         {v1.8h}, [%1], #16            \n"  // 12345678 (16b)
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "ushll2      v4.4s, v0.8h, #0              \n"  // 2233 (1u1v, 32b)
+      "ushll2      v5.4s, v1.8h, #0              \n"  // 3344 (1u1v, 32b)
+
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (even)
+      "umlal2      v4.4s, v1.8h, v31.8h          \n"  // 3*near+far (odd)
+      "umlal2      v5.4s, v0.8h, v31.8h          \n"  // 3*near+far (even)
+
+      "rshrn       v2.4h, v2.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v1.4h, v3.4s, #2              \n"  // 3/4*near+1/4*far (even)
+      "rshrn       v4.4h, v4.4s, #2              \n"  // 3/4*near+1/4*far (odd)
+      "rshrn       v3.4h, v5.4s, #2              \n"  // 3/4*near+1/4*far (even)
+
+      "st2         {v1.2s, v2.2s}, [%2], #16     \n"  // store
+      "st2         {v3.2s, v4.2s}, [%2], #16     \n"  // store
+      "subs        %w3, %w3, #8                  \n"  // 4 uv -> 8 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),   // %0
+        "+r"(src_temp),  // %1
+        "+r"(dst_ptr),   // %2
+        "+r"(dst_width)  // %3
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+        "v31"  // Clobber List
+  );
+}
+
+void ScaleUVRowUp2_Bilinear_16_NEON(const uint16_t* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint16_t* dst_ptr,
+                                    ptrdiff_t dst_stride,
+                                    int dst_width) {
+  const uint16_t* src_ptr1 = src_ptr + src_stride;
+  uint16_t* dst_ptr1 = dst_ptr + dst_stride;
+  const uint16_t* src_temp = src_ptr + 2;
+  const uint16_t* src_temp1 = src_ptr1 + 2;
+
+  asm volatile(
+      "movi        v31.4h, #3                    \n"
+      "movi        v30.4s, #3                    \n"
+
+      "1:                                        \n"
+      "ldr         d0, [%0], #8                  \n"
+      "ldr         d1, [%2], #8                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v2.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v3.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "umlal       v2.4s, v1.4h, v31.4h          \n"  // 3*near+far (1, odd)
+      "umlal       v3.4s, v0.4h, v31.4h          \n"  // 3*near+far (1, even)
+
+      "ldr         d0, [%1], #8                  \n"
+      "ldr         d1, [%3], #8                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "ushll       v4.4s, v0.4h, #0              \n"  // 0011 (1u1v, 32b)
+      "ushll       v5.4s, v1.4h, #0              \n"  // 1122 (1u1v, 32b)
+      "umlal       v4.4s, v1.4h, v31.4h          \n"  // 3*near+far (2, odd)
+      "umlal       v5.4s, v0.4h, v31.4h          \n"  // 3*near+far (2, even)
+
+      "mov         v0.16b, v4.16b                \n"
+      "mov         v1.16b, v5.16b                \n"
+      "mla         v4.4s, v2.4s, v30.4s          \n"  // 9 3 3 1 (1, odd)
+      "mla         v5.4s, v3.4s, v30.4s          \n"  // 9 3 3 1 (1, even)
+      "mla         v2.4s, v0.4s, v30.4s          \n"  // 9 3 3 1 (2, odd)
+      "mla         v3.4s, v1.4s, v30.4s          \n"  // 9 3 3 1 (2, even)
+
+      "rshrn       v1.4h, v2.4s, #4              \n"  // 2, odd
+      "rshrn       v0.4h, v3.4s, #4              \n"  // 2, even
+      "rshrn       v3.4h, v4.4s, #4              \n"  // 1, odd
+      "rshrn       v2.4h, v5.4s, #4              \n"  // 1, even
+
+      "st2         {v0.2s, v1.2s}, [%5], #16     \n"  // store 2
+      "st2         {v2.2s, v3.2s}, [%4], #16     \n"  // store 1
+      "subs        %w6, %w6, #4                  \n"  // 2 uv -> 4 uv
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),    // %0
+        "+r"(src_ptr1),   // %1
+        "+r"(src_temp),   // %2
+        "+r"(src_temp1),  // %3
+        "+r"(dst_ptr),    // %4
+        "+r"(dst_ptr1),   // %5
+        "+r"(dst_width)   // %6
+      :
+      : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v30",
+        "v31"  // Clobber List
+  );
+}
+
 // Add a row of bytes to a row of shorts.  Used for box filter.
 // Reads 16 bytes and accumulates to 16 shorts at a time.
 void ScaleAddRow_NEON(const uint8_t* src_ptr,
@@ -522,13 +1024,14 @@ void ScaleAddRow_NEON(const uint8_t* src_ptr,
                       int src_width) {
   asm volatile(
       "1:                                        \n"
-      "ld1      {v1.8h, v2.8h}, [%1]             \n"  // load accumulator
-      "ld1      {v0.16b}, [%0], #16              \n"  // load 16 bytes
-      "uaddw2   v2.8h, v2.8h, v0.16b             \n"  // add
-      "uaddw    v1.8h, v1.8h, v0.8b              \n"
-      "st1      {v1.8h, v2.8h}, [%1], #32        \n"  // store accumulator
-      "subs     %w2, %w2, #16                    \n"  // 16 processed per loop
-      "b.gt     1b                               \n"
+      "ld1         {v1.8h, v2.8h}, [%1]          \n"  // load accumulator
+      "ld1         {v0.16b}, [%0], #16           \n"  // load 16 bytes
+      "uaddw2      v2.8h, v2.8h, v0.16b          \n"  // add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddw       v1.8h, v1.8h, v0.8b           \n"
+      "st1         {v1.8h, v2.8h}, [%1], #32     \n"  // store accumulator
+      "subs        %w2, %w2, #16                 \n"  // 16 processed per loop
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst_ptr),   // %1
         "+r"(src_width)  // %2
@@ -560,17 +1063,17 @@ void ScaleFilterCols_NEON(uint8_t* dst_ptr,
   int64_t x64 = (int64_t)x;    // NOLINT
   int64_t dx64 = (int64_t)dx;  // NOLINT
   asm volatile (
-    "dup        v0.4s, %w3                     \n"  // x
-    "dup        v1.4s, %w4                     \n"  // dx
-    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
-    "shl        v3.4s, v1.4s, #2               \n"  // 4 * dx
-    "mul        v1.4s, v1.4s, v2.4s            \n"
+      "dup         v0.4s, %w3                    \n"  // x
+      "dup         v1.4s, %w4                    \n"  // dx
+      "ld1         {v2.4s}, [%5]                 \n"  // 0 1 2 3
+      "shl         v3.4s, v1.4s, #2              \n"  // 4 * dx
+      "mul         v1.4s, v1.4s, v2.4s           \n"
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "add        v1.4s, v1.4s, v0.4s            \n"
+      "add         v1.4s, v1.4s, v0.4s           \n"
     // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
-    "add        v2.4s, v1.4s, v3.4s            \n"
-    "shl        v0.4s, v3.4s, #1               \n"  // 8 * dx
-  "1:                                          \n"
+      "add         v2.4s, v1.4s, v3.4s           \n"
+      "shl         v0.4s, v3.4s, #1              \n"  // 8 * dx
+      "1:                                        \n"
     LOAD2_DATA8_LANE(0)
     LOAD2_DATA8_LANE(1)
     LOAD2_DATA8_LANE(2)
@@ -579,27 +1082,27 @@ void ScaleFilterCols_NEON(uint8_t* dst_ptr,
     LOAD2_DATA8_LANE(5)
     LOAD2_DATA8_LANE(6)
     LOAD2_DATA8_LANE(7)
-    "mov       v6.16b, v1.16b                  \n"
-    "mov       v7.16b, v2.16b                  \n"
-    "uzp1      v6.8h, v6.8h, v7.8h             \n"
-    "ushll     v4.8h, v4.8b, #0                \n"
-    "ushll     v5.8h, v5.8b, #0                \n"
-    "ssubl     v16.4s, v5.4h, v4.4h            \n"
-    "ssubl2    v17.4s, v5.8h, v4.8h            \n"
-    "ushll     v7.4s, v6.4h, #0                \n"
-    "ushll2    v6.4s, v6.8h, #0                \n"
-    "mul       v16.4s, v16.4s, v7.4s           \n"
-    "mul       v17.4s, v17.4s, v6.4s           \n"
-    "rshrn     v6.4h, v16.4s, #16              \n"
-    "rshrn2    v6.8h, v17.4s, #16              \n"
-    "add       v4.8h, v4.8h, v6.8h             \n"
-    "xtn       v4.8b, v4.8h                    \n"
-
-    "st1       {v4.8b}, [%0], #8               \n"  // store pixels
-    "add       v1.4s, v1.4s, v0.4s             \n"
-    "add       v2.4s, v2.4s, v0.4s             \n"
-    "subs      %w2, %w2, #8                    \n"  // 8 processed per loop
-    "b.gt      1b                              \n"
+      "mov         v6.16b, v1.16b                \n"
+      "mov         v7.16b, v2.16b                \n"
+      "uzp1        v6.8h, v6.8h, v7.8h           \n"
+      "ushll       v4.8h, v4.8b, #0              \n"
+      "ushll       v5.8h, v5.8b, #0              \n"
+      "ssubl       v16.4s, v5.4h, v4.4h          \n"
+      "ssubl2      v17.4s, v5.8h, v4.8h          \n"
+      "ushll       v7.4s, v6.4h, #0              \n"
+      "ushll2      v6.4s, v6.8h, #0              \n"
+      "mul         v16.4s, v16.4s, v7.4s         \n"
+      "mul         v17.4s, v17.4s, v6.4s         \n"
+      "rshrn       v6.4h, v16.4s, #16            \n"
+      "rshrn2      v6.8h, v17.4s, #16            \n"
+      "add         v4.8h, v4.8h, v6.8h           \n"
+      "xtn         v4.8b, v4.8h                  \n"
+
+      "st1         {v4.8b}, [%0], #8             \n"  // store pixels
+      "add         v1.4s, v1.4s, v0.4s           \n"
+      "add         v2.4s, v2.4s, v0.4s           \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "b.gt        1b                            \n"
   : "+r"(dst_ptr),          // %0
     "+r"(src_ptr),          // %1
     "+r"(dst_width),        // %2
@@ -623,74 +1126,83 @@ void ScaleFilterRows_NEON(uint8_t* dst_ptr,
                           int source_y_fraction) {
   int y_fraction = 256 - source_y_fraction;
   asm volatile(
-      "cmp          %w4, #0                      \n"
-      "b.eq         100f                         \n"
-      "add          %2, %2, %1                   \n"
-      "cmp          %w4, #64                     \n"
-      "b.eq         75f                          \n"
-      "cmp          %w4, #128                    \n"
-      "b.eq         50f                          \n"
-      "cmp          %w4, #192                    \n"
-      "b.eq         25f                          \n"
-
-      "dup          v5.8b, %w4                   \n"
-      "dup          v4.8b, %w5                   \n"
+      "cmp         %w4, #0                       \n"
+      "b.eq        100f                          \n"
+      "add         %2, %2, %1                    \n"
+      "cmp         %w4, #64                      \n"
+      "b.eq        75f                           \n"
+      "cmp         %w4, #128                     \n"
+      "b.eq        50f                           \n"
+      "cmp         %w4, #192                     \n"
+      "b.eq        25f                           \n"
+
+      "dup         v5.8b, %w4                    \n"
+      "dup         v4.8b, %w5                    \n"
       // General purpose row blend.
       "1:                                        \n"
-      "ld1          {v0.16b}, [%1], #16          \n"
-      "ld1          {v1.16b}, [%2], #16          \n"
-      "subs         %w3, %w3, #16                \n"
-      "umull        v6.8h, v0.8b, v4.8b          \n"
-      "umull2       v7.8h, v0.16b, v4.16b        \n"
-      "umlal        v6.8h, v1.8b, v5.8b          \n"
-      "umlal2       v7.8h, v1.16b, v5.16b        \n"
-      "rshrn        v0.8b, v6.8h, #8             \n"
-      "rshrn2       v0.16b, v7.8h, #8            \n"
-      "st1          {v0.16b}, [%0], #16          \n"
-      "b.gt         1b                           \n"
-      "b            99f                          \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "umull       v6.8h, v0.8b, v4.8b           \n"
+      "umull2      v7.8h, v0.16b, v4.16b         \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "umlal       v6.8h, v1.8b, v5.8b           \n"
+      "umlal2      v7.8h, v1.16b, v5.16b         \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "rshrn       v0.8b, v6.8h, #8              \n"
+      "rshrn2      v0.16b, v7.8h, #8             \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        1b                            \n"
+      "b           99f                           \n"
 
       // Blend 25 / 75.
       "25:                                       \n"
-      "ld1          {v0.16b}, [%1], #16          \n"
-      "ld1          {v1.16b}, [%2], #16          \n"
-      "subs         %w3, %w3, #16                \n"
-      "urhadd       v0.16b, v0.16b, v1.16b       \n"
-      "urhadd       v0.16b, v0.16b, v1.16b       \n"
-      "st1          {v0.16b}, [%0], #16          \n"
-      "b.gt         25b                          \n"
-      "b            99f                          \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        25b                           \n"
+      "b           99f                           \n"
 
       // Blend 50 / 50.
       "50:                                       \n"
-      "ld1          {v0.16b}, [%1], #16          \n"
-      "ld1          {v1.16b}, [%2], #16          \n"
-      "subs         %w3, %w3, #16                \n"
-      "urhadd       v0.16b, v0.16b, v1.16b       \n"
-      "st1          {v0.16b}, [%0], #16          \n"
-      "b.gt         50b                          \n"
-      "b            99f                          \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "ld1         {v1.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        50b                           \n"
+      "b           99f                           \n"
 
       // Blend 75 / 25.
       "75:                                       \n"
-      "ld1          {v1.16b}, [%1], #16          \n"
-      "ld1          {v0.16b}, [%2], #16          \n"
-      "subs         %w3, %w3, #16                \n"
-      "urhadd       v0.16b, v0.16b, v1.16b       \n"
-      "urhadd       v0.16b, v0.16b, v1.16b       \n"
-      "st1          {v0.16b}, [%0], #16          \n"
-      "b.gt         75b                          \n"
-      "b            99f                          \n"
+      "ld1         {v1.16b}, [%1], #16           \n"
+      "ld1         {v0.16b}, [%2], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"
+      "prfm        pldl1keep, [%2, 448]          \n"
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        75b                           \n"
+      "b           99f                           \n"
 
       // Blend 100 / 0 - Copy row unchanged.
       "100:                                      \n"
-      "ld1          {v0.16b}, [%1], #16          \n"
-      "subs         %w3, %w3, #16                \n"
-      "st1          {v0.16b}, [%0], #16          \n"
-      "b.gt         100b                         \n"
+      "ld1         {v0.16b}, [%1], #16           \n"
+      "subs        %w3, %w3, #16                 \n"
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%0], #16           \n"
+      "b.gt        100b                          \n"
 
       "99:                                       \n"
-      "st1          {v0.b}[15], [%0]             \n"
+      "st1         {v0.b}[15], [%0]              \n"
       : "+r"(dst_ptr),            // %0
         "+r"(src_ptr),            // %1
         "+r"(src_stride),         // %2
@@ -709,11 +1221,12 @@ void ScaleARGBRowDown2_NEON(const uint8_t* src_ptr,
   asm volatile(
       "1:                                        \n"
       // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3
-      "ld4        {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "mov        v2.16b, v3.16b                 \n"
-      "st2        {v1.4s,v2.4s}, [%1], #32       \n"  // store 8 odd pixels
-      "b.gt       1b                             \n"
+      "ld4         {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "mov         v2.16b, v3.16b                \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st2         {v1.4s,v2.4s}, [%1], #32      \n"  // store 8 odd pixels
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),   // %0
         "+r"(dst),       // %1
         "+r"(dst_width)  // %2
@@ -730,13 +1243,14 @@ void ScaleARGBRowDown2Linear_NEON(const uint8_t* src_argb,
   asm volatile(
       "1:                                        \n"
       // load 16 ARGB pixels with even pixels into q0/q2, odd into q1/q3
-      "ld4        {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-
-      "urhadd     v0.16b, v0.16b, v1.16b         \n"  // rounding half add
-      "urhadd     v1.16b, v2.16b, v3.16b         \n"
-      "st2        {v0.4s,v1.4s}, [%1], #32       \n"  // store 8 pixels
-      "b.gt       1b                             \n"
+      "ld4         {v0.4s,v1.4s,v2.4s,v3.4s}, [%0], #64 \n"
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+
+      "urhadd      v0.16b, v0.16b, v1.16b        \n"  // rounding half add
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "urhadd      v1.16b, v2.16b, v3.16b        \n"
+      "st2         {v0.4s,v1.4s}, [%1], #32      \n"  // store 8 pixels
+      "b.gt        1b                            \n"
       : "+r"(src_argb),  // %0
         "+r"(dst_argb),  // %1
         "+r"(dst_width)  // %2
@@ -751,25 +1265,27 @@ void ScaleARGBRowDown2Box_NEON(const uint8_t* src_ptr,
                                int dst_width) {
   asm volatile(
       // change the stride to row 2 pointer
-      "add        %1, %1, %0                     \n"
+      "add         %1, %1, %0                    \n"
       "1:                                        \n"
-      "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 8 ARGB
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
-      "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
-      "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
-      "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
-      "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
-      "ld4        {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8
-      "uadalp     v0.8h, v16.16b                 \n"  // B 16 bytes -> 8 shorts.
-      "uadalp     v1.8h, v17.16b                 \n"  // G 16 bytes -> 8 shorts.
-      "uadalp     v2.8h, v18.16b                 \n"  // R 16 bytes -> 8 shorts.
-      "uadalp     v3.8h, v19.16b                 \n"  // A 16 bytes -> 8 shorts.
-      "rshrn      v0.8b, v0.8h, #2               \n"  // round and pack
-      "rshrn      v1.8b, v1.8h, #2               \n"
-      "rshrn      v2.8b, v2.8h, #2               \n"
-      "rshrn      v3.8b, v3.8h, #2               \n"
-      "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32     \n"
-      "b.gt       1b                             \n"
+      "ld4         {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 ARGB
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uaddlp      v0.8h, v0.16b                 \n"  // B 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v1.16b                 \n"  // G 16 bytes -> 8 shorts.
+      "uaddlp      v2.8h, v2.16b                 \n"  // R 16 bytes -> 8 shorts.
+      "uaddlp      v3.8h, v3.16b                 \n"  // A 16 bytes -> 8 shorts.
+      "ld4         {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8
+      "uadalp      v0.8h, v16.16b                \n"  // B 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v17.16b                \n"  // G 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v2.8h, v18.16b                \n"  // R 16 bytes -> 8 shorts.
+      "uadalp      v3.8h, v19.16b                \n"  // A 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "rshrn       v1.8b, v1.8h, #2              \n"
+      "rshrn       v2.8b, v2.8h, #2              \n"
+      "rshrn       v3.8b, v3.8h, #2              \n"
+      "st4         {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -788,13 +1304,14 @@ void ScaleARGBRowDownEven_NEON(const uint8_t* src_argb,
   (void)src_stride;
   asm volatile(
       "1:                                        \n"
-      "ld1        {v0.s}[0], [%0], %3            \n"
-      "ld1        {v0.s}[1], [%0], %3            \n"
-      "ld1        {v0.s}[2], [%0], %3            \n"
-      "ld1        {v0.s}[3], [%0], %3            \n"
-      "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
-      "st1        {v0.16b}, [%1], #16            \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.s}[0], [%0], %3           \n"
+      "ld1         {v0.s}[1], [%0], %3           \n"
+      "ld1         {v0.s}[2], [%0], %3           \n"
+      "ld1         {v0.s}[3], [%0], %3           \n"
+      "subs        %w2, %w2, #4                  \n"  // 4 pixels per loop.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "st1         {v0.16b}, [%1], #16           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_argb),                // %0
         "+r"(dst_argb),                // %1
         "+r"(dst_width)                // %2
@@ -812,33 +1329,35 @@ void ScaleARGBRowDownEvenBox_NEON(const uint8_t* src_argb,
                                   uint8_t* dst_argb,
                                   int dst_width) {
   asm volatile(
-      "add        %1, %1, %0                     \n"
+      "add         %1, %1, %0                    \n"
       "1:                                        \n"
-      "ld1        {v0.8b}, [%0], %4              \n"  // Read 4 2x2 -> 2x1
-      "ld1        {v1.8b}, [%1], %4              \n"
-      "ld1        {v2.8b}, [%0], %4              \n"
-      "ld1        {v3.8b}, [%1], %4              \n"
-      "ld1        {v4.8b}, [%0], %4              \n"
-      "ld1        {v5.8b}, [%1], %4              \n"
-      "ld1        {v6.8b}, [%0], %4              \n"
-      "ld1        {v7.8b}, [%1], %4              \n"
-      "uaddl      v0.8h, v0.8b, v1.8b            \n"
-      "uaddl      v2.8h, v2.8b, v3.8b            \n"
-      "uaddl      v4.8h, v4.8b, v5.8b            \n"
-      "uaddl      v6.8h, v6.8b, v7.8b            \n"
-      "mov        v16.d[1], v0.d[1]              \n"  // ab_cd -> ac_bd
-      "mov        v0.d[1], v2.d[0]               \n"
-      "mov        v2.d[0], v16.d[1]              \n"
-      "mov        v16.d[1], v4.d[1]              \n"  // ef_gh -> eg_fh
-      "mov        v4.d[1], v6.d[0]               \n"
-      "mov        v6.d[0], v16.d[1]              \n"
-      "add        v0.8h, v0.8h, v2.8h            \n"  // (a+b)_(c+d)
-      "add        v4.8h, v4.8h, v6.8h            \n"  // (e+f)_(g+h)
-      "rshrn      v0.8b, v0.8h, #2               \n"  // first 2 pixels.
-      "rshrn2     v0.16b, v4.8h, #2              \n"  // next 2 pixels.
-      "subs       %w3, %w3, #4                   \n"  // 4 pixels per loop.
-      "st1     {v0.16b}, [%2], #16               \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.8b}, [%0], %4             \n"  // Read 4 2x2 -> 2x1
+      "ld1         {v1.8b}, [%1], %4             \n"
+      "ld1         {v2.8b}, [%0], %4             \n"
+      "ld1         {v3.8b}, [%1], %4             \n"
+      "ld1         {v4.8b}, [%0], %4             \n"
+      "ld1         {v5.8b}, [%1], %4             \n"
+      "ld1         {v6.8b}, [%0], %4             \n"
+      "ld1         {v7.8b}, [%1], %4             \n"
+      "uaddl       v0.8h, v0.8b, v1.8b           \n"
+      "uaddl       v2.8h, v2.8b, v3.8b           \n"
+      "uaddl       v4.8h, v4.8b, v5.8b           \n"
+      "uaddl       v6.8h, v6.8b, v7.8b           \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "mov         v16.d[1], v0.d[1]             \n"  // ab_cd -> ac_bd
+      "mov         v0.d[1], v2.d[0]              \n"
+      "mov         v2.d[0], v16.d[1]             \n"
+      "mov         v16.d[1], v4.d[1]             \n"  // ef_gh -> eg_fh
+      "mov         v4.d[1], v6.d[0]              \n"
+      "mov         v6.d[0], v16.d[1]             \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "add         v0.8h, v0.8h, v2.8h           \n"  // (a+b)_(c+d)
+      "add         v4.8h, v4.8h, v6.8h           \n"  // (e+f)_(g+h)
+      "rshrn       v0.8b, v0.8h, #2              \n"  // first 2 pixels.
+      "rshrn2      v0.16b, v4.8h, #2             \n"  // next 2 pixels.
+      "subs        %w3, %w3, #4                  \n"  // 4 pixels per loop.
+      "st1         {v0.16b}, [%2], #16           \n"
+      "b.gt        1b                            \n"
       : "+r"(src_argb),                // %0
         "+r"(src_stride),              // %1
         "+r"(dst_argb),                // %2
@@ -875,10 +1394,11 @@ void ScaleARGBCols_NEON(uint8_t* dst_argb,
       LOAD1_DATA32_LANE(v1, 1)
       LOAD1_DATA32_LANE(v1, 2)
       LOAD1_DATA32_LANE(v1, 3)
+      "prfm        pldl1keep, [%1, 448]          \n"  // prefetch 7 lines ahead
       // clang-format on
-      "st1        {v0.4s, v1.4s}, [%0], #32      \n"  // store pixels
-      "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
-      "b.gt       1b                             \n"
+      "st1         {v0.4s, v1.4s}, [%0], #32     \n"  // store pixels
+      "subs        %w2, %w2, #8                  \n"  // 8 processed per loop
+      "b.gt        1b                            \n"
       : "+r"(dst_argb),   // %0
         "+r"(src_argb),   // %1
         "+r"(dst_width),  // %2
@@ -911,16 +1431,16 @@ void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
   int64_t x64 = (int64_t)x;    // NOLINT
   int64_t dx64 = (int64_t)dx;  // NOLINT
   asm volatile (
-    "dup        v0.4s, %w3                     \n"  // x
-    "dup        v1.4s, %w4                     \n"  // dx
-    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
-    "shl        v6.4s, v1.4s, #2               \n"  // 4 * dx
-    "mul        v1.4s, v1.4s, v2.4s            \n"
-    "movi       v3.16b, #0x7f                  \n"  // 0x7F
-    "movi       v4.8h, #0x7f                   \n"  // 0x7F
+      "dup         v0.4s, %w3                    \n"  // x
+      "dup         v1.4s, %w4                    \n"  // dx
+      "ld1         {v2.4s}, [%5]                 \n"  // 0 1 2 3
+      "shl         v6.4s, v1.4s, #2              \n"  // 4 * dx
+      "mul         v1.4s, v1.4s, v2.4s           \n"
+      "movi        v3.16b, #0x7f                 \n"  // 0x7F
+      "movi        v4.8h, #0x7f                  \n"  // 0x7F
     // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
-    "add        v5.4s, v1.4s, v0.4s            \n"
-  "1:                                          \n"
+      "add         v5.4s, v1.4s, v0.4s           \n"
+      "1:                                        \n"
     // d0, d1: a
     // d2, d3: b
     LOAD2_DATA32_LANE(v0, v1, 0)
@@ -941,15 +1461,15 @@ void ScaleARGBFilterCols_NEON(uint8_t* dst_argb,
     "umull2     v17.8h, v0.16b, v7.16b         \n"
     "umull      v18.8h, v1.8b, v2.8b           \n"
     "umull2     v19.8h, v1.16b, v2.16b         \n"
+    "prfm       pldl1keep, [%1, 448]           \n"  // prefetch 7 lines ahead
     "add        v16.8h, v16.8h, v18.8h         \n"
     "add        v17.8h, v17.8h, v19.8h         \n"
     "shrn       v0.8b, v16.8h, #7              \n"
     "shrn2      v0.16b, v17.8h, #7             \n"
-
     "st1     {v0.4s}, [%0], #16                \n"  // store pixels
     "add     v5.4s, v5.4s, v6.4s               \n"
     "subs    %w2, %w2, #4                      \n"  // 4 processed per loop
-    "b.gt    1b                                \n"
+    "b.gt       1b                             \n"
   : "+r"(dst_argb),         // %0
     "+r"(src_argb),         // %1
     "+r"(dst_width),        // %2
@@ -972,19 +1492,21 @@ void ScaleRowDown2Box_16_NEON(const uint16_t* src_ptr,
                               int dst_width) {
   asm volatile(
       // change the stride to row 2 pointer
-      "add        %1, %0, %1, lsl #1             \n"  // ptr + stide * 2
+      "add         %1, %0, %1, lsl #1            \n"  // ptr + stide * 2
       "1:                                        \n"
-      "ld1        {v0.8h, v1.8h}, [%0], #32      \n"  // load row 1 and post inc
-      "ld1        {v2.8h, v3.8h}, [%1], #32      \n"  // load row 2 and post inc
-      "subs       %w3, %w3, #8                   \n"  // 8 processed per loop
-      "uaddlp     v0.4s, v0.8h                   \n"  // row 1 add adjacent
-      "uaddlp     v1.4s, v1.8h                   \n"
-      "uadalp     v0.4s, v2.8h                   \n"  // +row 2 add adjacent
-      "uadalp     v1.4s, v3.8h                   \n"
-      "rshrn      v0.4h, v0.4s, #2               \n"  // round and pack
-      "rshrn2     v0.8h, v1.4s, #2               \n"
-      "st1        {v0.8h}, [%2], #16             \n"
-      "b.gt       1b                             \n"
+      "ld1         {v0.8h, v1.8h}, [%0], #32     \n"  // load row 1 and post inc
+      "ld1         {v2.8h, v3.8h}, [%1], #32     \n"  // load row 2 and post inc
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop
+      "uaddlp      v0.4s, v0.8h                  \n"  // row 1 add adjacent
+      "uaddlp      v1.4s, v1.8h                  \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uadalp      v0.4s, v2.8h                  \n"  // +row 2 add adjacent
+      "uadalp      v1.4s, v3.8h                  \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v0.4h, v0.4s, #2              \n"  // round and pack
+      "rshrn2      v0.8h, v1.4s, #2              \n"
+      "st1         {v0.8h}, [%2], #16            \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -1001,38 +1523,40 @@ void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
                          uint16_t* dst,
                          int dst_width) {
   asm volatile(
-      "add        %1, %0, %1, lsl #1             \n"  // ptr + stide * 2
-      "movi       v0.8h, #9                      \n"  // constants
-      "movi       v1.4s, #3                      \n"
+      "add         %1, %0, %1, lsl #1            \n"  // ptr + stide * 2
+      "movi        v0.8h, #9                     \n"  // constants
+      "movi        v1.4s, #3                     \n"
 
       "1:                                        \n"
-      "ld1        {v3.8h}, [%0], %4              \n"  // TL read first 8
-      "ld1        {v4.8h}, [%0], %5              \n"  // TR read 8 offset by 1
-      "ld1        {v5.8h}, [%1], %4              \n"  // BL read 8 from next row
-      "ld1        {v6.8h}, [%1], %5              \n"  // BR offset by 1
-      "subs       %w3, %w3, #16                  \n"  // 16 dst pixels per loop
-      "umull      v16.4s, v3.4h, v0.4h           \n"
-      "umull2     v7.4s, v3.8h, v0.8h            \n"
-      "umull      v18.4s, v4.4h, v0.4h           \n"
-      "umull2     v17.4s, v4.8h, v0.8h           \n"
-      "uaddw      v16.4s, v16.4s, v6.4h          \n"
-      "uaddl2     v19.4s, v6.8h, v3.8h           \n"
-      "uaddl      v3.4s, v6.4h, v3.4h            \n"
-      "uaddw2     v6.4s, v7.4s, v6.8h            \n"
-      "uaddl2     v7.4s, v5.8h, v4.8h            \n"
-      "uaddl      v4.4s, v5.4h, v4.4h            \n"
-      "uaddw      v18.4s, v18.4s, v5.4h          \n"
-      "mla        v16.4s, v4.4s, v1.4s           \n"
-      "mla        v18.4s, v3.4s, v1.4s           \n"
-      "mla        v6.4s, v7.4s, v1.4s            \n"
-      "uaddw2     v4.4s, v17.4s, v5.8h           \n"
-      "uqrshrn    v16.4h,  v16.4s, #4            \n"
-      "mla        v4.4s, v19.4s, v1.4s           \n"
-      "uqrshrn2   v16.8h, v6.4s, #4              \n"
-      "uqrshrn    v17.4h, v18.4s, #4             \n"
-      "uqrshrn2   v17.8h, v4.4s, #4              \n"
-      "st2        {v16.8h-v17.8h}, [%2], #32     \n"
-      "b.gt       1b                             \n"
+      "ld1         {v3.8h}, [%0], %4             \n"  // TL read first 8
+      "ld1         {v4.8h}, [%0], %5             \n"  // TR read 8 offset by 1
+      "ld1         {v5.8h}, [%1], %4             \n"  // BL read 8 from next row
+      "ld1         {v6.8h}, [%1], %5             \n"  // BR offset by 1
+      "subs        %w3, %w3, #16                 \n"  // 16 dst pixels per loop
+      "umull       v16.4s, v3.4h, v0.4h          \n"
+      "umull2      v7.4s, v3.8h, v0.8h           \n"
+      "umull       v18.4s, v4.4h, v0.4h          \n"
+      "umull2      v17.4s, v4.8h, v0.8h          \n"
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "uaddw       v16.4s, v16.4s, v6.4h         \n"
+      "uaddl2      v19.4s, v6.8h, v3.8h          \n"
+      "uaddl       v3.4s, v6.4h, v3.4h           \n"
+      "uaddw2      v6.4s, v7.4s, v6.8h           \n"
+      "uaddl2      v7.4s, v5.8h, v4.8h           \n"
+      "uaddl       v4.4s, v5.4h, v4.4h           \n"
+      "uaddw       v18.4s, v18.4s, v5.4h         \n"
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "mla         v16.4s, v4.4s, v1.4s          \n"
+      "mla         v18.4s, v3.4s, v1.4s          \n"
+      "mla         v6.4s, v7.4s, v1.4s           \n"
+      "uaddw2      v4.4s, v17.4s, v5.8h          \n"
+      "uqrshrn     v16.4h,  v16.4s, #4           \n"
+      "mla         v4.4s, v19.4s, v1.4s          \n"
+      "uqrshrn2    v16.8h, v6.4s, #4             \n"
+      "uqrshrn     v17.4h, v18.4s, #4            \n"
+      "uqrshrn2    v17.8h, v4.4s, #4             \n"
+      "st2         {v16.8h-v17.8h}, [%2], #32    \n"
+      "b.gt        1b                            \n"
       : "+r"(src_ptr),     // %0
         "+r"(src_stride),  // %1
         "+r"(dst),         // %2
@@ -1044,6 +1568,64 @@ void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
   );
 }
 
+void ScaleUVRowDown2Box_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             uint8_t* dst,
+                             int dst_width) {
+  asm volatile(
+      // change the stride to row 2 pointer
+      "add         %1, %1, %0                    \n"
+      "1:                                        \n"
+      "ld2         {v0.16b,v1.16b}, [%0], #32    \n"  // load 16 UV
+      "subs        %w3, %w3, #8                  \n"  // 8 processed per loop.
+      "uaddlp      v0.8h, v0.16b                 \n"  // U 16 bytes -> 8 shorts.
+      "uaddlp      v1.8h, v1.16b                 \n"  // V 16 bytes -> 8 shorts.
+      "ld2         {v16.16b,v17.16b}, [%1], #32  \n"  // load 16
+      "uadalp      v0.8h, v16.16b                \n"  // U 16 bytes -> 8 shorts.
+      "uadalp      v1.8h, v17.16b                \n"  // V 16 bytes -> 8 shorts.
+      "prfm        pldl1keep, [%0, 448]          \n"  // prefetch 7 lines ahead
+      "rshrn       v0.8b, v0.8h, #2              \n"  // round and pack
+      "prfm        pldl1keep, [%1, 448]          \n"
+      "rshrn       v1.8b, v1.8h, #2              \n"
+      "st2         {v0.8b,v1.8b}, [%2], #16      \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),     // %0
+        "+r"(src_stride),  // %1
+        "+r"(dst),         // %2
+        "+r"(dst_width)    // %3
+      :
+      : "memory", "cc", "v0", "v1", "v16", "v17");
+}
+
+// Reads 4 pixels at a time.
+void ScaleUVRowDownEven_NEON(const uint8_t* src_ptr,
+                             ptrdiff_t src_stride,
+                             int src_stepx,  // pixel step
+                             uint8_t* dst_ptr,
+                             int dst_width) {
+  const uint8_t* src1_ptr = src_ptr + src_stepx * 2;
+  const uint8_t* src2_ptr = src_ptr + src_stepx * 4;
+  const uint8_t* src3_ptr = src_ptr + src_stepx * 6;
+  (void)src_stride;
+  asm volatile(
+      "1:                                        \n"
+      "ld1         {v0.h}[0], [%0], %6           \n"
+      "ld1         {v1.h}[0], [%1], %6           \n"
+      "ld1         {v2.h}[0], [%2], %6           \n"
+      "ld1         {v3.h}[0], [%3], %6           \n"
+      "subs        %w5, %w5, #4                  \n"  // 4 pixels per loop.
+      "st4         {v0.h, v1.h, v2.h, v3.h}[0], [%4], #8 \n"
+      "b.gt        1b                            \n"
+      : "+r"(src_ptr),                 // %0
+        "+r"(src1_ptr),                // %1
+        "+r"(src2_ptr),                // %2
+        "+r"(src3_ptr),                // %3
+        "+r"(dst_ptr),                 // %4
+        "+r"(dst_width)                // %5
+      : "r"((int64_t)(src_stepx * 8))  // %6
+      : "memory", "cc", "v0", "v1", "v2", "v3");
+}
+
 #endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
 
 #ifdef __cplusplus
diff --git a/files/source/scale_rgb.cc b/files/source/scale_rgb.cc
new file mode 100644
index 00000000..8db59b56
--- /dev/null
+++ b/files/source/scale_rgb.cc
@@ -0,0 +1,66 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h" /* For FilterMode */
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/convert_argb.h"
+#include "libyuv/convert_from_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/scale_argb.h"
+#include "libyuv/scale_rgb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Scale a 24 bit image.
+// Converts to ARGB as intermediate step
+
+LIBYUV_API
+int RGBScale(const uint8_t* src_rgb,
+             int src_stride_rgb,
+             int src_width,
+             int src_height,
+             uint8_t* dst_rgb,
+             int dst_stride_rgb,
+             int dst_width,
+             int dst_height,
+             enum FilterMode filtering) {
+  int r;
+  uint8_t* src_argb =
+      (uint8_t*)malloc(src_width * src_height * 4 + dst_width * dst_height * 4);
+  uint8_t* dst_argb = src_argb + src_width * src_height * 4;
+
+  if (!src_argb) {
+    return 1;
+  }
+
+  r = RGB24ToARGB(src_rgb, src_stride_rgb, src_argb, src_width * 4, src_width,
+                  src_height);
+  if (!r) {
+    r = ARGBScale(src_argb, src_width * 4, src_width, src_height, dst_argb,
+                  dst_width * 4, dst_width, dst_height, filtering);
+    if (!r) {
+      r = ARGBToRGB24(dst_argb, dst_width * 4, dst_rgb, dst_stride_rgb,
+                      dst_width, dst_height);
+    }
+  }
+  free(src_argb);
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/files/source/scale_uv.cc b/files/source/scale_uv.cc
new file mode 100644
index 00000000..3b3d7b8e
--- /dev/null
+++ b/files/source/scale_uv.cc
@@ -0,0 +1,1161 @@
+/*
+ *  Copyright 2020 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyUV
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Macros to enable specialized scalers
+
+#ifndef HAS_SCALEUVDOWN2
+#define HAS_SCALEUVDOWN2 1
+#endif
+#ifndef HAS_SCALEUVDOWN4BOX
+#define HAS_SCALEUVDOWN4BOX 1
+#endif
+#ifndef HAS_SCALEUVDOWNEVEN
+#define HAS_SCALEUVDOWNEVEN 1
+#endif
+#ifndef HAS_SCALEUVBILINEARDOWN
+#define HAS_SCALEUVBILINEARDOWN 1
+#endif
+#ifndef HAS_SCALEUVBILINEARUP
+#define HAS_SCALEUVBILINEARUP 1
+#endif
+#ifndef HAS_UVCOPY
+#define HAS_UVCOPY 1
+#endif
+#ifndef HAS_SCALEPLANEVERTICAL
+#define HAS_SCALEPLANEVERTICAL 1
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleUV, 1/2
+// This is an optimized version for scaling down a UV to 1/2 of
+// its original size.
+#if HAS_SCALEUVDOWN2
+static void ScaleUVDown2(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         int src_stride,
+                         int dst_stride,
+                         const uint8_t* src_uv,
+                         uint8_t* dst_uv,
+                         int x,
+                         int dx,
+                         int y,
+                         int dy,
+                         enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleUVRowDown2)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                          uint8_t* dst_uv, int dst_width) =
+      filtering == kFilterNone
+          ? ScaleUVRowDown2_C
+          : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_C
+                                        : ScaleUVRowDown2Box_C);
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 2);      // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_uv += (y >> 16) * (int64_t)src_stride + (x >> 16) * 2;
+  } else {
+    src_uv += (y >> 16) * (int64_t)src_stride + ((x >> 16) - 1) * 2;
+  }
+
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && filtering) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_NEON;
+    }
+  }
+#endif
+
+// This code is not enabled.  Only box filter is available at this time.
+#if defined(HAS_SCALEUVROWDOWN2_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_SSSE3
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_SSSE3
+                                          : ScaleUVRowDown2Box_Any_SSSE3);
+    if (IS_ALIGNED(dst_width, 2)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_SSSE3
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_SSSE3
+                                            : ScaleUVRowDown2Box_SSSE3);
+    }
+  }
+#endif
+// This code is not enabled.  Only box filter is available at this time.
+#if defined(HAS_SCALEUVROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_NEON
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_NEON
+                                          : ScaleUVRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_NEON
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_NEON
+                                            : ScaleUVRowDown2Box_NEON);
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVRowDown2 =
+        filtering == kFilterNone
+            ? ScaleUVRowDown2_Any_MSA
+            : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_Any_MSA
+                                          : ScaleUVRowDown2Box_Any_MSA);
+    if (IS_ALIGNED(dst_width, 2)) {
+      ScaleUVRowDown2 =
+          filtering == kFilterNone
+              ? ScaleUVRowDown2_MSA
+              : (filtering == kFilterLinear ? ScaleUVRowDown2Linear_MSA
+                                            : ScaleUVRowDown2Box_MSA);
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDown2(src_uv, src_stride, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+}
+#endif  // HAS_SCALEUVDOWN2
+
+// ScaleUV, 1/4
+// This is an optimized version for scaling down a UV to 1/4 of
+// its original size.
+#if HAS_SCALEUVDOWN4BOX
+static void ScaleUVDown4Box(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_uv,
+                            uint8_t* dst_uv,
+                            int x,
+                            int dx,
+                            int y,
+                            int dy) {
+  int j;
+  // Allocate 2 rows of UV.
+  const int kRowSize = (dst_width * 2 * 2 + 15) & ~15;
+  align_buffer_64(row, kRowSize * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleUVRowDown2)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                          uint8_t* dst_uv, int dst_width) =
+      ScaleUVRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_uv += (y >> 16) * (int64_t)src_stride + (x >> 16) * 2;
+  (void)src_width;
+  (void)src_height;
+  (void)dx;
+  assert(dx == 65536 * 4);      // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+
+#if defined(HAS_SCALEUVROWDOWN2BOX_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWN2BOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDown2 = ScaleUVRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVRowDown2 = ScaleUVRowDown2Box_NEON;
+    }
+  }
+#endif
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDown2(src_uv, src_stride, row, dst_width * 2);
+    ScaleUVRowDown2(src_uv + src_stride * 2, src_stride, row + kRowSize,
+                    dst_width * 2);
+    ScaleUVRowDown2(row, kRowSize, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+#endif  // HAS_SCALEUVDOWN4BOX
+
+// ScaleUV Even
+// This is an optimized version for scaling down a UV to even
+// multiple of its original size.
+#if HAS_SCALEUVDOWNEVEN
+static void ScaleUVDownEven(int src_width,
+                            int src_height,
+                            int dst_width,
+                            int dst_height,
+                            int src_stride,
+                            int dst_stride,
+                            const uint8_t* src_uv,
+                            uint8_t* dst_uv,
+                            int x,
+                            int dx,
+                            int y,
+                            int dy,
+                            enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  int row_stride = (dy >> 16) * (int64_t)src_stride;
+  void (*ScaleUVRowDownEven)(const uint8_t* src_uv, ptrdiff_t src_stride,
+                             int src_step, uint8_t* dst_uv, int dst_width) =
+      filtering ? ScaleUVRowDownEvenBox_C : ScaleUVRowDownEven_C;
+  (void)src_width;
+  (void)src_height;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_uv += (y >> 16) * (int64_t)src_stride + (x >> 16) * 2;
+#if defined(HAS_SCALEUVROWDOWNEVEN_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_SSSE3
+                                   : ScaleUVRowDownEven_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_SSE2 : ScaleUVRowDownEven_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && !filtering) {
+    ScaleUVRowDownEven = ScaleUVRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven = ScaleUVRowDownEven_NEON;
+    }
+  }
+#endif  // TODO(fbarchard): Enable Box filter
+#if defined(HAS_SCALEUVROWDOWNEVENBOX_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVRowDownEven = filtering ? ScaleUVRowDownEvenBox_Any_NEON
+                                   : ScaleUVRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_NEON : ScaleUVRowDownEven_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVROWDOWNEVEN_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVRowDownEven =
+        filtering ? ScaleUVRowDownEvenBox_Any_MSA : ScaleUVRowDownEven_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVRowDownEven =
+          filtering ? ScaleUVRowDownEvenBox_MSA : ScaleUVRowDownEven_MSA;
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVRowDownEven(src_uv, src_stride, col_step, dst_uv, dst_width);
+    src_uv += row_stride;
+    dst_uv += dst_stride;
+  }
+}
+#endif
+
+// Scale UV down with bilinear interpolation.
+#if HAS_SCALEUVBILINEARDOWN
+static void ScaleUVBilinearDown(int src_width,
+                                int src_height,
+                                int dst_width,
+                                int dst_height,
+                                int src_stride,
+                                int dst_stride,
+                                const uint8_t* src_uv,
+                                uint8_t* dst_uv,
+                                int x,
+                                int dx,
+                                int y,
+                                int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8_t * dst_uv, const uint8_t* src_uv,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleUVFilterCols)(uint8_t * dst_uv, const uint8_t* src_uv,
+                            int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleUVFilterCols64_C : ScaleUVFilterCols_C;
+  int64_t xlast = x + (int64_t)(dst_width - 1) * dx;
+  int64_t xl = (dx >= 0) ? x : xlast;
+  int64_t xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;    // Left edge aligned.
+  xr = (xr >> 16) + 1;     // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 2;  // Width aligned to 2.
+  src_uv += xl * 2;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_MSA;
+    }
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of UV.
+  {
+    align_buffer_64(row, clip_src_width * 2);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8_t* src = src_uv + yi * (int64_t)src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleUVFilterCols(dst_uv, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleUVFilterCols(dst_uv, row, dst_width, x, dx);
+      }
+      dst_uv += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+#endif
+
+// Scale UV up with bilinear interpolation.
+#if HAS_SCALEUVBILINEARUP
+static void ScaleUVBilinearUp(int src_width,
+                              int src_height,
+                              int dst_width,
+                              int dst_height,
+                              int src_stride,
+                              int dst_stride,
+                              const uint8_t* src_uv,
+                              uint8_t* dst_uv,
+                              int x,
+                              int dx,
+                              int y,
+                              int dy,
+                              enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8_t * dst_uv, const uint8_t* src_uv,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  void (*ScaleUVFilterCols)(uint8_t * dst_uv, const uint8_t* src_uv,
+                            int dst_width, int x, int dx) =
+      filtering ? ScaleUVFilterCols_C : ScaleUVCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    InterpolateRow = InterpolateRow_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_MSA;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    InterpolateRow = InterpolateRow_Any_LSX;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_LSX;
+    }
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleUVFilterCols = filtering ? ScaleUVFilterCols64_C : ScaleUVCols64_C;
+  }
+#if defined(HAS_SCALEUVFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVFILTERCOLS_MSA)
+  if (filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVFilterCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleUVFilterCols = ScaleUVFilterCols_MSA;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_SSSE3)
+  if (!filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVFilterCols = ScaleUVCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVFilterCols = ScaleUVCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleUVFilterCols = ScaleUVCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_MSA)
+  if (!filtering && TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVFilterCols = ScaleUVCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVCols_MSA;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleUVFilterCols = ScaleUVColsUp2_C;
+#if defined(HAS_SCALEUVCOLSUP2_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(dst_width, 8)) {
+      ScaleUVFilterCols = ScaleUVColsUp2_SSSE3;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8_t* src = src_uv + yi * (int64_t)src_stride;
+
+    // Allocate 2 rows of UV.
+    const int kRowSize = (dst_width * 2 + 15) & ~15;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8_t* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleUVFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleUVFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    if (src_height > 2) {
+      src += src_stride;
+    }
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_uv + yi * (int64_t)src_stride;
+        }
+        if (yi != lasty) {
+          ScaleUVFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          if ((y + 65536) < max_y) {
+            src += src_stride;
+          }
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_uv, rowptr, 0, dst_width * 2, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_uv, rowptr, rowstride, dst_width * 2, yf);
+      }
+      dst_uv += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+#endif  // HAS_SCALEUVBILINEARUP
+
+// Scale UV, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of NV16 to NV24.
+void ScaleUVLinearUp2(int src_width,
+                      int src_height,
+                      int dst_width,
+                      int dst_height,
+                      int src_stride,
+                      int dst_stride,
+                      const uint8_t* src_uv,
+                      uint8_t* dst_uv) {
+  void (*ScaleRowUp)(const uint8_t* src_uv, uint8_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_uv + ((src_height - 1) / 2) * (int64_t)src_stride, dst_uv,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_uv + (y >> 16) * (int64_t)src_stride, dst_uv, dst_width);
+      dst_uv += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale plane, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of NV12 to NV24.
+void ScaleUVBilinearUp2(int src_width,
+                        int src_height,
+                        int dst_width,
+                        int dst_height,
+                        int src_stride,
+                        int dst_stride,
+                        const uint8_t* src_ptr,
+                        uint8_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint8_t* src_ptr, ptrdiff_t src_stride,
+                      uint8_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_SSSE3
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_SSSE3;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale 16 bit UV, horizontally up by 2 times.
+// Uses linear filter horizontally, nearest vertically.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original width, using linear interpolation.
+// This is used to scale U and V planes of P210 to P410.
+void ScaleUVLinearUp2_16(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         int src_stride,
+                         int dst_stride,
+                         const uint16_t* src_uv,
+                         uint16_t* dst_uv) {
+  void (*ScaleRowUp)(const uint16_t* src_uv, uint16_t* dst_uv, int dst_width) =
+      ScaleUVRowUp2_Linear_16_Any_C;
+  int i;
+  int y;
+  int dy;
+
+  // This function can only scale up by 2 times horizontally.
+  assert(src_width == ((dst_width + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2LINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowUp = ScaleUVRowUp2_Linear_16_Any_NEON;
+  }
+#endif
+
+  if (dst_height == 1) {
+    ScaleRowUp(src_uv + ((src_height - 1) / 2) * (int64_t)src_stride, dst_uv,
+               dst_width);
+  } else {
+    dy = FixedDiv(src_height - 1, dst_height - 1);
+    y = (1 << 15) - 1;
+    for (i = 0; i < dst_height; ++i) {
+      ScaleRowUp(src_uv + (y >> 16) * (int64_t)src_stride, dst_uv, dst_width);
+      dst_uv += dst_stride;
+      y += dy;
+    }
+  }
+}
+
+// Scale 16 bit UV, up by 2 times.
+// This is an optimized version for scaling up a plane to 2 times of
+// its original size, using bilinear interpolation.
+// This is used to scale U and V planes of P010 to P410.
+void ScaleUVBilinearUp2_16(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           int src_stride,
+                           int dst_stride,
+                           const uint16_t* src_ptr,
+                           uint16_t* dst_ptr) {
+  void (*Scale2RowUp)(const uint16_t* src_ptr, ptrdiff_t src_stride,
+                      uint16_t* dst_ptr, ptrdiff_t dst_stride, int dst_width) =
+      ScaleUVRowUp2_Bilinear_16_Any_C;
+  int x;
+
+  // This function can only scale up by 2 times.
+  assert(src_width == ((dst_width + 1) / 2));
+  assert(src_height == ((dst_height + 1) / 2));
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_SSE41
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_SSE41;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_AVX2
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_AVX2;
+  }
+#endif
+
+#ifdef HAS_SCALEUVROWUP2BILINEAR_16_NEON
+  if (TestCpuFlag(kCpuHasNEON)) {
+    Scale2RowUp = ScaleUVRowUp2_Bilinear_16_Any_NEON;
+  }
+#endif
+
+  Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  dst_ptr += dst_stride;
+  for (x = 0; x < src_height - 1; ++x) {
+    Scale2RowUp(src_ptr, src_stride, dst_ptr, dst_stride, dst_width);
+    src_ptr += src_stride;
+    // TODO(fbarchard): Test performance of writing one row of destination at a
+    // time.
+    dst_ptr += 2 * dst_stride;
+  }
+  if (!(dst_height & 1)) {
+    Scale2RowUp(src_ptr, 0, dst_ptr, 0, dst_width);
+  }
+}
+
+// Scale UV to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleUVSimple(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          int src_stride,
+                          int dst_stride,
+                          const uint8_t* src_uv,
+                          uint8_t* dst_uv,
+                          int x,
+                          int dx,
+                          int y,
+                          int dy) {
+  int j;
+  void (*ScaleUVCols)(uint8_t * dst_uv, const uint8_t* src_uv, int dst_width,
+                      int x, int dx) =
+      (src_width >= 32768) ? ScaleUVCols64_C : ScaleUVCols_C;
+  (void)src_height;
+#if defined(HAS_SCALEUVCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleUVCols = ScaleUVCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleUVCols = ScaleUVCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleUVCols = ScaleUVCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEUVCOLS_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ScaleUVCols = ScaleUVCols_Any_MSA;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleUVCols = ScaleUVCols_MSA;
+    }
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleUVCols = ScaleUVColsUp2_C;
+#if defined(HAS_SCALEUVCOLSUP2_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(dst_width, 8)) {
+      ScaleUVCols = ScaleUVColsUp2_SSSE3;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleUVCols(dst_uv, src_uv + (y >> 16) * (int64_t)src_stride, dst_width, x,
+                dx);
+    dst_uv += dst_stride;
+    y += dy;
+  }
+}
+
+// Copy UV with optional flipping
+#if HAS_UVCOPY
+static int UVCopy(const uint8_t* src_uv,
+                  int src_stride_uv,
+                  uint8_t* dst_uv,
+                  int dst_stride_uv,
+                  int width,
+                  int height) {
+  if (!src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * (int64_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  CopyPlane(src_uv, src_stride_uv, dst_uv, dst_stride_uv, width * 2, height);
+  return 0;
+}
+
+static int UVCopy_16(const uint16_t* src_uv,
+                     int src_stride_uv,
+                     uint16_t* dst_uv,
+                     int dst_stride_uv,
+                     int width,
+                     int height) {
+  if (!src_uv || !dst_uv || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uv = src_uv + (height - 1) * (int64_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  CopyPlane_16(src_uv, src_stride_uv, dst_uv, dst_stride_uv, width * 2, height);
+  return 0;
+}
+#endif  // HAS_UVCOPY
+
+// Scale a UV plane (from NV12)
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleUV(const uint8_t* src,
+                    int src_stride,
+                    int src_width,
+                    int src_height,
+                    uint8_t* dst,
+                    int dst_stride,
+                    int dst_width,
+                    int dst_height,
+                    int clip_x,
+                    int clip_y,
+                    int clip_width,
+                    int clip_height,
+                    enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // UV does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * (int64_t)src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering, &x, &y,
+             &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64_t clipf = (int64_t)(clip_x)*dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 2;
+    dst += clip_x * 2;
+  }
+  if (clip_y) {
+    int64_t clipf = (int64_t)(clip_y)*dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * (int64_t)src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+#if HAS_SCALEUVDOWN2
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleUVDown2(src_width, src_height, clip_width, clip_height,
+                       src_stride, dst_stride, src, dst, x, dx, y, dy,
+                       filtering);
+          return;
+        }
+#endif
+#if HAS_SCALEUVDOWN4BOX
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleUVDown4Box(src_width, src_height, clip_width, clip_height,
+                          src_stride, dst_stride, src, dst, x, dx, y, dy);
+          return;
+        }
+#endif
+#if HAS_SCALEUVDOWNEVEN
+        ScaleUVDownEven(src_width, src_height, clip_width, clip_height,
+                        src_stride, dst_stride, src, dst, x, dx, y, dy,
+                        filtering);
+        return;
+#endif
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+#ifdef HAS_UVCOPY
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          UVCopy(src + (y >> 16) * (int64_t)src_stride + (x >> 16) * 2,
+                 src_stride, dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+#endif
+      }
+    }
+  }
+  // HAS_SCALEPLANEVERTICAL
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical(src_height, clip_width, clip_height, src_stride,
+                       dst_stride, src, dst, x, y, dy, /*bpp=*/2, filtering);
+    return;
+  }
+  if (filtering && (dst_width + 1) / 2 == src_width) {
+    ScaleUVLinearUp2(src_width, src_height, clip_width, clip_height, src_stride,
+                     dst_stride, src, dst);
+    return;
+  }
+  if ((clip_height + 1) / 2 == src_height &&
+      (clip_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScaleUVBilinearUp2(src_width, src_height, clip_width, clip_height,
+                       src_stride, dst_stride, src, dst);
+    return;
+  }
+#if HAS_SCALEUVBILINEARUP
+  if (filtering && dy < 65536) {
+    ScaleUVBilinearUp(src_width, src_height, clip_width, clip_height,
+                      src_stride, dst_stride, src, dst, x, dx, y, dy,
+                      filtering);
+    return;
+  }
+#endif
+#if HAS_SCALEUVBILINEARDOWN
+  if (filtering) {
+    ScaleUVBilinearDown(src_width, src_height, clip_width, clip_height,
+                        src_stride, dst_stride, src, dst, x, dx, y, dy,
+                        filtering);
+    return;
+  }
+#endif
+  ScaleUVSimple(src_width, src_height, clip_width, clip_height, src_stride,
+                dst_stride, src, dst, x, dx, y, dy);
+}
+
+// Scale an UV image.
+LIBYUV_API
+int UVScale(const uint8_t* src_uv,
+            int src_stride_uv,
+            int src_width,
+            int src_height,
+            uint8_t* dst_uv,
+            int dst_stride_uv,
+            int dst_width,
+            int dst_height,
+            enum FilterMode filtering) {
+  if (!src_uv || src_width <= 0 || src_height == 0 || src_width > 32768 ||
+      src_height > 32768 || !dst_uv || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleUV(src_uv, src_stride_uv, src_width, src_height, dst_uv, dst_stride_uv,
+          dst_width, dst_height, 0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+// Scale a 16 bit UV image.
+// This function is currently incomplete, it can't handle all cases.
+LIBYUV_API
+int UVScale_16(const uint16_t* src_uv,
+               int src_stride_uv,
+               int src_width,
+               int src_height,
+               uint16_t* dst_uv,
+               int dst_stride_uv,
+               int dst_width,
+               int dst_height,
+               enum FilterMode filtering) {
+  int dy = 0;
+
+  if (!src_uv || src_width <= 0 || src_height == 0 || src_width > 32768 ||
+      src_height > 32768 || !dst_uv || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  // UV does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height, dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src_uv = src_uv + (src_height - 1) * (int64_t)src_stride_uv;
+    src_stride_uv = -src_stride_uv;
+  }
+  src_width = Abs(src_width);
+
+#ifdef HAS_UVCOPY
+  if (!filtering && src_width == dst_width && (src_height % dst_height == 0)) {
+    if (dst_height == 1) {
+      UVCopy_16(src_uv + ((src_height - 1) / 2) * (int64_t)src_stride_uv,
+                src_stride_uv, dst_uv, dst_stride_uv, dst_width, dst_height);
+    } else {
+      dy = src_height / dst_height;
+      UVCopy_16(src_uv + ((dy - 1) / 2) * (int64_t)src_stride_uv,
+                dy * (int64_t)src_stride_uv, dst_uv, dst_stride_uv, dst_width,
+                dst_height);
+    }
+
+    return 0;
+  }
+#endif
+
+  if (filtering && (dst_width + 1) / 2 == src_width) {
+    ScaleUVLinearUp2_16(src_width, src_height, dst_width, dst_height,
+                        src_stride_uv, dst_stride_uv, src_uv, dst_uv);
+    return 0;
+  }
+
+  if ((dst_height + 1) / 2 == src_height && (dst_width + 1) / 2 == src_width &&
+      (filtering == kFilterBilinear || filtering == kFilterBox)) {
+    ScaleUVBilinearUp2_16(src_width, src_height, dst_width, dst_height,
+                          src_stride_uv, dst_stride_uv, src_uv, dst_uv);
+    return 0;
+  }
+
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/files/source/scale_win.cc b/files/source/scale_win.cc
index c5fc86f3..ea1f95c6 100644
--- a/files/source/scale_win.cc
+++ b/files/source/scale_win.cc
@@ -16,8 +16,9 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// This module is for 32 bit Visual C x86 and clangcl
-#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+// This module is for 32 bit Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && \
+    !defined(__clang__) && defined(_M_IX86)
 
 // Offsets for source bytes 0 to 9
 static const uvec8 kShuf0 = {0,   1,   3,   4,   5,   7,   8,   9,
diff --git a/files/source/test.sh b/files/source/test.sh
new file mode 100755
index 00000000..7f12c3c1
--- /dev/null
+++ b/files/source/test.sh
@@ -0,0 +1,35 @@
+#!/bin/bash
+set -x
+
+function runbenchmark1 {
+  perf record /google/src/cloud/fbarchard/clean/google3/blaze-bin/third_party/libyuv/libyuv_test --gunit_filter=*$1 --libyuv_width=1280 --libyuv_height=720 --libyuv_repeat=1000 --libyuv_flags=-1 --libyuv_cpu_info=-1
+  perf report | grep AVX
+}
+
+runbenchmark1 ABGRToI420
+runbenchmark1 Android420ToI420
+runbenchmark1 ARGBToI420
+runbenchmark1 Convert16To8Plane
+runbenchmark1 ConvertToARGB
+runbenchmark1 ConvertToI420
+runbenchmark1 CopyPlane
+runbenchmark1 H010ToAB30
+runbenchmark1 H010ToAR30
+runbenchmark1 HalfFloatPlane
+runbenchmark1 I010ToAB30
+runbenchmark1 I010ToAR30
+runbenchmark1 I420Copy
+runbenchmark1 I420Psnr
+runbenchmark1 I420Scale
+runbenchmark1 I420Ssim
+runbenchmark1 I420ToARGB
+runbenchmark1 I420ToNV12
+runbenchmark1 I420ToUYVY
+runbenchmark1 I422ToI420
+runbenchmark1 InitCpuFlags
+runbenchmark1 J420ToARGB
+runbenchmark1 NV12ToARGB
+runbenchmark1 NV12ToI420
+runbenchmark1 NV12ToI420Rotate
+runbenchmark1 SetCpuFlags
+runbenchmark1 YUY2ToI420
diff --git a/files/sync_chromium.py b/files/sync_chromium.py
deleted file mode 100755
index 4e51b6bd..00000000
--- a/files/sync_chromium.py
+++ /dev/null
@@ -1,154 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Script to download a Chromium checkout into the workspace.
-
-The script downloads a full Chromium Git clone and its DEPS.
-
-The following environment variable can be used to alter the behavior:
-* CHROMIUM_NO_HISTORY - If set to 1, a Git checkout with no history will be
-  downloaded. This is consumes less bandwidth and disk space but is known to be
-  slower in general if you have a high-speed connection.
-
-After a successful sync has completed, a .last_sync_chromium file is written to
-the chromium directory. While it exists, no more gclient sync operations will be
-performed until the --target-revision changes or the SCRIPT_VERSION constant is
-incremented. The file can be removed manually to force a new sync.
-"""
-
-import argparse
-import os
-import subprocess
-import sys
-
-# Bump this whenever the algorithm changes and you need bots/devs to re-sync,
-# ignoring the .last_sync_chromium file
-SCRIPT_VERSION = 4
-
-ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-CHROMIUM_NO_HISTORY = 'CHROMIUM_NO_HISTORY'
-
-def _parse_gclient_dict():
-  gclient_dict = {}
-  try:
-    main_gclient = os.path.join(os.path.dirname(ROOT_DIR), '.gclient')
-    with open(main_gclient, 'rb') as deps_content:
-      exec(deps_content, gclient_dict)
-  except Exception as e:
-    print >> sys.stderr, 'error while parsing .gclient:', e
-  return gclient_dict
-
-
-def get_cache_dir():
-  return _parse_gclient_dict().get('cache_dir')
-
-
-def get_target_os_list():
-  return ','.join(_parse_gclient_dict().get('target_os', []))
-
-
-def main():
-  CR_DIR = os.path.join(ROOT_DIR, 'chromium')
-
-  p = argparse.ArgumentParser()
-  p.add_argument('--target-revision', required=True,
-                 help='The target chromium git revision [REQUIRED]')
-  p.add_argument('--chromium-dir', default=CR_DIR,
-                 help=('The path to the chromium directory to sync '
-                       '(default: %(default)r)'))
-  opts = p.parse_args()
-  opts.chromium_dir = os.path.abspath(opts.chromium_dir)
-
-  target_os_list = get_target_os_list()
-
-  # Do a quick check to see if we were successful last time to make runhooks
-  # sooper fast.
-  flag_file = os.path.join(opts.chromium_dir, '.last_sync_chromium')
-  flag_file_content = '\n'.join([
-    str(SCRIPT_VERSION),
-    opts.target_revision,
-    repr(target_os_list),
-  ])
-  if (os.path.exists(os.path.join(opts.chromium_dir, 'src')) and
-      os.path.exists(flag_file)):
-    with open(flag_file, 'r') as f:
-      if f.read() == flag_file_content:
-        print 'Chromium already up to date: ', opts.target_revision
-        return 0
-    os.unlink(flag_file)
-
-  env = os.environ.copy()
-
-  # Avoid downloading NaCl toolchain as part of the Chromium hooks.
-  env['GYP_CHROMIUM_NO_ACTION'] = '1'
-  gclient_cmd = 'gclient.bat' if sys.platform.startswith('win') else 'gclient'
-  args = [
-      gclient_cmd, 'sync', '--force', '--revision', 'src@'+opts.target_revision
-  ]
-
-  if os.environ.get('CHROME_HEADLESS') == '1':
-    # Running on a buildbot.
-    args.append('-vvv')
-
-    if sys.platform.startswith('win'):
-      cache_path = os.path.join(os.path.splitdrive(ROOT_DIR)[0] + os.path.sep,
-                                'b', 'git-cache')
-    else:
-      cache_path = '/b/git-cache'
-  else:
-    # Support developers setting the cache_dir in .gclient.
-    cache_path = get_cache_dir()
-
-  # Allow for users with poor internet connections to download a Git clone
-  # without history (saves several gigs but is generally slower and doesn't work
-  # with the Git cache).
-  if os.environ.get(CHROMIUM_NO_HISTORY) == '1':
-    if cache_path:
-      print >> sys.stderr, (
-          'You cannot use "no-history" mode for syncing Chrome (i.e. set the '
-          '%s environment variable to 1) when you have cache_dir configured in '
-          'your .gclient.' % CHROMIUM_NO_HISTORY)
-      return 1
-    args.append('--no-history')
-    gclient_entries_file = os.path.join(opts.chromium_dir, '.gclient_entries')
-  else:
-    # Write a temporary .gclient file that has the cache_dir variable added.
-    gclientfile = os.path.join(opts.chromium_dir, '.gclient')
-    with open(gclientfile, 'rb') as spec:
-      spec = spec.read().splitlines()
-      spec[-1] = 'cache_dir = %r' % (cache_path,)
-    with open(gclientfile + '.tmp', 'wb') as f:
-      f.write('\n'.join(spec))
-
-    args += [
-      '--gclientfile', '.gclient.tmp',
-      '--delete_unversioned_trees', '--reset', '--upstream'
-    ]
-    gclient_entries_file = os.path.join(opts.chromium_dir,
-                                        '.gclient.tmp_entries')
-
-  # To avoid gclient sync problems when DEPS entries have been removed we must
-  # wipe the gclient's entries file that contains cached URLs for all DEPS.
-  if os.path.exists(gclient_entries_file):
-    os.unlink(gclient_entries_file)
-
-  if target_os_list:
-    args += ['--deps=' + target_os_list]
-
-  print 'Running "%s" in %s' % (' '.join(args), opts.chromium_dir)
-  ret = subprocess.call(args, cwd=opts.chromium_dir, env=env)
-  if ret == 0:
-    with open(flag_file, 'wb') as f:
-      f.write(flag_file_content)
-
-  return ret
-
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/files/third_party/gflags/BUILD.gn b/files/third_party/gflags/BUILD.gn
deleted file mode 100644
index af41b7ec..00000000
--- a/files/third_party/gflags/BUILD.gn
+++ /dev/null
@@ -1,73 +0,0 @@
-# Copyright (c) 2014 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-if (is_win) {
-  gflags_gen_arch_root = "gen/win"
-} else {
-  gflags_gen_arch_root = "gen/posix"
-}
-
-config("gflags_config") {
-  include_dirs = [
-    "$gflags_gen_arch_root/include",  # For configured files.
-    "src/src",  # For everything else.
-  ]
-
-  defines = [
-    # These macros exist so flags and symbols are properly exported when
-    # building DLLs. Since we don't build DLLs, we need to disable them.
-    "GFLAGS_DLL_DECL=",
-    "GFLAGS_DLL_DECLARE_FLAG=",
-    "GFLAGS_DLL_DEFINE_FLAG=",
-  ]
-
-  # GN orders flags on a target before flags from configs. The default config
-  # adds -Wall, and this flag have to be after -Wall -- so they need to
-  # come from a config and can't be on the target directly.
-  if (is_clang) {
-    cflags = [ "-Wno-unused-local-typedef" ]
-  }
-}
-
-source_set("gflags") {
-  cflags = []
-  sources = [
-    "src/src/gflags.cc",
-    "src/src/gflags_completions.cc",
-    "src/src/gflags_reporting.cc",
-  ]
-  if (is_win) {
-    sources += [ "src/src/windows_port.cc" ]
-
-    cflags += [
-      "/wd4005",  # WIN32_LEAN_AND_MEAN.
-      "/wd4267",  # Conversion from size_t to "type".
-    ]
-  }
-
-  include_dirs = [
-    "$gflags_gen_arch_root/include/gflags",  # For configured files.
-    "$gflags_gen_arch_root/include/private",  # For config.h
-  ]
-
-  public_configs = [ ":gflags_config" ]
-
-  configs -= [ "//build/config/compiler:chromium_code" ]
-  configs += [ "//build/config/compiler:no_chromium_code" ]
-
-  if (is_win) {
-    configs -= [ "//build/config/win:unicode" ]
-  }
-
-  if (is_clang) {
-    # TODO(andrew): Look into fixing this warning upstream:
-    # http://code.google.com/p/webrtc/issues/detail?id=760
-    configs -= [ "//build/config/clang:extra_warnings" ]
-    cflags += [ "-Wno-microsoft-include" ]
-  }
-}
diff --git a/files/third_party/gflags/LICENSE b/files/third_party/gflags/LICENSE
deleted file mode 100644
index d15b0c24..00000000
--- a/files/third_party/gflags/LICENSE
+++ /dev/null
@@ -1,28 +0,0 @@
-Copyright (c) 2006, Google Inc.
-All rights reserved.
-
-Redistribution and use in source and binary forms, with or without
-modification, are permitted provided that the following conditions are
-met:
-
-    * Redistributions of source code must retain the above copyright
-notice, this list of conditions and the following disclaimer.
-    * Redistributions in binary form must reproduce the above
-copyright notice, this list of conditions and the following disclaimer
-in the documentation and/or other materials provided with the
-distribution.
-    * Neither the name of Google Inc. nor the names of its
-contributors may be used to endorse or promote products derived from
-this software without specific prior written permission.
-
-THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/files/third_party/gflags/README.libyuv b/files/third_party/gflags/README.libyuv
deleted file mode 100644
index 5b3bc2db..00000000
--- a/files/third_party/gflags/README.libyuv
+++ /dev/null
@@ -1,28 +0,0 @@
-URL: https://github.com/gflags/gflags
-Version: 2.1.2
-License: New BSD
-License File: LICENSE
-
-Description:
-The gflags package contains a library that implements commandline
-flags processing. As such it's a replacement for getopt(). It has
-increased flexibility, including built-in support for C++ types like
-string, and the ability to define flags in the source file in which
-they're used.
-
-Local Modifications: None
-
-
-How to update platform configuration files:
-The gen/ directory contains pre-generated configuration header files.
-Historically, all operating systems and architectures have generated
-similar configurations except for Windows. This is why there's only
-posix and win directories below gen/.
-When rolling gflags to a newer version, it's a good idea to check if
-new configuration files needs to be generated as well.
-Do this by running ./configure in the newly checked out version of
-gflags. Then diff the generated files with the ones below gen/.
-If you notice a diff, update the files with the updated ones.
-If you suspect platform dependend changes other than Windows, you'll
-have to checkout gflags on the other platforms as well and run
-./configure there too.
diff --git a/files/third_party/gflags/gen/posix/include/gflags/gflags.h b/files/third_party/gflags/gen/posix/include/gflags/gflags.h
deleted file mode 100644
index 0db38f5c..00000000
--- a/files/third_party/gflags/gen/posix/include/gflags/gflags.h
+++ /dev/null
@@ -1,573 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// or defines a command line flag or wants to parse command line flags
-// or print a program usage message (which will include information about
-// flags).  Executive summary, in the form of an example foo.cc file:
-//
-//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
-//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
-//
-//    DEFINE_int32(end, 1000, "The last record to read");
-//
-//    DEFINE_string(filename, "my_file.txt", "The file to read");
-//    // Crash if the specified file does not exist.
-//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
-//                                              &ValidateIsFile);
-//
-//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
-//
-//    void MyFunc() {
-//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
-//    }
-//
-//    Then, at the command-line:
-//       ./foo --noverbose --start=5 --end=100
-//
-// For more details, see
-//    doc/gflags.html
-//
-// --- A note about thread-safety:
-//
-// We describe many functions in this routine as being thread-hostile,
-// thread-compatible, or thread-safe.  Here are the meanings we use:
-//
-// thread-safe: it is safe for multiple threads to call this routine
-//   (or, when referring to a class, methods of this class)
-//   concurrently.
-// thread-hostile: it is not safe for multiple threads to call this
-//   routine (or methods of this class) concurrently.  In gflags,
-//   most thread-hostile routines are intended to be called early in,
-//   or even before, main() -- that is, before threads are spawned.
-// thread-compatible: it is safe for multiple threads to read from
-//   this variable (when applied to variables), or to call const
-//   methods of this class (when applied to classes), as long as no
-//   other thread is writing to the variable or calling non-const
-//   methods of this class.
-
-#ifndef GFLAGS_GFLAGS_H_
-#define GFLAGS_GFLAGS_H_
-
-#include <string>
-#include <vector>
-
-#include "gflags_declare.h" // IWYU pragma: export
-
-
-// We always want to export variables defined in user code
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DEFINE_FLAG
-#  endif
-#endif
-
-
-namespace GFLAGS_NAMESPACE {
-
-
-// --------------------------------------------------------------------
-// To actually define a flag in a file, use DEFINE_bool,
-// DEFINE_string, etc. at the bottom of this file.  You may also find
-// it useful to register a validator with the flag.  This ensures that
-// when the flag is parsed from the commandline, or is later set via
-// SetCommandLineOption, we call the validation function. It is _not_
-// called when you assign the value to the flag directly using the = operator.
-//
-// The validation function should return true if the flag value is valid, and
-// false otherwise. If the function returns false for the new setting of the
-// flag, the flag will retain its current value. If it returns false for the
-// default value, ParseCommandLineFlags() will die.
-//
-// This function is safe to call at global construct time (as in the
-// example below).
-//
-// Example use:
-//    static bool ValidatePort(const char* flagname, int32 value) {
-//       if (value > 0 && value < 32768)   // value is ok
-//         return true;
-//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
-//       return false;
-//    }
-//    DEFINE_int32(port, 0, "What port to listen on");
-//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
-
-// Returns true if successfully registered, false if not (because the
-// first argument doesn't point to a command-line flag, or because a
-// validator is already registered for this flag).
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
-
-// Convenience macro for the registration of a flag validator
-#define DEFINE_validator(name, validator) \
-    static const bool name##_validator_registered = \
-            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
-
-
-// --------------------------------------------------------------------
-// These methods are the best way to get access to info about the
-// list of commandline flags.  Note that these routines are pretty slow.
-//   GetAllFlags: mostly-complete info about the list, sorted by file.
-//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
-//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
-//
-// In addition to accessing flags, you can also access argv[0] (the program
-// name) and argv (the entire commandline), which we sock away a copy of.
-// These variables are static, so you should only set them once.
-//
-// No need to export this data only structure from DLL, avoiding VS warning 4251.
-struct CommandLineFlagInfo {
-  std::string name;            // the name of the flag
-  std::string type;            // the type of the flag: int32, etc
-  std::string description;     // the "help text" associated with the flag
-  std::string current_value;   // the current value, as a string
-  std::string default_value;   // the default value, as a string
-  std::string filename;        // 'cleaned' version of filename holding the flag
-  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
-  bool is_default;             // true if the flag has the default value and
-                               // has not been set explicitly from the cmdline
-                               // or via SetCommandLineOption
-  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
-};
-
-// Using this inside of a validator is a recipe for a deadlock.
-// TODO(user) Fix locking when validators are running, to make it safe to
-// call validators during ParseAllFlags.
-// Also make sure then to uncomment the corresponding unit test in
-// gflags_unittest.sh
-extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
-// These two are actually defined in gflags_reporting.cc.
-extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
-extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
-
-// Create a descriptive string for a flag.
-// Goes to some trouble to make pretty line breaks.
-extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
-
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
-
-// The following functions are thread-safe as long as SetArgv() is
-// only called before any threads start.
-extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
-extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
-extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
-extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
-extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
-extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
-
-// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
-
-// VersionString() is thread-safe as long as SetVersionString() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
-
-
-
-// --------------------------------------------------------------------
-// Normally you access commandline flags by just saying "if (FLAGS_foo)"
-// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
-// commonly, via the DEFINE_foo macro).  But if you need a bit more
-// control, we have programmatic ways to get/set the flags as well.
-// These programmatic ways to access flags are thread-safe, but direct
-// access is only thread-compatible.
-
-// Return true iff the flagname was found.
-// OUTPUT is set to the flag's value, or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
-
-// Return true iff the flagname was found. OUTPUT is set to the flag's
-// CommandLineFlagInfo or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
-
-// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
-// Example usage, to check if a flag's value is currently the default value:
-//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
-extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
-
-enum GFLAGS_DLL_DECL FlagSettingMode {
-  // update the flag's value (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not yet updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Set a particular flag ("command line option").  Returns a string
-// describing the new value that the option has been set to.  The
-// return value API is not well-specified, so basically just depend on
-// it to be empty if the setting failed for some reason -- the name is
-// not a valid flag name, or the value is not a valid value -- and
-// non-empty else.
-
-// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
-extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
-extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
-
-
-// --------------------------------------------------------------------
-// Saves the states (value, default value, whether the user has set
-// the flag, registered validators, etc) of all flags, and restores
-// them when the FlagSaver is destroyed.  This is very useful in
-// tests, say, when you want to let your tests change the flags, but
-// make sure that they get reverted to the original states when your
-// test is complete.
-//
-// Example usage:
-//   void TestFoo() {
-//     FlagSaver s1;
-//     FLAG_foo = false;
-//     FLAG_bar = "some value";
-//
-//     // test happens here.  You can return at any time
-//     // without worrying about restoring the FLAG values.
-//   }
-//
-// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
-// the work is done in the constructor and destructor, so in the standard
-// usage example above, the compiler would complain that it's an
-// unused variable.
-//
-// This class is thread-safe.  However, its destructor writes to
-// exactly the set of flags that have changed value during its
-// lifetime, so concurrent _direct_ access to those flags
-// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
-
-class GFLAGS_DLL_DECL FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
- private:
-  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
-
-  FlagSaver(const FlagSaver&);  // no copying!
-  void operator=(const FlagSaver&);
-}__attribute((unused));
-
-// --------------------------------------------------------------------
-// Some deprecated or hopefully-soon-to-be-deprecated functions.
-
-// This is often used for logging.  TODO(csilvers): figure out a better way
-extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
-// Usually where this is used, a FlagSaver should be used instead.
-extern GFLAGS_DLL_DECL
-bool ReadFlagsFromString(const std::string& flagfilecontents,
-                         const char* prog_name,
-                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
-
-// These let you manually implement --flagfile functionality.
-// DEPRECATED.
-extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
-extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
-
-
-// --------------------------------------------------------------------
-// Useful routines for initializing flags from the environment.
-// In each case, if 'varname' does not exist in the environment
-// return defval.  If 'varname' does exist but is not valid
-// (e.g., not a number for an int32 flag), abort with an error.
-// Otherwise, return the value.  NOTE: for booleans, for true use
-// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
-
-extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
-extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
-extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
-extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
-extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
-extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
-
-
-// --------------------------------------------------------------------
-// The next two functions parse gflags from main():
-
-// Set the "usage" message for this program.  For example:
-//   string usage("This program does nothing.  Sample usage:\n");
-//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
-//   SetUsageMessage(usage);
-// Do not include commandline flags in the usage: we do that for you!
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
-
-// Sets the version string, which is emitted with --version.
-// For instance: SetVersionString("1.3");
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
-
-
-// Looks for flags in argv and parses them.  Rearranges argv to put
-// flags first, or removes them entirely if remove_flags is true.
-// If a flag is defined more than once in the command line or flag
-// file, the last definition is used.  Returns the index (into argv)
-// of the first non-flag argument.
-// See top-of-file for more details on this function.
-#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
-#endif
-
-
-// Calls to ParseCommandLineNonHelpFlags and then to
-// HandleCommandLineHelpFlags can be used instead of a call to
-// ParseCommandLineFlags during initialization, in order to allow for
-// changing default values for some FLAGS (via
-// e.g. SetCommandLineOptionWithMode calls) between the time of
-// command line parsing and the time of dumping help information for
-// the flags as a result of command line parsing.  If a flag is
-// defined more than once in the command line or flag file, the last
-// definition is used.  Returns the index (into argv) of the first
-// non-flag argument.  (If remove_flags is true, will always return 1.)
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
-
-// This is actually defined in gflags_reporting.cc.
-// This function is misnamed (it also handles --version, etc.), but
-// it's too late to change that now. :-(
-extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
-
-// Allow command line reparsing.  Disables the error normally
-// generated when an unknown flag is found, since it may be found in a
-// later parse.  Thread-hostile; meant to be called before any threads
-// are spawned.
-extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
-
-// Reparse the flags that have not yet been recognized.  Only flags
-// registered since the last parse will be recognized.  Any flag value
-// must be provided as part of the argument using "=", not as a
-// separate command line argument that follows the flag argument.
-// Intended for handling flags from dynamically loaded libraries,
-// since their flags are not registered until they are loaded.
-extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
-
-// Clean up memory allocated by flags.  This is only needed to reduce
-// the quantity of "potentially leaked" reports emitted by memory
-// debugging tools such as valgrind.  It is not required for normal
-// operation, or for the google perftools heap-checker.  It must only
-// be called when the process is about to exit, and all threads that
-// might access flags are quiescent.  Referencing flags after this is
-// called will have unexpected consequences.  This is not safe to run
-// when multiple threads might be running: the function is
-// thread-hostile.
-extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
-
-
-// --------------------------------------------------------------------
-// Now come the command line flag declaration/definition macros that
-// will actually be used.  They're kind of hairy.  A major reason
-// for this is initialization: we want people to be able to access
-// variables in global constructors and have that not crash, even if
-// their global constructor runs before the global constructor here.
-// (Obviously, we can't guarantee the flags will have the correct
-// default value in that case, but at least accessing them is safe.)
-// The only way to do that is have flags point to a static buffer.
-// So we make one, using a union to ensure proper alignment, and
-// then use placement-new to actually set up the flag with the
-// correct default value.  In the same vein, we have to worry about
-// flag access in global destructors, so FlagRegisterer has to be
-// careful never to destroy the flag-values it constructs.
-//
-// Note that when we define a flag variable FLAGS_<name>, we also
-// preemptively define a junk variable, FLAGS_no<name>.  This is to
-// cause a link-time error if someone tries to define 2 flags with
-// names like "logging" and "nologging".  We do this because a bool
-// flag FLAG can be set from the command line to true with a "-FLAG"
-// argument, and to false with a "-noFLAG" argument, and so this can
-// potentially avert confusion.
-//
-// We also put flags into their own namespace.  It is purposefully
-// named in an opaque way that people should have trouble typing
-// directly.  The idea is that DEFINE puts the flag in the weird
-// namespace, and DECLARE imports the flag from there into the current
-// namespace.  The net result is to force people to use DECLARE to get
-// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
-// or some such instead.  We want this so we can put extra
-// functionality (like sanity-checking) in DECLARE if we want, and
-// make sure it is picked up everywhere.
-//
-// We also put the type of the variable in the namespace, so that
-// people can't DECLARE_int32 something that they DEFINE_bool'd
-// elsewhere.
-
-class GFLAGS_DLL_DECL FlagRegisterer {
- public:
-  FlagRegisterer(const char* name, const char* type,
-                 const char* help, const char* filename,
-                 void* current_storage, void* defvalue_storage);
-};
-
-// If your application #defines STRIP_FLAG_HELP to a non-zero value
-// before #including this file, we remove the help message from the
-// binary file. This can reduce the size of the resulting binary
-// somewhat, and may also be useful for security reasons.
-
-extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
-
-
-} // namespace GFLAGS_NAMESPACE
-
-
-#ifndef SWIG  // In swig, ignore the main flag declarations
-
-#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
-// Need this construct to avoid the 'defined but not used' warning.
-#define MAYBE_STRIPPED_HELP(txt) \
-   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
-#else
-#define MAYBE_STRIPPED_HELP(txt) txt
-#endif
-
-// Each command-line flag has two variables associated with it: one
-// with the current value, and one with the default value.  However,
-// we have a third variable, which is where value is assigned; it's a
-// constant.  This guarantees that FLAG_##value is initialized at
-// static initialization time (e.g. before program-start) rather than
-// than global construction time (which is after program-start but
-// before main), at least when 'value' is a compile-time constant.  We
-// use a small trick for the "default value" variable, and call it
-// FLAGS_no<name>.  This serves the second purpose of assuring a
-// compile error if someone tries to define a flag named no<name>
-// which is illegal (--foo and --nofoo both affect the "foo" flag).
-#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
-  namespace fL##shorttype {                                             \
-    static const type FLAGS_nono##name = value;                         \
-    /* We always want to export defined variables, dll or no */         \
-    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
-    type FLAGS_no##name = FLAGS_nono##name;                             \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
-      #name, #type, MAYBE_STRIPPED_HELP(help), __FILE__,                \
-      &FLAGS_##name, &FLAGS_no##name);                                  \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// For DEFINE_bool, we want to do the extra check that the passed-in
-// value is actually a bool, and not a string or something that can be
-// coerced to a bool.  These declarations (no definition needed!) will
-// help us do that, and never evaluate From, which is important.
-// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
-// that the compiler have different sizes for bool & double. Since
-// this is not guaranteed by the standard, we check it with a
-// COMPILE_ASSERT.
-namespace fLB {
-struct CompileAssert {};
-typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
-                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
-template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
-GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
-}  // namespace fLB
-
-// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
-// are in a separate include, gflags_declare.h, for reducing
-// the physical transitive size for DECLARE use.
-#define DEFINE_bool(name, val, txt)                                     \
-  namespace fLB {                                                       \
-    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
-            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
-  }                                                                     \
-  DEFINE_VARIABLE(bool, B, name, val, txt)
-
-#define DEFINE_int32(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
-                   name, val, txt)
-
-#define DEFINE_int64(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
-                   name, val, txt)
-
-#define DEFINE_uint64(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
-                   name, val, txt)
-
-#define DEFINE_double(name, val, txt) \
-   DEFINE_VARIABLE(double, D, name, val, txt)
-
-// Strings are trickier, because they're not a POD, so we can't
-// construct them at static-initialization time (instead they get
-// constructed at global-constructor time, which is much later).  To
-// try to avoid crashes in that case, we use a char buffer to store
-// the string, which we can static-initialize, and then placement-new
-// into it later.  It's not perfect, but the best we can do.
-
-namespace fLS {
-
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const char *value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const clstring &value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           int value);
-}  // namespace fLS
-
-// We need to define a var named FLAGS_no##name so people don't define
-// --string and --nostring.  And we need a temporary place to put val
-// so we don't have to evaluate it twice.  Two great needs that go
-// great together!
-// The weird 'using' + 'extern' inside the fLS namespace is to work around
-// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
-//    http://code.google.com/p/google-gflags/issues/detail?id=20
-#define DEFINE_string(name, val, txt)                                       \
-  namespace fLS {                                                           \
-    using ::fLS::clstring;                                                  \
-    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
-    clstring* const FLAGS_no##name = ::fLS::                                \
-                                   dont_pass0toDEFINE_string(s_##name[0].s, \
-                                                             val);          \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
-        #name, "string", MAYBE_STRIPPED_HELP(txt), __FILE__,                \
-        s_##name[0].s, new (s_##name[1].s) clstring(*FLAGS_no##name));      \
-    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
-    using fLS::FLAGS_##name;                                                \
-    clstring& FLAGS_##name = *FLAGS_no##name;                               \
-  }                                                                         \
-  using fLS::FLAGS_##name
-
-#endif  // SWIG
-
-
-// Import gflags library symbols into alternative/deprecated namespace(s)
-#include "gflags_gflags.h"
-
-
-#endif  // GFLAGS_GFLAGS_H_
diff --git a/files/third_party/gflags/gen/posix/include/gflags/gflags_completions.h b/files/third_party/gflags/gen/posix/include/gflags/gflags_completions.h
deleted file mode 100644
index f951c1e0..00000000
--- a/files/third_party/gflags/gen/posix/include/gflags/gflags_completions.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright (c) 2008, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// ---
-
-//
-// Implement helpful bash-style command line flag completions
-//
-// ** Functional API:
-// HandleCommandLineCompletions() should be called early during
-// program startup, but after command line flag code has been
-// initialized, such as the beginning of HandleCommandLineHelpFlags().
-// It checks the value of the flag --tab_completion_word.  If this
-// flag is empty, nothing happens here.  If it contains a string,
-// however, then HandleCommandLineCompletions() will hijack the
-// process, attempting to identify the intention behind this
-// completion.  Regardless of the outcome of this deduction, the
-// process will be terminated, similar to --helpshort flag
-// handling.
-//
-// ** Overview of Bash completions:
-// Bash can be told to programatically determine completions for the
-// current 'cursor word'.  It does this by (in this case) invoking a
-// command with some additional arguments identifying the command
-// being executed, the word being completed, and the previous word
-// (if any).  Bash then expects a sequence of output lines to be
-// printed to stdout.  If these lines all contain a common prefix
-// longer than the cursor word, bash will replace the cursor word
-// with that common prefix, and display nothing.  If there isn't such
-// a common prefix, bash will display the lines in pages using 'more'.
-//
-// ** Strategy taken for command line completions:
-// If we can deduce either the exact flag intended, or a common flag
-// prefix, we'll output exactly that.  Otherwise, if information
-// must be displayed to the user, we'll take the opportunity to add
-// some helpful information beyond just the flag name (specifically,
-// we'll include the default flag value and as much of the flag's
-// description as can fit on a single terminal line width, as specified
-// by the flag --tab_completion_columns).  Furthermore, we'll try to
-// make bash order the output such that the most useful or relevent
-// flags are the most likely to be shown at the top.
-//
-// ** Additional features:
-// To assist in finding that one really useful flag, substring matching
-// was implemented.  Before pressing a <TAB> to get completion for the
-// current word, you can append one or more '?' to the flag to do
-// substring matching.  Here's the semantics:
-//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
-//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
-//   --foo??<TAB>   Same as prior case, but also search in module
-//                  definition path for 'foo'
-//   --foo???<TAB>  Same as prior case, but also search in flag
-//                  descriptions for 'foo'
-// Finally, we'll trim the output to a relatively small number of
-// flags to keep bash quiet about the verbosity of output.  If one
-// really wanted to see all possible matches, appending a '+' to the
-// search word will force the exhaustive list of matches to be printed.
-//
-// ** How to have bash accept completions from a binary:
-// Bash requires that it be informed about each command that programmatic
-// completion should be enabled for.  Example addition to a .bashrc
-// file would be (your path to gflags_completions.sh file may differ):
-
-/*
-$ complete -o bashdefault -o default -o nospace -C                            \
- '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
-  time  env  binary_name  another_binary  [...]
-*/
-
-// This would allow the following to work:
-//   $ /path/to/binary_name --vmodule<TAB>
-// Or:
-//   $ ./bin/path/another_binary --gfs_u<TAB>
-// (etc)
-//
-// Sadly, it appears that bash gives no easy way to force this behavior for
-// all commands.  That's where the "time" in the above example comes in.
-// If you haven't specifically added a command to the list of completion
-// supported commands, you can still get completions by prefixing the
-// entire command with "env".
-//   $ env /some/brand/new/binary --vmod<TAB>
-// Assuming that "binary" is a newly compiled binary, this should still
-// produce the expected completion output.
-
-
-#ifndef GFLAGS_COMPLETIONS_H_
-#define GFLAGS_COMPLETIONS_H_
-
-namespace google {
-
-extern void HandleCommandLineCompletions(void);
-
-}
-
-#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/files/third_party/gflags/gen/posix/include/gflags/gflags_declare.h b/files/third_party/gflags/gen/posix/include/gflags/gflags_declare.h
deleted file mode 100644
index 935a20e7..00000000
--- a/files/third_party/gflags/gen/posix/include/gflags/gflags_declare.h
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-//
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// command line flag.
-
-#ifndef GFLAGS_DECLARE_H_
-#define GFLAGS_DECLARE_H_
-
-
-// ---------------------------------------------------------------------------
-// Namespace of gflags library symbols.
-#define GFLAGS_NAMESPACE google
-
-// ---------------------------------------------------------------------------
-// Windows DLL import/export.
-
-// We always want to import the symbols of the gflags library
-#ifndef GFLAGS_DLL_DECL
-#  if 0 && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-
-// We always want to import variables declared in user code
-#ifndef GFLAGS_DLL_DECLARE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECLARE_FLAG
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Flag types
-#include <string>
-#if 1
-#  include <stdint.h>                   // the normal place uint32_t is defined
-#elif 1
-#  include <sys/types.h>                // the normal place u_int32_t is defined
-#elif 1
-#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
-#endif
-
-namespace GFLAGS_NAMESPACE {
-
-#if 1 // C99
-typedef int32_t          int32;
-typedef uint32_t         uint32;
-typedef int64_t          int64;
-typedef uint64_t         uint64;
-#elif 0 // BSD
-typedef int32_t          int32;
-typedef u_int32_t        uint32;
-typedef int64_t          int64;
-typedef u_int64_t        uint64;
-#elif 0 // Windows
-typedef __int32          int32;
-typedef unsigned __int32 uint32;
-typedef __int64          int64;
-typedef unsigned __int64 uint64;
-#else
-#  error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-} // namespace GFLAGS_NAMESPACE
-
-
-namespace fLS {
-
-// The meaning of "string" might be different between now and when the
-// macros below get invoked (e.g., if someone is experimenting with
-// other string implementations that get defined after this file is
-// included).  Save the current meaning now and use it in the macros.
-typedef std::string clstring;
-
-} // namespace fLS
-
-
-#define DECLARE_VARIABLE(type, shorttype, name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
-  using fL##shorttype::FLAGS_##name
-
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name)
-
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
-
-#define DECLARE_int64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
-
-#define DECLARE_uint64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
-
-#define DECLARE_double(name) \
-  DECLARE_VARIABLE(double, D, name)
-
-#define DECLARE_string(name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fLS { \
-  using ::fLS::clstring; \
-  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
-  } \
-  using fLS::FLAGS_##name
-
-
-#endif  // GFLAGS_DECLARE_H_
diff --git a/files/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h b/files/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h
deleted file mode 100644
index 0c17825d..00000000
--- a/files/third_party/gflags/gen/posix/include/gflags/gflags_gflags.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Copyright (c) 2014, Andreas Schuh
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// -----------------------------------------------------------------------------
-// Imports the gflags library symbols into an alternative/deprecated namespace.
-
-#ifndef GFLAGS_GFLAGS_H_
-#  error The internal header gflags_gflags.h may only be included by gflags.h
-#endif
-
-#ifndef GFLAGS_NS_GFLAGS_H_
-#define GFLAGS_NS_GFLAGS_H_
-
-
-namespace gflags {
-
-
-using GFLAGS_NAMESPACE::int32;
-using GFLAGS_NAMESPACE::uint32;
-using GFLAGS_NAMESPACE::int64;
-using GFLAGS_NAMESPACE::uint64;
-
-using GFLAGS_NAMESPACE::RegisterFlagValidator;
-using GFLAGS_NAMESPACE::CommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetAllFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlagsRestrict;
-using GFLAGS_NAMESPACE::DescribeOneFlag;
-using GFLAGS_NAMESPACE::SetArgv;
-using GFLAGS_NAMESPACE::GetArgvs;
-using GFLAGS_NAMESPACE::GetArgv;
-using GFLAGS_NAMESPACE::GetArgv0;
-using GFLAGS_NAMESPACE::GetArgvSum;
-using GFLAGS_NAMESPACE::ProgramInvocationName;
-using GFLAGS_NAMESPACE::ProgramInvocationShortName;
-using GFLAGS_NAMESPACE::ProgramUsage;
-using GFLAGS_NAMESPACE::VersionString;
-using GFLAGS_NAMESPACE::GetCommandLineOption;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfoOrDie;
-using GFLAGS_NAMESPACE::FlagSettingMode;
-using GFLAGS_NAMESPACE::SET_FLAGS_VALUE;
-using GFLAGS_NAMESPACE::SET_FLAG_IF_DEFAULT;
-using GFLAGS_NAMESPACE::SET_FLAGS_DEFAULT;
-using GFLAGS_NAMESPACE::SetCommandLineOption;
-using GFLAGS_NAMESPACE::SetCommandLineOptionWithMode;
-using GFLAGS_NAMESPACE::FlagSaver;
-using GFLAGS_NAMESPACE::CommandlineFlagsIntoString;
-using GFLAGS_NAMESPACE::ReadFlagsFromString;
-using GFLAGS_NAMESPACE::AppendFlagsIntoFile;
-using GFLAGS_NAMESPACE::ReadFromFlagsFile;
-using GFLAGS_NAMESPACE::BoolFromEnv;
-using GFLAGS_NAMESPACE::Int32FromEnv;
-using GFLAGS_NAMESPACE::Int64FromEnv;
-using GFLAGS_NAMESPACE::Uint64FromEnv;
-using GFLAGS_NAMESPACE::DoubleFromEnv;
-using GFLAGS_NAMESPACE::StringFromEnv;
-using GFLAGS_NAMESPACE::SetUsageMessage;
-using GFLAGS_NAMESPACE::SetVersionString;
-using GFLAGS_NAMESPACE::ParseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::HandleCommandLineHelpFlags;
-using GFLAGS_NAMESPACE::AllowCommandLineReparsing;
-using GFLAGS_NAMESPACE::ReparseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::ShutDownCommandLineFlags;
-using GFLAGS_NAMESPACE::FlagRegisterer;
-
-#ifndef SWIG
-using GFLAGS_NAMESPACE::ParseCommandLineFlags;
-#endif
-
-
-} // namespace gflags
-
-
-#endif  // GFLAGS_NS_GFLAGS_H_
diff --git a/files/third_party/gflags/gen/posix/include/private/config.h b/files/third_party/gflags/gen/posix/include/private/config.h
deleted file mode 100644
index 592d61c4..00000000
--- a/files/third_party/gflags/gen/posix/include/private/config.h
+++ /dev/null
@@ -1,112 +0,0 @@
-/* Generated from config.h.in during build configuration using CMake. */
-
-// Note: This header file is only used internally. It is not part of public interface!
-
-// ---------------------------------------------------------------------------
-// System checks
-
-// Define if you build this library for a MS Windows OS.
-/* #undef OS_WINDOWS */
-
-// Define if you have the <stdint.h> header file.
-#define HAVE_STDINT_H
-
-// Define if you have the <sys/types.h> header file.
-#define HAVE_SYS_TYPES_H
-
-// Define if you have the <inttypes.h> header file.
-#define HAVE_INTTYPES_H
-
-// Define if you have the <sys/stat.h> header file.
-#define HAVE_SYS_STAT_H
-
-// Define if you have the <unistd.h> header file.
-#define HAVE_UNISTD_H
-
-// Define if you have the <fnmatch.h> header file.
-#define HAVE_FNMATCH_H
-
-// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
-/* #undef HAVE_SHLWAPI_H */
-
-// Define if you have the strtoll function.
-#define HAVE_STRTOLL
-
-// Define if you have the strtoq function.
-/* #undef HAVE_STRTOQ */
-
-// Define if you have the <pthread.h> header file.
-#define HAVE_PTHREAD
-
-// Define if your pthread library defines the type pthread_rwlock_t
-#define HAVE_RWLOCK
-
-// gcc requires this to get PRId64, etc.
-#if defined(HAVE_INTTYPES_H) && !defined(__STDC_FORMAT_MACROS)
-#  define __STDC_FORMAT_MACROS 1
-#endif
-
-// ---------------------------------------------------------------------------
-// Package information
-
-// Name of package.
-#define PACKAGE gflags
-
-// Define to the full name of this package.
-#define PACKAGE_NAME gflags
-
-// Define to the full name and version of this package.
-#define PACKAGE_STRING gflags 2.2.0
-
-// Define to the one symbol short name of this package.
-#define PACKAGE_TARNAME gflags-2.2.0
-
-// Define to the version of this package.
-#define PACKAGE_VERSION 2.2.0
-
-// Version number of package.
-#define VERSION PACKAGE_VERSION
-
-// Define to the address where bug reports for this package should be sent.
-#define PACKAGE_BUGREPORT https://github.com/schuhschuh/gflags/issues
-
-// ---------------------------------------------------------------------------
-// Path separator
-#ifndef PATH_SEPARATOR
-#  ifdef OS_WINDOWS
-#    define PATH_SEPARATOR  '\\'
-#  else
-#    define PATH_SEPARATOR  '/'
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Windows
-
-// Whether gflags library is a DLL.
-#ifndef GFLAGS_IS_A_DLL
-#  define GFLAGS_IS_A_DLL 0
-#endif
-
-// Always export symbols when compiling a shared library as this file is only
-// included by internal modules when building the gflags library itself.
-// The gflags_declare.h header file will set it to import these symbols otherwise.
-#ifndef GFLAGS_DLL_DECL
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-// Flags defined by the gflags library itself must be exported
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  define GFLAGS_DLL_DEFINE_FLAG GFLAGS_DLL_DECL
-#endif
-
-#ifdef OS_WINDOWS
-// The unittests import the symbols of the shared gflags library
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL_FOR_UNITTESTS __declspec(dllimport)
-#  endif
-#  include "windows_port.h"
-#endif
diff --git a/files/third_party/gflags/gen/win/include/gflags/gflags.h b/files/third_party/gflags/gen/win/include/gflags/gflags.h
deleted file mode 100644
index 357eec6b..00000000
--- a/files/third_party/gflags/gen/win/include/gflags/gflags.h
+++ /dev/null
@@ -1,573 +0,0 @@
-// Copyright (c) 2006, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// or defines a command line flag or wants to parse command line flags
-// or print a program usage message (which will include information about
-// flags).  Executive summary, in the form of an example foo.cc file:
-//
-//    #include "foo.h"         // foo.h has a line "DECLARE_int32(start);"
-//    #include "validators.h"  // hypothetical file defining ValidateIsFile()
-//
-//    DEFINE_int32(end, 1000, "The last record to read");
-//
-//    DEFINE_string(filename, "my_file.txt", "The file to read");
-//    // Crash if the specified file does not exist.
-//    static bool dummy = RegisterFlagValidator(&FLAGS_filename,
-//                                              &ValidateIsFile);
-//
-//    DECLARE_bool(verbose); // some other file has a DEFINE_bool(verbose, ...)
-//
-//    void MyFunc() {
-//      if (FLAGS_verbose) printf("Records %d-%d\n", FLAGS_start, FLAGS_end);
-//    }
-//
-//    Then, at the command-line:
-//       ./foo --noverbose --start=5 --end=100
-//
-// For more details, see
-//    doc/gflags.html
-//
-// --- A note about thread-safety:
-//
-// We describe many functions in this routine as being thread-hostile,
-// thread-compatible, or thread-safe.  Here are the meanings we use:
-//
-// thread-safe: it is safe for multiple threads to call this routine
-//   (or, when referring to a class, methods of this class)
-//   concurrently.
-// thread-hostile: it is not safe for multiple threads to call this
-//   routine (or methods of this class) concurrently.  In gflags,
-//   most thread-hostile routines are intended to be called early in,
-//   or even before, main() -- that is, before threads are spawned.
-// thread-compatible: it is safe for multiple threads to read from
-//   this variable (when applied to variables), or to call const
-//   methods of this class (when applied to classes), as long as no
-//   other thread is writing to the variable or calling non-const
-//   methods of this class.
-
-#ifndef GFLAGS_GFLAGS_H_
-#define GFLAGS_GFLAGS_H_
-
-#include <string>
-#include <vector>
-
-#include "gflags_declare.h" // IWYU pragma: export
-
-
-// We always want to export variables defined in user code
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DEFINE_FLAG __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DEFINE_FLAG
-#  endif
-#endif
-
-
-namespace GFLAGS_NAMESPACE {
-
-
-// --------------------------------------------------------------------
-// To actually define a flag in a file, use DEFINE_bool,
-// DEFINE_string, etc. at the bottom of this file.  You may also find
-// it useful to register a validator with the flag.  This ensures that
-// when the flag is parsed from the commandline, or is later set via
-// SetCommandLineOption, we call the validation function. It is _not_
-// called when you assign the value to the flag directly using the = operator.
-//
-// The validation function should return true if the flag value is valid, and
-// false otherwise. If the function returns false for the new setting of the
-// flag, the flag will retain its current value. If it returns false for the
-// default value, ParseCommandLineFlags() will die.
-//
-// This function is safe to call at global construct time (as in the
-// example below).
-//
-// Example use:
-//    static bool ValidatePort(const char* flagname, int32 value) {
-//       if (value > 0 && value < 32768)   // value is ok
-//         return true;
-//       printf("Invalid value for --%s: %d\n", flagname, (int)value);
-//       return false;
-//    }
-//    DEFINE_int32(port, 0, "What port to listen on");
-//    static bool dummy = RegisterFlagValidator(&FLAGS_port, &ValidatePort);
-
-// Returns true if successfully registered, false if not (because the
-// first argument doesn't point to a command-line flag, or because a
-// validator is already registered for this flag).
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const bool*        flag, bool (*validate_fn)(const char*, bool));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int32*       flag, bool (*validate_fn)(const char*, int32));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const int64*       flag, bool (*validate_fn)(const char*, int64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const uint64*      flag, bool (*validate_fn)(const char*, uint64));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const double*      flag, bool (*validate_fn)(const char*, double));
-extern GFLAGS_DLL_DECL bool RegisterFlagValidator(const std::string* flag, bool (*validate_fn)(const char*, const std::string&));
-
-// Convenience macro for the registration of a flag validator
-#define DEFINE_validator(name, validator) \
-    static const bool name##_validator_registered = \
-            GFLAGS_NAMESPACE::RegisterFlagValidator(&FLAGS_##name, validator)
-
-
-// --------------------------------------------------------------------
-// These methods are the best way to get access to info about the
-// list of commandline flags.  Note that these routines are pretty slow.
-//   GetAllFlags: mostly-complete info about the list, sorted by file.
-//   ShowUsageWithFlags: pretty-prints the list to stdout (what --help does)
-//   ShowUsageWithFlagsRestrict: limit to filenames with restrict as a substr
-//
-// In addition to accessing flags, you can also access argv[0] (the program
-// name) and argv (the entire commandline), which we sock away a copy of.
-// These variables are static, so you should only set them once.
-//
-// No need to export this data only structure from DLL, avoiding VS warning 4251.
-struct CommandLineFlagInfo {
-  std::string name;            // the name of the flag
-  std::string type;            // the type of the flag: int32, etc
-  std::string description;     // the "help text" associated with the flag
-  std::string current_value;   // the current value, as a string
-  std::string default_value;   // the default value, as a string
-  std::string filename;        // 'cleaned' version of filename holding the flag
-  bool has_validator_fn;       // true if RegisterFlagValidator called on this flag
-  bool is_default;             // true if the flag has the default value and
-                               // has not been set explicitly from the cmdline
-                               // or via SetCommandLineOption
-  const void* flag_ptr;        // pointer to the flag's current value (i.e. FLAGS_foo)
-};
-
-// Using this inside of a validator is a recipe for a deadlock.
-// TODO(user) Fix locking when validators are running, to make it safe to
-// call validators during ParseAllFlags.
-// Also make sure then to uncomment the corresponding unit test in
-// gflags_unittest.sh
-extern GFLAGS_DLL_DECL void GetAllFlags(std::vector<CommandLineFlagInfo>* OUTPUT);
-// These two are actually defined in gflags_reporting.cc.
-extern GFLAGS_DLL_DECL void ShowUsageWithFlags(const char *argv0);  // what --help does
-extern GFLAGS_DLL_DECL void ShowUsageWithFlagsRestrict(const char *argv0, const char *restrict);
-
-// Create a descriptive string for a flag.
-// Goes to some trouble to make pretty line breaks.
-extern GFLAGS_DLL_DECL std::string DescribeOneFlag(const CommandLineFlagInfo& flag);
-
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetArgv(int argc, const char** argv);
-
-// The following functions are thread-safe as long as SetArgv() is
-// only called before any threads start.
-extern GFLAGS_DLL_DECL const std::vector<std::string>& GetArgvs();
-extern GFLAGS_DLL_DECL const char* GetArgv();                      // all of argv as a string
-extern GFLAGS_DLL_DECL const char* GetArgv0();                     // only argv0
-extern GFLAGS_DLL_DECL uint32 GetArgvSum();                        // simple checksum of argv
-extern GFLAGS_DLL_DECL const char* ProgramInvocationName();        // argv0, or "UNKNOWN" if not set
-extern GFLAGS_DLL_DECL const char* ProgramInvocationShortName();   // basename(argv0)
-
-// ProgramUsage() is thread-safe as long as SetUsageMessage() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* ProgramUsage();                 // string set by SetUsageMessage()
-
-// VersionString() is thread-safe as long as SetVersionString() is only
-// called before any threads start.
-extern GFLAGS_DLL_DECL const char* VersionString();                // string set by SetVersionString()
-
-
-
-// --------------------------------------------------------------------
-// Normally you access commandline flags by just saying "if (FLAGS_foo)"
-// or whatever, and set them by calling "FLAGS_foo = bar" (or, more
-// commonly, via the DEFINE_foo macro).  But if you need a bit more
-// control, we have programmatic ways to get/set the flags as well.
-// These programmatic ways to access flags are thread-safe, but direct
-// access is only thread-compatible.
-
-// Return true iff the flagname was found.
-// OUTPUT is set to the flag's value, or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineOption(const char* name, std::string* OUTPUT);
-
-// Return true iff the flagname was found. OUTPUT is set to the flag's
-// CommandLineFlagInfo or unchanged if we return false.
-extern GFLAGS_DLL_DECL bool GetCommandLineFlagInfo(const char* name, CommandLineFlagInfo* OUTPUT);
-
-// Return the CommandLineFlagInfo of the flagname.  exit() if name not found.
-// Example usage, to check if a flag's value is currently the default value:
-//   if (GetCommandLineFlagInfoOrDie("foo").is_default) ...
-extern GFLAGS_DLL_DECL CommandLineFlagInfo GetCommandLineFlagInfoOrDie(const char* name);
-
-enum GFLAGS_DLL_DECL FlagSettingMode {
-  // update the flag's value (can call this multiple times).
-  SET_FLAGS_VALUE,
-  // update the flag's value, but *only if* it has not yet been updated
-  // with SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef".
-  SET_FLAG_IF_DEFAULT,
-  // set the flag's default value to this.  If the flag has not yet updated
-  // yet (via SET_FLAGS_VALUE, SET_FLAG_IF_DEFAULT, or "FLAGS_xxx = nondef")
-  // change the flag's current value to the new default value as well.
-  SET_FLAGS_DEFAULT
-};
-
-// Set a particular flag ("command line option").  Returns a string
-// describing the new value that the option has been set to.  The
-// return value API is not well-specified, so basically just depend on
-// it to be empty if the setting failed for some reason -- the name is
-// not a valid flag name, or the value is not a valid value -- and
-// non-empty else.
-
-// SetCommandLineOption uses set_mode == SET_FLAGS_VALUE (the common case)
-extern GFLAGS_DLL_DECL std::string SetCommandLineOption        (const char* name, const char* value);
-extern GFLAGS_DLL_DECL std::string SetCommandLineOptionWithMode(const char* name, const char* value, FlagSettingMode set_mode);
-
-
-// --------------------------------------------------------------------
-// Saves the states (value, default value, whether the user has set
-// the flag, registered validators, etc) of all flags, and restores
-// them when the FlagSaver is destroyed.  This is very useful in
-// tests, say, when you want to let your tests change the flags, but
-// make sure that they get reverted to the original states when your
-// test is complete.
-//
-// Example usage:
-//   void TestFoo() {
-//     FlagSaver s1;
-//     FLAG_foo = false;
-//     FLAG_bar = "some value";
-//
-//     // test happens here.  You can return at any time
-//     // without worrying about restoring the FLAG values.
-//   }
-//
-// Note: This class is marked with GFLAGS_ATTRIBUTE_UNUSED because all
-// the work is done in the constructor and destructor, so in the standard
-// usage example above, the compiler would complain that it's an
-// unused variable.
-//
-// This class is thread-safe.  However, its destructor writes to
-// exactly the set of flags that have changed value during its
-// lifetime, so concurrent _direct_ access to those flags
-// (i.e. FLAGS_foo instead of {Get,Set}CommandLineOption()) is unsafe.
-
-class GFLAGS_DLL_DECL FlagSaver {
- public:
-  FlagSaver();
-  ~FlagSaver();
-
- private:
-  class FlagSaverImpl* impl_;   // we use pimpl here to keep API steady
-
-  FlagSaver(const FlagSaver&);  // no copying!
-  void operator=(const FlagSaver&);
-};
-
-// --------------------------------------------------------------------
-// Some deprecated or hopefully-soon-to-be-deprecated functions.
-
-// This is often used for logging.  TODO(csilvers): figure out a better way
-extern GFLAGS_DLL_DECL std::string CommandlineFlagsIntoString();
-// Usually where this is used, a FlagSaver should be used instead.
-extern GFLAGS_DLL_DECL
-bool ReadFlagsFromString(const std::string& flagfilecontents,
-                         const char* prog_name,
-                         bool errors_are_fatal);  // uses SET_FLAGS_VALUE
-
-// These let you manually implement --flagfile functionality.
-// DEPRECATED.
-extern GFLAGS_DLL_DECL bool AppendFlagsIntoFile(const std::string& filename, const char* prog_name);
-extern GFLAGS_DLL_DECL bool ReadFromFlagsFile(const std::string& filename, const char* prog_name, bool errors_are_fatal);   // uses SET_FLAGS_VALUE
-
-
-// --------------------------------------------------------------------
-// Useful routines for initializing flags from the environment.
-// In each case, if 'varname' does not exist in the environment
-// return defval.  If 'varname' does exist but is not valid
-// (e.g., not a number for an int32 flag), abort with an error.
-// Otherwise, return the value.  NOTE: for booleans, for true use
-// 't' or 'T' or 'true' or '1', for false 'f' or 'F' or 'false' or '0'.
-
-extern GFLAGS_DLL_DECL bool BoolFromEnv(const char *varname, bool defval);
-extern GFLAGS_DLL_DECL int32 Int32FromEnv(const char *varname, int32 defval);
-extern GFLAGS_DLL_DECL int64 Int64FromEnv(const char *varname, int64 defval);
-extern GFLAGS_DLL_DECL uint64 Uint64FromEnv(const char *varname, uint64 defval);
-extern GFLAGS_DLL_DECL double DoubleFromEnv(const char *varname, double defval);
-extern GFLAGS_DLL_DECL const char *StringFromEnv(const char *varname, const char *defval);
-
-
-// --------------------------------------------------------------------
-// The next two functions parse gflags from main():
-
-// Set the "usage" message for this program.  For example:
-//   string usage("This program does nothing.  Sample usage:\n");
-//   usage += argv[0] + " <uselessarg1> <uselessarg2>";
-//   SetUsageMessage(usage);
-// Do not include commandline flags in the usage: we do that for you!
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetUsageMessage(const std::string& usage);
-
-// Sets the version string, which is emitted with --version.
-// For instance: SetVersionString("1.3");
-// Thread-hostile; meant to be called before any threads are spawned.
-extern GFLAGS_DLL_DECL void SetVersionString(const std::string& version);
-
-
-// Looks for flags in argv and parses them.  Rearranges argv to put
-// flags first, or removes them entirely if remove_flags is true.
-// If a flag is defined more than once in the command line or flag
-// file, the last definition is used.  Returns the index (into argv)
-// of the first non-flag argument.
-// See top-of-file for more details on this function.
-#ifndef SWIG   // In swig, use ParseCommandLineFlagsScript() instead.
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineFlags(int *argc, char*** argv, bool remove_flags);
-#endif
-
-
-// Calls to ParseCommandLineNonHelpFlags and then to
-// HandleCommandLineHelpFlags can be used instead of a call to
-// ParseCommandLineFlags during initialization, in order to allow for
-// changing default values for some FLAGS (via
-// e.g. SetCommandLineOptionWithMode calls) between the time of
-// command line parsing and the time of dumping help information for
-// the flags as a result of command line parsing.  If a flag is
-// defined more than once in the command line or flag file, the last
-// definition is used.  Returns the index (into argv) of the first
-// non-flag argument.  (If remove_flags is true, will always return 1.)
-extern GFLAGS_DLL_DECL uint32 ParseCommandLineNonHelpFlags(int *argc, char*** argv, bool remove_flags);
-
-// This is actually defined in gflags_reporting.cc.
-// This function is misnamed (it also handles --version, etc.), but
-// it's too late to change that now. :-(
-extern GFLAGS_DLL_DECL void HandleCommandLineHelpFlags();   // in gflags_reporting.cc
-
-// Allow command line reparsing.  Disables the error normally
-// generated when an unknown flag is found, since it may be found in a
-// later parse.  Thread-hostile; meant to be called before any threads
-// are spawned.
-extern GFLAGS_DLL_DECL void AllowCommandLineReparsing();
-
-// Reparse the flags that have not yet been recognized.  Only flags
-// registered since the last parse will be recognized.  Any flag value
-// must be provided as part of the argument using "=", not as a
-// separate command line argument that follows the flag argument.
-// Intended for handling flags from dynamically loaded libraries,
-// since their flags are not registered until they are loaded.
-extern GFLAGS_DLL_DECL void ReparseCommandLineNonHelpFlags();
-
-// Clean up memory allocated by flags.  This is only needed to reduce
-// the quantity of "potentially leaked" reports emitted by memory
-// debugging tools such as valgrind.  It is not required for normal
-// operation, or for the google perftools heap-checker.  It must only
-// be called when the process is about to exit, and all threads that
-// might access flags are quiescent.  Referencing flags after this is
-// called will have unexpected consequences.  This is not safe to run
-// when multiple threads might be running: the function is
-// thread-hostile.
-extern GFLAGS_DLL_DECL void ShutDownCommandLineFlags();
-
-
-// --------------------------------------------------------------------
-// Now come the command line flag declaration/definition macros that
-// will actually be used.  They're kind of hairy.  A major reason
-// for this is initialization: we want people to be able to access
-// variables in global constructors and have that not crash, even if
-// their global constructor runs before the global constructor here.
-// (Obviously, we can't guarantee the flags will have the correct
-// default value in that case, but at least accessing them is safe.)
-// The only way to do that is have flags point to a static buffer.
-// So we make one, using a union to ensure proper alignment, and
-// then use placement-new to actually set up the flag with the
-// correct default value.  In the same vein, we have to worry about
-// flag access in global destructors, so FlagRegisterer has to be
-// careful never to destroy the flag-values it constructs.
-//
-// Note that when we define a flag variable FLAGS_<name>, we also
-// preemptively define a junk variable, FLAGS_no<name>.  This is to
-// cause a link-time error if someone tries to define 2 flags with
-// names like "logging" and "nologging".  We do this because a bool
-// flag FLAG can be set from the command line to true with a "-FLAG"
-// argument, and to false with a "-noFLAG" argument, and so this can
-// potentially avert confusion.
-//
-// We also put flags into their own namespace.  It is purposefully
-// named in an opaque way that people should have trouble typing
-// directly.  The idea is that DEFINE puts the flag in the weird
-// namespace, and DECLARE imports the flag from there into the current
-// namespace.  The net result is to force people to use DECLARE to get
-// access to a flag, rather than saying "extern GFLAGS_DLL_DECL bool FLAGS_whatever;"
-// or some such instead.  We want this so we can put extra
-// functionality (like sanity-checking) in DECLARE if we want, and
-// make sure it is picked up everywhere.
-//
-// We also put the type of the variable in the namespace, so that
-// people can't DECLARE_int32 something that they DEFINE_bool'd
-// elsewhere.
-
-class GFLAGS_DLL_DECL FlagRegisterer {
- public:
-  FlagRegisterer(const char* name, const char* type,
-                 const char* help, const char* filename,
-                 void* current_storage, void* defvalue_storage);
-};
-
-// If your application #defines STRIP_FLAG_HELP to a non-zero value
-// before #including this file, we remove the help message from the
-// binary file. This can reduce the size of the resulting binary
-// somewhat, and may also be useful for security reasons.
-
-extern GFLAGS_DLL_DECL const char kStrippedFlagHelp[];
-
-
-} // namespace GFLAGS_NAMESPACE
-
-
-#ifndef SWIG  // In swig, ignore the main flag declarations
-
-#if defined(STRIP_FLAG_HELP) && STRIP_FLAG_HELP > 0
-// Need this construct to avoid the 'defined but not used' warning.
-#define MAYBE_STRIPPED_HELP(txt) \
-   (false ? (txt) : GFLAGS_NAMESPACE::kStrippedFlagHelp)
-#else
-#define MAYBE_STRIPPED_HELP(txt) txt
-#endif
-
-// Each command-line flag has two variables associated with it: one
-// with the current value, and one with the default value.  However,
-// we have a third variable, which is where value is assigned; it's a
-// constant.  This guarantees that FLAG_##value is initialized at
-// static initialization time (e.g. before program-start) rather than
-// than global construction time (which is after program-start but
-// before main), at least when 'value' is a compile-time constant.  We
-// use a small trick for the "default value" variable, and call it
-// FLAGS_no<name>.  This serves the second purpose of assuring a
-// compile error if someone tries to define a flag named no<name>
-// which is illegal (--foo and --nofoo both affect the "foo" flag).
-#define DEFINE_VARIABLE(type, shorttype, name, value, help)             \
-  namespace fL##shorttype {                                             \
-    static const type FLAGS_nono##name = value;                         \
-    /* We always want to export defined variables, dll or no */         \
-    GFLAGS_DLL_DEFINE_FLAG type FLAGS_##name = FLAGS_nono##name;        \
-    type FLAGS_no##name = FLAGS_nono##name;                             \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                   \
-      #name, #type, MAYBE_STRIPPED_HELP(help), __FILE__,                \
-      &FLAGS_##name, &FLAGS_no##name);                                  \
-  }                                                                     \
-  using fL##shorttype::FLAGS_##name
-
-// For DEFINE_bool, we want to do the extra check that the passed-in
-// value is actually a bool, and not a string or something that can be
-// coerced to a bool.  These declarations (no definition needed!) will
-// help us do that, and never evaluate From, which is important.
-// We'll use 'sizeof(IsBool(val))' to distinguish. This code requires
-// that the compiler have different sizes for bool & double. Since
-// this is not guaranteed by the standard, we check it with a
-// COMPILE_ASSERT.
-namespace fLB {
-struct CompileAssert {};
-typedef CompileAssert expected_sizeof_double_neq_sizeof_bool[
-                      (sizeof(double) != sizeof(bool)) ? 1 : -1];
-template<typename From> double GFLAGS_DLL_DECL IsBoolFlag(const From& from);
-GFLAGS_DLL_DECL bool IsBoolFlag(bool from);
-}  // namespace fLB
-
-// Here are the actual DEFINE_*-macros. The respective DECLARE_*-macros
-// are in a separate include, gflags_declare.h, for reducing
-// the physical transitive size for DECLARE use.
-#define DEFINE_bool(name, val, txt)                                     \
-  namespace fLB {                                                       \
-    typedef ::fLB::CompileAssert FLAG_##name##_value_is_not_a_bool[     \
-            (sizeof(::fLB::IsBoolFlag(val)) != sizeof(double))? 1: -1]; \
-  }                                                                     \
-  DEFINE_VARIABLE(bool, B, name, val, txt)
-
-#define DEFINE_int32(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int32, I, \
-                   name, val, txt)
-
-#define DEFINE_int64(name, val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::int64, I64, \
-                   name, val, txt)
-
-#define DEFINE_uint64(name,val, txt) \
-   DEFINE_VARIABLE(GFLAGS_NAMESPACE::uint64, U64, \
-                   name, val, txt)
-
-#define DEFINE_double(name, val, txt) \
-   DEFINE_VARIABLE(double, D, name, val, txt)
-
-// Strings are trickier, because they're not a POD, so we can't
-// construct them at static-initialization time (instead they get
-// constructed at global-constructor time, which is much later).  To
-// try to avoid crashes in that case, we use a char buffer to store
-// the string, which we can static-initialize, and then placement-new
-// into it later.  It's not perfect, but the best we can do.
-
-namespace fLS {
-
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const char *value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           const clstring &value) {
-  return new(stringspot) clstring(value);
-}
-inline clstring* dont_pass0toDEFINE_string(char *stringspot,
-                                           int value);
-}  // namespace fLS
-
-// We need to define a var named FLAGS_no##name so people don't define
-// --string and --nostring.  And we need a temporary place to put val
-// so we don't have to evaluate it twice.  Two great needs that go
-// great together!
-// The weird 'using' + 'extern' inside the fLS namespace is to work around
-// an unknown compiler bug/issue with the gcc 4.2.1 on SUSE 10.  See
-//    http://code.google.com/p/google-gflags/issues/detail?id=20
-#define DEFINE_string(name, val, txt)                                       \
-  namespace fLS {                                                           \
-    using ::fLS::clstring;                                                  \
-    static union { void* align; char s[sizeof(clstring)]; } s_##name[2];    \
-    clstring* const FLAGS_no##name = ::fLS::                                \
-                                   dont_pass0toDEFINE_string(s_##name[0].s, \
-                                                             val);          \
-    static GFLAGS_NAMESPACE::FlagRegisterer o_##name(                       \
-        #name, "string", MAYBE_STRIPPED_HELP(txt), __FILE__,                \
-        s_##name[0].s, new (s_##name[1].s) clstring(*FLAGS_no##name));      \
-    extern GFLAGS_DLL_DEFINE_FLAG clstring& FLAGS_##name;                   \
-    using fLS::FLAGS_##name;                                                \
-    clstring& FLAGS_##name = *FLAGS_no##name;                               \
-  }                                                                         \
-  using fLS::FLAGS_##name
-
-#endif  // SWIG
-
-
-// Import gflags library symbols into alternative/deprecated namespace(s)
-#include "gflags_gflags.h"
-
-
-#endif  // GFLAGS_GFLAGS_H_
diff --git a/files/third_party/gflags/gen/win/include/gflags/gflags_completions.h b/files/third_party/gflags/gen/win/include/gflags/gflags_completions.h
deleted file mode 100644
index f951c1e0..00000000
--- a/files/third_party/gflags/gen/win/include/gflags/gflags_completions.h
+++ /dev/null
@@ -1,121 +0,0 @@
-// Copyright (c) 2008, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-//
-// ---
-
-//
-// Implement helpful bash-style command line flag completions
-//
-// ** Functional API:
-// HandleCommandLineCompletions() should be called early during
-// program startup, but after command line flag code has been
-// initialized, such as the beginning of HandleCommandLineHelpFlags().
-// It checks the value of the flag --tab_completion_word.  If this
-// flag is empty, nothing happens here.  If it contains a string,
-// however, then HandleCommandLineCompletions() will hijack the
-// process, attempting to identify the intention behind this
-// completion.  Regardless of the outcome of this deduction, the
-// process will be terminated, similar to --helpshort flag
-// handling.
-//
-// ** Overview of Bash completions:
-// Bash can be told to programatically determine completions for the
-// current 'cursor word'.  It does this by (in this case) invoking a
-// command with some additional arguments identifying the command
-// being executed, the word being completed, and the previous word
-// (if any).  Bash then expects a sequence of output lines to be
-// printed to stdout.  If these lines all contain a common prefix
-// longer than the cursor word, bash will replace the cursor word
-// with that common prefix, and display nothing.  If there isn't such
-// a common prefix, bash will display the lines in pages using 'more'.
-//
-// ** Strategy taken for command line completions:
-// If we can deduce either the exact flag intended, or a common flag
-// prefix, we'll output exactly that.  Otherwise, if information
-// must be displayed to the user, we'll take the opportunity to add
-// some helpful information beyond just the flag name (specifically,
-// we'll include the default flag value and as much of the flag's
-// description as can fit on a single terminal line width, as specified
-// by the flag --tab_completion_columns).  Furthermore, we'll try to
-// make bash order the output such that the most useful or relevent
-// flags are the most likely to be shown at the top.
-//
-// ** Additional features:
-// To assist in finding that one really useful flag, substring matching
-// was implemented.  Before pressing a <TAB> to get completion for the
-// current word, you can append one or more '?' to the flag to do
-// substring matching.  Here's the semantics:
-//   --foo<TAB>     Show me all flags with names prefixed by 'foo'
-//   --foo?<TAB>    Show me all flags with 'foo' somewhere in the name
-//   --foo??<TAB>   Same as prior case, but also search in module
-//                  definition path for 'foo'
-//   --foo???<TAB>  Same as prior case, but also search in flag
-//                  descriptions for 'foo'
-// Finally, we'll trim the output to a relatively small number of
-// flags to keep bash quiet about the verbosity of output.  If one
-// really wanted to see all possible matches, appending a '+' to the
-// search word will force the exhaustive list of matches to be printed.
-//
-// ** How to have bash accept completions from a binary:
-// Bash requires that it be informed about each command that programmatic
-// completion should be enabled for.  Example addition to a .bashrc
-// file would be (your path to gflags_completions.sh file may differ):
-
-/*
-$ complete -o bashdefault -o default -o nospace -C                            \
- '/home/build/eng/bash/bash_completions.sh --tab_completion_columns $COLUMNS' \
-  time  env  binary_name  another_binary  [...]
-*/
-
-// This would allow the following to work:
-//   $ /path/to/binary_name --vmodule<TAB>
-// Or:
-//   $ ./bin/path/another_binary --gfs_u<TAB>
-// (etc)
-//
-// Sadly, it appears that bash gives no easy way to force this behavior for
-// all commands.  That's where the "time" in the above example comes in.
-// If you haven't specifically added a command to the list of completion
-// supported commands, you can still get completions by prefixing the
-// entire command with "env".
-//   $ env /some/brand/new/binary --vmod<TAB>
-// Assuming that "binary" is a newly compiled binary, this should still
-// produce the expected completion output.
-
-
-#ifndef GFLAGS_COMPLETIONS_H_
-#define GFLAGS_COMPLETIONS_H_
-
-namespace google {
-
-extern void HandleCommandLineCompletions(void);
-
-}
-
-#endif  // GFLAGS_COMPLETIONS_H_
diff --git a/files/third_party/gflags/gen/win/include/gflags/gflags_declare.h b/files/third_party/gflags/gen/win/include/gflags/gflags_declare.h
deleted file mode 100644
index fbc8466f..00000000
--- a/files/third_party/gflags/gen/win/include/gflags/gflags_declare.h
+++ /dev/null
@@ -1,141 +0,0 @@
-// Copyright (c) 1999, Google Inc.
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// ---
-//
-// Revamped and reorganized by Craig Silverstein
-//
-// This is the file that should be included by any file which declares
-// command line flag.
-
-#ifndef GFLAGS_DECLARE_H_
-#define GFLAGS_DECLARE_H_
-
-
-// ---------------------------------------------------------------------------
-// Namespace of gflags library symbols.
-#define GFLAGS_NAMESPACE google
-
-// ---------------------------------------------------------------------------
-// Windows DLL import/export.
-
-// We always want to import the symbols of the gflags library
-#ifndef GFLAGS_DLL_DECL
-#  if 0 && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-
-// We always want to import variables declared in user code
-#ifndef GFLAGS_DLL_DECLARE_FLAG
-#  ifdef _MSC_VER
-#    define GFLAGS_DLL_DECLARE_FLAG __declspec(dllimport)
-#  else
-#    define GFLAGS_DLL_DECLARE_FLAG
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Flag types
-#include <string>
-#if 1
-#  include <stdint.h>                   // the normal place uint32_t is defined
-#elif 1
-#  include <sys/types.h>                // the normal place u_int32_t is defined
-#elif 0
-#  include <inttypes.h>                 // a third place for uint32_t or u_int32_t
-#endif
-
-namespace GFLAGS_NAMESPACE {
-
-#if 0 // C99
-typedef int32_t          int32;
-typedef uint32_t         uint32;
-typedef int64_t          int64;
-typedef uint64_t         uint64;
-#elif 0 // BSD
-typedef int32_t          int32;
-typedef u_int32_t        uint32;
-typedef int64_t          int64;
-typedef u_int64_t        uint64;
-#elif 1 // Windows
-typedef __int32          int32;
-typedef unsigned __int32 uint32;
-typedef __int64          int64;
-typedef unsigned __int64 uint64;
-#else
-#  error Do not know how to define a 32-bit integer quantity on your system
-#endif
-
-} // namespace GFLAGS_NAMESPACE
-
-
-namespace fLS {
-
-// The meaning of "string" might be different between now and when the
-// macros below get invoked (e.g., if someone is experimenting with
-// other string implementations that get defined after this file is
-// included).  Save the current meaning now and use it in the macros.
-typedef std::string clstring;
-
-} // namespace fLS
-
-
-#define DECLARE_VARIABLE(type, shorttype, name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fL##shorttype { extern GFLAGS_DLL_DECLARE_FLAG type FLAGS_##name; } \
-  using fL##shorttype::FLAGS_##name
-
-#define DECLARE_bool(name) \
-  DECLARE_VARIABLE(bool, B, name)
-
-#define DECLARE_int32(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int32, I, name)
-
-#define DECLARE_int64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::int64, I64, name)
-
-#define DECLARE_uint64(name) \
-  DECLARE_VARIABLE(::GFLAGS_NAMESPACE::uint64, U64, name)
-
-#define DECLARE_double(name) \
-  DECLARE_VARIABLE(double, D, name)
-
-#define DECLARE_string(name) \
-  /* We always want to import declared variables, dll or no */ \
-  namespace fLS { \
-  using ::fLS::clstring; \
-  extern GFLAGS_DLL_DECLARE_FLAG ::fLS::clstring& FLAGS_##name; \
-  } \
-  using fLS::FLAGS_##name
-
-
-#endif  // GFLAGS_DECLARE_H_
diff --git a/files/third_party/gflags/gen/win/include/gflags/gflags_gflags.h b/files/third_party/gflags/gen/win/include/gflags/gflags_gflags.h
deleted file mode 100644
index 0c17825d..00000000
--- a/files/third_party/gflags/gen/win/include/gflags/gflags_gflags.h
+++ /dev/null
@@ -1,101 +0,0 @@
-// Copyright (c) 2014, Andreas Schuh
-// All rights reserved.
-//
-// Redistribution and use in source and binary forms, with or without
-// modification, are permitted provided that the following conditions are
-// met:
-//
-//     * Redistributions of source code must retain the above copyright
-// notice, this list of conditions and the following disclaimer.
-//     * Redistributions in binary form must reproduce the above
-// copyright notice, this list of conditions and the following disclaimer
-// in the documentation and/or other materials provided with the
-// distribution.
-//     * Neither the name of Google Inc. nor the names of its
-// contributors may be used to endorse or promote products derived from
-// this software without specific prior written permission.
-//
-// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
-// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
-// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
-// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
-// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
-// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
-// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
-// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
-// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
-// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-// -----------------------------------------------------------------------------
-// Imports the gflags library symbols into an alternative/deprecated namespace.
-
-#ifndef GFLAGS_GFLAGS_H_
-#  error The internal header gflags_gflags.h may only be included by gflags.h
-#endif
-
-#ifndef GFLAGS_NS_GFLAGS_H_
-#define GFLAGS_NS_GFLAGS_H_
-
-
-namespace gflags {
-
-
-using GFLAGS_NAMESPACE::int32;
-using GFLAGS_NAMESPACE::uint32;
-using GFLAGS_NAMESPACE::int64;
-using GFLAGS_NAMESPACE::uint64;
-
-using GFLAGS_NAMESPACE::RegisterFlagValidator;
-using GFLAGS_NAMESPACE::CommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetAllFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlags;
-using GFLAGS_NAMESPACE::ShowUsageWithFlagsRestrict;
-using GFLAGS_NAMESPACE::DescribeOneFlag;
-using GFLAGS_NAMESPACE::SetArgv;
-using GFLAGS_NAMESPACE::GetArgvs;
-using GFLAGS_NAMESPACE::GetArgv;
-using GFLAGS_NAMESPACE::GetArgv0;
-using GFLAGS_NAMESPACE::GetArgvSum;
-using GFLAGS_NAMESPACE::ProgramInvocationName;
-using GFLAGS_NAMESPACE::ProgramInvocationShortName;
-using GFLAGS_NAMESPACE::ProgramUsage;
-using GFLAGS_NAMESPACE::VersionString;
-using GFLAGS_NAMESPACE::GetCommandLineOption;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfo;
-using GFLAGS_NAMESPACE::GetCommandLineFlagInfoOrDie;
-using GFLAGS_NAMESPACE::FlagSettingMode;
-using GFLAGS_NAMESPACE::SET_FLAGS_VALUE;
-using GFLAGS_NAMESPACE::SET_FLAG_IF_DEFAULT;
-using GFLAGS_NAMESPACE::SET_FLAGS_DEFAULT;
-using GFLAGS_NAMESPACE::SetCommandLineOption;
-using GFLAGS_NAMESPACE::SetCommandLineOptionWithMode;
-using GFLAGS_NAMESPACE::FlagSaver;
-using GFLAGS_NAMESPACE::CommandlineFlagsIntoString;
-using GFLAGS_NAMESPACE::ReadFlagsFromString;
-using GFLAGS_NAMESPACE::AppendFlagsIntoFile;
-using GFLAGS_NAMESPACE::ReadFromFlagsFile;
-using GFLAGS_NAMESPACE::BoolFromEnv;
-using GFLAGS_NAMESPACE::Int32FromEnv;
-using GFLAGS_NAMESPACE::Int64FromEnv;
-using GFLAGS_NAMESPACE::Uint64FromEnv;
-using GFLAGS_NAMESPACE::DoubleFromEnv;
-using GFLAGS_NAMESPACE::StringFromEnv;
-using GFLAGS_NAMESPACE::SetUsageMessage;
-using GFLAGS_NAMESPACE::SetVersionString;
-using GFLAGS_NAMESPACE::ParseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::HandleCommandLineHelpFlags;
-using GFLAGS_NAMESPACE::AllowCommandLineReparsing;
-using GFLAGS_NAMESPACE::ReparseCommandLineNonHelpFlags;
-using GFLAGS_NAMESPACE::ShutDownCommandLineFlags;
-using GFLAGS_NAMESPACE::FlagRegisterer;
-
-#ifndef SWIG
-using GFLAGS_NAMESPACE::ParseCommandLineFlags;
-#endif
-
-
-} // namespace gflags
-
-
-#endif  // GFLAGS_NS_GFLAGS_H_
diff --git a/files/third_party/gflags/gen/win/include/private/config.h b/files/third_party/gflags/gen/win/include/private/config.h
deleted file mode 100644
index d541580e..00000000
--- a/files/third_party/gflags/gen/win/include/private/config.h
+++ /dev/null
@@ -1,112 +0,0 @@
-/* Generated from config.h.in during build configuration using CMake. */
-
-// Note: This header file is only used internally. It is not part of public interface!
-
-// ---------------------------------------------------------------------------
-// System checks
-
-// Define if you build this library for a MS Windows OS.
-#define OS_WINDOWS
-
-// Define if you have the <stdint.h> header file.
-#define HAVE_STDINT_H
-
-// Define if you have the <sys/types.h> header file.
-#define HAVE_SYS_TYPES_H
-
-// Define if you have the <inttypes.h> header file.
-/* #undef HAVE_INTTYPES_H */
-
-// Define if you have the <sys/stat.h> header file.
-#define HAVE_SYS_STAT_H
-
-// Define if you have the <unistd.h> header file.
-/* #undef HAVE_UNISTD_H */
-
-// Define if you have the <fnmatch.h> header file.
-/* #undef HAVE_FNMATCH_H */
-
-// Define if you have the <shlwapi.h> header file (Windows 2000/XP).
-#define HAVE_SHLWAPI_H
-
-// Define if you have the strtoll function.
-/* #undef HAVE_STRTOLL */
-
-// Define if you have the strtoq function.
-/* #undef HAVE_STRTOQ */
-
-// Define if you have the <pthread.h> header file.
-/* #undef HAVE_PTHREAD */
-
-// Define if your pthread library defines the type pthread_rwlock_t
-/* #undef HAVE_RWLOCK */
-
-// gcc requires this to get PRId64, etc.
-#if defined(HAVE_INTTYPES_H) && !defined(__STDC_FORMAT_MACROS)
-#  define __STDC_FORMAT_MACROS 1
-#endif
-
-// ---------------------------------------------------------------------------
-// Package information
-
-// Name of package.
-#define PACKAGE gflags
-
-// Define to the full name of this package.
-#define PACKAGE_NAME gflags
-
-// Define to the full name and version of this package.
-#define PACKAGE_STRING gflags 2.2.0
-
-// Define to the one symbol short name of this package.
-#define PACKAGE_TARNAME gflags-2.2.0
-
-// Define to the version of this package.
-#define PACKAGE_VERSION 2.2.0
-
-// Version number of package.
-#define VERSION PACKAGE_VERSION
-
-// Define to the address where bug reports for this package should be sent.
-#define PACKAGE_BUGREPORT https://github.com/schuhschuh/gflags/issues
-
-// ---------------------------------------------------------------------------
-// Path separator
-#ifndef PATH_SEPARATOR
-#  ifdef OS_WINDOWS
-#    define PATH_SEPARATOR  '\\'
-#  else
-#    define PATH_SEPARATOR  '/'
-#  endif
-#endif
-
-// ---------------------------------------------------------------------------
-// Windows
-
-// Whether gflags library is a DLL.
-#ifndef GFLAGS_IS_A_DLL
-#  define GFLAGS_IS_A_DLL 0
-#endif
-
-// Always export symbols when compiling a shared library as this file is only
-// included by internal modules when building the gflags library itself.
-// The gflags_declare.h header file will set it to import these symbols otherwise.
-#ifndef GFLAGS_DLL_DECL
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL __declspec(dllexport)
-#  else
-#    define GFLAGS_DLL_DECL
-#  endif
-#endif
-// Flags defined by the gflags library itself must be exported
-#ifndef GFLAGS_DLL_DEFINE_FLAG
-#  define GFLAGS_DLL_DEFINE_FLAG GFLAGS_DLL_DECL
-#endif
-
-#ifdef OS_WINDOWS
-// The unittests import the symbols of the shared gflags library
-#  if GFLAGS_IS_A_DLL && defined(_MSC_VER)
-#    define GFLAGS_DLL_DECL_FOR_UNITTESTS __declspec(dllimport)
-#  endif
-#  include "windows_port.h"
-#endif
diff --git a/files/third_party/gflags/gflags.gyp b/files/third_party/gflags/gflags.gyp
deleted file mode 100644
index 37f2815a..00000000
--- a/files/third_party/gflags/gflags.gyp
+++ /dev/null
@@ -1,92 +0,0 @@
-#
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This is a copy of WebRTC's gflags.gyp.
-
-{
-  'variables': {
-    'gflags_root': '<(DEPTH)/third_party/gflags',
-    'conditions': [
-      ['OS=="win"', {
-        'gflags_gen_arch_root': '<(gflags_root)/gen/win',
-      }, {
-        'gflags_gen_arch_root': '<(gflags_root)/gen/posix',
-      }],
-    ],
-  },
-  'targets': [
-    {
-      'target_name': 'gflags',
-      'type': 'static_library',
-      'include_dirs': [
-        '<(gflags_gen_arch_root)/include/gflags',  # For configured files.
-        '<(gflags_gen_arch_root)/include/private',  # For config.h
-        '<(gflags_root)/src/src',  # For everything else.
-      ],
-      'defines': [
-        # These macros exist so flags and symbols are properly
-        # exported when building DLLs. Since we don't build DLLs, we
-        # need to disable them.
-        'GFLAGS_DLL_DECL=',
-        'GFLAGS_DLL_DECLARE_FLAG=',
-        'GFLAGS_DLL_DEFINE_FLAG=',
-      ],
-      'direct_dependent_settings': {
-        'include_dirs': [
-          '<(gflags_gen_arch_root)/include',  # For configured files.
-          '<(gflags_root)/src/src',  # For everything else.
-        ],
-        'defines': [
-          'GFLAGS_DLL_DECL=',
-          'GFLAGS_DLL_DECLARE_FLAG=',
-          'GFLAGS_DLL_DEFINE_FLAG=',
-        ],
-      },
-      'sources': [
-        'src/src/gflags.cc',
-        'src/src/gflags_completions.cc',
-        'src/src/gflags_reporting.cc',
-      ],
-      'conditions': [
-        ['OS=="win"', {
-          'sources': [
-            'src/src/windows_port.cc',
-          ],
-          'msvs_disabled_warnings': [
-            4005,  # WIN32_LEAN_AND_MEAN redefinition.
-            4267,  # Conversion from size_t to "type".
-          ],
-          'configurations': {
-            'Common_Base': {
-              'msvs_configuration_attributes': {
-                'CharacterSet': '2',  # Use Multi-byte Character Set.
-              },
-            },
-          },
-        }],
-        # TODO(andrew): Look into fixing this warning upstream:
-        # http://code.google.com/p/webrtc/issues/detail?id=760
-        ['OS=="win" and clang==1', {
-          'msvs_settings': {
-            'VCCLCompilerTool': {
-              'AdditionalOptions': [
-                '-Wno-microsoft-include',
-              ],
-            },
-          },
-        }],
-        ['clang==1', {
-          'cflags': [
-            '-Wno-microsoft-include',
-          ],
-        }],
-      ],
-    },
-  ],
-}
diff --git a/files/tools/OWNERS b/files/tools/OWNERS
deleted file mode 100644
index f0963525..00000000
--- a/files/tools/OWNERS
+++ /dev/null
@@ -1,61 +0,0 @@
-# You can add new small tools to this directory at your desire, feel free
-# to owners-TBR new folders (assuming you have a regular review already,
-# of course). Include an OWNERS file with at least two people for your new
-# folder.
-# If you're changing existing tools, have your change reviewed by the
-# OWNERS of the existing tool.
-
-dpranke@chromium.org
-scottmg@chromium.org
-thakis@chromium.org
-
-# These aren't actually great contact points for this directory, but
-# changes in this directory are rare and most changes happen in better-owned
-# subdirectories.
-#
-# TEAM: infra-dev@chromium.org
-# COMPONENT: Build
-
-per-file bisect*.py=anantha@chromium.org
-per-file bisect*.py=prasadv@chromium.org
-per-file bisect*.py=robertocn@chromium.org
-per-file run-bisect*.py=prasadv@chromium.org
-per-file run-bisect*.py=robertocn@chromium.org
-per-file prepare-bisect*.py=prasadv@chromium.org
-per-file prepare-bisect*.py=robertocn@chromium.org
-
-per-file boilerplate.py=rsesek@chromium.org
-
-per-file check_git_config.py=iannucci@chromium.org
-per-file check_git_config.py=vadimsh@chromium.org
-
-per-file check_grd_for_unused_strings.py=estade@chromium.org
-
-per-file gyp-explain.py=thakis@chromium.org
-
-per-file gypv8shy.py=jochen@chromium.org
-
-per-file include_tracer.py=thakis@chromium.org
-
-per-file ipc_messages_log.py=yfriedman@chromium.org
-
-per-file licenses.py=file://tools/copyright_scanner/OWNERS
-
-per-file remove_stale_pyc_files.py=dtu@chromium.org
-
-per-file roll_angle.py=kbr@chromium.org
-per-file roll_angle.py=kjellander@chromium.org
-per-file roll_angle.py=geofflang@chromium.org
-per-file roll_webgl_conformance.py=bajones@chromium.org
-per-file roll_webgl_conformance.py=kbr@chromium.org
-per-file roll_webgl_conformance.py=kjellander@chromium.org
-per-file roll_webgl_conformance.py=geofflang@chromium.org
-per-file roll_webgl_conformance.py=zmo@chromium.org
-per-file roll_webrtc.py=kjellander@chromium.org
-
-per-file safely-roll-deps.py=borenet@chromium.org
-
-per-file sort-headers.py=satorux@chromium.org
-per-file sort-sources.py=satorux@chromium.org
-per-file yes_no.py=satorux@chromium.org
-
diff --git a/files/tools/msan/OWNERS b/files/tools/msan/OWNERS
deleted file mode 100644
index ab97cb0f..00000000
--- a/files/tools/msan/OWNERS
+++ /dev/null
@@ -1,3 +0,0 @@
-# pbos@chromium.org
-fbarchard@google.com
-kjellander@google.com
diff --git a/files/tools/msan/blacklist.txt b/files/tools/msan/blacklist.txt
deleted file mode 100644
index 40ea4b83..00000000
--- a/files/tools/msan/blacklist.txt
+++ /dev/null
@@ -1,24 +0,0 @@
-# The rules in this file are only applied at compile time. If you can modify the
-# source in question, consider function attributes to disable instrumentation.
-#
-# Please think twice before you add or remove these rules.
-
-# False positive in ffmpeg due to assembly code. http://crbug.com/344505
-fun:ff_get_cpu_flags_x86
-
-# Benign uninits in zlib.
-# http://crbug.com/116277
-fun:*MOZ_Z_deflate*
-# http://crbug.com/418383
-fun:longest_match
-
-# Uninit in zlib with SIMD intrinsic http://crbug.com/426868
-fun:crc_fold512_to_32
-
-# Uninit in OSMesa. http://crbug.com/347967
-fun:unpack_RGBA8888
-fun:unpack_RGB888
-
-# False positives due to use of linux_syscall_support. http://crbug.com/394028
-src:*/breakpad/src/*
-src:*/components/crash/content/app/breakpad_linux.cc
diff --git a/files/tools/ubsan/OWNERS b/files/tools/ubsan/OWNERS
deleted file mode 100644
index 32b7466f..00000000
--- a/files/tools/ubsan/OWNERS
+++ /dev/null
@@ -1,3 +0,0 @@
-# pbos@webrtc.org
-kjellander@google.com
-fbarchard@google.com
diff --git a/files/tools/ubsan/blacklist.txt b/files/tools/ubsan/blacklist.txt
deleted file mode 100644
index e1e3c08a..00000000
--- a/files/tools/ubsan/blacklist.txt
+++ /dev/null
@@ -1,77 +0,0 @@
-#############################################################################
-# UBSan blacklist.
-
-#############################################################################
-# YASM does some funny things that UBsan doesn't like.
-# https://crbug.com/489901
-src:*/third_party/yasm/*
-
-#############################################################################
-# V8 gives too many false positives. Ignore them for now.
-src:*/v8/*
-
-#############################################################################
-# Ignore system libraries.
-src:*/usr/*
-
-#############################################################################
-# V8 UBsan supressions, commented out for now since we are ignorning v8
-# completely.
-# fun:*v8*internal*FastD2I*
-# fun:*v8*internal*ComputeIntegerHash*
-# fun:*v8*internal*ComputeLongHash*
-# fun:*v8*internal*ComputePointerHash*
-# src:*/v8/src/base/bits.cc
-# src:*/v8/src/base/functional.cc
-# Undefined behaviour (integer overflow) is expected but ignored in this
-# function.
-# fun:*JsonParser*ParseJsonNumber*
-
-# Runtime numeric functions.
-# src:*/v8/src/runtime/runtime-numbers.cc
-
-# Shifts of negative numbers
-# fun:*v8*internal*HPositionInfo*TagPosition*
-# fun:*v8*internal*Range*Shl*
-# fun:*v8*internal*RelocInfoWriter*WriteTaggedData*
-
-#############################################################################
-# Undefined arithmetic that can be safely ignored.
-src:*/base/numerics/saturated_arithmetic.h
-src:*/ppapi/shared_impl/id_assignment.h
-
-#############################################################################
-# ICU supressions. Mostly hash functions where integer overflow is OK.
-fun:*hashEntry*
-fun:*LocaleCacheKey*hashCode*
-fun:*google*protobuf*hash*
-fun:*(hash|Hash)*
-
-#############################################################################
-# Bounds blacklist.
-# Array at the end of struct pattern:
-# Maybe UBSan itself can be improved here?
-# e.g.
-# struct blah {
-#   int a;
-#   char foo[2]; // not actually 2
-# }
-src:*/net/disk_cache/blockfile/backend_impl.cc
-src:*/net/disk_cache/blockfile/entry_impl.cc
-src:*/third_party/icu/source/common/rbbi.cpp
-src:*/third_party/icu/source/common/rbbitblb.cpp
-src:*/third_party/icu/source/common/ucmndata.c
-
-#############################################################################
-# Delete in destructor on a this where this == nullptr
-fun:*re2*RegexpD*
-
-#############################################################################
-# Harmless float division by zero.
-fun:*RendererFrameManager*CullUnlockedFrames*
-
-#############################################################################
-# libc++ __tree and map are not UBSAN clean
-# https://llvm.org/bugs/show_bug.cgi?id=19302
-src:*/third_party/libc\+\+/trunk/include/__tree
-src:*/third_party/libc\+\+/trunk/include/map
diff --git a/files/tools/ubsan/vptr_blacklist.txt b/files/tools/ubsan/vptr_blacklist.txt
deleted file mode 100644
index e8382039..00000000
--- a/files/tools/ubsan/vptr_blacklist.txt
+++ /dev/null
@@ -1,128 +0,0 @@
-#############################################################################
-# UBSan vptr blacklist.
-# Function and type based blacklisting use a mangled name, and it is especially
-# tricky to represent C++ types. For now, any possible changes by name manglings
-# are simply represented as wildcard expressions of regexp, and thus it might be
-# over-blacklisted.
-
-#############################################################################
-# Identical layouts.
-# If base and derived classes have identifical memory layouts (i.e., the same
-# object size) and both have no virtual functions, we blacklist them as there
-# would be not much security implications.
-
-fun:*LifecycleNotifier*addObserver*
-fun:*LifecycleNotifier*removeObserver*
-fun:*toWebInputElement*
-type:*base*MessageLoopForIO*
-type:*BlockRefType*
-type:*SkAutoTUnref*
-type:*WDResult*
-type:*ExecutionContext*
-type:*WebInputElement*
-type:*WebFormControlElement*
-
-# Avoid identical layout cases for 86 different classes in InspectorTypeBuilder,
-# all of which are guarded using COMPILER_ASSERT on the object size. Two more
-# types are also blacklisted due to the template class (JSONArray <-> Array<T>).
-
-src:*InspectorTypeBuilder.h*
-type:*TypeBuilder*
-type:*JSONArray*
-
-#############################################################################
-# Base class's constructor accesses a derived class's member.
-
-fun:*DoublyLinkedListNode*
-type:*content*WebUIExtensionData*
-
-# RenderFrameObserverTracker<T>::RenderFrameObserverTracker()
-fun:*content*RenderFrameObserverTracker*RenderFrame*
-
-# RenderViewObserverTracker<T>::RenderViewObserverTracker()
-fun:*content*RenderViewObserverTracker*RenderView*
-
-#############################################################################
-# Base class's destructor accesses a derived class.
-
-fun:*DatabaseContext*contextDestroyed*
-
-# FIXME: Cannot handle template function LifecycleObserver<>::setContext,
-# so exclude source file for now. 
-src:*LifecycleObserver.h*
-
-#############################################################################
-# static_cast into itself in the constructor.
-
-fun:*RefCountedGarbageCollected*makeKeepAlive*
-fun:*ThreadSafeRefCountedGarbageCollected*makeKeepAlive*
-
-#############################################################################
-# Accessing data in destructors where the class has virtual inheritances.
-
-type:*content*RenderWidgetHost*
-
-# Match mangled name for X::~X().
-fun:*content*RenderThreadImplD*
-fun:*content*RenderViewHostImplD*
-fun:*content*UtilityThreadImplD*
-
-#############################################################################
-# Using raw pointer values.
-#
-# A raw pointer value (16) is used to infer the field offset by
-# GOOGLE_PROTOBUF_GENERATED_MESSAGE_FIELD_OFFSET.
-
-src:*/third_party/protobuf/src/google/protobuf/compiler/plugin.pb.cc
-src:*/third_party/protobuf/src/google/protobuf/compiler/cpp/cpp_message.cc
-src:*/third_party/protobuf/src/google/protobuf/descriptor.pb.cc
-
-#############################################################################
-# Avoid link errors.
-# Ubsan vptr needs typeinfo on the target class, but it looks like typeinfo is
-# not avaiable if the class is not exported.  For now, simply blacklisted to
-# avoid link errors; e.g., undefined reference to 'typeinfo for [CLASS_NAME]'.
-
-# obj/ppapi/libppapi_proxy.a(obj/ppapi/proxy/ppapi_proxy.proxy_channel.o):../../ppapi/proxy/proxy_channel.cc:__unnamed_53: error: undefined reference to 'typeinfo for IPC::TestSink'
-src:*/ppapi/proxy/proxy_channel.cc
-
-# obj/chrome/libbrowser.a(obj/chrome/browser/net/browser.predictor.o):../../chrome/browser/net/predictor.cc:__unnamed_577: error: undefined reference to 'typeinfo for ProxyAdvisor'
-src:*/chrome/browser/net/predictor.cc
-
-# obj/third_party/pdfium/libfpdfapi.a(obj/third_party/pdfium/core/src/fpdfapi/fpdf_render/fpdfapi.fpdf_render_text.o):../../third_party/pdfium/core/src/fpdfapi/fpdf_render/:__unnamed_360: error: undefined reference to 'typeinfo for CPDF_InlineImages'
-src:*/third_party/pdfium/core/src/fpdfapi/fpdf_render/fpdf_render_text.cpp
-
-# obj/third_party/libwebm/libwebm.a(obj/third_party/libwebm/source/libwebm.mkvmuxer.o)(.data.rel..L__unnamed_2+0x18): error: undefined reference to 'typeinfo for mkvparser::IMkvReader'
-src:*/third_party/libwebm/source/mkvmuxer.cpp
-
-#############################################################################
-# LLVM is not UBSan vptr clean.
-src:*third_party/swiftshader/third_party/LLVM*
-
-#############################################################################
-# UBSan seems to be emit false positives when virtual base classes are
-# involved, see e.g. crbug.com/448102.
-
-type:*v8*internal*OFStream*
-
-#############################################################################
-# UBsan is unable to handle static_cast<A*>(nullptr) and crashes on SIGSEGV.
-#
-
-# static_cast<StartPageService*> in StartPageServiceFactory::GetForProfile.
-type:*StartPageService*
-
-# Remove once function attribute level blacklisting is implemented.
-# See crbug.com/476063.
-fun:*forbidGCDuringConstruction*
-
-#############################################################################
-# UBsan goes into an infinite recursion when __dynamic_cast instrumented with
-# "vptr". See crbug.com/609786.
-
-src:*/third_party/libc\+\+abi/trunk/src/private_typeinfo.cpp
-
-#############################################################################
-# invalid downcasts for IPC messages
-# https://crbug.com/520760
-src:*nacl_message_scanner.cc
diff --git a/files/tools_libyuv/autoroller/roll_deps.py b/files/tools_libyuv/autoroller/roll_deps.py
index 8359d309..977c86de 100755
--- a/files/tools_libyuv/autoroller/roll_deps.py
+++ b/files/tools_libyuv/autoroller/roll_deps.py
@@ -1,4 +1,5 @@
-#!/usr/bin/env python
+#!/usr/bin/env vpython3
+
 # Copyright 2017 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -11,7 +12,6 @@
 # https://webrtc.googlesource.com/src/+/master/tools_webrtc/autoroller/roll_deps.py
 # customized for libyuv.
 
-
 """Script to automatically roll dependencies in the libyuv DEPS file."""
 
 import argparse
@@ -22,7 +22,7 @@ import os
 import re
 import subprocess
 import sys
-import urllib2
+import urllib.request
 
 
 # Skip these dependencies (list without solution name prefix).
@@ -37,7 +37,7 @@ CHROMIUM_LOG_TEMPLATE = CHROMIUM_SRC_URL + '/+log/%s'
 CHROMIUM_FILE_TEMPLATE = CHROMIUM_SRC_URL + '/+/%s/%s'
 
 COMMIT_POSITION_RE = re.compile('^Cr-Commit-Position: .*#([0-9]+).*$')
-CLANG_REVISION_RE = re.compile(r'^CLANG_REVISION = \'(\d+)\'$')
+CLANG_REVISION_RE = re.compile(r'^CLANG_REVISION = \'([0-9a-z-]+)\'$')
 ROLL_BRANCH_NAME = 'roll_chromium_revision'
 
 SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
@@ -46,7 +46,7 @@ CHECKOUT_SRC_DIR = os.path.realpath(os.path.join(SCRIPT_DIR, os.pardir,
 CHECKOUT_ROOT_DIR = os.path.realpath(os.path.join(CHECKOUT_SRC_DIR, os.pardir))
 
 sys.path.append(os.path.join(CHECKOUT_SRC_DIR, 'build'))
-import find_depot_tools
+import find_depot_tools  # pylint: disable=wrong-import-position
 find_depot_tools.add_depot_tools_to_path()
 
 CLANG_UPDATE_SCRIPT_URL_PATH = 'tools/clang/scripts/update.py'
@@ -69,6 +69,7 @@ def ParseDepsDict(deps_content):
   local_scope = {}
   global_scope = {
     'Var': VarLookup(local_scope),
+    'Str': lambda s: s,
     'deps_os': {},
   }
   exec(deps_content, global_scope, local_scope)
@@ -77,7 +78,7 @@ def ParseDepsDict(deps_content):
 
 def ParseLocalDepsFile(filename):
   with open(filename, 'rb') as f:
-    deps_content = f.read()
+    deps_content = f.read().decode('utf-8')
   return ParseDepsDict(deps_content)
 
 
@@ -97,7 +98,7 @@ def ParseCommitPosition(commit_message):
 
 
 def _RunCommand(command, working_dir=None, ignore_exit_code=False,
-                extra_env=None):
+                extra_env=None, input_data=None):
   """Runs a command and returns the output from that command.
 
   If the command fails (exit code != 0), the function will exit the process.
@@ -112,12 +113,14 @@ def _RunCommand(command, working_dir=None, ignore_exit_code=False,
     assert all(isinstance(value, str) for value in extra_env.values())
     logging.debug('extra env: %s', extra_env)
     env.update(extra_env)
-  p = subprocess.Popen(command, stdout=subprocess.PIPE,
-                       stderr=subprocess.PIPE, env=env,
-                       cwd=working_dir, universal_newlines=True)
-  std_output = p.stdout.read()
-  err_output = p.stderr.read()
-  p.wait()
+  p = subprocess.Popen(command,
+                       stdin=subprocess.PIPE,
+                       stdout=subprocess.PIPE,
+                       stderr=subprocess.PIPE,
+                       env=env,
+                       cwd=working_dir,
+                       universal_newlines=True)
+  std_output, err_output = p.communicate(input_data)
   p.stdout.close()
   p.stderr.close()
   if not ignore_exit_code and p.returncode != 0:
@@ -153,7 +156,7 @@ def _ReadGitilesContent(url):
   # Download and decode BASE64 content until
   # https://code.google.com/p/gitiles/issues/detail?id=7 is fixed.
   base64_content = ReadUrlContent(url + '?format=TEXT')
-  return base64.b64decode(base64_content[0])
+  return base64.b64decode(base64_content[0]).decode('utf-8')
 
 
 def ReadRemoteCrFile(path_below_src, revision):
@@ -169,7 +172,7 @@ def ReadRemoteCrCommit(revision):
 
 def ReadUrlContent(url):
   """Connect to a remote host and read the contents. Returns a list of lines."""
-  conn = urllib2.urlopen(url)
+  conn = urllib.request.urlopen(url)
   try:
     return conn.readlines()
   except IOError as e:
@@ -192,7 +195,7 @@ def GetMatchingDepsEntries(depsentry_dict, dir_path):
     A list of DepsEntry objects.
   """
   result = []
-  for path, depsentry in depsentry_dict.iteritems():
+  for path, depsentry in depsentry_dict.items():
     if path == dir_path:
       result.append(depsentry)
     else:
@@ -202,26 +205,24 @@ def GetMatchingDepsEntries(depsentry_dict, dir_path):
         result.append(depsentry)
   return result
 
-
 def BuildDepsentryDict(deps_dict):
-  """Builds a dict of paths to DepsEntry objects from a raw parsed deps dict."""
+  """Builds a dict of paths to DepsEntry objects from a raw deps dict."""
   result = {}
+
   def AddDepsEntries(deps_subdict):
-    for path, deps_url_spec in deps_subdict.iteritems():
-      # The deps url is either an URL and a condition, or just the URL.
+    for path, deps_url_spec in deps_subdict.items():
       if isinstance(deps_url_spec, dict):
         if deps_url_spec.get('dep_type') == 'cipd':
           continue
         deps_url = deps_url_spec['url']
       else:
         deps_url = deps_url_spec
-
-      if not result.has_key(path):
+      if not path in result:
         url, revision = deps_url.split('@') if deps_url else (None, None)
         result[path] = DepsEntry(path, url, revision)
 
   AddDepsEntries(deps_dict['deps'])
-  for deps_os in ['win', 'mac', 'unix', 'android', 'ios', 'unix']:
+  for deps_os in ['win', 'mac', 'linux', 'android', 'ios', 'unix']:
     AddDepsEntries(deps_dict.get('deps_os', {}).get(deps_os, {}))
   return result
 
@@ -244,7 +245,7 @@ def CalculateChangedDeps(libyuv_deps, new_cr_deps):
   result = []
   libyuv_entries = BuildDepsentryDict(libyuv_deps)
   new_cr_entries = BuildDepsentryDict(new_cr_deps)
-  for path, libyuv_deps_entry in libyuv_entries.iteritems():
+  for path, libyuv_deps_entry in libyuv_entries.items():
     if path in DONT_AUTOROLL_THESE:
       continue
     cr_deps_entry = new_cr_entries.get(path)
@@ -274,9 +275,9 @@ def CalculateChangedClang(new_cr_rev):
       match = CLANG_REVISION_RE.match(line)
       if match:
         return match.group(1)
-    raise RollError('Could not parse Clang revision!')
+    raise RollError('Could not parse Clang revision from:\n' + '\n'.join('  ' + l for l in lines))
 
-  with open(CLANG_UPDATE_SCRIPT_LOCAL_PATH, 'rb') as f:
+  with open(CLANG_UPDATE_SCRIPT_LOCAL_PATH, 'r') as f:
     current_lines = f.readlines()
   current_rev = GetClangRev(current_lines)
 
@@ -334,10 +335,10 @@ def UpdateDepsFile(deps_filename, old_cr_revision, new_cr_revision,
 
   # Update the chromium_revision variable.
   with open(deps_filename, 'rb') as deps_file:
-    deps_content = deps_file.read()
+    deps_content = deps_file.read().decode('utf-8')
   deps_content = deps_content.replace(old_cr_revision, new_cr_revision)
   with open(deps_filename, 'wb') as deps_file:
-    deps_file.write(deps_content)
+    deps_file.write(deps_content.encode('utf-8'))
 
   # Update each individual DEPS entry.
   for dep in changed_deps:
@@ -365,12 +366,12 @@ def _IsTreeClean():
 def _EnsureUpdatedMasterBranch(dry_run):
   current_branch = _RunCommand(
       ['git', 'rev-parse', '--abbrev-ref', 'HEAD'])[0].splitlines()[0]
-  if current_branch != 'master':
-    logging.error('Please checkout the master branch and re-run this script.')
+  if current_branch != 'main':
+    logging.error('Please checkout the main branch and re-run this script.')
     if not dry_run:
       sys.exit(-1)
 
-  logging.info('Updating master branch...')
+  logging.info('Updating main branch...')
   _RunCommand(['git', 'pull'])
 
 
@@ -383,7 +384,7 @@ def _CreateRollBranch(dry_run):
 def _RemovePreviousRollBranch(dry_run):
   active_branch, branches = _GetBranches()
   if active_branch == ROLL_BRANCH_NAME:
-    active_branch = 'master'
+    active_branch = 'main'
   if ROLL_BRANCH_NAME in branches:
     logging.info('Removing previous roll branch (%s)', ROLL_BRANCH_NAME)
     if not dry_run:
@@ -417,10 +418,11 @@ def _UploadCL(commit_queue_mode):
   cmd = ['git', 'cl', 'upload', '--force', '--bypass-hooks', '--send-mail']
   if commit_queue_mode >= 2:
     logging.info('Sending the CL to the CQ...')
-    cmd.extend(['--use-commit-queue'])
+    cmd.extend(['-o', 'label=Bot-Commit+1'])
+    cmd.extend(['-o', 'label=Commit-Queue+2'])
   elif commit_queue_mode >= 1:
     logging.info('Starting CQ dry run...')
-    cmd.extend(['--cq-dry-run'])
+    cmd.extend(['-o', 'label=Commit-Queue+1'])
   extra_env = {
       'EDITOR': 'true',
       'SKIP_GCE_AUTH_FOR_GIT': '1',
@@ -443,7 +445,7 @@ def main():
                        'tryjobs.'))
   p.add_argument('-i', '--ignore-unclean-workdir', action='store_true',
                  default=False,
-                 help=('Ignore if the current branch is not master or if there '
+                 help=('Ignore if the current branch is not main or if there '
                        'are uncommitted changes (default: %(default)s).'))
   grp = p.add_mutually_exclusive_group()
   grp.add_argument('--skip-cq', action='store_true', default=False,
diff --git a/files/tools_libyuv/autoroller/unittests/.DS_Store b/files/tools_libyuv/autoroller/unittests/.DS_Store
deleted file mode 100644
index 70369d69..00000000
--- a/files/tools_libyuv/autoroller/unittests/.DS_Store
+++ /dev/null
diff --git a/files/tools_libyuv/autoroller/unittests/roll_deps_test.py b/files/tools_libyuv/autoroller/unittests/roll_deps_test.py
index 477b6e40..af86bdd5 100755
--- a/files/tools_libyuv/autoroller/unittests/roll_deps_test.py
+++ b/files/tools_libyuv/autoroller/unittests/roll_deps_test.py
@@ -1,4 +1,5 @@
-#!/usr/bin/env python
+#!/usr/bin/env vpython3
+
 # Copyright 2017 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -14,14 +15,13 @@ import sys
 import tempfile
 import unittest
 
-
-SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
-PARENT_DIR = os.path.join(SCRIPT_DIR, os.pardir)
-sys.path.append(PARENT_DIR)
 import roll_deps
 from roll_deps import CalculateChangedDeps, GetMatchingDepsEntries, \
   ParseDepsDict, ParseLocalDepsFile, UpdateDepsFile
 
+SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+PARENT_DIR = os.path.join(SCRIPT_DIR, os.pardir)
+sys.path.append(PARENT_DIR)
 
 TEST_DATA_VARS = {
   'chromium_git': 'https://chromium.googlesource.com',
@@ -45,7 +45,7 @@ class TestError(Exception):
   pass
 
 
-class FakeCmd(object):
+class FakeCmd():
   def __init__(self):
     self.expectations = []
 
@@ -85,43 +85,43 @@ class TestRollChromiumRevision(unittest.TestCase):
   def testVarLookup(self):
     local_scope = {'foo': 'wrong', 'vars': {'foo': 'bar'}}
     lookup = roll_deps.VarLookup(local_scope)
-    self.assertEquals(lookup('foo'), 'bar')
+    self.assertEqual(lookup('foo'), 'bar')
 
   def testUpdateDepsFile(self):
     new_rev = 'aaaaabbbbbcccccdddddeeeeefffff0000011111'
 
     current_rev = TEST_DATA_VARS['chromium_revision']
     UpdateDepsFile(self._libyuv_depsfile, current_rev, new_rev, [])
-    with open(self._libyuv_depsfile) as deps_file:
+    with open(self._libyuv_depsfile, 'r') as deps_file:
       deps_contents = deps_file.read()
       self.assertTrue(new_rev in deps_contents,
                       'Failed to find %s in\n%s' % (new_rev, deps_contents))
 
   def testParseDepsDict(self):
-    with open(self._libyuv_depsfile) as deps_file:
+    with open(self._libyuv_depsfile, 'r') as deps_file:
       deps_contents = deps_file.read()
     local_scope = ParseDepsDict(deps_contents)
     vars_dict = local_scope['vars']
 
     def assertVar(variable_name):
-      self.assertEquals(vars_dict[variable_name], TEST_DATA_VARS[variable_name])
+      self.assertEqual(vars_dict[variable_name], TEST_DATA_VARS[variable_name])
     assertVar('chromium_git')
     assertVar('chromium_revision')
-    self.assertEquals(len(local_scope['deps']), 3)
+    self.assertEqual(len(local_scope['deps']), 3)
 
   def testGetMatchingDepsEntriesReturnsPathInSimpleCase(self):
     entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/testing/gtest')
-    self.assertEquals(len(entries), 1)
-    self.assertEquals(entries[0], DEPS_ENTRIES['src/testing/gtest'])
+    self.assertEqual(len(entries), 1)
+    self.assertEqual(entries[0], DEPS_ENTRIES['src/testing/gtest'])
 
   def testGetMatchingDepsEntriesHandlesSimilarStartingPaths(self):
     entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/testing')
-    self.assertEquals(len(entries), 2)
+    self.assertEqual(len(entries), 2)
 
   def testGetMatchingDepsEntriesHandlesTwoPathsWithIdenticalFirstParts(self):
     entries = GetMatchingDepsEntries(DEPS_ENTRIES, 'src/build')
-    self.assertEquals(len(entries), 1)
-    self.assertEquals(entries[0], DEPS_ENTRIES['src/build'])
+    self.assertEqual(len(entries), 1)
+    self.assertEqual(entries[0], DEPS_ENTRIES['src/build'])
 
   def testCalculateChangedDeps(self):
     _SetupGitLsRemoteCall(self.fake,
@@ -129,14 +129,14 @@ class TestRollChromiumRevision(unittest.TestCase):
     libyuv_deps = ParseLocalDepsFile(self._libyuv_depsfile)
     new_cr_deps = ParseLocalDepsFile(self._new_cr_depsfile)
     changed_deps = CalculateChangedDeps(libyuv_deps, new_cr_deps)
-    self.assertEquals(len(changed_deps), 2)
-    self.assertEquals(changed_deps[0].path, 'src/build')
-    self.assertEquals(changed_deps[0].current_rev, BUILD_OLD_REV)
-    self.assertEquals(changed_deps[0].new_rev, BUILD_NEW_REV)
-
-    self.assertEquals(changed_deps[1].path, 'src/buildtools')
-    self.assertEquals(changed_deps[1].current_rev, BUILDTOOLS_OLD_REV)
-    self.assertEquals(changed_deps[1].new_rev, BUILDTOOLS_NEW_REV)
+    self.assertEqual(len(changed_deps), 2)
+    self.assertEqual(changed_deps[0].path, 'src/build')
+    self.assertEqual(changed_deps[0].current_rev, BUILD_OLD_REV)
+    self.assertEqual(changed_deps[0].new_rev, BUILD_NEW_REV)
+
+    self.assertEqual(changed_deps[1].path, 'src/buildtools')
+    self.assertEqual(changed_deps[1].current_rev, BUILDTOOLS_OLD_REV)
+    self.assertEqual(changed_deps[1].new_rev, BUILDTOOLS_NEW_REV)
 
 
 def _SetupGitLsRemoteCall(cmd_fake, url, revision):
diff --git a/files/tools_libyuv/autoroller/unittests/testdata/DEPS b/files/tools_libyuv/autoroller/unittests/testdata/DEPS
index 9fbb48a7..4f45860c 100644
--- a/files/tools_libyuv/autoroller/unittests/testdata/DEPS
+++ b/files/tools_libyuv/autoroller/unittests/testdata/DEPS
@@ -3,6 +3,7 @@
 vars = {
   'chromium_git': 'https://chromium.googlesource.com',
   'chromium_revision': '1b9c098a08e40114e44b6c1ec33ddf95c40b901d',
+  'ignored_str': Str(''),
 }
 
 deps = {
diff --git a/files/tools_libyuv/get_landmines.py b/files/tools_libyuv/get_landmines.py
index c554f04a..8b33483e 100755
--- a/files/tools_libyuv/get_landmines.py
+++ b/files/tools_libyuv/get_landmines.py
@@ -1,4 +1,5 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
+
 # Copyright 2016 The LibYuv Project Authors. All rights reserved.
 #
 # Use of this source code is governed by a BSD-style license
@@ -25,8 +26,8 @@ def print_landmines():
   # dependency problems, fix the dependency problems instead of adding a
   # landmine.
   # See the Chromium version in src/build/get_landmines.py for usage examples.
-  print 'Clobber to remove GYP artifacts after switching bots to GN.'
-  print 'Another try to remove GYP artifacts after switching bots to GN.'
+  print('Clobber to remove GYP artifacts after switching bots to GN.')
+  print('Another try to remove GYP artifacts after switching bots to GN.')
 
 
 def main():
diff --git a/files/tools_libyuv/valgrind/chrome_tests.bat b/files/tools_libyuv/valgrind/chrome_tests.bat
deleted file mode 100755
index 9d4c8ca8..00000000
--- a/files/tools_libyuv/valgrind/chrome_tests.bat
+++ /dev/null
@@ -1,53 +0,0 @@
-@echo off
-:: Copyright (c) 2011 The Chromium Authors. All rights reserved.
-:: Use of this source code is governed by a BSD-style license that can be
-:: found in the LICENSE file.
-
-setlocal
-
-set THISDIR=%~dp0
-set TOOL_NAME="unknown"
-
-:: Get the tool name and put it into TOOL_NAME {{{1
-:: NB: SHIFT command doesn't modify %*
-:PARSE_ARGS_LOOP
-  if %1 == () GOTO:TOOLNAME_NOT_FOUND
-  if %1 == --tool GOTO:TOOLNAME_FOUND
-  SHIFT
-  goto :PARSE_ARGS_LOOP
-
-:TOOLNAME_NOT_FOUND
-echo "Please specify a tool (e.g. drmemory) by using --tool flag"
-exit /B 1
-
-:TOOLNAME_FOUND
-SHIFT
-set TOOL_NAME=%1
-:: }}}
-if "%TOOL_NAME%" == "drmemory"          GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_light"    GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_full"     GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_pattern"  GOTO :SETUP_DRMEMORY
-echo "Unknown tool: `%TOOL_NAME%`! Only drmemory is supported right now"
-exit /B 1
-
-:SETUP_DRMEMORY
-:: Set up DRMEMORY_COMMAND to invoke Dr. Memory {{{1
-set DRMEMORY_PATH=%THISDIR%..\..\third_party\drmemory
-set DRMEMORY_SFX=%DRMEMORY_PATH%\drmemory-windows-sfx.exe
-if EXIST %DRMEMORY_SFX% GOTO DRMEMORY_BINARY_OK
-echo "Can't find Dr. Memory executables."
-echo "See http://www.chromium.org/developers/how-tos/using-valgrind/dr-memory"
-echo "for the instructions on how to get them."
-exit /B 1
-
-:DRMEMORY_BINARY_OK
-%DRMEMORY_SFX% -o%DRMEMORY_PATH%\unpacked -y
-set DRMEMORY_COMMAND=%DRMEMORY_PATH%\unpacked\bin\drmemory.exe
-:: }}}
-goto :RUN_TESTS
-
-:RUN_TESTS
-set PYTHONPATH=%THISDIR%../python/google
-set RUNNING_ON_VALGRIND=yes
-python %THISDIR%/chrome_tests.py %*
diff --git a/files/tools_libyuv/valgrind/chrome_tests.py b/files/tools_libyuv/valgrind/chrome_tests.py
deleted file mode 100755
index fe899bce..00000000
--- a/files/tools_libyuv/valgrind/chrome_tests.py
+++ /dev/null
@@ -1,869 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-''' Runs various chrome tests through valgrind_test.py.'''
-
-import glob
-import logging
-import multiprocessing
-import optparse
-import os
-import stat
-import subprocess
-import sys
-
-import logging_utils
-import path_utils
-
-import common
-import valgrind_test
-
-class TestNotFound(Exception): pass
-
-class MultipleGTestFiltersSpecified(Exception): pass
-
-class BuildDirNotFound(Exception): pass
-
-class BuildDirAmbiguous(Exception): pass
-
-class ExecutableNotFound(Exception): pass
-
-class BadBinary(Exception): pass
-
-class ChromeTests:
-  SLOW_TOOLS = ["memcheck", "drmemory"]
-  LAYOUT_TESTS_DEFAULT_CHUNK_SIZE = 300
-
-  def __init__(self, options, args, test):
-    if ':' in test:
-      (self._test, self._gtest_filter) = test.split(':', 1)
-    else:
-      self._test = test
-      self._gtest_filter = options.gtest_filter
-
-    if self._test not in self._test_list:
-      raise TestNotFound("Unknown test: %s" % test)
-
-    if options.gtest_filter and options.gtest_filter != self._gtest_filter:
-      raise MultipleGTestFiltersSpecified("Can not specify both --gtest_filter "
-                                          "and --test %s" % test)
-
-    self._options = options
-    self._args = args
-
-    script_dir = path_utils.ScriptDir()
-    # Compute the top of the tree (the "source dir") from the script dir (where
-    # this script lives).  We assume that the script dir is in tools/valgrind/
-    # relative to the top of the tree.
-    self._source_dir = os.path.dirname(os.path.dirname(script_dir))
-    # since this path is used for string matching, make sure it's always
-    # an absolute Unix-style path
-    self._source_dir = os.path.abspath(self._source_dir).replace('\\', '/')
-    valgrind_test_script = os.path.join(script_dir, "valgrind_test.py")
-    self._command_preamble = ["--source-dir=%s" % (self._source_dir)]
-
-    if not self._options.build_dir:
-      dirs = [
-        os.path.join(self._source_dir, "xcodebuild", "Debug"),
-        os.path.join(self._source_dir, "out", "Debug"),
-        os.path.join(self._source_dir, "build", "Debug"),
-      ]
-      build_dir = [d for d in dirs if os.path.isdir(d)]
-      if len(build_dir) > 1:
-        raise BuildDirAmbiguous("Found more than one suitable build dir:\n"
-                                "%s\nPlease specify just one "
-                                "using --build-dir" % ", ".join(build_dir))
-      elif build_dir:
-        self._options.build_dir = build_dir[0]
-      else:
-        self._options.build_dir = None
-
-    if self._options.build_dir:
-      build_dir = os.path.abspath(self._options.build_dir)
-      self._command_preamble += ["--build-dir=%s" % (self._options.build_dir)]
-
-  def _EnsureBuildDirFound(self):
-    if not self._options.build_dir:
-      raise BuildDirNotFound("Oops, couldn't find a build dir, please "
-                             "specify it manually using --build-dir")
-
-  def _DefaultCommand(self, tool, exe=None, valgrind_test_args=None):
-    '''Generates the default command array that most tests will use.'''
-    if exe and common.IsWindows():
-      exe += '.exe'
-
-    cmd = list(self._command_preamble)
-
-    # Find all suppressions matching the following pattern:
-    # tools/valgrind/TOOL/suppressions[_PLATFORM].txt
-    # and list them with --suppressions= prefix.
-    script_dir = path_utils.ScriptDir()
-    tool_name = tool.ToolName();
-    suppression_file = os.path.join(script_dir, tool_name, "suppressions.txt")
-    if os.path.exists(suppression_file):
-      cmd.append("--suppressions=%s" % suppression_file)
-    # Platform-specific suppression
-    for platform in common.PlatformNames():
-      platform_suppression_file = \
-          os.path.join(script_dir, tool_name, 'suppressions_%s.txt' % platform)
-      if os.path.exists(platform_suppression_file):
-        cmd.append("--suppressions=%s" % platform_suppression_file)
-
-    if tool_name == "drmemory":
-      if self._options.drmemory_ops:
-        # prepending " " to avoid Dr. Memory's option confusing optparse
-        cmd += ["--drmemory_ops", " " + self._options.drmemory_ops]
-
-    if self._options.valgrind_tool_flags:
-      cmd += self._options.valgrind_tool_flags.split(" ")
-    if self._options.keep_logs:
-      cmd += ["--keep_logs"]
-    if valgrind_test_args != None:
-      for arg in valgrind_test_args:
-        cmd.append(arg)
-    if exe:
-      self._EnsureBuildDirFound()
-      exe_path = os.path.join(self._options.build_dir, exe)
-      if not os.path.exists(exe_path):
-        raise ExecutableNotFound("Couldn't find '%s'" % exe_path)
-
-      # Make sure we don't try to test ASan-built binaries
-      # with other dynamic instrumentation-based tools.
-      # TODO(timurrrr): also check TSan and MSan?
-      # `nm` might not be available, so use try-except.
-      try:
-        # Do not perform this check on OS X, as 'nm' on 10.6 can't handle
-        # binaries built with Clang 3.5+.
-        if not common.IsMac():
-          nm_output = subprocess.check_output(["nm", exe_path])
-          if nm_output.find("__asan_init") != -1:
-            raise BadBinary("You're trying to run an executable instrumented "
-                            "with AddressSanitizer under %s. Please provide "
-                            "an uninstrumented executable." % tool_name)
-      except OSError:
-        pass
-
-      cmd.append(exe_path)
-      # Valgrind runs tests slowly, so slow tests hurt more; show elapased time
-      # so we can find the slowpokes.
-      cmd.append("--gtest_print_time")
-      # Built-in test launcher for gtest-based executables runs tests using
-      # multiple process by default. Force the single-process mode back.
-      cmd.append("--single-process-tests")
-    if self._options.gtest_repeat:
-      cmd.append("--gtest_repeat=%s" % self._options.gtest_repeat)
-    if self._options.gtest_shuffle:
-      cmd.append("--gtest_shuffle")
-    if self._options.gtest_break_on_failure:
-      cmd.append("--gtest_break_on_failure")
-    if self._options.test_launcher_bot_mode:
-      cmd.append("--test-launcher-bot-mode")
-    if self._options.test_launcher_total_shards is not None:
-      cmd.append("--test-launcher-total-shards=%d"
-                 % self._options.test_launcher_total_shards)
-    if self._options.test_launcher_shard_index is not None:
-      cmd.append("--test-launcher-shard-index=%d"
-                 % self._options.test_launcher_shard_index)
-    return cmd
-
-  def Run(self):
-    ''' Runs the test specified by command-line argument --test '''
-    logging.info("running test %s" % (self._test))
-    return self._test_list[self._test](self)
-
-  def _AppendGtestFilter(self, tool, name, cmd):
-    '''Append an appropriate --gtest_filter flag to the googletest binary
-       invocation.
-       If the user passed their own filter mentioning only one test, just use
-       it. Otherwise, filter out tests listed in the appropriate gtest_exclude
-       files.
-    '''
-    if (self._gtest_filter and
-        ":" not in self._gtest_filter and
-        "?" not in self._gtest_filter and
-        "*" not in self._gtest_filter):
-      cmd.append("--gtest_filter=%s" % self._gtest_filter)
-      return
-
-    filters = []
-    gtest_files_dir = os.path.join(path_utils.ScriptDir(), "gtest_exclude")
-
-    gtest_filter_files = [
-        os.path.join(gtest_files_dir, name + ".gtest-%s.txt" % tool.ToolName())]
-    # Use ".gtest.txt" files only for slow tools, as they now contain
-    # Valgrind- and Dr.Memory-specific filters.
-    # TODO(glider): rename the files to ".gtest_slow.txt"
-    if tool.ToolName() in ChromeTests.SLOW_TOOLS:
-      gtest_filter_files += [os.path.join(gtest_files_dir, name + ".gtest.txt")]
-    for platform_suffix in common.PlatformNames():
-      gtest_filter_files += [
-        os.path.join(gtest_files_dir, name + ".gtest_%s.txt" % platform_suffix),
-        os.path.join(gtest_files_dir, name + ".gtest-%s_%s.txt" % \
-            (tool.ToolName(), platform_suffix))]
-    logging.info("Reading gtest exclude filter files:")
-    for filename in gtest_filter_files:
-      # strip the leading absolute path (may be very long on the bot)
-      # and the following / or \.
-      readable_filename = filename.replace("\\", "/")  # '\' on Windows
-      readable_filename = readable_filename.replace(self._source_dir, "")[1:]
-      if not os.path.exists(filename):
-        logging.info("  \"%s\" - not found" % readable_filename)
-        continue
-      logging.info("  \"%s\" - OK" % readable_filename)
-      f = open(filename, 'r')
-      for line in f.readlines():
-        if line.startswith("#") or line.startswith("//") or line.isspace():
-          continue
-        line = line.rstrip()
-        test_prefixes = ["FLAKY", "FAILS"]
-        for p in test_prefixes:
-          # Strip prefixes from the test names.
-          line = line.replace(".%s_" % p, ".")
-        # Exclude the original test name.
-        filters.append(line)
-        if line[-2:] != ".*":
-          # List all possible prefixes if line doesn't end with ".*".
-          for p in test_prefixes:
-            filters.append(line.replace(".", ".%s_" % p))
-    # Get rid of duplicates.
-    filters = set(filters)
-    gtest_filter = self._gtest_filter
-    if len(filters):
-      if gtest_filter:
-        gtest_filter += ":"
-        if gtest_filter.find("-") < 0:
-          gtest_filter += "-"
-      else:
-        gtest_filter = "-"
-      gtest_filter += ":".join(filters)
-    if gtest_filter:
-      cmd.append("--gtest_filter=%s" % gtest_filter)
-
-  @staticmethod
-  def ShowTests():
-    test_to_names = {}
-    for name, test_function in ChromeTests._test_list.iteritems():
-      test_to_names.setdefault(test_function, []).append(name)
-
-    name_to_aliases = {}
-    for names in test_to_names.itervalues():
-      names.sort(key=lambda name: len(name))
-      name_to_aliases[names[0]] = names[1:]
-
-    print
-    print "Available tests:"
-    print "----------------"
-    for name, aliases in sorted(name_to_aliases.iteritems()):
-      if aliases:
-        print "   {} (aka {})".format(name, ', '.join(aliases))
-      else:
-        print "   {}".format(name)
-
-  def SetupLdPath(self, requires_build_dir):
-    if requires_build_dir:
-      self._EnsureBuildDirFound()
-    elif not self._options.build_dir:
-      return
-
-    # Append build_dir to LD_LIBRARY_PATH so external libraries can be loaded.
-    if (os.getenv("LD_LIBRARY_PATH")):
-      os.putenv("LD_LIBRARY_PATH", "%s:%s" % (os.getenv("LD_LIBRARY_PATH"),
-                                              self._options.build_dir))
-    else:
-      os.putenv("LD_LIBRARY_PATH", self._options.build_dir)
-
-  def SimpleTest(self, module, name, valgrind_test_args=None, cmd_args=None):
-    tool = valgrind_test.CreateTool(self._options.valgrind_tool)
-    cmd = self._DefaultCommand(tool, name, valgrind_test_args)
-    self._AppendGtestFilter(tool, name, cmd)
-    cmd.extend(['--test-tiny-timeout=1000'])
-    if cmd_args:
-      cmd.extend(cmd_args)
-
-    self.SetupLdPath(True)
-    return tool.Run(cmd, module)
-
-  def RunCmdLine(self):
-    tool = valgrind_test.CreateTool(self._options.valgrind_tool)
-    cmd = self._DefaultCommand(tool, None, self._args)
-    self.SetupLdPath(False)
-    return tool.Run(cmd, None)
-
-  def TestAccessibility(self):
-    return self.SimpleTest("accessibility", "accessibility_unittests")
-
-  def TestAddressInput(self):
-    return self.SimpleTest("addressinput", "libaddressinput_unittests")
-
-  def TestAngle(self):
-    return self.SimpleTest("angle", "angle_unittests")
-
-  def TestAppList(self):
-    return self.SimpleTest("app_list", "app_list_unittests")
-
-  def TestAsh(self):
-    return self.SimpleTest("ash", "ash_unittests")
-
-  def TestAura(self):
-    return self.SimpleTest("aura", "aura_unittests")
-
-  def TestBase(self):
-    return self.SimpleTest("base", "base_unittests")
-
-  def TestBlinkHeap(self):
-    return self.SimpleTest("blink_heap", "blink_heap_unittests")
-
-  def TestBlinkPlatform(self):
-    return self.SimpleTest("blink_platform", "blink_platform_unittests")
-
-  def TestCacheInvalidation(self):
-    return self.SimpleTest("cacheinvalidation", "cacheinvalidation_unittests")
-
-  def TestCast(self):
-    return self.SimpleTest("chrome", "cast_unittests")
-
-  def TestCC(self):
-    return self.SimpleTest("cc", "cc_unittests",
-                           cmd_args=[
-                               "--cc-layer-tree-test-long-timeout"])
-
-  def TestChromeApp(self):
-    return self.SimpleTest("chrome_app", "chrome_app_unittests")
-
-  def TestChromeElf(self):
-    return self.SimpleTest("chrome_elf", "chrome_elf_unittests")
-
-  def TestChromeDriver(self):
-    return self.SimpleTest("chromedriver", "chromedriver_unittests")
-
-  def TestChromeOS(self):
-    return self.SimpleTest("chromeos", "chromeos_unittests")
-
-  def TestComponents(self):
-    return self.SimpleTest("components", "components_unittests")
-
-  def TestCompositor(self):
-    return self.SimpleTest("compositor", "compositor_unittests")
-
-  def TestContent(self):
-    return self.SimpleTest("content", "content_unittests")
-
-  def TestCourgette(self):
-    return self.SimpleTest("courgette", "courgette_unittests")
-
-  def TestCrypto(self):
-    return self.SimpleTest("crypto", "crypto_unittests")
-
-  def TestDevice(self):
-    return self.SimpleTest("device", "device_unittests")
-
-  def TestDisplay(self):
-    return self.SimpleTest("display", "display_unittests")
-
-  def TestEvents(self):
-    return self.SimpleTest("events", "events_unittests")
-
-  def TestExtensions(self):
-    return self.SimpleTest("extensions", "extensions_unittests")
-
-  def TestFFmpegRegressions(self):
-    return self.SimpleTest("chrome", "ffmpeg_regression_tests")
-
-  def TestGCM(self):
-    return self.SimpleTest("gcm", "gcm_unit_tests")
-
-  def TestGfx(self):
-    return self.SimpleTest("gfx", "gfx_unittests")
-
-  def TestGin(self):
-    return self.SimpleTest("gin", "gin_unittests")
-
-  def TestGoogleApis(self):
-    return self.SimpleTest("google_apis", "google_apis_unittests")
-
-  def TestGPU(self):
-    return self.SimpleTest("gpu", "gpu_unittests")
-
-  def TestIpc(self):
-    return self.SimpleTest("ipc", "ipc_tests",
-                           valgrind_test_args=["--trace_children"])
-
-  def TestInstallerUtil(self):
-    return self.SimpleTest("installer_util", "installer_util_unittests")
-
-  def TestInstallStatic(self):
-    return self.SimpleTest("install_static", "install_static_unittests")
-
-  def TestJingle(self):
-    return self.SimpleTest("chrome", "jingle_unittests")
-
-  def TestKeyboard(self):
-    return self.SimpleTest("keyboard", "keyboard_unittests")
-
-  def TestLatency(self):
-    return self.SimpleTest("latency", "latency_unittests")
-
-  def TestMedia(self):
-    return self.SimpleTest("chrome", "media_unittests")
-
-  def TestMessageCenter(self):
-    return self.SimpleTest("message_center", "message_center_unittests")
-
-  def TestMidi(self):
-    return self.SimpleTest("chrome", "midi_unittests")
-
-  def TestMojoCommon(self):
-    return self.SimpleTest("mojo_common", "mojo_common_unittests")
-
-  def TestMojoPublicBindings(self):
-    return self.SimpleTest("mojo_public_bindings",
-                           "mojo_public_bindings_unittests")
-
-  def TestMojoPublicSystem(self):
-    return self.SimpleTest("mojo_public_system",
-                           "mojo_public_system_unittests")
-
-  def TestMojoPublicSysPerf(self):
-    return self.SimpleTest("mojo_public_sysperf",
-                           "mojo_public_system_perftests")
-
-  def TestMojoSystem(self):
-    return self.SimpleTest("mojo_system", "mojo_system_unittests")
-
-  def TestNet(self):
-    return self.SimpleTest("net", "net_unittests")
-
-  def TestNetPerf(self):
-    return self.SimpleTest("net", "net_perftests")
-
-  def TestPhoneNumber(self):
-    return self.SimpleTest("phonenumber", "libphonenumber_unittests")
-
-  def TestPPAPI(self):
-    return self.SimpleTest("chrome", "ppapi_unittests")
-
-  def TestPrinting(self):
-    return self.SimpleTest("chrome", "printing_unittests")
-
-  def TestRemoting(self):
-    return self.SimpleTest("chrome", "remoting_unittests",
-                           cmd_args=[
-                               "--ui-test-action-timeout=60000",
-                               "--ui-test-action-max-timeout=150000"])
-
-  def TestSkia(self):
-    return self.SimpleTest("skia", "skia_unittests")
-
-  def TestSql(self):
-    return self.SimpleTest("chrome", "sql_unittests")
-
-  def TestStorage(self):
-    return self.SimpleTest("storage", "storage_unittests")
-
-  def TestLinuxSandbox(self):
-    return self.SimpleTest("sandbox", "sandbox_linux_unittests")
-
-  def TestUnit(self):
-    # http://crbug.com/51716
-    # Disabling all unit tests
-    # Problems reappeared after r119922
-    if common.IsMac() and (self._options.valgrind_tool == "memcheck"):
-      logging.warning("unit_tests are disabled for memcheck on MacOS.")
-      return 0;
-    return self.SimpleTest("chrome", "unit_tests")
-
-  def TestUIBaseUnit(self):
-    return self.SimpleTest("chrome", "ui_base_unittests")
-
-  def TestUIChromeOS(self):
-    return self.SimpleTest("chrome", "ui_chromeos_unittests")
-
-  def TestURL(self):
-    return self.SimpleTest("chrome", "url_unittests")
-
-  def TestViews(self):
-    return self.SimpleTest("views", "views_unittests")
-
-
-  # Valgrind timeouts are in seconds.
-  UI_VALGRIND_ARGS = ["--timeout=14400", "--trace_children", "--indirect"]
-  # UI test timeouts are in milliseconds.
-  UI_TEST_ARGS = ["--ui-test-action-timeout=60000",
-                  "--ui-test-action-max-timeout=150000",
-                  "--no-sandbox"]
-
-  # TODO(thestig) fine-tune these values.
-  # Valgrind timeouts are in seconds.
-  BROWSER_VALGRIND_ARGS = ["--timeout=50000", "--trace_children", "--indirect"]
-  # Browser test timeouts are in milliseconds.
-  BROWSER_TEST_ARGS = ["--ui-test-action-timeout=400000",
-                       "--ui-test-action-max-timeout=800000",
-                       "--no-sandbox"]
-
-  def TestBrowser(self):
-    return self.SimpleTest("chrome", "browser_tests",
-                           valgrind_test_args=self.BROWSER_VALGRIND_ARGS,
-                           cmd_args=self.BROWSER_TEST_ARGS)
-
-  def TestContentBrowser(self):
-    return self.SimpleTest("content", "content_browsertests",
-                           valgrind_test_args=self.BROWSER_VALGRIND_ARGS,
-                           cmd_args=self.BROWSER_TEST_ARGS)
-
-  def TestInteractiveUI(self):
-    return self.SimpleTest("chrome", "interactive_ui_tests",
-                           valgrind_test_args=self.UI_VALGRIND_ARGS,
-                           cmd_args=self.UI_TEST_ARGS)
-
-  def TestSyncIntegration(self):
-    return self.SimpleTest("chrome", "sync_integration_tests",
-                           valgrind_test_args=self.UI_VALGRIND_ARGS,
-                           cmd_args=(["--ui-test-action-max-timeout=450000"]))
-
-  def TestLayoutChunk(self, chunk_num, chunk_size):
-    # Run tests [chunk_num*chunk_size .. (chunk_num+1)*chunk_size) from the
-    # list of tests.  Wrap around to beginning of list at end.
-    # If chunk_size is zero, run all tests in the list once.
-    # If a text file is given as argument, it is used as the list of tests.
-    assert((chunk_size == 0) != (len(self._args) == 0))
-    # Build the ginormous commandline in 'cmd'.
-    # It's going to be roughly
-    #  python valgrind_test.py ...
-    # but we'll use the --indirect flag to valgrind_test.py
-    # to avoid valgrinding python.
-    # Start by building the valgrind_test.py commandline.
-    tool = valgrind_test.CreateTool(self._options.valgrind_tool)
-    cmd = self._DefaultCommand(tool)
-    cmd.append("--trace_children")
-    cmd.append("--indirect_webkit_layout")
-    cmd.append("--ignore_exit_code")
-    # Now build script_cmd, the run-webkits-tests commandline.
-    # Store each chunk in its own directory so that we can find the data later
-    chunk_dir = os.path.join("layout", "chunk_%05d" % chunk_num)
-    out_dir = os.path.join(path_utils.ScriptDir(), "latest")
-    out_dir = os.path.join(out_dir, chunk_dir)
-    if os.path.exists(out_dir):
-      old_files = glob.glob(os.path.join(out_dir, "*.txt"))
-      for f in old_files:
-        os.remove(f)
-    else:
-      os.makedirs(out_dir)
-    script = os.path.join(self._source_dir, "third_party", "WebKit", "Tools",
-                          "Scripts", "run-webkit-tests")
-    # http://crbug.com/260627: After the switch to content_shell from DRT, each
-    # test now brings up 3 processes.  Under Valgrind, they become memory bound
-    # and can eventually OOM if we don't reduce the total count.
-    # It'd be nice if content_shell automatically throttled the startup of new
-    # tests if we're low on memory.
-    jobs = max(1, int(multiprocessing.cpu_count() * 0.3))
-    script_cmd = ["python", script, "-v",
-                  # run a separate DumpRenderTree for each test
-                  "--batch-size=1",
-                  "--fully-parallel",
-                  "--child-processes=%d" % jobs,
-                  "--time-out-ms=800000",
-                  "--no-retry-failures",  # retrying takes too much time
-                  # http://crbug.com/176908: Don't launch a browser when done.
-                  "--no-show-results",
-                  "--nocheck-sys-deps",
-                  "--additional-driver-flag=--no-sandbox"]
-    # Pass build mode to run-webkit-tests.  We aren't passed it directly,
-    # so parse it out of build_dir.  run-webkit-tests can only handle
-    # the two values "Release" and "Debug".
-    # TODO(Hercules): unify how all our scripts pass around build mode
-    # (--mode / --target / --build-dir / --debug)
-    if self._options.build_dir:
-      build_root, mode = os.path.split(self._options.build_dir)
-      script_cmd.extend(["--build-directory", build_root, "--target", mode])
-    if (chunk_size > 0):
-      script_cmd.append("--run-chunk=%d:%d" % (chunk_num, chunk_size))
-    if len(self._args):
-      # if the arg is a txt file, then treat it as a list of tests
-      if os.path.isfile(self._args[0]) and self._args[0][-4:] == ".txt":
-        script_cmd.append("--test-list=%s" % self._args[0])
-      else:
-        script_cmd.extend(self._args)
-    self._AppendGtestFilter(tool, "layout", script_cmd)
-    # Now run script_cmd with the wrapper in cmd
-    cmd.extend(["--"])
-    cmd.extend(script_cmd)
-
-    # Layout tests often times fail quickly, but the buildbot remains green.
-    # Detect this situation when running with the default chunk size.
-    if chunk_size == self.LAYOUT_TESTS_DEFAULT_CHUNK_SIZE:
-      min_runtime_in_seconds=120
-    else:
-      min_runtime_in_seconds=0
-    ret = tool.Run(cmd, "layout", min_runtime_in_seconds=min_runtime_in_seconds)
-    return ret
-
-
-  def TestLayout(self):
-    # A "chunk file" is maintained in the local directory so that each test
-    # runs a slice of the layout tests of size chunk_size that increments with
-    # each run.  Since tests can be added and removed from the layout tests at
-    # any time, this is not going to give exact coverage, but it will allow us
-    # to continuously run small slices of the layout tests under valgrind rather
-    # than having to run all of them in one shot.
-    chunk_size = self._options.num_tests
-    if chunk_size == 0 or len(self._args):
-      return self.TestLayoutChunk(0, 0)
-    chunk_num = 0
-    chunk_file = os.path.join("valgrind_layout_chunk.txt")
-    logging.info("Reading state from " + chunk_file)
-    try:
-      f = open(chunk_file)
-      if f:
-        chunk_str = f.read()
-        if len(chunk_str):
-          chunk_num = int(chunk_str)
-        # This should be enough so that we have a couple of complete runs
-        # of test data stored in the archive (although note that when we loop
-        # that we almost guaranteed won't be at the end of the test list)
-        if chunk_num > 10000:
-          chunk_num = 0
-        f.close()
-    except IOError, (errno, strerror):
-      logging.error("error reading from file %s (%d, %s)" % (chunk_file,
-                    errno, strerror))
-    # Save the new chunk size before running the tests. Otherwise if a
-    # particular chunk hangs the bot, the chunk number will never get
-    # incremented and the bot will be wedged.
-    logging.info("Saving state to " + chunk_file)
-    try:
-      f = open(chunk_file, "w")
-      chunk_num += 1
-      f.write("%d" % chunk_num)
-      f.close()
-    except IOError, (errno, strerror):
-      logging.error("error writing to file %s (%d, %s)" % (chunk_file, errno,
-                    strerror))
-    # Since we're running small chunks of the layout tests, it's important to
-    # mark the ones that have errors in them.  These won't be visible in the
-    # summary list for long, but will be useful for someone reviewing this bot.
-    return self.TestLayoutChunk(chunk_num, chunk_size)
-
-  # The known list of tests.
-  # Recognise the original abbreviations as well as full executable names.
-  _test_list = {
-    "cmdline" : RunCmdLine,
-    "addressinput": TestAddressInput,
-    "libaddressinput_unittests": TestAddressInput,
-    "accessibility": TestAccessibility,
-    "angle": TestAngle,          "angle_unittests": TestAngle,
-    "app_list": TestAppList,     "app_list_unittests": TestAppList,
-    "ash": TestAsh,              "ash_unittests": TestAsh,
-    "aura": TestAura,            "aura_unittests": TestAura,
-    "base": TestBase,            "base_unittests": TestBase,
-    "blink_heap": TestBlinkHeap,
-    "blink_platform": TestBlinkPlatform,
-    "browser": TestBrowser,      "browser_tests": TestBrowser,
-    "cacheinvalidation": TestCacheInvalidation,
-    "cacheinvalidation_unittests": TestCacheInvalidation,
-    "cast": TestCast,            "cast_unittests": TestCast,
-    "cc": TestCC,                "cc_unittests": TestCC,
-    "chrome_app": TestChromeApp,
-    "chrome_elf": TestChromeElf,
-    "chromedriver": TestChromeDriver,
-    "chromeos": TestChromeOS,    "chromeos_unittests": TestChromeOS,
-    "components": TestComponents,"components_unittests": TestComponents,
-    "compositor": TestCompositor,"compositor_unittests": TestCompositor,
-    "content": TestContent,      "content_unittests": TestContent,
-    "content_browsertests": TestContentBrowser,
-    "courgette": TestCourgette,  "courgette_unittests": TestCourgette,
-    "crypto": TestCrypto,        "crypto_unittests": TestCrypto,
-    "device": TestDevice,        "device_unittests": TestDevice,
-    "display": TestDisplay,      "display_unittests": TestDisplay,
-    "events": TestEvents,        "events_unittests": TestEvents,
-    "extensions": TestExtensions, "extensions_unittests": TestExtensions,
-    "ffmpeg_regression_tests": TestFFmpegRegressions,
-    "gcm": TestGCM,              "gcm_unit_tests": TestGCM,
-    "gin": TestGin,              "gin_unittests": TestGin,
-    "gfx": TestGfx,              "gfx_unittests": TestGfx,
-    "google_apis": TestGoogleApis,
-    "gpu": TestGPU,              "gpu_unittests": TestGPU,
-    "ipc": TestIpc,              "ipc_tests": TestIpc,
-    "installer_util": TestInstallerUtil,
-    "installer_util_unittests": TestInstallerUtil,
-    "install_static_unittests": TestInstallStatic,
-    "interactive_ui": TestInteractiveUI,
-    "jingle": TestJingle,        "jingle_unittests": TestJingle,
-    "keyboard": TestKeyboard,    "keyboard_unittests": TestKeyboard,
-    "latency": TestLatency,      "latency_unittests": TestLatency,
-    "layout": TestLayout,        "layout_tests": TestLayout,
-    "media": TestMedia,          "media_unittests": TestMedia,
-    "message_center": TestMessageCenter,
-    "message_center_unittests" : TestMessageCenter,
-    "midi": TestMidi,             "midi_unittests": TestMidi,
-    "mojo_common": TestMojoCommon,
-    "mojo_common_unittests": TestMojoCommon,
-    "mojo_system": TestMojoSystem,
-    "mojo_system_unittests": TestMojoSystem,
-    "mojo_public_system": TestMojoPublicSystem,
-    "mojo_public_system_unittests": TestMojoPublicSystem,
-    "mojo_public_bindings": TestMojoPublicBindings,
-    "mojo_public_bindings_unittests": TestMojoPublicBindings,
-    "mojo_public_sysperf": TestMojoPublicSysPerf,
-    "net": TestNet,              "net_unittests": TestNet,
-    "net_perf": TestNetPerf,     "net_perftests": TestNetPerf,
-    "phonenumber": TestPhoneNumber,
-    "libphonenumber_unittests": TestPhoneNumber,
-    "ppapi": TestPPAPI,          "ppapi_unittests": TestPPAPI,
-    "printing": TestPrinting,    "printing_unittests": TestPrinting,
-    "remoting": TestRemoting,    "remoting_unittests": TestRemoting,
-    "sandbox": TestLinuxSandbox, "sandbox_linux_unittests": TestLinuxSandbox,
-    "skia": TestSkia,            "skia_unittests": TestSkia,
-    "sql": TestSql,              "sql_unittests": TestSql,
-    "storage": TestStorage,      "storage_unittests": TestStorage,
-    "sync_integration_tests": TestSyncIntegration,
-    "sync_integration": TestSyncIntegration,
-    "ui_base_unit": TestUIBaseUnit,       "ui_base_unittests": TestUIBaseUnit,
-    "ui_chromeos": TestUIChromeOS, "ui_chromeos_unittests": TestUIChromeOS,
-    "unit": TestUnit,            "unit_tests": TestUnit,
-    "url": TestURL,              "url_unittests": TestURL,
-    "views": TestViews,          "views_unittests": TestViews,
-    "webkit": TestLayout,
-  }
-
-
-def _main():
-  parser = optparse.OptionParser("usage: %prog -b <dir> -t <test> "
-                                 "[-t <test> ...]")
-
-  parser.add_option("--help-tests", dest="help_tests", action="store_true",
-                    default=False, help="List all available tests")
-  parser.add_option("-b", "--build-dir",
-                    help="the location of the compiler output")
-  parser.add_option("--target", help="Debug or Release")
-  parser.add_option("-t", "--test", action="append", default=[],
-                    help="which test to run, supports test:gtest_filter format "
-                         "as well.")
-  parser.add_option("--baseline", action="store_true", default=False,
-                    help="generate baseline data instead of validating")
-  parser.add_option("-f", "--force", action="store_true", default=False,
-                    help="run a broken test anyway")
-  parser.add_option("--gtest_filter",
-                    help="additional arguments to --gtest_filter")
-  parser.add_option("--gtest_repeat", help="argument for --gtest_repeat")
-  parser.add_option("--gtest_shuffle", action="store_true", default=False,
-                    help="Randomize tests' orders on every iteration.")
-  parser.add_option("--gtest_break_on_failure", action="store_true",
-                    default=False,
-                    help="Drop in to debugger on assertion failure. Also "
-                         "useful for forcing tests to exit with a stack dump "
-                         "on the first assertion failure when running with "
-                         "--gtest_repeat=-1")
-  parser.add_option("-v", "--verbose", action="store_true", default=False,
-                    help="verbose output - enable debug log messages")
-  parser.add_option("--tool", dest="valgrind_tool", default="memcheck",
-                    help="specify a valgrind tool to run the tests under")
-  parser.add_option("--tool_flags", dest="valgrind_tool_flags", default="",
-                    help="specify custom flags for the selected valgrind tool")
-  parser.add_option("--keep_logs", action="store_true", default=False,
-                    help="store memory tool logs in the <tool>.logs directory "
-                         "instead of /tmp.\nThis can be useful for tool "
-                         "developers/maintainers.\nPlease note that the <tool>"
-                         ".logs directory will be clobbered on tool startup.")
-  parser.add_option("-n", "--num_tests", type="int",
-                    default=ChromeTests.LAYOUT_TESTS_DEFAULT_CHUNK_SIZE,
-                    help="for layout tests: # of subtests per run.  0 for all.")
-  parser.add_option("--test-launcher-bot-mode", action="store_true",
-                    help="run the tests with --test-launcher-bot-mode")
-  parser.add_option("--test-launcher-total-shards", type=int,
-                    help="run the tests with --test-launcher-total-shards")
-  parser.add_option("--test-launcher-shard-index", type=int,
-                    help="run the tests with --test-launcher-shard-index")
-  parser.add_option("--drmemory_ops",
-                    help="extra options passed to Dr. Memory")
-
-  options, args = parser.parse_args()
-
-  # Bake target into build_dir.
-  if options.target and options.build_dir:
-    assert (options.target !=
-            os.path.basename(os.path.dirname(options.build_dir)))
-    options.build_dir = os.path.join(os.path.abspath(options.build_dir),
-                                     options.target)
-
-  if options.verbose:
-    logging_utils.config_root(logging.DEBUG)
-  else:
-    logging_utils.config_root()
-
-  if options.help_tests:
-    ChromeTests.ShowTests()
-    return 0
-
-  if not options.test:
-    parser.error("--test not specified")
-
-  if len(options.test) != 1 and options.gtest_filter:
-    parser.error("--gtest_filter and multiple tests don't make sense together")
-
-  BROKEN_TESTS = {
-    'drmemory_light': [
-      'addressinput',
-      'aura',
-      'base_unittests',
-      'cc',
-      'components', # x64 only?
-      'content',
-      'gfx',
-      'mojo_public_bindings',
-    ],
-    'drmemory_full': [
-      'addressinput',
-      'aura',
-      'base_unittests',
-      'blink_heap',
-      'blink_platform',
-      'browser_tests',
-      'cast',
-      'cc',
-      'chromedriver',
-      'compositor',
-      'content',
-      'content_browsertests',
-      'device',
-      'events',
-      'extensions',
-      'gfx',
-      'google_apis',
-      'gpu',
-      'ipc_tests',
-      'jingle',
-      'keyboard',
-      'media',
-      'midi',
-      'mojo_common',
-      'mojo_public_bindings',
-      'mojo_public_sysperf',
-      'mojo_public_system',
-      'mojo_system',
-      'net',
-      'remoting',
-      'unit',
-      'url',
-    ],
-  }
-
-  for t in options.test:
-    if t in BROKEN_TESTS[options.valgrind_tool] and not options.force:
-      logging.info("Skipping broken %s test %s -- see crbug.com/633693" %
-                   (options.valgrind_tool, t))
-      return 0
-
-    tests = ChromeTests(options, args, t)
-    ret = tests.Run()
-    if ret: return ret
-  return 0
-
-
-if __name__ == "__main__":
-  sys.exit(_main())
diff --git a/files/tools_libyuv/valgrind/chrome_tests.sh b/files/tools_libyuv/valgrind/chrome_tests.sh
deleted file mode 100755
index dc17684f..00000000
--- a/files/tools_libyuv/valgrind/chrome_tests.sh
+++ /dev/null
@@ -1,94 +0,0 @@
-#!/bin/bash
-
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# Set up some paths and re-direct the arguments to chrome_tests.py
-
-export THISDIR=`dirname $0`
-ARGV_COPY="$@"
-
-# We need to set CHROME_VALGRIND iff using Memcheck:
-#   tools/valgrind/chrome_tests.sh --tool memcheck
-# or
-#   tools/valgrind/chrome_tests.sh --tool=memcheck
-tool="memcheck"  # Default to memcheck.
-while (( "$#" ))
-do
-  if [[ "$1" == "--tool" ]]
-  then
-    tool="$2"
-    shift
-  elif [[ "$1" =~ --tool=(.*) ]]
-  then
-    tool="${BASH_REMATCH[1]}"
-  fi
-  shift
-done
-
-NEEDS_VALGRIND=0
-NEEDS_DRMEMORY=0
-
-case "$tool" in
-  "memcheck")
-    NEEDS_VALGRIND=1
-    ;;
-  "drmemory" | "drmemory_light" | "drmemory_full" | "drmemory_pattern")
-    NEEDS_DRMEMORY=1
-    ;;
-esac
-
-if [ "$NEEDS_VALGRIND" == "1" ]
-then
-  export CHROME_VALGRIND=`sh $THISDIR/locate_valgrind.sh`
-  if [ "$CHROME_VALGRIND" = "" ]
-  then
-    # locate_valgrind.sh failed
-    exit 1
-  fi
-  echo "Using valgrind binaries from ${CHROME_VALGRIND}"
-
-  PATH="${CHROME_VALGRIND}/bin:$PATH"
-  # We need to set these variables to override default lib paths hard-coded into
-  # Valgrind binary.
-  export VALGRIND_LIB="$CHROME_VALGRIND/lib/valgrind"
-  export VALGRIND_LIB_INNER="$CHROME_VALGRIND/lib/valgrind"
-
-  # Clean up some /tmp directories that might be stale due to interrupted
-  # chrome_tests.py execution.
-  # FYI:
-  #   -mtime +1  <- only print files modified more than 24h ago,
-  #   -print0/-0 are needed to handle possible newlines in the filenames.
-  echo "Cleanup /tmp from Valgrind stuff"
-  find /tmp -maxdepth 1 \(\
-        -name "vgdb-pipe-*" -or -name "vg_logs_*" -or -name "valgrind.*" \
-      \) -mtime +1 -print0 | xargs -0 rm -rf
-fi
-
-if [ "$NEEDS_DRMEMORY" == "1" ]
-then
-  if [ -z "$DRMEMORY_COMMAND" ]
-  then
-    DRMEMORY_PATH="$THISDIR/../../third_party/drmemory"
-    DRMEMORY_SFX="$DRMEMORY_PATH/drmemory-windows-sfx.exe"
-    if [ ! -f "$DRMEMORY_SFX" ]
-    then
-      echo "Can't find Dr. Memory executables."
-      echo "See http://www.chromium.org/developers/how-tos/using-valgrind/dr-memory"
-      echo "for the instructions on how to get them."
-      exit 1
-    fi
-
-    chmod +x "$DRMEMORY_SFX"  # Cygwin won't run it without +x.
-    "$DRMEMORY_SFX" -o"$DRMEMORY_PATH/unpacked" -y
-    export DRMEMORY_COMMAND="$DRMEMORY_PATH/unpacked/bin/drmemory.exe"
-  fi
-fi
-
-PYTHONPATH=$THISDIR/../python/google python \
-           "$THISDIR/chrome_tests.py" $ARGV_COPY
diff --git a/files/tools_libyuv/valgrind/common.py b/files/tools_libyuv/valgrind/common.py
deleted file mode 100644
index e9ee51e4..00000000
--- a/files/tools_libyuv/valgrind/common.py
+++ /dev/null
@@ -1,256 +0,0 @@
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-import logging
-import platform
-import os
-import signal
-import subprocess
-import sys
-import time
-
-
-class NotImplementedError(Exception):
-  pass
-
-
-class TimeoutError(Exception):
-  pass
-
-
-def RunSubprocessInBackground(proc):
-  """Runs a subprocess in the background. Returns a handle to the process."""
-  logging.info("running %s in the background" % " ".join(proc))
-  return subprocess.Popen(proc)
-
-
-def RunSubprocess(proc, timeout=0):
-  """ Runs a subprocess, until it finishes or |timeout| is exceeded and the
-  process is killed with taskkill.  A |timeout| <= 0  means no timeout.
-
-  Args:
-    proc: list of process components (exe + args)
-    timeout: how long to wait before killing, <= 0 means wait forever
-  """
-
-  logging.info("running %s, timeout %d sec" % (" ".join(proc), timeout))
-  sys.stdout.flush()
-  sys.stderr.flush()
-
-  # Manually read and print out stdout and stderr.
-  # By default, the subprocess is supposed to inherit these from its parent,
-  # however when run under buildbot, it seems unable to read data from a
-  # grandchild process, so we have to read the child and print the data as if
-  # it came from us for buildbot to read it.  We're not sure why this is
-  # necessary.
-  # TODO(erikkay): should we buffer stderr and stdout separately?
-  p = subprocess.Popen(proc, universal_newlines=True,
-                       bufsize=0,  # unbuffered
-                       stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
-
-  logging.info("started subprocess")
-
-  did_timeout = False
-  if timeout > 0:
-    wait_until = time.time() + timeout
-  while p.poll() is None and not did_timeout:
-    # Have to use readline rather than readlines() or "for line in p.stdout:",
-    # otherwise we get buffered even with bufsize=0.
-    line = p.stdout.readline()
-    while line and not did_timeout:
-      sys.stdout.write(line)
-      sys.stdout.flush()
-      line = p.stdout.readline()
-      if timeout > 0:
-        did_timeout = time.time() > wait_until
-
-  if did_timeout:
-    logging.info("process timed out")
-  else:
-    logging.info("process ended, did not time out")
-
-  if did_timeout:
-    if IsWindows():
-      subprocess.call(["taskkill", "/T", "/F", "/PID", str(p.pid)])
-    else:
-      # Does this kill all children, too?
-      os.kill(p.pid, signal.SIGINT)
-    logging.error("KILLED %d" % p.pid)
-    # Give the process a chance to actually die before continuing
-    # so that cleanup can happen safely.
-    time.sleep(1.0)
-    logging.error("TIMEOUT waiting for %s" % proc[0])
-    raise TimeoutError(proc[0])
-  else:
-    for line in p.stdout:
-      sys.stdout.write(line)
-    if not IsMac():   # stdout flush fails on Mac
-      logging.info("flushing stdout")
-      sys.stdout.flush()
-
-  logging.info("collecting result code")
-  result = p.poll()
-  if result:
-    logging.error("%s exited with non-zero result code %d" % (proc[0], result))
-  return result
-
-
-def IsLinux():
-  return sys.platform.startswith('linux')
-
-
-def IsMac():
-  return sys.platform.startswith('darwin')
-
-
-def IsWindows():
-  return sys.platform == 'cygwin' or sys.platform.startswith('win')
-
-
-def WindowsVersionName():
-  """Returns the name of the Windows version if it is known, or None.
-
-  Possible return values are: xp, vista, 7, 8, or None
-  """
-  if sys.platform == 'cygwin':
-    # Windows version number is hiding in system name.  Looks like:
-    # CYGWIN_NT-6.1-WOW64
-    try:
-      version_str = platform.uname()[0].split('-')[1]
-    except:
-      return None
-  elif sys.platform.startswith('win'):
-    # Normal Windows version string.  Mine: 6.1.7601
-    version_str = platform.version()
-  else:
-    return None
-
-  parts = version_str.split('.')
-  try:
-    major = int(parts[0])
-    minor = int(parts[1])
-  except:
-    return None  # Can't parse, unknown version.
-
-  if major == 5:
-    return 'xp'
-  elif major == 6 and minor == 0:
-    return 'vista'
-  elif major == 6 and minor == 1:
-    return '7'
-  elif major == 6 and minor == 2:
-    return '8'  # Future proof.  ;)
-  return None
-
-
-def PlatformNames():
-  """Return an array of string to be used in paths for the platform
-  (e.g. suppressions, gtest filters, ignore files etc.)
-  The first element of the array describes the 'main' platform
-  """
-  if IsLinux():
-    return ['linux']
-  if IsMac():
-    return ['mac']
-  if IsWindows():
-    names = ['win32']
-    version_name = WindowsVersionName()
-    if version_name is not None:
-      names.append('win-%s' % version_name)
-    return names
-  raise NotImplementedError('Unknown platform "%s".' % sys.platform)
-
-
-def PutEnvAndLog(env_name, env_value):
-  os.putenv(env_name, env_value)
-  logging.info('export %s=%s', env_name, env_value)
-
-def BoringCallers(mangled, use_re_wildcards):
-  """Return a list of 'boring' function names (optinally mangled)
-  with */? wildcards (optionally .*/.).
-  Boring = we drop off the bottom of stack traces below such functions.
-  """
-
-  need_mangling = [
-    # Don't show our testing framework:
-    ("testing::Test::Run",     "_ZN7testing4Test3RunEv"),
-    ("testing::TestInfo::Run", "_ZN7testing8TestInfo3RunEv"),
-    ("testing::internal::Handle*ExceptionsInMethodIfSupported*",
-     "_ZN7testing8internal3?Handle*ExceptionsInMethodIfSupported*"),
-
-    # Depend on scheduling:
-    ("MessageLoop::Run",     "_ZN11MessageLoop3RunEv"),
-    ("MessageLoop::RunTask", "_ZN11MessageLoop7RunTask*"),
-    ("RunnableMethod*",      "_ZN14RunnableMethod*"),
-    ("DispatchToMethod*",    "_Z*16DispatchToMethod*"),
-    ("base::internal::Invoker*::DoInvoke*",
-     "_ZN4base8internal8Invoker*DoInvoke*"),  # Invoker{1,2,3}
-    ("base::internal::RunnableAdapter*::Run*",
-     "_ZN4base8internal15RunnableAdapter*Run*"),
-  ]
-
-  ret = []
-  for pair in need_mangling:
-    ret.append(pair[1 if mangled else 0])
-
-  ret += [
-    # Also don't show the internals of libc/pthread.
-    "start_thread",
-    "main",
-    "BaseThreadInitThunk",
-  ]
-
-  if use_re_wildcards:
-    for i in range(0, len(ret)):
-      ret[i] = ret[i].replace('*', '.*').replace('?', '.')
-
-  return ret
-
-def NormalizeWindowsPath(path):
-  """If we're using Cygwin Python, turn the path into a Windows path.
-
-  Don't turn forward slashes into backslashes for easier copy-pasting and
-  escaping.
-
-  TODO(rnk): If we ever want to cut out the subprocess invocation, we can use
-  _winreg to get the root Cygwin directory from the registry key:
-  HKEY_LOCAL_MACHINE\SOFTWARE\Cygwin\setup\rootdir.
-  """
-  if sys.platform.startswith("cygwin"):
-    p = subprocess.Popen(["cygpath", "-m", path],
-                         stdout=subprocess.PIPE,
-                         stderr=subprocess.PIPE)
-    (out, err) = p.communicate()
-    if err:
-      logging.warning("WARNING: cygpath error: %s", err)
-    return out.strip()
-  else:
-    return path
-
-############################
-# Common output format code
-
-def PrintUsedSuppressionsList(suppcounts):
-  """ Prints out the list of used suppressions in a format common to all the
-      memory tools. If the list is empty, prints nothing and returns False,
-      otherwise True.
-
-      suppcounts: a dictionary of used suppression counts,
-                  Key -> name, Value -> count.
-  """
-  if not suppcounts:
-    return False
-
-  print "-----------------------------------------------------"
-  print "Suppressions used:"
-  print "  count name"
-  for (name, count) in sorted(suppcounts.items(), key=lambda (k,v): (v,k)):
-    print "%7d %s" % (count, name)
-  print "-----------------------------------------------------"
-  sys.stdout.flush()
-  return True
diff --git a/files/tools_libyuv/valgrind/gdb_helper.py b/files/tools_libyuv/valgrind/gdb_helper.py
deleted file mode 100644
index d127f760..00000000
--- a/files/tools_libyuv/valgrind/gdb_helper.py
+++ /dev/null
@@ -1,91 +0,0 @@
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-''' A bunch of helper functions for querying gdb.'''
-
-import logging
-import os
-import re
-import tempfile
-
-GDB_LINE_RE = re.compile(r'Line ([0-9]*) of "([^"]*)".*')
-
-def _GdbOutputToFileLine(output_line):
-  ''' Parse the gdb output line, return a pair (file, line num) '''
-  match =  GDB_LINE_RE.match(output_line)
-  if match:
-    return match.groups()[1], match.groups()[0]
-  else:
-    return None
-
-def ResolveAddressesWithinABinary(binary_name, load_address, address_list):
-  ''' For each address, return a pair (file, line num) '''
-  commands = tempfile.NamedTemporaryFile()
-  commands.write('add-symbol-file "%s" %s\n' % (binary_name, load_address))
-  for addr in address_list:
-    commands.write('info line *%s\n' % addr)
-  commands.write('quit\n')
-  commands.flush()
-  gdb_commandline = 'gdb -batch -x %s 2>/dev/null' % commands.name
-  gdb_pipe = os.popen(gdb_commandline)
-  result = gdb_pipe.readlines()
-
-  address_count = 0
-  ret = {}
-  for line in result:
-    if line.startswith('Line'):
-      ret[address_list[address_count]] = _GdbOutputToFileLine(line)
-      address_count += 1
-    if line.startswith('No line'):
-      ret[address_list[address_count]] = (None, None)
-      address_count += 1
-  gdb_pipe.close()
-  commands.close()
-  return ret
-
-class AddressTable(object):
-  ''' Object to do batched line number lookup. '''
-  def __init__(self):
-    self._load_addresses = {}
-    self._binaries = {}
-    self._all_resolved = False
-
-  def AddBinaryAt(self, binary, load_address):
-    ''' Register a new shared library or executable. '''
-    self._load_addresses[binary] = load_address
-
-  def Add(self, binary, address):
-    ''' Register a lookup request. '''
-    if binary == '':
-      logging.warn('adding address %s in empty binary?' % address)
-    if binary in self._binaries:
-      self._binaries[binary].append(address)
-    else:
-      self._binaries[binary] = [address]
-    self._all_resolved = False
-
-  def ResolveAll(self):
-    ''' Carry out all lookup requests. '''
-    self._translation = {}
-    for binary in self._binaries.keys():
-      if binary != '' and binary in self._load_addresses:
-        load_address = self._load_addresses[binary]
-        addr = ResolveAddressesWithinABinary(
-            binary, load_address, self._binaries[binary])
-        self._translation[binary] = addr
-    self._all_resolved = True
-
-  def GetFileLine(self, binary, addr):
-    ''' Get the (filename, linenum) result of a previously-registered lookup
-    request.
-    '''
-    if self._all_resolved:
-      if binary in self._translation:
-        if addr in self._translation[binary]:
-          return self._translation[binary][addr]
-    return (None, None)
diff --git a/files/tools_libyuv/valgrind/libyuv_tests.bat b/files/tools_libyuv/valgrind/libyuv_tests.bat
deleted file mode 100644
index 5fceca67..00000000
--- a/files/tools_libyuv/valgrind/libyuv_tests.bat
+++ /dev/null
@@ -1,79 +0,0 @@
-@echo off
-:: Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-::
-:: Use of this source code is governed by a BSD-style license
-:: that can be found in the LICENSE file in the root of the source
-:: tree. An additional intellectual property rights grant can be found
-:: in the file PATENTS.  All contributing project authors may
-:: be found in the AUTHORS file in the root of the source tree.
-
-:: This script is a copy of chrome_tests.bat with the following changes:
-:: - Invokes libyuv_tests.py instead of chrome_tests.py
-:: - Chromium's Valgrind scripts directory is added to the PYTHONPATH to make
-::   it possible to execute the Python scripts properly.
-
-:: TODO(timurrrr): batch files 'export' all the variables to the parent shell
-set THISDIR=%~dp0
-set TOOL_NAME="unknown"
-
-:: Get the tool name and put it into TOOL_NAME {{{1
-:: NB: SHIFT command doesn't modify %*
-:PARSE_ARGS_LOOP
-  if %1 == () GOTO:TOOLNAME_NOT_FOUND
-  if %1 == --tool GOTO:TOOLNAME_FOUND
-  SHIFT
-  goto :PARSE_ARGS_LOOP
-
-:TOOLNAME_NOT_FOUND
-echo "Please specify a tool (tsan or drmemory) by using --tool flag"
-exit /B 1
-
-:TOOLNAME_FOUND
-SHIFT
-set TOOL_NAME=%1
-:: }}}
-if "%TOOL_NAME%" == "drmemory"          GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_light"    GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_full"     GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "drmemory_pattern"  GOTO :SETUP_DRMEMORY
-if "%TOOL_NAME%" == "tsan"     GOTO :SETUP_TSAN
-echo "Unknown tool: `%TOOL_NAME%`! Only tsan and drmemory are supported."
-exit /B 1
-
-:SETUP_DRMEMORY
-if NOT "%DRMEMORY_COMMAND%"=="" GOTO :RUN_TESTS
-:: Set up DRMEMORY_COMMAND to invoke Dr. Memory {{{1
-set DRMEMORY_PATH=%THISDIR%..\..\third_party\drmemory
-set DRMEMORY_SFX=%DRMEMORY_PATH%\drmemory-windows-sfx.exe
-if EXIST %DRMEMORY_SFX% GOTO DRMEMORY_BINARY_OK
-echo "Can't find Dr. Memory executables."
-echo "See http://www.chromium.org/developers/how-tos/using-valgrind/dr-memory"
-echo "for the instructions on how to get them."
-exit /B 1
-
-:DRMEMORY_BINARY_OK
-%DRMEMORY_SFX% -o%DRMEMORY_PATH%\unpacked -y
-set DRMEMORY_COMMAND=%DRMEMORY_PATH%\unpacked\bin\drmemory.exe
-:: }}}
-goto :RUN_TESTS
-
-:SETUP_TSAN
-:: Set up PIN_COMMAND to invoke TSan {{{1
-set TSAN_PATH=%THISDIR%..\..\third_party\tsan
-set TSAN_SFX=%TSAN_PATH%\tsan-x86-windows-sfx.exe
-if EXIST %TSAN_SFX% GOTO TSAN_BINARY_OK
-echo "Can't find ThreadSanitizer executables."
-echo "See http://www.chromium.org/developers/how-tos/using-valgrind/threadsanitizer/threadsanitizer-on-windows"
-echo "for the instructions on how to get them."
-exit /B 1
-
-:TSAN_BINARY_OK
-%TSAN_SFX% -o%TSAN_PATH%\unpacked -y
-set PIN_COMMAND=%TSAN_PATH%\unpacked\tsan-x86-windows\tsan.bat
-:: }}}
-goto :RUN_TESTS
-
-:RUN_TESTS
-set PYTHONPATH=%THISDIR%..\python\google;%THISDIR%..\valgrind
-set RUNNING_ON_VALGRIND=yes
-python %THISDIR%libyuv_tests.py %*
diff --git a/files/tools_libyuv/valgrind/libyuv_tests.py b/files/tools_libyuv/valgrind/libyuv_tests.py
deleted file mode 100755
index e780bd95..00000000
--- a/files/tools_libyuv/valgrind/libyuv_tests.py
+++ /dev/null
@@ -1,139 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Runs various libyuv tests through valgrind_test.py.
-
-This script inherits the chrome_tests.py in Chrome, but allows running any test
-instead of only the hard-coded ones. It uses the -t cmdline flag to do this, and
-only supports specifying a single test for each run.
-
-Suppression files:
-The Chrome valgrind directory we use as a DEPS dependency contains the following
-suppression files:
-  valgrind/memcheck/suppressions.txt
-  valgrind/memcheck/suppressions_mac.txt
-  valgrind/tsan/suppressions.txt
-  valgrind/tsan/suppressions_mac.txt
-  valgrind/tsan/suppressions_win32.txt
-Since they're referenced from the chrome_tests.py script, we have similar files
-below the directory of this script. When executing, this script will setup both
-Chrome's suppression files and our own, so we can easily maintain libyuv
-specific suppressions in our own files.
-"""
-
-import logging
-import optparse
-import os
-import sys
-
-import logging_utils
-import path_utils
-
-import chrome_tests
-
-
-class LibyuvTest(chrome_tests.ChromeTests):
-  """Class that handles setup of suppressions for libyuv.
-
-  Everything else is inherited from chrome_tests.ChromeTests.
-  """
-
-  def _DefaultCommand(self, tool, exe=None, valgrind_test_args=None):
-    """Override command-building method so we can add more suppressions."""
-    cmd = chrome_tests.ChromeTests._DefaultCommand(self, tool, exe,
-                                                   valgrind_test_args)
-    # When ChromeTests._DefaultCommand has executed, it has setup suppression
-    # files based on what's found in the memcheck/ or tsan/ subdirectories of
-    # this script's location. If Mac or Windows is executing, additional
-    # platform specific files have also been added.
-    # Since only the ones located below this directory is added, we must also
-    # add the ones maintained by Chrome, located in ../../tools/valgrind.
-
-    # The idea is to look for --suppression arguments in the cmd list and add a
-    # modified copy of each suppression file, for the corresponding file in
-    # ../../tools/valgrind.
-    script_dir = path_utils.ScriptDir()
-    old_base, _ = os.path.split(script_dir)
-
-    checkout_src = os.path.abspath(os.path.join(script_dir, os.pardir,
-                                                os.pardir))
-    new_dir = os.path.join(checkout_src, 'tools', 'valgrind')
-    add_suppressions = []
-    for token in cmd:
-      if '--suppressions' in token:
-        add_suppressions.append(token.replace(script_dir, new_dir))
-    return add_suppressions + cmd
-
-
-def main(_):
-  parser = optparse.OptionParser('usage: %prog -b <dir> -t <test> <test args>')
-  parser.disable_interspersed_args()
-  parser.add_option('-b', '--build-dir',
-                    help=('Location of the compiler output. Can only be used '
-                          'when the test argument does not contain this path.'))
-  parser.add_option("--target", help="Debug or Release")
-  parser.add_option('-t', '--test', help='Test to run.')
-  parser.add_option('', '--baseline', action='store_true', default=False,
-                    help='Generate baseline data instead of validating')
-  parser.add_option('', '--gtest_filter',
-                    help='Additional arguments to --gtest_filter')
-  parser.add_option('', '--gtest_repeat',
-                    help='Argument for --gtest_repeat')
-  parser.add_option("--gtest_shuffle", action="store_true", default=False,
-                    help="Randomize tests' orders on every iteration.")
-  parser.add_option("--gtest_break_on_failure", action="store_true",
-                    default=False,
-                    help="Drop in to debugger on assertion failure. Also "
-                         "useful for forcing tests to exit with a stack dump "
-                         "on the first assertion failure when running with "
-                         "--gtest_repeat=-1")
-  parser.add_option('-v', '--verbose', action='store_true', default=False,
-                    help='Verbose output - enable debug log messages')
-  parser.add_option('', '--tool', dest='valgrind_tool', default='memcheck',
-                    help='Specify a valgrind tool to run the tests under')
-  parser.add_option('', '--tool_flags', dest='valgrind_tool_flags', default='',
-                    help='Specify custom flags for the selected valgrind tool')
-  parser.add_option('', '--keep_logs', action='store_true', default=False,
-                    help=('Store memory tool logs in the <tool>.logs directory '
-                          'instead of /tmp.\nThis can be useful for tool '
-                          'developers/maintainers.\nPlease note that the <tool>'
-                          '.logs directory will be clobbered on tool startup.'))
-  parser.add_option("--test-launcher-bot-mode", action="store_true",
-                    help="run the tests with --test-launcher-bot-mode")
-  parser.add_option("--test-launcher-total-shards", type=int,
-                    help="run the tests with --test-launcher-total-shards")
-  parser.add_option("--test-launcher-shard-index", type=int,
-                    help="run the tests with --test-launcher-shard-index")
-  options, args = parser.parse_args()
-
-  if options.verbose:
-    logging_utils.config_root(logging.DEBUG)
-  else:
-    logging_utils.config_root()
-
-  if not options.test:
-    parser.error('--test not specified')
-
-  # Support build dir both with and without the target.
-  if (options.target and options.build_dir and
-      not options.build_dir.endswith(options.target)):
-    options.build_dir = os.path.join(options.build_dir, options.target)
-
-  # If --build_dir is provided, prepend it to the test executable if needed.
-  test_executable = options.test
-  if options.build_dir and not test_executable.startswith(options.build_dir):
-    test_executable = os.path.join(options.build_dir, test_executable)
-  args = [test_executable] + args
-
-  test = LibyuvTest(options, args, 'cmdline')
-  return test.Run()
-
-if __name__ == '__main__':
-  return_code = main(sys.argv)
-  sys.exit(return_code)
diff --git a/files/tools_libyuv/valgrind/libyuv_tests.sh b/files/tools_libyuv/valgrind/libyuv_tests.sh
deleted file mode 100755
index 249032ca..00000000
--- a/files/tools_libyuv/valgrind/libyuv_tests.sh
+++ /dev/null
@@ -1,101 +0,0 @@
-#!/bin/bash
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# Set up some paths and re-direct the arguments to libyuv_tests.py
-
-# This script is a copy of the chrome_tests.sh wrapper script with the following
-# changes:
-# - The locate_valgrind.sh of Chromium's Valgrind scripts dir is used to locate
-#   the Valgrind framework install. If it fails a fallback path is used instead
-#   (../../chromium/src/third_party/valgrind/linux_x64) and a warning message
-#   is showed by |show_locate_valgrind_failed_warning|.
-# - libyuv_tests.py is invoked instead of chrome_tests.py.
-# - Chromium's Valgrind scripts directory is added to the PYTHONPATH to make it
-#   possible to execute the Python scripts properly.
-
-export THISDIR=`dirname $0`
-ARGV_COPY="$@"
-
-# We need to set CHROME_VALGRIND iff using Memcheck:
-#   tools_libyuv/valgrind/libyuv_tests.sh --tool memcheck
-# or
-#   tools_libyuv/valgrind/libyuv_tests.sh --tool=memcheck
-tool="memcheck"  # Default to memcheck.
-while (( "$#" ))
-do
-  if [[ "$1" == "--tool" ]]
-  then
-    tool="$2"
-    shift
-  elif [[ "$1" =~ --tool=(.*) ]]
-  then
-    tool="${BASH_REMATCH[1]}"
-  fi
-  shift
-done
-
-NEEDS_VALGRIND=0
-
-case "$tool" in
-  "memcheck")
-    NEEDS_VALGRIND=1
-    ;;
-esac
-
-# For libyuv, we'll use the locate_valgrind.sh script in Chromium's Valgrind
-# scripts dir to locate the Valgrind framework install
-CHROME_VALGRIND_SCRIPTS=$THISDIR/../../tools/valgrind
-
-if [ "$NEEDS_VALGRIND" == "1" ]
-then
-  CHROME_VALGRIND=`sh $THISDIR/locate_valgrind.sh`
-  if [ "$CHROME_VALGRIND" = "" ]
-  then
-    CHROME_VALGRIND=../../src/third_party/valgrind/linux_x64
-    echo
-    echo "-------------------- WARNING ------------------------"
-    echo "locate_valgrind.sh failed."
-    echo "Using $CHROME_VALGRIND as a fallback location."
-    echo "This might be because:"
-    echo "1) This is a swarming bot"
-    echo "2) You haven't set up the valgrind binaries correctly."
-    echo "In this case, please make sure you have followed the instructions at"
-    echo "http://www.chromium.org/developers/how-tos/using-valgrind/get-valgrind"
-    echo "Notice: In the .gclient file, you need to add this for the 'libyuv'"
-    echo "solution since our directory structure is different from Chromium's:"
-    echo "\"custom_deps\": {"
-    echo "  \"libyuv/third_party/valgrind\":"
-    echo "      \"https://chromium.googlesource.com/chromium/deps/valgrind/binaries\","
-    echo "},"
-    echo "-----------------------------------------------------"
-    echo
-  fi
-  echo "Using valgrind binaries from ${CHROME_VALGRIND}"
-
-  PATH="${CHROME_VALGRIND}/bin:$PATH"
-  # We need to set these variables to override default lib paths hard-coded into
-  # Valgrind binary.
-  export VALGRIND_LIB="$CHROME_VALGRIND/lib/valgrind"
-  export VALGRIND_LIB_INNER="$CHROME_VALGRIND/lib/valgrind"
-
-  # Clean up some /tmp directories that might be stale due to interrupted
-  # chrome_tests.py execution.
-  # FYI:
-  #   -mtime +1  <- only print files modified more than 24h ago,
-  #   -print0/-0 are needed to handle possible newlines in the filenames.
-  echo "Cleanup /tmp from Valgrind stuff"
-  find /tmp -maxdepth 1 \(\
-        -name "vgdb-pipe-*" -or -name "vg_logs_*" -or -name "valgrind.*" \
-      \) -mtime +1 -print0 | xargs -0 rm -rf
-fi
-
-# Add Chrome's Valgrind scripts dir to the PYTHON_PATH since it contains
-# the scripts that are needed for this script to run
-PYTHONPATH=$THISDIR/../../tools/python/google:$CHROME_VALGRIND_SCRIPTS python \
-           "$THISDIR/libyuv_tests.py" $ARGV_COPY
diff --git a/files/tools_libyuv/valgrind/locate_valgrind.sh b/files/tools_libyuv/valgrind/locate_valgrind.sh
deleted file mode 100755
index d9594f48..00000000
--- a/files/tools_libyuv/valgrind/locate_valgrind.sh
+++ /dev/null
@@ -1,73 +0,0 @@
-#!/bin/bash
-
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# Prints a path to Valgrind binaries to be used for Chromium.
-# Select the valgrind from third_party/valgrind by default,
-# but allow users to override this default without editing scripts and
-# without specifying a commandline option
-
-export THISDIR=`dirname $0`
-
-# User may use their own valgrind by giving its path with CHROME_VALGRIND env.
-if [ "$CHROME_VALGRIND" = "" ]
-then
-  # Guess which binaries we should use by uname
-  case "$(uname -a)" in
-  *Linux*x86_64*)
-    PLATFORM="linux_x64"
-    ;;
-  *Linux*86*)
-    PLATFORM="linux_x86"
-    ;;
-  *Darwin*9.[678].[01]*i386*)
-    # Didn't test other kernels.
-    PLATFORM="mac"
-    ;;
-  *Darwin*10.[0-9].[0-9]*i386*)
-    PLATFORM="mac_10.6"
-    ;;
-  *Darwin*10.[0-9].[0-9]*x86_64*)
-    PLATFORM="mac_10.6"
-    ;;
-  *Darwin*11.[0-9].[0-9]*x86_64*)
-    PLATFORM="mac_10.7"
-    ;;
-  *)
-    (echo "Sorry, your platform is not supported:" &&
-     uname -a
-     echo
-     echo "If you're on Mac OS X, please see http://crbug.com/441425") >&2
-    exit 42
-  esac
-
-  # The binaries should be in third_party/valgrind
-  # (checked out from deps/third_party/valgrind/binaries).
-  CHROME_VALGRIND="$THISDIR/../../third_party/valgrind/$PLATFORM"
-
-  # TODO(timurrrr): readlink -f is not present on Mac...
-  if [ "$PLATFORM" != "mac" ] && \
-    [ "$PLATFORM" != "mac_10.6" ] && \
-    [ "$PLATFORM" != "mac_10.7" ]
-  then
-    # Get rid of all "../" dirs
-    CHROME_VALGRIND=$(readlink -f $CHROME_VALGRIND)
-  fi
-fi
-
-if ! test -x $CHROME_VALGRIND/bin/valgrind
-then
-  echo "Oops, could not find Valgrind binaries in your checkout." >&2
-  echo "Please see" >&2
-  echo "  http://dev.chromium.org/developers/how-tos/using-valgrind/get-valgrind" >&2
-  echo "for the instructions on how to download pre-built binaries." >&2
-  exit 1
-fi
-
-echo $CHROME_VALGRIND
diff --git a/files/tools_libyuv/valgrind/memcheck/OWNERS b/files/tools_libyuv/valgrind/memcheck/OWNERS
deleted file mode 100644
index 72e8ffc0..00000000
--- a/files/tools_libyuv/valgrind/memcheck/OWNERS
+++ /dev/null
@@ -1 +0,0 @@
-*
diff --git a/files/tools_libyuv/valgrind/memcheck/PRESUBMIT.py b/files/tools_libyuv/valgrind/memcheck/PRESUBMIT.py
deleted file mode 100644
index 03329214..00000000
--- a/files/tools_libyuv/valgrind/memcheck/PRESUBMIT.py
+++ /dev/null
@@ -1,99 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2012 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""
-Copied from Chrome's src/tools/valgrind/memcheck/PRESUBMIT.py
-
-See http://dev.chromium.org/developers/how-tos/depottools/presubmit-scripts
-for more details on the presubmit API built into gcl.
-"""
-
-import os
-import re
-import sys
-
-def CheckChange(input_api, output_api):
-  """Checks the memcheck suppressions files for bad data."""
-
-  # Add the path to the Chrome valgrind dir to the import path:
-  tools_vg_path = os.path.join(input_api.PresubmitLocalPath(), '..', '..', '..',
-                               'tools', 'valgrind')
-  sys.path.append(tools_vg_path)
-  import suppressions
-
-  sup_regex = re.compile('suppressions.*\.txt$')
-  suppressions = {}
-  errors = []
-  check_for_memcheck = False
-  # skip_next_line has 3 possible values:
-  # - False: don't skip the next line.
-  # - 'skip_suppression_name': the next line is a suppression name, skip.
-  # - 'skip_param': the next line is a system call parameter error, skip.
-  skip_next_line = False
-  for f in filter(lambda x: sup_regex.search(x.LocalPath()),
-                  input_api.AffectedFiles()):
-    for line, line_num in zip(f.NewContents(),
-                              xrange(1, len(f.NewContents()) + 1)):
-      line = line.lstrip()
-      if line.startswith('#') or not line:
-        continue
-
-      if skip_next_line:
-        if skip_next_line == 'skip_suppression_name':
-          if 'insert_a_suppression_name_here' in line:
-            errors.append('"insert_a_suppression_name_here" is not a valid '
-                          'suppression name')
-          if suppressions.has_key(line):
-            if f.LocalPath() == suppressions[line][1]:
-              errors.append('suppression with name "%s" at %s line %s '
-                            'has already been defined at line %s' %
-                            (line, f.LocalPath(), line_num,
-                             suppressions[line][1]))
-            else:
-              errors.append('suppression with name "%s" at %s line %s '
-                            'has already been defined at %s line %s' %
-                            (line, f.LocalPath(), line_num,
-                             suppressions[line][0], suppressions[line][1]))
-          else:
-            suppressions[line] = (f, line_num)
-            check_for_memcheck = True;
-        skip_next_line = False
-        continue
-      if check_for_memcheck:
-        if not line.startswith('Memcheck:'):
-          errors.append('"%s" should be "Memcheck:..." in %s line %s' %
-                        (line, f.LocalPath(), line_num))
-        check_for_memcheck = False;
-      if line == '{':
-        skip_next_line = 'skip_suppression_name'
-        continue
-      if line == "Memcheck:Param":
-        skip_next_line = 'skip_param'
-        continue
-
-      if (line.startswith('fun:') or line.startswith('obj:') or
-          line.startswith('Memcheck:') or line == '}' or
-          line == '...'):
-        continue
-      errors.append('"%s" is probably wrong: %s line %s' % (line, f.LocalPath(),
-                                                            line_num))
-  if errors:
-    return [output_api.PresubmitError('\n'.join(errors))]
-  return []
-
-def CheckChangeOnUpload(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def CheckChangeOnCommit(input_api, output_api):
-  return CheckChange(input_api, output_api)
-
-def GetPreferredTrySlaves():
-  # We don't have any memcheck slaves yet, so there's no use for this method.
-  # When we have, the slave name(s) should be put into this list.
-  return []
diff --git a/files/tools_libyuv/valgrind/memcheck/suppressions.txt b/files/tools_libyuv/valgrind/memcheck/suppressions.txt
deleted file mode 100644
index 3f0f6d44..00000000
--- a/files/tools_libyuv/valgrind/memcheck/suppressions.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-# There are two of suppressions in this file.
-# 1. third_party libraries
-# 2. libyuv stuff
-# 3. libjingle stuff (talk folder)
-#-----------------------------------------------------------------------
-
-# third_party libraries
-{
-   bug_729
-   Memcheck:Free
-   fun:_ZdaPv
-   ...
-   fun:_ZN7testing8internal12UnitTestImplD1Ev
-   ...
-}
-
-# libyuv (empty so far)
diff --git a/files/tools_libyuv/valgrind/memcheck/suppressions_mac.txt b/files/tools_libyuv/valgrind/memcheck/suppressions_mac.txt
deleted file mode 100644
index 3ad0c8cc..00000000
--- a/files/tools_libyuv/valgrind/memcheck/suppressions_mac.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/files/tools_libyuv/valgrind/memcheck/suppressions_win32.txt b/files/tools_libyuv/valgrind/memcheck/suppressions_win32.txt
deleted file mode 100644
index 3ad0c8cc..00000000
--- a/files/tools_libyuv/valgrind/memcheck/suppressions_win32.txt
+++ /dev/null
@@ -1,5 +0,0 @@
-# This file is used in addition to the one already maintained in Chrome.
-# It acts as a place holder for future additions for this project.
-# It must exist for the Python wrapper script to work properly.
-
-
diff --git a/files/tools_libyuv/valgrind/memcheck_analyze.py b/files/tools_libyuv/valgrind/memcheck_analyze.py
deleted file mode 100755
index 80e85eb4..00000000
--- a/files/tools_libyuv/valgrind/memcheck_analyze.py
+++ /dev/null
@@ -1,644 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# memcheck_analyze.py
-
-''' Given a valgrind XML file, parses errors and uniques them.'''
-
-import gdb_helper
-
-from collections import defaultdict
-import hashlib
-import logging
-import optparse
-import os
-import re
-import subprocess
-import sys
-import time
-from xml.dom.minidom import parse
-from xml.parsers.expat import ExpatError
-
-import common
-
-# Global symbol table (yuck)
-TheAddressTable = None
-
-# These are regexps that define functions (using C++ mangled names)
-# we don't want to see in stack traces while pretty printing
-# or generating suppressions.
-# Just stop printing the stack/suppression frames when the current one
-# matches any of these.
-_BORING_CALLERS = common.BoringCallers(mangled=True, use_re_wildcards=True)
-
-def getTextOf(top_node, name):
-  ''' Returns all text in all DOM nodes with a certain |name| that are children
-  of |top_node|.
-  '''
-
-  text = ""
-  for nodes_named in top_node.getElementsByTagName(name):
-    text += "".join([node.data for node in nodes_named.childNodes
-                     if node.nodeType == node.TEXT_NODE])
-  return text
-
-def getCDATAOf(top_node, name):
-  ''' Returns all CDATA in all DOM nodes with a certain |name| that are children
-  of |top_node|.
-  '''
-
-  text = ""
-  for nodes_named in top_node.getElementsByTagName(name):
-    text += "".join([node.data for node in nodes_named.childNodes
-                     if node.nodeType == node.CDATA_SECTION_NODE])
-  if (text == ""):
-    return None
-  return text
-
-def shortenFilePath(source_dir, directory):
-  '''Returns a string with the string prefix |source_dir| removed from
-  |directory|.'''
-  prefixes_to_cut = ["build/src/", "valgrind/coregrind/", "out/Release/../../"]
-
-  if source_dir:
-    prefixes_to_cut.append(source_dir)
-
-  for p in prefixes_to_cut:
-    index = directory.rfind(p)
-    if index != -1:
-      directory = directory[index + len(p):]
-
-  return directory
-
-# Constants that give real names to the abbreviations in valgrind XML output.
-INSTRUCTION_POINTER = "ip"
-OBJECT_FILE = "obj"
-FUNCTION_NAME = "fn"
-SRC_FILE_DIR = "dir"
-SRC_FILE_NAME = "file"
-SRC_LINE = "line"
-
-def gatherFrames(node, source_dir):
-  frames = []
-  for frame in node.getElementsByTagName("frame"):
-    frame_dict = {
-      INSTRUCTION_POINTER : getTextOf(frame, INSTRUCTION_POINTER),
-      OBJECT_FILE         : getTextOf(frame, OBJECT_FILE),
-      FUNCTION_NAME       : getTextOf(frame, FUNCTION_NAME),
-      SRC_FILE_DIR        : shortenFilePath(
-          source_dir, getTextOf(frame, SRC_FILE_DIR)),
-      SRC_FILE_NAME       : getTextOf(frame, SRC_FILE_NAME),
-      SRC_LINE            : getTextOf(frame, SRC_LINE)
-    }
-
-    # Ignore this frame and all the following if it's a "boring" function.
-    enough_frames = False
-    for regexp in _BORING_CALLERS:
-      if re.match("^%s$" % regexp, frame_dict[FUNCTION_NAME]):
-        enough_frames = True
-        break
-    if enough_frames:
-      break
-
-    frames += [frame_dict]
-
-    global TheAddressTable
-    if TheAddressTable != None and frame_dict[SRC_LINE] == "":
-      # Try using gdb
-      TheAddressTable.Add(frame_dict[OBJECT_FILE],
-                          frame_dict[INSTRUCTION_POINTER])
-  return frames
-
-class ValgrindError:
-  ''' Takes a <DOM Element: error> node and reads all the data from it. A
-  ValgrindError is immutable and is hashed on its pretty printed output.
-  '''
-
-  def __init__(self, source_dir, error_node, commandline, testcase):
-    ''' Copies all the relevant information out of the DOM and into object
-    properties.
-
-    Args:
-      error_node: The <error></error> DOM node we're extracting from.
-      source_dir: Prefix that should be stripped from the <dir> node.
-      commandline: The command that was run under valgrind
-      testcase: The test case name, if known.
-    '''
-
-    # Valgrind errors contain one <what><stack> pair, plus an optional
-    # <auxwhat><stack> pair, plus an optional <origin><what><stack></origin>,
-    # plus (since 3.5.0) a <suppression></suppression> pair.
-    # (Origin is nicely enclosed; too bad the other two aren't.)
-    # The most common way to see all three in one report is
-    # a syscall with a parameter that points to uninitialized memory, e.g.
-    # Format:
-    # <error>
-    #   <unique>0x6d</unique>
-    #   <tid>1</tid>
-    #   <kind>SyscallParam</kind>
-    #   <what>Syscall param write(buf) points to uninitialised byte(s)</what>
-    #   <stack>
-    #     <frame>
-    #     ...
-    #     </frame>
-    #   </stack>
-    #   <auxwhat>Address 0x5c9af4f is 7 bytes inside a block of ...</auxwhat>
-    #   <stack>
-    #     <frame>
-    #     ...
-    #     </frame>
-    #   </stack>
-    #   <origin>
-    #   <what>Uninitialised value was created by a heap allocation</what>
-    #   <stack>
-    #     <frame>
-    #     ...
-    #     </frame>
-    #   </stack>
-    #   </origin>
-    #   <suppression>
-    #     <sname>insert_a_suppression_name_here</sname>
-    #     <skind>Memcheck:Param</skind>
-    #     <skaux>write(buf)</skaux>
-    #     <sframe> <fun>__write_nocancel</fun> </sframe>
-    #     ...
-    #     <sframe> <fun>main</fun> </sframe>
-    #     <rawtext>
-    # <![CDATA[
-    # {
-    #    <insert_a_suppression_name_here>
-    #    Memcheck:Param
-    #    write(buf)
-    #    fun:__write_nocancel
-    #    ...
-    #    fun:main
-    # }
-    # ]]>
-    #     </rawtext>
-    #   </suppression>
-    # </error>
-    #
-    # Each frame looks like this:
-    #  <frame>
-    #    <ip>0x83751BC</ip>
-    #    <obj>/data/dkegel/chrome-build/src/out/Release/base_unittests</obj>
-    #    <fn>_ZN7testing8internal12TestInfoImpl7RunTestEPNS_8TestInfoE</fn>
-    #    <dir>/data/dkegel/chrome-build/src/testing/gtest/src</dir>
-    #    <file>gtest-internal-inl.h</file>
-    #    <line>655</line>
-    #  </frame>
-    # although the dir, file, and line elements are missing if there is
-    # no debug info.
-
-    self._kind = getTextOf(error_node, "kind")
-    self._backtraces = []
-    self._suppression = None
-    self._commandline = commandline
-    self._testcase = testcase
-    self._additional = []
-
-    # Iterate through the nodes, parsing <what|auxwhat><stack> pairs.
-    description = None
-    for node in error_node.childNodes:
-      if node.localName == "what" or node.localName == "auxwhat":
-        description = "".join([n.data for n in node.childNodes
-                              if n.nodeType == n.TEXT_NODE])
-      elif node.localName == "xwhat":
-        description = getTextOf(node, "text")
-      elif node.localName == "stack":
-        assert description
-        self._backtraces.append([description, gatherFrames(node, source_dir)])
-        description = None
-      elif node.localName == "origin":
-        description = getTextOf(node, "what")
-        stack = node.getElementsByTagName("stack")[0]
-        frames = gatherFrames(stack, source_dir)
-        self._backtraces.append([description, frames])
-        description = None
-        stack = None
-        frames = None
-      elif description and node.localName != None:
-        # The lastest description has no stack, e.g. "Address 0x28 is unknown"
-        self._additional.append(description)
-        description = None
-
-      if node.localName == "suppression":
-        self._suppression = getCDATAOf(node, "rawtext");
-
-  def __str__(self):
-    ''' Pretty print the type and backtrace(s) of this specific error,
-        including suppression (which is just a mangled backtrace).'''
-    output = ""
-    output += "\n" # Make sure the ### is at the beginning of line.
-    output += "### BEGIN MEMORY TOOL REPORT (error hash=#%016X#)\n" % \
-        self.ErrorHash()
-    if (self._commandline):
-      output += self._commandline + "\n"
-
-    output += self._kind + "\n"
-    for backtrace in self._backtraces:
-      output += backtrace[0] + "\n"
-      filter = subprocess.Popen("c++filt -n", stdin=subprocess.PIPE,
-                                stdout=subprocess.PIPE,
-                                stderr=subprocess.STDOUT,
-                                shell=True,
-                                close_fds=True)
-      buf = ""
-      for frame in backtrace[1]:
-        buf +=  (frame[FUNCTION_NAME] or frame[INSTRUCTION_POINTER]) + "\n"
-      (stdoutbuf, stderrbuf) = filter.communicate(buf.encode('latin-1'))
-      demangled_names = stdoutbuf.split("\n")
-
-      i = 0
-      for frame in backtrace[1]:
-        output += ("  " + demangled_names[i])
-        i = i + 1
-
-        global TheAddressTable
-        if TheAddressTable != None and frame[SRC_FILE_DIR] == "":
-           # Try using gdb
-           foo = TheAddressTable.GetFileLine(frame[OBJECT_FILE],
-                                             frame[INSTRUCTION_POINTER])
-           if foo[0] != None:
-             output += (" (" + foo[0] + ":" + foo[1] + ")")
-        elif frame[SRC_FILE_DIR] != "":
-          output += (" (" + frame[SRC_FILE_DIR] + "/" + frame[SRC_FILE_NAME] +
-                     ":" + frame[SRC_LINE] + ")")
-        else:
-          output += " (" + frame[OBJECT_FILE] + ")"
-        output += "\n"
-
-    for additional in self._additional:
-      output += additional + "\n"
-
-    assert self._suppression != None, "Your Valgrind doesn't generate " \
-                                      "suppressions - is it too old?"
-
-    if self._testcase:
-      output += "The report came from the `%s` test.\n" % self._testcase
-    output += "Suppression (error hash=#%016X#):\n" % self.ErrorHash()
-    output += ("  For more info on using suppressions see "
-               "http://dev.chromium.org/developers/tree-sheriffs/sheriff-details-chromium/memory-sheriff#TOC-Suppressing-memory-reports")
-
-    # Widen suppression slightly to make portable between mac and linux
-    # TODO(timurrrr): Oops, these transformations should happen
-    # BEFORE calculating the hash!
-    supp = self._suppression;
-    supp = supp.replace("fun:_Znwj", "fun:_Znw*")
-    supp = supp.replace("fun:_Znwm", "fun:_Znw*")
-    supp = supp.replace("fun:_Znaj", "fun:_Zna*")
-    supp = supp.replace("fun:_Znam", "fun:_Zna*")
-
-    # Make suppressions even less platform-dependent.
-    for sz in [1, 2, 4, 8]:
-      supp = supp.replace("Memcheck:Addr%d" % sz, "Memcheck:Unaddressable")
-      supp = supp.replace("Memcheck:Value%d" % sz, "Memcheck:Uninitialized")
-    supp = supp.replace("Memcheck:Cond", "Memcheck:Uninitialized")
-
-    # Split into lines so we can enforce length limits
-    supplines = supp.split("\n")
-    supp = None  # to avoid re-use
-
-    # Truncate at line 26 (VG_MAX_SUPP_CALLERS plus 2 for name and type)
-    # or at the first 'boring' caller.
-    # (https://bugs.kde.org/show_bug.cgi?id=199468 proposes raising
-    # VG_MAX_SUPP_CALLERS, but we're probably fine with it as is.)
-    newlen = min(26, len(supplines));
-
-    # Drop boring frames and all the following.
-    enough_frames = False
-    for frameno in range(newlen):
-      for boring_caller in _BORING_CALLERS:
-        if re.match("^ +fun:%s$" % boring_caller, supplines[frameno]):
-          newlen = frameno
-          enough_frames = True
-          break
-      if enough_frames:
-        break
-    if (len(supplines) > newlen):
-      supplines = supplines[0:newlen]
-      supplines.append("}")
-
-    for frame in range(len(supplines)):
-      # Replace the always-changing anonymous namespace prefix with "*".
-      m = re.match("( +fun:)_ZN.*_GLOBAL__N_.*\.cc_" +
-                   "[0-9a-fA-F]{8}_[0-9a-fA-F]{8}(.*)",
-                   supplines[frame])
-      if m:
-        supplines[frame] = "*".join(m.groups())
-
-    output += "\n".join(supplines) + "\n"
-    output += "### END MEMORY TOOL REPORT (error hash=#%016X#)\n" % \
-        self.ErrorHash()
-
-    return output
-
-  def UniqueString(self):
-    ''' String to use for object identity. Don't print this, use str(obj)
-    instead.'''
-    rep = self._kind + " "
-    for backtrace in self._backtraces:
-      for frame in backtrace[1]:
-        rep += frame[FUNCTION_NAME]
-
-        if frame[SRC_FILE_DIR] != "":
-          rep += frame[SRC_FILE_DIR] + "/" + frame[SRC_FILE_NAME]
-        else:
-          rep += frame[OBJECT_FILE]
-
-    return rep
-
-  # This is a device-independent hash identifying the suppression.
-  # By printing out this hash we can find duplicate reports between tests and
-  # different shards running on multiple buildbots
-  def ErrorHash(self):
-    return int(hashlib.md5(self.UniqueString()).hexdigest()[:16], 16)
-
-  def __hash__(self):
-    return hash(self.UniqueString())
-  def __eq__(self, rhs):
-    return self.UniqueString() == rhs
-
-def log_is_finished(f, force_finish):
-  f.seek(0)
-  prev_line = ""
-  while True:
-    line = f.readline()
-    if line == "":
-      if not force_finish:
-        return False
-      # Okay, the log is not finished but we can make it up to be parseable:
-      if prev_line.strip() in ["</error>", "</errorcounts>", "</status>"]:
-        f.write("</valgrindoutput>\n")
-        return True
-      return False
-    if '</valgrindoutput>' in line:
-      # Valgrind often has garbage after </valgrindoutput> upon crash.
-      f.truncate()
-      return True
-    prev_line = line
-
-class MemcheckAnalyzer:
-  ''' Given a set of Valgrind XML files, parse all the errors out of them,
-  unique them and output the results.'''
-
-  SANITY_TEST_SUPPRESSIONS = {
-      "Memcheck sanity test 01 (memory leak).": 1,
-      "Memcheck sanity test 02 (malloc/read left).": 1,
-      "Memcheck sanity test 03 (malloc/read right).": 1,
-      "Memcheck sanity test 04 (malloc/write left).": 1,
-      "Memcheck sanity test 05 (malloc/write right).": 1,
-      "Memcheck sanity test 06 (new/read left).": 1,
-      "Memcheck sanity test 07 (new/read right).": 1,
-      "Memcheck sanity test 08 (new/write left).": 1,
-      "Memcheck sanity test 09 (new/write right).": 1,
-      "Memcheck sanity test 10 (write after free).": 1,
-      "Memcheck sanity test 11 (write after delete).": 1,
-      "Memcheck sanity test 12 (array deleted without []).": 1,
-      "Memcheck sanity test 13 (single element deleted with []).": 1,
-      "Memcheck sanity test 14 (malloc/read uninit).": 1,
-      "Memcheck sanity test 15 (new/read uninit).": 1,
-  }
-
-  # Max time to wait for memcheck logs to complete.
-  LOG_COMPLETION_TIMEOUT = 180.0
-
-  def __init__(self, source_dir, show_all_leaks=False, use_gdb=False):
-    '''Create a parser for Memcheck logs.
-
-    Args:
-      source_dir: Path to top of source tree for this build
-      show_all_leaks: Whether to show even less important leaks
-      use_gdb: Whether to use gdb to resolve source filenames and line numbers
-               in the report stacktraces
-    '''
-    self._source_dir = source_dir
-    self._show_all_leaks = show_all_leaks
-    self._use_gdb = use_gdb
-
-    # Contains the set of unique errors
-    self._errors = set()
-
-    # Contains the time when the we started analyzing the first log file.
-    # This variable is used to skip incomplete logs after some timeout.
-    self._analyze_start_time = None
-
-
-  def Report(self, files, testcase, check_sanity=False):
-    '''Reads in a set of files and prints Memcheck report.
-
-    Args:
-      files: A list of filenames.
-      check_sanity: if true, search for SANITY_TEST_SUPPRESSIONS
-    '''
-    # Beyond the detailed errors parsed by ValgrindError above,
-    # the xml file contain records describing suppressions that were used:
-    # <suppcounts>
-    #  <pair>
-    #    <count>28</count>
-    #    <name>pango_font_leak_todo</name>
-    #  </pair>
-    #  <pair>
-    #    <count>378</count>
-    #    <name>bug_13243</name>
-    #  </pair>
-    # </suppcounts
-    # Collect these and print them at the end.
-    #
-    # With our patch for https://bugs.kde.org/show_bug.cgi?id=205000 in,
-    # the file also includes records of the form
-    # <load_obj><obj>/usr/lib/libgcc_s.1.dylib</obj><ip>0x27000</ip></load_obj>
-    # giving the filename and load address of each binary that was mapped
-    # into the process.
-
-    global TheAddressTable
-    if self._use_gdb:
-      TheAddressTable = gdb_helper.AddressTable()
-    else:
-      TheAddressTable = None
-    cur_report_errors = set()
-    suppcounts = defaultdict(int)
-    badfiles = set()
-
-    if self._analyze_start_time == None:
-      self._analyze_start_time = time.time()
-    start_time = self._analyze_start_time
-
-    parse_failed = False
-    for file in files:
-      # Wait up to three minutes for valgrind to finish writing all files,
-      # but after that, just skip incomplete files and warn.
-      f = open(file, "r+")
-      pid = re.match(".*\.([0-9]+)$", file)
-      if pid:
-        pid = pid.groups()[0]
-      found = False
-      running = True
-      firstrun = True
-      skip = False
-      origsize = os.path.getsize(file)
-      while (running and not found and not skip and
-             (firstrun or
-              ((time.time() - start_time) < self.LOG_COMPLETION_TIMEOUT))):
-        firstrun = False
-        f.seek(0)
-        if pid:
-          # Make sure the process is still running so we don't wait for
-          # 3 minutes if it was killed. See http://crbug.com/17453
-          ps_out = subprocess.Popen("ps p %s" % pid, shell=True,
-                                    stdout=subprocess.PIPE).stdout
-          if len(ps_out.readlines()) < 2:
-            running = False
-        else:
-          skip = True
-          running = False
-        found = log_is_finished(f, False)
-        if not running and not found:
-          logging.warn("Valgrind process PID = %s is not running but its "
-                       "XML log has not been finished correctly.\n"
-                       "Make it up by adding some closing tags manually." % pid)
-          found = log_is_finished(f, not running)
-        if running and not found:
-          time.sleep(1)
-      f.close()
-      if not found:
-        badfiles.add(file)
-      else:
-        newsize = os.path.getsize(file)
-        if origsize > newsize+1:
-          logging.warn(str(origsize - newsize) +
-                       " bytes of junk were after </valgrindoutput> in %s!" %
-                       file)
-        try:
-          parsed_file = parse(file);
-        except ExpatError, e:
-          parse_failed = True
-          logging.warn("could not parse %s: %s" % (file, e))
-          lineno = e.lineno - 1
-          context_lines = 5
-          context_start = max(0, lineno - context_lines)
-          context_end = lineno + context_lines + 1
-          context_file = open(file, "r")
-          for i in range(0, context_start):
-            context_file.readline()
-          for i in range(context_start, context_end):
-            context_data = context_file.readline().rstrip()
-            if i != lineno:
-              logging.warn("  %s" % context_data)
-            else:
-              logging.warn("> %s" % context_data)
-          context_file.close()
-          continue
-        if TheAddressTable != None:
-          load_objs = parsed_file.getElementsByTagName("load_obj")
-          for load_obj in load_objs:
-            obj = getTextOf(load_obj, "obj")
-            ip = getTextOf(load_obj, "ip")
-            TheAddressTable.AddBinaryAt(obj, ip)
-
-        commandline = None
-        preamble = parsed_file.getElementsByTagName("preamble")[0];
-        for node in preamble.getElementsByTagName("line"):
-          if node.localName == "line":
-            for x in node.childNodes:
-              if x.nodeType == node.TEXT_NODE and "Command" in x.data:
-                commandline = x.data
-                break
-
-        raw_errors = parsed_file.getElementsByTagName("error")
-        for raw_error in raw_errors:
-          # Ignore "possible" leaks for now by default.
-          if (self._show_all_leaks or
-              getTextOf(raw_error, "kind") != "Leak_PossiblyLost"):
-            error = ValgrindError(self._source_dir,
-                                  raw_error, commandline, testcase)
-            if error not in cur_report_errors:
-              # We haven't seen such errors doing this report yet...
-              if error in self._errors:
-                # ... but we saw it in earlier reports, e.g. previous UI test
-                cur_report_errors.add("This error was already printed in "
-                                      "some other test, see 'hash=#%016X#'" % \
-                                      error.ErrorHash())
-              else:
-                # ... and we haven't seen it in other tests as well
-                self._errors.add(error)
-                cur_report_errors.add(error)
-
-        suppcountlist = parsed_file.getElementsByTagName("suppcounts")
-        if len(suppcountlist) > 0:
-          suppcountlist = suppcountlist[0]
-          for node in suppcountlist.getElementsByTagName("pair"):
-            count = getTextOf(node, "count");
-            name = getTextOf(node, "name");
-            suppcounts[name] += int(count)
-
-    if len(badfiles) > 0:
-      logging.warn("valgrind didn't finish writing %d files?!" % len(badfiles))
-      for file in badfiles:
-        logging.warn("Last 20 lines of %s :" % file)
-        os.system("tail -n 20 '%s' 1>&2" % file)
-
-    if parse_failed:
-      logging.error("FAIL! Couldn't parse Valgrind output file")
-      return -2
-
-    common.PrintUsedSuppressionsList(suppcounts)
-
-    retcode = 0
-    if cur_report_errors:
-      logging.error("FAIL! There were %s errors: " % len(cur_report_errors))
-
-      if TheAddressTable != None:
-        TheAddressTable.ResolveAll()
-
-      for error in cur_report_errors:
-        logging.error(error)
-
-      retcode = -1
-
-    # Report tool's insanity even if there were errors.
-    if check_sanity:
-      remaining_sanity_supp = MemcheckAnalyzer.SANITY_TEST_SUPPRESSIONS
-      for (name, count) in suppcounts.iteritems():
-        # Workaround for http://crbug.com/334074
-        if (name in remaining_sanity_supp and
-            remaining_sanity_supp[name] <= count):
-          del remaining_sanity_supp[name]
-      if remaining_sanity_supp:
-        logging.error("FAIL! Sanity check failed!")
-        logging.info("The following test errors were not handled: ")
-        for (name, count) in remaining_sanity_supp.iteritems():
-          logging.info("  * %dx %s" % (count, name))
-        retcode = -3
-
-    if retcode != 0:
-      return retcode
-
-    logging.info("PASS! No errors found!")
-    return 0
-
-
-def _main():
-  '''For testing only. The MemcheckAnalyzer class should be imported instead.'''
-  parser = optparse.OptionParser("usage: %prog [options] <files to analyze>")
-  parser.add_option("", "--source-dir",
-                    help="path to top of source tree for this build"
-                    "(used to normalize source paths in baseline)")
-
-  (options, args) = parser.parse_args()
-  if len(args) == 0:
-    parser.error("no filename specified")
-  filenames = args
-
-  analyzer = MemcheckAnalyzer(options.source_dir, use_gdb=True)
-  return analyzer.Report(filenames, None)
-
-
-if __name__ == "__main__":
-  sys.exit(_main())
diff --git a/files/tools_libyuv/valgrind/valgrind.sh b/files/tools_libyuv/valgrind/valgrind.sh
deleted file mode 100755
index 7f3f7926..00000000
--- a/files/tools_libyuv/valgrind/valgrind.sh
+++ /dev/null
@@ -1,110 +0,0 @@
-#!/bin/bash
-
-# Copyright (c) 2017 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-# This is a small script for manually launching valgrind, along with passing
-# it the suppression file, and some helpful arguments (automatically attaching
-# the debugger on failures, etc).  Run it from your repo root, something like:
-#  $ sh ./tools/valgrind/valgrind.sh ./out/Debug/chrome
-#
-# This is mostly intended for running the chrome browser interactively.
-# To run unit tests, you probably want to run chrome_tests.sh instead.
-# That's the script used by the valgrind buildbot.
-
-export THISDIR=`dirname $0`
-
-setup_memcheck() {
-  RUN_COMMAND="valgrind"
-
-  # Prompt to attach gdb when there was an error detected.
-  DEFAULT_TOOL_FLAGS=("--db-command=gdb -nw %f %p" "--db-attach=yes" \
-                      # Keep the registers in gdb in sync with the code.
-                      "--vex-iropt-register-updates=allregs-at-mem-access" \
-                      # Overwrite newly allocated or freed objects
-                      # with 0x41 to catch inproper use.
-                      "--malloc-fill=41" "--free-fill=41" \
-                      # Increase the size of stacks being tracked.
-                      "--num-callers=30")
-}
-
-setup_unknown() {
-  echo "Unknown tool \"$TOOL_NAME\" specified, the result is not guaranteed"
-  DEFAULT_TOOL_FLAGS=()
-}
-
-set -e
-
-if [ $# -eq 0 ]; then
-  echo "usage: <command to run> <arguments ...>"
-  exit 1
-fi
-
-TOOL_NAME="memcheck"
-declare -a DEFAULT_TOOL_FLAGS[0]
-
-# Select a tool different from memcheck with --tool=TOOL as a first argument
-TMP_STR=`echo $1 | sed 's/^\-\-tool=//'`
-if [ "$TMP_STR" != "$1" ]; then
-  TOOL_NAME="$TMP_STR"
-  shift
-fi
-
-if echo "$@" | grep "\-\-tool" ; then
-  echo "--tool=TOOL must be the first argument" >&2
-  exit 1
-fi
-
-case $TOOL_NAME in
-  memcheck*)  setup_memcheck "$1";;
-  *)          setup_unknown;;
-esac
-
-
-SUPPRESSIONS="$THISDIR/$TOOL_NAME/suppressions.txt"
-
-CHROME_VALGRIND=`sh $THISDIR/locate_valgrind.sh`
-if [ "$CHROME_VALGRIND" = "" ]
-then
-  # locate_valgrind.sh failed
-  exit 1
-fi
-echo "Using valgrind binaries from ${CHROME_VALGRIND}"
-
-set -x
-PATH="${CHROME_VALGRIND}/bin:$PATH"
-# We need to set these variables to override default lib paths hard-coded into
-# Valgrind binary.
-export VALGRIND_LIB="$CHROME_VALGRIND/lib/valgrind"
-export VALGRIND_LIB_INNER="$CHROME_VALGRIND/lib/valgrind"
-
-# G_SLICE=always-malloc: make glib use system malloc
-# NSS_DISABLE_UNLOAD=1: make nss skip dlclosing dynamically loaded modules,
-# which would result in "obj:*" in backtraces.
-# NSS_DISABLE_ARENA_FREE_LIST=1: make nss use system malloc
-# G_DEBUG=fatal_warnings: make  GTK abort on any critical or warning assertions.
-# If it crashes on you in the Options menu, you hit bug 19751,
-# comment out the G_DEBUG=fatal_warnings line.
-#
-# GTEST_DEATH_TEST_USE_FORK=1: make gtest death tests valgrind-friendly
-#
-# When everyone has the latest valgrind, we might want to add
-#  --show-possibly-lost=no
-# to ignore possible but not definite leaks.
-
-G_SLICE=always-malloc \
-NSS_DISABLE_UNLOAD=1 \
-NSS_DISABLE_ARENA_FREE_LIST=1 \
-G_DEBUG=fatal_warnings \
-GTEST_DEATH_TEST_USE_FORK=1 \
-$RUN_COMMAND \
-  --trace-children=yes \
-  --leak-check=yes \
-  --suppressions="$SUPPRESSIONS" \
-  "${DEFAULT_TOOL_FLAGS[@]}" \
-  "$@"
diff --git a/files/tools_libyuv/valgrind/valgrind_test.py b/files/tools_libyuv/valgrind/valgrind_test.py
deleted file mode 100755
index 0fd3d97f..00000000
--- a/files/tools_libyuv/valgrind/valgrind_test.py
+++ /dev/null
@@ -1,517 +0,0 @@
-#!/usr/bin/env python
-# Copyright (c) 2017 The WebRTC project authors. All Rights Reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS.  All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""Runs an exe through Valgrind and puts the intermediate files in a
-directory.
-"""
-
-import datetime
-import glob
-import logging
-import optparse
-import os
-import re
-import shutil
-import stat
-import subprocess
-import sys
-import tempfile
-
-import common
-
-import memcheck_analyze
-
-class BaseTool(object):
-  """Abstract class for running dynamic error detection tools.
-
-  Always subclass this and implement ToolCommand with framework- and
-  tool-specific stuff.
-  """
-
-  def __init__(self):
-    temp_parent_dir = None
-    self.log_parent_dir = ""
-    if common.IsWindows():
-      # gpu process on Windows Vista+ runs at Low Integrity and can only
-      # write to certain directories (http://crbug.com/119131)
-      #
-      # TODO(bruening): if scripts die in middle and don't clean up temp
-      # dir, we'll accumulate files in profile dir.  should remove
-      # really old files automatically.
-      profile = os.getenv("USERPROFILE")
-      if profile:
-        self.log_parent_dir = profile + "\\AppData\\LocalLow\\"
-        if os.path.exists(self.log_parent_dir):
-          self.log_parent_dir = common.NormalizeWindowsPath(self.log_parent_dir)
-          temp_parent_dir = self.log_parent_dir
-    # Generated every time (even when overridden)
-    self.temp_dir = tempfile.mkdtemp(prefix="vg_logs_", dir=temp_parent_dir)
-    self.log_dir = self.temp_dir # overridable by --keep_logs
-    self.option_parser_hooks = []
-    # TODO(glider): we may not need some of the env vars on some of the
-    # platforms.
-    self._env = {
-      "G_SLICE" : "always-malloc",
-      "NSS_DISABLE_UNLOAD" : "1",
-      "NSS_DISABLE_ARENA_FREE_LIST" : "1",
-      "GTEST_DEATH_TEST_USE_FORK": "1",
-    }
-
-  def ToolName(self):
-    raise NotImplementedError, "This method should be implemented " \
-                               "in the tool-specific subclass"
-
-  def Analyze(self, check_sanity=False):
-    raise NotImplementedError, "This method should be implemented " \
-                               "in the tool-specific subclass"
-
-  def RegisterOptionParserHook(self, hook):
-    # Frameworks and tools can add their own flags to the parser.
-    self.option_parser_hooks.append(hook)
-
-  def CreateOptionParser(self):
-    # Defines Chromium-specific flags.
-    self._parser = optparse.OptionParser("usage: %prog [options] <program to "
-                                         "test>")
-    self._parser.disable_interspersed_args()
-    self._parser.add_option("-t", "--timeout",
-                      dest="timeout", metavar="TIMEOUT", default=10000,
-                      help="timeout in seconds for the run (default 10000)")
-    self._parser.add_option("", "--build-dir",
-                            help="the location of the compiler output")
-    self._parser.add_option("", "--source-dir",
-                            help="path to top of source tree for this build"
-                                 "(used to normalize source paths in baseline)")
-    self._parser.add_option("", "--gtest_filter", default="",
-                            help="which test case to run")
-    self._parser.add_option("", "--gtest_repeat",
-                            help="how many times to run each test")
-    self._parser.add_option("", "--gtest_print_time", action="store_true",
-                            default=False,
-                            help="show how long each test takes")
-    self._parser.add_option("", "--ignore_exit_code", action="store_true",
-                            default=False,
-                            help="ignore exit code of the test "
-                                 "(e.g. test failures)")
-    self._parser.add_option("", "--keep_logs", action="store_true",
-                            default=False,
-                            help="store memory tool logs in the <tool>.logs "
-                                 "directory instead of /tmp.\nThis can be "
-                                 "useful for tool developers/maintainers.\n"
-                                 "Please note that the <tool>.logs directory "
-                                 "will be clobbered on tool startup.")
-
-    # To add framework- or tool-specific flags, please add a hook using
-    # RegisterOptionParserHook in the corresponding subclass.
-    # See ValgrindTool for an example.
-    for hook in self.option_parser_hooks:
-      hook(self, self._parser)
-
-  def ParseArgv(self, args):
-    self.CreateOptionParser()
-
-    # self._tool_flags will store those tool flags which we don't parse
-    # manually in this script.
-    self._tool_flags = []
-    known_args = []
-
-    """ We assume that the first argument not starting with "-" is a program
-    name and all the following flags should be passed to the program.
-    TODO(timurrrr): customize optparse instead
-    """
-    while len(args) > 0 and args[0][:1] == "-":
-      arg = args[0]
-      if (arg == "--"):
-        break
-      if self._parser.has_option(arg.split("=")[0]):
-        known_args += [arg]
-      else:
-        self._tool_flags += [arg]
-      args = args[1:]
-
-    if len(args) > 0:
-      known_args += args
-
-    self._options, self._args = self._parser.parse_args(known_args)
-
-    self._timeout = int(self._options.timeout)
-    self._source_dir = self._options.source_dir
-    if self._options.keep_logs:
-      # log_parent_dir has trailing slash if non-empty
-      self.log_dir = self.log_parent_dir + "%s.logs" % self.ToolName()
-      if os.path.exists(self.log_dir):
-        shutil.rmtree(self.log_dir)
-      os.mkdir(self.log_dir)
-      logging.info("Logs are in " + self.log_dir)
-
-    self._ignore_exit_code = self._options.ignore_exit_code
-    if self._options.gtest_filter != "":
-      self._args.append("--gtest_filter=%s" % self._options.gtest_filter)
-    if self._options.gtest_repeat:
-      self._args.append("--gtest_repeat=%s" % self._options.gtest_repeat)
-    if self._options.gtest_print_time:
-      self._args.append("--gtest_print_time")
-
-    return True
-
-  def Setup(self, args):
-    return self.ParseArgv(args)
-
-  def ToolCommand(self):
-    raise NotImplementedError, "This method should be implemented " \
-                               "in the tool-specific subclass"
-
-  def Cleanup(self):
-    # You may override it in the tool-specific subclass
-    pass
-
-  def Execute(self):
-    """ Execute the app to be tested after successful instrumentation.
-    Full execution command-line provided by subclassers via proc."""
-    logging.info("starting execution...")
-    proc = self.ToolCommand()
-    for var in self._env:
-      common.PutEnvAndLog(var, self._env[var])
-    return common.RunSubprocess(proc, self._timeout)
-
-  def RunTestsAndAnalyze(self, check_sanity):
-    exec_retcode = self.Execute()
-    analyze_retcode = self.Analyze(check_sanity)
-
-    if analyze_retcode:
-      logging.error("Analyze failed.")
-      logging.info("Search the log for '[ERROR]' to see the error reports.")
-      return analyze_retcode
-
-    if exec_retcode:
-      if self._ignore_exit_code:
-        logging.info("Test execution failed, but the exit code is ignored.")
-      else:
-        logging.error("Test execution failed.")
-        return exec_retcode
-    else:
-      logging.info("Test execution completed successfully.")
-
-    if not analyze_retcode:
-      logging.info("Analysis completed successfully.")
-
-    return 0
-
-  def Main(self, args, check_sanity, min_runtime_in_seconds):
-    """Call this to run through the whole process: Setup, Execute, Analyze"""
-    start_time = datetime.datetime.now()
-    retcode = -1
-    if self.Setup(args):
-      retcode = self.RunTestsAndAnalyze(check_sanity)
-      shutil.rmtree(self.temp_dir, ignore_errors=True)
-      self.Cleanup()
-    else:
-      logging.error("Setup failed")
-    end_time = datetime.datetime.now()
-    runtime_in_seconds = (end_time - start_time).seconds
-    hours = runtime_in_seconds / 3600
-    seconds = runtime_in_seconds % 3600
-    minutes = seconds / 60
-    seconds = seconds % 60
-    logging.info("elapsed time: %02d:%02d:%02d" % (hours, minutes, seconds))
-    if (min_runtime_in_seconds > 0 and
-        runtime_in_seconds < min_runtime_in_seconds):
-      logging.error("Layout tests finished too quickly. "
-                    "It should have taken at least %d seconds. "
-                    "Something went wrong?" % min_runtime_in_seconds)
-      retcode = -1
-    return retcode
-
-  def Run(self, args, module, min_runtime_in_seconds=0):
-    MODULES_TO_SANITY_CHECK = ["base"]
-
-    check_sanity = module in MODULES_TO_SANITY_CHECK
-    return self.Main(args, check_sanity, min_runtime_in_seconds)
-
-
-class ValgrindTool(BaseTool):
-  """Abstract class for running Valgrind tools.
-
-  Always subclass this and implement ToolSpecificFlags() and
-  ExtendOptionParser() for tool-specific stuff.
-  """
-  def __init__(self):
-    super(ValgrindTool, self).__init__()
-    self.RegisterOptionParserHook(ValgrindTool.ExtendOptionParser)
-
-  def UseXML(self):
-    # Override if tool prefers nonxml output
-    return True
-
-  def ExtendOptionParser(self, parser):
-    parser.add_option("", "--suppressions", default=[],
-                            action="append",
-                            help="path to a valgrind suppression file")
-    parser.add_option("", "--indirect", action="store_true",
-                            default=False,
-                            help="set BROWSER_WRAPPER rather than "
-                                 "running valgrind directly")
-    parser.add_option("", "--indirect_webkit_layout", action="store_true",
-                            default=False,
-                            help="set --wrapper rather than running Dr. Memory "
-                                 "directly.")
-    parser.add_option("", "--trace_children", action="store_true",
-                            default=False,
-                            help="also trace child processes")
-    parser.add_option("", "--num-callers",
-                            dest="num_callers", default=30,
-                            help="number of callers to show in stack traces")
-    parser.add_option("", "--generate_dsym", action="store_true",
-                          default=False,
-                          help="Generate .dSYM file on Mac if needed. Slow!")
-
-  def Setup(self, args):
-    if not BaseTool.Setup(self, args):
-      return False
-    return True
-
-  def ToolCommand(self):
-    """Get the valgrind command to run."""
-    # Note that self._args begins with the exe to be run.
-    tool_name = self.ToolName()
-
-    # Construct the valgrind command.
-    if 'CHROME_VALGRIND' in os.environ:
-      path = os.path.join(os.environ['CHROME_VALGRIND'], "bin", "valgrind")
-    else:
-      path = "valgrind"
-    proc = [path, "--tool=%s" % tool_name]
-
-    proc += ["--num-callers=%i" % int(self._options.num_callers)]
-
-    if self._options.trace_children:
-      proc += ["--trace-children=yes"]
-      proc += ["--trace-children-skip='*dbus-daemon*'"]
-      proc += ["--trace-children-skip='*dbus-launch*'"]
-      proc += ["--trace-children-skip='*perl*'"]
-      proc += ["--trace-children-skip='*python*'"]
-      # This is really Python, but for some reason Valgrind follows it.
-      proc += ["--trace-children-skip='*lsb_release*'"]
-
-    proc += self.ToolSpecificFlags()
-    proc += self._tool_flags
-
-    suppression_count = 0
-    for suppression_file in self._options.suppressions:
-      if os.path.exists(suppression_file):
-        suppression_count += 1
-        proc += ["--suppressions=%s" % suppression_file]
-
-    if not suppression_count:
-      logging.warning("WARNING: NOT USING SUPPRESSIONS!")
-
-    logfilename = self.log_dir + ("/%s." % tool_name) + "%p"
-    if self.UseXML():
-      proc += ["--xml=yes", "--xml-file=" + logfilename]
-    else:
-      proc += ["--log-file=" + logfilename]
-
-    # The Valgrind command is constructed.
-
-    # Handle --indirect_webkit_layout separately.
-    if self._options.indirect_webkit_layout:
-      # Need to create the wrapper before modifying |proc|.
-      wrapper = self.CreateBrowserWrapper(proc, webkit=True)
-      proc = self._args
-      proc.append("--wrapper")
-      proc.append(wrapper)
-      return proc
-
-    if self._options.indirect:
-      wrapper = self.CreateBrowserWrapper(proc)
-      os.environ["BROWSER_WRAPPER"] = wrapper
-      logging.info('export BROWSER_WRAPPER=' + wrapper)
-      proc = []
-    proc += self._args
-    return proc
-
-  def ToolSpecificFlags(self):
-    raise NotImplementedError, "This method should be implemented " \
-                               "in the tool-specific subclass"
-
-  def CreateBrowserWrapper(self, proc, webkit=False):
-    """The program being run invokes Python or something else that can't stand
-    to be valgrinded, and also invokes the Chrome browser. In this case, use a
-    magic wrapper to only valgrind the Chrome browser. Build the wrapper here.
-    Returns the path to the wrapper. It's up to the caller to use the wrapper
-    appropriately.
-    """
-    command = " ".join(proc)
-    # Add the PID of the browser wrapper to the logfile names so we can
-    # separate log files for different UI tests at the analyze stage.
-    command = command.replace("%p", "$$.%p")
-
-    (fd, indirect_fname) = tempfile.mkstemp(dir=self.log_dir,
-                                            prefix="browser_wrapper.",
-                                            text=True)
-    f = os.fdopen(fd, "w")
-    f.write('#!/bin/bash\n'
-            'echo "Started Valgrind wrapper for this test, PID=$$" >&2\n')
-
-    f.write('DIR=`dirname $0`\n'
-            'TESTNAME_FILE=$DIR/testcase.$$.name\n\n')
-
-    if webkit:
-      # Webkit layout_tests pass the URL as the first line of stdin.
-      f.write('tee $TESTNAME_FILE | %s "$@"\n' % command)
-    else:
-      # Try to get the test case name by looking at the program arguments.
-      # i.e. Chromium ui_tests used --test-name arg.
-      # TODO(timurrrr): This doesn't handle "--test-name Test.Name"
-      # TODO(timurrrr): ui_tests are dead. Where do we use the non-webkit
-      # wrapper now? browser_tests? What do they do?
-      f.write('for arg in $@\ndo\n'
-              '  if [[ "$arg" =~ --test-name=(.*) ]]\n  then\n'
-              '    echo ${BASH_REMATCH[1]} >$TESTNAME_FILE\n'
-              '  fi\n'
-              'done\n\n'
-              '%s "$@"\n' % command)
-
-    f.close()
-    os.chmod(indirect_fname, stat.S_IRUSR|stat.S_IXUSR)
-    return indirect_fname
-
-  def CreateAnalyzer(self):
-    raise NotImplementedError, "This method should be implemented " \
-                               "in the tool-specific subclass"
-
-  def GetAnalyzeResults(self, check_sanity=False):
-    # Glob all the files in the log directory
-    filenames = glob.glob(self.log_dir + "/" + self.ToolName() + ".*")
-
-    # If we have browser wrapper, the logfiles are named as
-    # "toolname.wrapper_PID.valgrind_PID".
-    # Let's extract the list of wrapper_PIDs and name it ppids
-    ppids = set([int(f.split(".")[-2]) \
-                for f in filenames if re.search("\.[0-9]+\.[0-9]+$", f)])
-
-    analyzer = self.CreateAnalyzer()
-    if len(ppids) == 0:
-      # Fast path - no browser wrapper was set.
-      return analyzer.Report(filenames, None, check_sanity)
-
-    ret = 0
-    for ppid in ppids:
-      testcase_name = None
-      try:
-        f = open(self.log_dir + ("/testcase.%d.name" % ppid))
-        testcase_name = f.read().strip()
-        f.close()
-        wk_layout_prefix="third_party/WebKit/LayoutTests/"
-        wk_prefix_at = testcase_name.rfind(wk_layout_prefix)
-        if wk_prefix_at != -1:
-          testcase_name = testcase_name[wk_prefix_at + len(wk_layout_prefix):]
-      except IOError:
-        pass
-      print "====================================================="
-      print " Below is the report for valgrind wrapper PID=%d." % ppid
-      if testcase_name:
-        print " It was used while running the `%s` test." % testcase_name
-      else:
-        print " You can find the corresponding test"
-        print " by searching the above log for 'PID=%d'" % ppid
-      sys.stdout.flush()
-
-      ppid_filenames = [f for f in filenames \
-                        if re.search("\.%d\.[0-9]+$" % ppid, f)]
-      # check_sanity won't work with browser wrappers
-      assert check_sanity == False
-      ret |= analyzer.Report(ppid_filenames, testcase_name)
-      print "====================================================="
-      sys.stdout.flush()
-
-    if ret != 0:
-      print ""
-      print "The Valgrind reports are grouped by test names."
-      print "Each test has its PID printed in the log when the test was run"
-      print "and at the beginning of its Valgrind report."
-      print "Hint: you can search for the reports by Ctrl+F -> `=#`"
-      sys.stdout.flush()
-
-    return ret
-
-
-# TODO(timurrrr): Split into a separate file.
-class Memcheck(ValgrindTool):
-  """Memcheck
-  Dynamic memory error detector for Linux & Mac
-
-  http://valgrind.org/info/tools.html#memcheck
-  """
-
-  def __init__(self):
-    super(Memcheck, self).__init__()
-    self.RegisterOptionParserHook(Memcheck.ExtendOptionParser)
-
-  def ToolName(self):
-    return "memcheck"
-
-  def ExtendOptionParser(self, parser):
-    parser.add_option("--leak-check", "--leak_check", type="string",
-                      default="yes",  # --leak-check=yes is equivalent of =full
-                      help="perform leak checking at the end of the run")
-    parser.add_option("", "--show_all_leaks", action="store_true",
-                      default=False,
-                      help="also show less blatant leaks")
-    parser.add_option("", "--track_origins", action="store_true",
-                      default=False,
-                      help="Show whence uninitialized bytes came. 30% slower.")
-
-  def ToolSpecificFlags(self):
-    ret = ["--gen-suppressions=all", "--demangle=no"]
-    ret += ["--leak-check=%s" % self._options.leak_check]
-
-    if self._options.show_all_leaks:
-      ret += ["--show-reachable=yes"]
-    else:
-      ret += ["--show-possibly-lost=no"]
-
-    if self._options.track_origins:
-      ret += ["--track-origins=yes"]
-
-    # TODO(glider): this is a temporary workaround for http://crbug.com/51716
-    # Let's see whether it helps.
-    if common.IsMac():
-      ret += ["--smc-check=all"]
-
-    return ret
-
-  def CreateAnalyzer(self):
-    use_gdb = common.IsMac()
-    return memcheck_analyze.MemcheckAnalyzer(self._source_dir,
-                                            self._options.show_all_leaks,
-                                            use_gdb=use_gdb)
-
-  def Analyze(self, check_sanity=False):
-    ret = self.GetAnalyzeResults(check_sanity)
-
-    if ret != 0:
-      logging.info("Please see http://dev.chromium.org/developers/how-tos/"
-                   "using-valgrind for the info on Memcheck/Valgrind")
-    return ret
-
-
-class ToolFactory:
-  def Create(self, tool_name):
-    if tool_name == "memcheck":
-      return Memcheck()
-    try:
-      platform_name = common.PlatformNames()[0]
-    except common.NotImplementedError:
-      platform_name = sys.platform + "(Unknown)"
-    raise RuntimeError, "Unknown tool (tool=%s, platform=%s)" % (tool_name,
-                                                                 platform_name)
-
-def CreateTool(tool):
-  return ToolFactory().Create(tool)
diff --git a/files/unit_test/color_test.cc b/files/unit_test/color_test.cc
index 4bb448d5..01267ff1 100644
--- a/files/unit_test/color_test.cc
+++ b/files/unit_test/color_test.cc
@@ -20,20 +20,22 @@
 
 namespace libyuv {
 
-// TODO(fbarchard): Port high accuracy YUV to RGB to Neon.
-#if !defined(LIBYUV_DISABLE_NEON) && \
-    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
-#define ERROR_R 1
-#define ERROR_G 1
-#define ERROR_B 3
-#define ERROR_FULL 6
-#define ERROR_J420 5
+// TODO(fbarchard): clang x86 has a higher accuracy YUV to RGB.
+// Port to Visual C and other CPUs
+#if !defined(LIBYUV_BIT_EXACT) && !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || defined(__i386__))
+#define ERROR_FULL 5
+#define ERROR_J420 4
 #else
+#define ERROR_FULL 6
+#define ERROR_J420 6
+#endif
 #define ERROR_R 1
 #define ERROR_G 1
-#define ERROR_B 3
-#define ERROR_FULL 5
-#define ERROR_J420 3
+#ifdef LIBYUV_UNLIMITED_DATA
+#define ERROR_B 1
+#else
+#define ERROR_B 18
 #endif
 
 #define TESTCS(TESTNAME, YUVTOARGB, ARGBTOYUV, HS1, HS, HN, DIFF)              \
@@ -187,6 +189,104 @@ static void YUVJToRGB(int y, int u, int v, int* r, int* g, int* b) {
   *r = orig_pixels[2];
 }
 
+static void YUVHToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  H422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+#define F422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+
+static void YUVFToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  F422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+static void YUVUToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  U422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
+#define V422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+
+static void YUVVToRGB(int y, int u, int v, int* r, int* g, int* b) {
+  const int kWidth = 16;
+  const int kHeight = 1;
+  const int kPixels = kWidth * kHeight;
+  const int kHalfPixels = ((kWidth + 1) / 2) * ((kHeight + 1) / 2);
+
+  SIMD_ALIGNED(uint8_t orig_y[16]);
+  SIMD_ALIGNED(uint8_t orig_u[8]);
+  SIMD_ALIGNED(uint8_t orig_v[8]);
+  SIMD_ALIGNED(uint8_t orig_pixels[16 * 4]);
+  memset(orig_y, y, kPixels);
+  memset(orig_u, u, kHalfPixels);
+  memset(orig_v, v, kHalfPixels);
+
+  /* YUV converted to ARGB. */
+  V422ToARGB(orig_y, kWidth, orig_u, (kWidth + 1) / 2, orig_v, (kWidth + 1) / 2,
+             orig_pixels, kWidth * 4, kWidth, kHeight);
+
+  *b = orig_pixels[0];
+  *g = orig_pixels[1];
+  *r = orig_pixels[2];
+}
+
 static void YToRGB(int y, int* r, int* g, int* b) {
   const int kWidth = 16;
   const int kHeight = 1;
@@ -335,18 +435,50 @@ TEST_F(LibYUVColorTest, TestRoundToByte) {
   EXPECT_LE(allb, 255);
 }
 
+// BT.601 limited range YUV to RGB reference
 static void YUVToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
   *r = RoundToByte((y - 16) * 1.164 - (v - 128) * -1.596);
   *g = RoundToByte((y - 16) * 1.164 - (u - 128) * 0.391 - (v - 128) * 0.813);
   *b = RoundToByte((y - 16) * 1.164 - (u - 128) * -2.018);
 }
 
+// BT.601 full range YUV to RGB reference (aka JPEG)
 static void YUVJToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
   *r = RoundToByte(y - (v - 128) * -1.40200);
   *g = RoundToByte(y - (u - 128) * 0.34414 - (v - 128) * 0.71414);
   *b = RoundToByte(y - (u - 128) * -1.77200);
 }
 
+// BT.709 limited range YUV to RGB reference
+// See also http://www.equasys.de/colorconversion.html
+static void YUVHToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte((y - 16) * 1.164 - (v - 128) * -1.793);
+  *g = RoundToByte((y - 16) * 1.164 - (u - 128) * 0.213 - (v - 128) * 0.533);
+  *b = RoundToByte((y - 16) * 1.164 - (u - 128) * -2.112);
+}
+
+// BT.709 full range YUV to RGB reference
+static void YUVFToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte(y - (v - 128) * -1.5748);
+  *g = RoundToByte(y - (u - 128) * 0.18732 - (v - 128) * 0.46812);
+  *b = RoundToByte(y - (u - 128) * -1.8556);
+}
+
+// BT.2020 limited range YUV to RGB reference
+static void YUVUToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte((y - 16) * 1.164384 - (v - 128) * -1.67867);
+  *g = RoundToByte((y - 16) * 1.164384 - (u - 128) * 0.187326 -
+                   (v - 128) * 0.65042);
+  *b = RoundToByte((y - 16) * 1.164384 - (u - 128) * -2.14177);
+}
+
+// BT.2020 full range YUV to RGB reference
+static void YUVVToRGBReference(int y, int u, int v, int* r, int* g, int* b) {
+  *r = RoundToByte(y + (v - 128) * 1.474600);
+  *g = RoundToByte(y - (u - 128) * 0.164553 - (v - 128) * 0.571353);
+  *b = RoundToByte(y + (u - 128) * 1.881400);
+}
+
 TEST_F(LibYUVColorTest, TestYUV) {
   int r0, g0, b0, r1, g1, b1;
 
@@ -370,7 +502,11 @@ TEST_F(LibYUVColorTest, TestYUV) {
   YUVToRGB(240, 0, 0, &r1, &g1, &b1);
   EXPECT_EQ(57, r1);
   EXPECT_EQ(255, g1);
+#ifdef LIBYUV_UNLIMITED_DATA
+  EXPECT_EQ(3, b1);
+#else
   EXPECT_EQ(5, b1);
+#endif
 
   for (int i = 0; i < 256; ++i) {
     YUVToRGBReference(i, 128, 128, &r0, &g0, &b0);
@@ -444,28 +580,28 @@ TEST_F(LibYUVColorTest, TestGreyYUV) {
 
 static void PrintHistogram(int rh[256], int gh[256], int bh[256]) {
   int i;
-  printf("hist");
+  printf("hist ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", i - 128);
+      printf(" %8d", i - 128);
     }
   }
-  printf("\nred");
+  printf("\nred  ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", rh[i]);
+      printf(" %8d", rh[i]);
     }
   }
   printf("\ngreen");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", gh[i]);
+      printf(" %8d", gh[i]);
     }
   }
-  printf("\nblue");
+  printf("\nblue ");
   for (i = 0; i < 256; ++i) {
     if (rh[i] || gh[i] || bh[i]) {
-      printf("\t%8d", bh[i]);
+      printf(" %8d", bh[i]);
     }
   }
   printf("\n");
@@ -473,7 +609,13 @@ static void PrintHistogram(int rh[256], int gh[256], int bh[256]) {
 
 // Step by 5 on inner loop goes from 0 to 255 inclusive.
 // Set to 1 for better converage.  3, 5 or 17 for faster testing.
+#ifdef DISABLE_SLOW_TESTS
 #define FASTSTEP 5
+#else
+#define FASTSTEP 1
+#endif
+
+// BT.601 limited range.
 TEST_F(LibYUVColorTest, TestFullYUV) {
   int rh[256] = {
       0,
@@ -503,6 +645,7 @@ TEST_F(LibYUVColorTest, TestFullYUV) {
   PrintHistogram(rh, gh, bh);
 }
 
+// BT.601 full range.
 TEST_F(LibYUVColorTest, TestFullYUVJ) {
   int rh[256] = {
       0,
@@ -520,9 +663,129 @@ TEST_F(LibYUVColorTest, TestFullYUVJ) {
         int y = RANDOM256(y2);
         YUVJToRGBReference(y, u, v, &r0, &g0, &b0);
         YUVJToRGB(y, u, v, &r1, &g1, &b1);
-        EXPECT_NEAR(r0, r1, 1);
-        EXPECT_NEAR(g0, g1, 1);
-        EXPECT_NEAR(b0, b1, 1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.709 limited range.
+TEST_F(LibYUVColorTest, TestFullYUVH) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVHToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVHToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.709 full range.
+TEST_F(LibYUVColorTest, TestFullYUVF) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVFToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVFToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.2020 limited range.
+TEST_F(LibYUVColorTest, TestFullYUVU) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVUToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVUToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, ERROR_G);
+        EXPECT_NEAR(b0, b1, ERROR_B);
+        ++rh[r1 - r0 + 128];
+        ++gh[g1 - g0 + 128];
+        ++bh[b1 - b0 + 128];
+      }
+    }
+  }
+  PrintHistogram(rh, gh, bh);
+}
+
+// BT.2020 full range.
+TEST_F(LibYUVColorTest, TestFullYUVV) {
+  int rh[256] = {
+      0,
+  };
+  int gh[256] = {
+      0,
+  };
+  int bh[256] = {
+      0,
+  };
+  for (int u = 0; u < 256; ++u) {
+    for (int v = 0; v < 256; ++v) {
+      for (int y2 = 0; y2 < 256; y2 += FASTSTEP) {
+        int r0, g0, b0, r1, g1, b1;
+        int y = RANDOM256(y2);
+        YUVVToRGBReference(y, u, v, &r0, &g0, &b0);
+        YUVVToRGB(y, u, v, &r1, &g1, &b1);
+        EXPECT_NEAR(r0, r1, ERROR_R);
+        EXPECT_NEAR(g0, g1, 2);
+        EXPECT_NEAR(b0, b1, ERROR_B);
         ++rh[r1 - r0 + 128];
         ++gh[g1 - g0 + 128];
         ++bh[b1 - b0 + 128];
diff --git a/files/unit_test/compare_test.cc b/files/unit_test/compare_test.cc
index 136254e1..c29562cb 100644
--- a/files/unit_test/compare_test.cc
+++ b/files/unit_test/compare_test.cc
@@ -15,10 +15,13 @@
 #include "../unit_test/unit_test.h"
 #include "libyuv/basic_types.h"
 #include "libyuv/compare.h"
-#include "libyuv/compare_row.h" /* For HammingDistance_C */
 #include "libyuv/cpu_id.h"
 #include "libyuv/video_common.h"
 
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/compare_row.h" /* For HammingDistance_C */
+#endif
+
 namespace libyuv {
 
 // hash seed of 5381 recommended.
@@ -206,6 +209,7 @@ TEST_F(LibYUVCompareTest, BenchmarkARGBDetect_Unaligned) {
   free_aligned_buffer_page_end(src_a);
 }
 
+#ifdef ENABLE_ROW_TESTS
 TEST_F(LibYUVCompareTest, BenchmarkHammingDistance_Opt) {
   const int kMaxWidth = 4096 * 3;
   align_buffer_page_end(src_a, kMaxWidth);
@@ -340,7 +344,7 @@ static const int kMaxOptCount = (1 << (32 - 3)) - 64;  // 536870848
 
 TEST_F(LibYUVCompareTest, TestHammingDistance_Opt) {
   uint32_t h1 = 0;
-  const int kMaxWidth = (benchmark_width_ * benchmark_height_ + 31) & ~31;
+  const int kMaxWidth = (benchmark_width_ * benchmark_height_ + 63) & ~63;
   align_buffer_page_end(src_a, kMaxWidth);
   align_buffer_page_end(src_b, kMaxWidth);
   memset(src_a, 255u, kMaxWidth);
@@ -403,6 +407,7 @@ TEST_F(LibYUVCompareTest, TestHammingDistance_Opt) {
   free_aligned_buffer_page_end(src_a);
   free_aligned_buffer_page_end(src_b);
 }
+#endif  // ENABLE_ROW_TESTS
 
 TEST_F(LibYUVCompareTest, TestHammingDistance) {
   align_buffer_page_end(src_a, benchmark_width_ * benchmark_height_);
diff --git a/files/unit_test/convert_test.cc b/files/unit_test/convert_test.cc
index 32a4cd1c..1f975825 100644
--- a/files/unit_test/convert_test.cc
+++ b/files/unit_test/convert_test.cc
@@ -12,8 +12,6 @@
 #include <stdlib.h>
 #include <time.h>
 
-#include "libyuv/row.h" /* For ARGBToAR30Row_AVX2 */
-
 #include "libyuv/basic_types.h"
 #include "libyuv/compare.h"
 #include "libyuv/convert.h"
@@ -29,12 +27,19 @@
 #include "libyuv/rotate.h"
 #include "libyuv/video_common.h"
 
-#if defined(__arm__) || defined(__aarch64__)
-// arm version subsamples by summing 4 pixels then multiplying by matrix with
-// 4x smaller coefficients which are rounded to nearest integer.
-#define ARM_YUV_ERROR 4
-#else
-#define ARM_YUV_ERROR 0
+#ifdef ENABLE_ROW_TESTS
+#include "libyuv/row.h" /* For ARGBToAR30Row_AVX2 */
+#endif
+
+// Some functions fail on big endian. Enable these tests on all cpus except
+// PowerPC, but they are not optimized so disabled by default.
+#if !defined(DISABLE_SLOW_TESTS) && !defined(__powerpc__)
+#define LITTLE_ENDIAN_ONLY_TEST 1
+#endif
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
 #endif
 
 namespace libyuv {
@@ -49,19 +54,20 @@ namespace libyuv {
 
 #define TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,         \
                        SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,             \
-                       DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF)      \
+                       DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,      \
+                       SRC_DEPTH)                                             \
   TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {              \
     static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");       \
     static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");       \
     static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                   \
-                  "DST SRC_SUBSAMP_X unsupported");                           \
+                  "SRC_SUBSAMP_X unsupported");                               \
     static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                   \
-                  "DST SRC_SUBSAMP_Y unsupported");                           \
+                  "SRC_SUBSAMP_Y unsupported");                               \
     static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                   \
-                  "DST DST_SUBSAMP_X unsupported");                           \
+                  "DST_SUBSAMP_X unsupported");                               \
     static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                   \
-                  "DST DST_SUBSAMP_Y unsupported");                           \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+                  "DST_SUBSAMP_Y unsupported");                               \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
     const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);               \
     const int kSrcHalfHeight = SUBSAMPLE(kHeight, SRC_SUBSAMP_Y);             \
@@ -81,6 +87,16 @@ namespace libyuv {
     MemRandomize(src_y + OFF, kWidth * kHeight * SRC_BPC);                    \
     MemRandomize(src_u + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
     MemRandomize(src_v + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                   \
+    SRC_T* src_u_p = reinterpret_cast<SRC_T*>(src_u + OFF);                   \
+    SRC_T* src_v_p = reinterpret_cast<SRC_T*>(src_v + OFF);                   \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y_p[i] = src_y_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    for (int i = 0; i < kSrcHalfWidth * kSrcHalfHeight; ++i) {                \
+      src_u_p[i] = src_u_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+      src_v_p[i] = src_v_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
     memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                             \
     memset(dst_u_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC);               \
     memset(dst_v_c, 3, kDstHalfWidth* kDstHalfHeight* DST_BPC);               \
@@ -89,9 +105,7 @@ namespace libyuv {
     memset(dst_v_opt, 103, kDstHalfWidth* kDstHalfHeight* DST_BPC);           \
     MaskCpuFlags(disable_cpu_flags_);                                         \
     SRC_FMT_PLANAR##To##FMT_PLANAR(                                           \
-        reinterpret_cast<SRC_T*>(src_y + OFF), kWidth,                        \
-        reinterpret_cast<SRC_T*>(src_u + OFF), kSrcHalfWidth,                 \
-        reinterpret_cast<SRC_T*>(src_v + OFF), kSrcHalfWidth,                 \
+        src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,      \
         reinterpret_cast<DST_T*>(dst_y_c), kWidth,                            \
         reinterpret_cast<DST_T*>(dst_u_c), kDstHalfWidth,                     \
         reinterpret_cast<DST_T*>(dst_v_c), kDstHalfWidth, kWidth,             \
@@ -99,9 +113,7 @@ namespace libyuv {
     MaskCpuFlags(benchmark_cpu_info_);                                        \
     for (int i = 0; i < benchmark_iterations_; ++i) {                         \
       SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
-          reinterpret_cast<SRC_T*>(src_y + OFF), kWidth,                      \
-          reinterpret_cast<SRC_T*>(src_u + OFF), kSrcHalfWidth,               \
-          reinterpret_cast<SRC_T*>(src_v + OFF), kSrcHalfWidth,               \
+          src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,    \
           reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                        \
           reinterpret_cast<DST_T*>(dst_u_opt), kDstHalfWidth,                 \
           reinterpret_cast<DST_T*>(dst_v_opt), kDstHalfWidth, kWidth,         \
@@ -127,41 +139,58 @@ namespace libyuv {
 
 #define TESTPLANARTOP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,           \
                       SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,               \
-                      DST_SUBSAMP_X, DST_SUBSAMP_Y)                            \
+                      DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH)                 \
   TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
                  FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
-                 benchmark_width_ - 4, _Any, +, 0)                             \
+                 benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH)                  \
   TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
                  FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
-                 benchmark_width_, _Unaligned, +, 1)                           \
+                 benchmark_width_, _Unaligned, +, 2, SRC_DEPTH)                \
   TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
                  FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
-                 benchmark_width_, _Invert, -, 0)                              \
+                 benchmark_width_, _Invert, -, 0, SRC_DEPTH)                   \
   TESTPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, \
                  FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X, DST_SUBSAMP_Y,     \
-                 benchmark_width_, _Opt, +, 0)
-
-TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2)
-TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I420, uint8_t, 1, 2, 2)
-TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I420, uint8_t, 1, 2, 2)
-TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I422, uint8_t, 1, 2, 1)
-TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I444, uint8_t, 1, 1, 1)
-TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420Mirror, uint8_t, 1, 2, 2)
-TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I422, uint8_t, 1, 2, 1)
-TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I444, uint8_t, 1, 1, 1)
-TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I010, uint16_t, 2, 2, 2)
-TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I420, uint8_t, 1, 2, 2)
-TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I010, uint16_t, 2, 2, 2)
-TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H010, uint16_t, 2, 2, 2)
-TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H420, uint8_t, 1, 2, 2)
-TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2)
+                 benchmark_width_, _Opt, +, 0, SRC_DEPTH)
+
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I420, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I422, uint8_t, 1, 2, 1, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I420Mirror, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I422, uint8_t, 1, 2, 1, 8)
+TESTPLANARTOP(I422, uint8_t, 1, 2, 1, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I444, uint8_t, 1, 1, 1, I444, uint8_t, 1, 1, 1, 8)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I010, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(I420, uint8_t, 1, 2, 2, I012, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(H010, uint16_t, 2, 2, 2, H420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H012, uint16_t, 2, 2, 2, 8)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I410, uint16_t, 2, 1, 1, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I410, uint16_t, 2, 1, 1, 10)
+TESTPLANARTOP(I012, uint16_t, 2, 2, 2, I412, uint16_t, 2, 1, 1, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I412, uint16_t, 2, 1, 1, 12)
+TESTPLANARTOP(I410, uint16_t, 2, 1, 1, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOP(I412, uint16_t, 2, 1, 1, I012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOP(I010, uint16_t, 2, 2, 2, I420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I420, uint8_t, 1, 2, 2, 10)
+TESTPLANARTOP(I210, uint16_t, 2, 2, 1, I422, uint8_t, 1, 2, 1, 10)
+TESTPLANARTOP(I410, uint16_t, 2, 1, 1, I444, uint8_t, 1, 1, 1, 10)
+TESTPLANARTOP(I012, uint16_t, 2, 2, 2, I420, uint8_t, 1, 2, 2, 12)
+TESTPLANARTOP(I212, uint16_t, 2, 2, 1, I422, uint8_t, 1, 2, 1, 12)
+TESTPLANARTOP(I412, uint16_t, 2, 1, 1, I444, uint8_t, 1, 1, 1, 12)
 
 // Test Android 420 to I420
 #define TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X,          \
                         SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
                         W1280, N, NEG, OFF, PN, OFF_U, OFF_V)                 \
-  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##_##PN##N) {       \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+  TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##To##PN##N) {      \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
     const int kSizeUV =                                                       \
         SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); \
@@ -216,41 +245,23 @@ TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2)
           dst_y_opt, kWidth, dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),         \
           dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight);      \
     }                                                                         \
-    int max_diff = 0;                                                         \
     for (int i = 0; i < kHeight; ++i) {                                       \
       for (int j = 0; j < kWidth; ++j) {                                      \
-        int abs_diff = abs(static_cast<int>(dst_y_c[i * kWidth + j]) -        \
-                           static_cast<int>(dst_y_opt[i * kWidth + j]));      \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
       }                                                                       \
     }                                                                         \
-    EXPECT_EQ(0, max_diff);                                                   \
     for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
       for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
-        int abs_diff = abs(                                                   \
-            static_cast<int>(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]) - \
-            static_cast<int>(                                                 \
-                dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
+        EXPECT_EQ(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
       }                                                                       \
     }                                                                         \
-    EXPECT_LE(max_diff, 3);                                                   \
     for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
       for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
-        int abs_diff = abs(                                                   \
-            static_cast<int>(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]) - \
-            static_cast<int>(                                                 \
-                dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]));            \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
+        EXPECT_EQ(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
       }                                                                       \
     }                                                                         \
-    EXPECT_LE(max_diff, 3);                                                   \
     free_aligned_buffer_page_end(dst_y_c);                                    \
     free_aligned_buffer_page_end(dst_u_c);                                    \
     free_aligned_buffer_page_end(dst_v_c);                                    \
@@ -265,11 +276,11 @@ TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2)
                        SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X,    \
                        SUBSAMP_Y)                                              \
   TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
-                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_ - 4,      \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_ + 1,      \
                   _Any, +, 0, PN, OFF_U, OFF_V)                                \
   TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_,          \
-                  _Unaligned, +, 1, PN, OFF_U, OFF_V)                          \
+                  _Unaligned, +, 2, PN, OFF_U, OFF_V)                          \
   TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
                   FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, \
                   -, 0, PN, OFF_U, OFF_V)                                      \
@@ -280,6 +291,8 @@ TESTPLANARTOP(H420, uint8_t, 1, 2, 2, H010, uint16_t, 2, 2, 2)
 TESTAPLANARTOP(Android420, I420, 1, 0, 0, 2, 2, I420, 2, 2)
 TESTAPLANARTOP(Android420, NV12, 2, 0, 1, 2, 2, I420, 2, 2)
 TESTAPLANARTOP(Android420, NV21, 2, 1, 0, 2, 2, I420, 2, 2)
+#undef TESTAPLANARTOP
+#undef TESTAPLANARTOPI
 
 // wrapper to keep API the same
 int I400ToNV21(const uint8_t* src_y,
@@ -298,77 +311,75 @@ int I400ToNV21(const uint8_t* src_y,
                     dst_stride_vu, width, height);
 }
 
-#define TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,         \
-                        FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF) \
+#define TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,        \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,            \
+                        DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,     \
+                        SRC_DEPTH)                                            \
   TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {              \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");       \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");       \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                   \
+                  "SRC_SUBSAMP_X unsupported");                               \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                   \
+                  "SRC_SUBSAMP_Y unsupported");                               \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                   \
+                  "DST_SUBSAMP_X unsupported");                               \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                   \
+                  "DST_SUBSAMP_Y unsupported");                               \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
-    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
-    align_buffer_page_end(src_u, SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *           \
-                                         SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) +  \
-                                     OFF);                                    \
-    align_buffer_page_end(src_v, SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *           \
-                                         SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) +  \
-                                     OFF);                                    \
-    align_buffer_page_end(dst_y_c, kWidth* kHeight);                          \
-    align_buffer_page_end(dst_uv_c, SUBSAMPLE(kWidth, SUBSAMP_X) * 2 *        \
-                                        SUBSAMPLE(kHeight, SUBSAMP_Y));       \
-    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                        \
-    align_buffer_page_end(dst_uv_opt, SUBSAMPLE(kWidth, SUBSAMP_X) * 2 *      \
-                                          SUBSAMPLE(kHeight, SUBSAMP_Y));     \
-    for (int i = 0; i < kHeight; ++i)                                         \
-      for (int j = 0; j < kWidth; ++j)                                        \
-        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                    \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {             \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {            \
-        src_u[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =             \
-            (fastrand() & 0xff);                                              \
-        src_v[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =             \
-            (fastrand() & 0xff);                                              \
-      }                                                                       \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);               \
+    const int kSrcHalfHeight = SUBSAMPLE(kHeight, SRC_SUBSAMP_Y);             \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);               \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);             \
+    align_buffer_page_end(src_y, kWidth* kHeight* SRC_BPC + OFF);             \
+    align_buffer_page_end(src_u,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(src_v,                                              \
+                          kSrcHalfWidth* kSrcHalfHeight* SRC_BPC + OFF);      \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                 \
+    align_buffer_page_end(dst_uv_c,                                           \
+                          kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);        \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);               \
+    align_buffer_page_end(dst_uv_opt,                                         \
+                          kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);        \
+    MemRandomize(src_y + OFF, kWidth * kHeight * SRC_BPC);                    \
+    MemRandomize(src_u + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    MemRandomize(src_v + OFF, kSrcHalfWidth * kSrcHalfHeight * SRC_BPC);      \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                   \
+    SRC_T* src_u_p = reinterpret_cast<SRC_T*>(src_u + OFF);                   \
+    SRC_T* src_v_p = reinterpret_cast<SRC_T*>(src_v + OFF);                   \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      src_y_p[i] = src_y_p[i] & ((1 << SRC_DEPTH) - 1);                       \
     }                                                                         \
-    memset(dst_y_c, 1, kWidth* kHeight);                                      \
-    memset(dst_uv_c, 2,                                                       \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \
-    memset(dst_y_opt, 101, kWidth* kHeight);                                  \
-    memset(dst_uv_opt, 102,                                                   \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y)); \
+    for (int i = 0; i < kSrcHalfWidth * kSrcHalfHeight; ++i) {                \
+      src_u_p[i] = src_u_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+      src_v_p[i] = src_v_p[i] & ((1 << SRC_DEPTH) - 1);                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                             \
+    memset(dst_uv_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);          \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                         \
+    memset(dst_uv_opt, 102, kDstHalfWidth* kDstHalfHeight* DST_BPC * 2);      \
     MaskCpuFlags(disable_cpu_flags_);                                         \
-    SRC_FMT_PLANAR##To##FMT_PLANAR(                                           \
-        src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X),   \
-        src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), dst_y_c, kWidth,       \
-        dst_uv_c, SUBSAMPLE(kWidth, SUBSAMP_X) * 2, kWidth, NEG kHeight);     \
+    SRC_FMT_PLANAR##To##FMT_PLANAR(src_y_p, kWidth, src_u_p, kSrcHalfWidth,   \
+                                   src_v_p, kSrcHalfWidth,                    \
+                                   reinterpret_cast<DST_T*>(dst_y_c), kWidth, \
+                                   reinterpret_cast<DST_T*>(dst_uv_c),        \
+                                   kDstHalfWidth * 2, kWidth, NEG kHeight);   \
     MaskCpuFlags(benchmark_cpu_info_);                                        \
     for (int i = 0; i < benchmark_iterations_; ++i) {                         \
       SRC_FMT_PLANAR##To##FMT_PLANAR(                                         \
-          src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), \
-          src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), dst_y_opt, kWidth,   \
-          dst_uv_opt, SUBSAMPLE(kWidth, SUBSAMP_X) * 2, kWidth, NEG kHeight); \
+          src_y_p, kWidth, src_u_p, kSrcHalfWidth, src_v_p, kSrcHalfWidth,    \
+          reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                        \
+          reinterpret_cast<DST_T*>(dst_uv_opt), kDstHalfWidth * 2, kWidth,    \
+          NEG kHeight);                                                       \
     }                                                                         \
-    int max_diff = 0;                                                         \
-    for (int i = 0; i < kHeight; ++i) {                                       \
-      for (int j = 0; j < kWidth; ++j) {                                      \
-        int abs_diff = abs(static_cast<int>(dst_y_c[i * kWidth + j]) -        \
-                           static_cast<int>(dst_y_opt[i * kWidth + j]));      \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
-      }                                                                       \
+    for (int i = 0; i < kHeight * kWidth * DST_BPC; ++i) {                    \
+      EXPECT_EQ(dst_y_c[i], dst_y_opt[i]);                                    \
     }                                                                         \
-    EXPECT_LE(max_diff, 1);                                                   \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X) * 2; ++j) {            \
-        int abs_diff =                                                        \
-            abs(static_cast<int>(                                             \
-                    dst_uv_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) * 2 + j]) -     \
-                static_cast<int>(                                             \
-                    dst_uv_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) * 2 + j]));   \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
-      }                                                                       \
+    for (int i = 0; i < kDstHalfWidth * kDstHalfHeight * DST_BPC * 2; ++i) {  \
+      EXPECT_EQ(dst_uv_c[i], dst_uv_opt[i]);                                  \
     }                                                                         \
-    EXPECT_LE(max_diff, 1);                                                   \
     free_aligned_buffer_page_end(dst_y_c);                                    \
     free_aligned_buffer_page_end(dst_uv_c);                                   \
     free_aligned_buffer_page_end(dst_y_opt);                                  \
@@ -378,92 +389,110 @@ int I400ToNV21(const uint8_t* src_y,
     free_aligned_buffer_page_end(src_v);                                      \
   }
 
-#define TESTPLANARTOBP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,        \
-                       FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                    \
-  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                  SUBSAMP_X, SUBSAMP_Y, benchmark_width_ - 4, _Any, +, 0)   \
-  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                  SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Unaligned, +, 1) \
-  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                  SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, -, 0)    \
-  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                  SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, 0)
-
-TESTPLANARTOBP(I420, 2, 2, NV12, 2, 2)
-TESTPLANARTOBP(I420, 2, 2, NV21, 2, 2)
-TESTPLANARTOBP(I422, 2, 1, NV21, 2, 2)
-TESTPLANARTOBP(I444, 1, 1, NV21, 2, 2)
-TESTPLANARTOBP(I400, 2, 2, NV21, 2, 2)
-
-#define TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,        \
-                          FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG,     \
-                          OFF)                                                 \
+#define TESTPLANARTOBP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,         \
+                       SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,             \
+                       DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH)               \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH) \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Unaligned, +, 2,          \
+                  SRC_DEPTH)                                                  \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Invert, -, 0, SRC_DEPTH)  \
+  TESTPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                  SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                  DST_SUBSAMP_Y, benchmark_width_, _Opt, +, 0, SRC_DEPTH)
+
+TESTPLANARTOBP(I420, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I420, uint8_t, 1, 2, 2, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I422, uint8_t, 1, 2, 1, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV12, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I444, uint8_t, 1, 1, 1, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I400, uint8_t, 1, 2, 2, NV21, uint8_t, 1, 2, 2, 8)
+TESTPLANARTOBP(I010, uint16_t, 2, 2, 2, P010, uint16_t, 2, 2, 2, 10)
+TESTPLANARTOBP(I210, uint16_t, 2, 2, 1, P210, uint16_t, 2, 2, 1, 10)
+TESTPLANARTOBP(I012, uint16_t, 2, 2, 2, P012, uint16_t, 2, 2, 2, 12)
+TESTPLANARTOBP(I212, uint16_t, 2, 2, 1, P212, uint16_t, 2, 2, 1, 12)
+
+#define TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,       \
+                          SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,           \
+                          DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,    \
+                          DOY, SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)             \
   TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {               \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");        \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");        \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                    \
+                  "SRC_SUBSAMP_X unsupported");                                \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                    \
+                  "SRC_SUBSAMP_Y unsupported");                                \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                    \
+                  "DST_SUBSAMP_X unsupported");                                \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                    \
+                  "DST_SUBSAMP_Y unsupported");                                \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
-    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
-    align_buffer_page_end(src_uv, SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * 2 *       \
-                                          SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) +  \
-                                      OFF);                                    \
-    align_buffer_page_end(dst_y_c, kWidth* kHeight);                           \
-    align_buffer_page_end(dst_uv_c, SUBSAMPLE(kWidth, SUBSAMP_X) * 2 *         \
-                                        SUBSAMPLE(kHeight, SUBSAMP_Y));        \
-    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                         \
-    align_buffer_page_end(dst_uv_opt, SUBSAMPLE(kWidth, SUBSAMP_X) * 2 *       \
-                                          SUBSAMPLE(kHeight, SUBSAMP_Y));      \
-    for (int i = 0; i < kHeight; ++i)                                          \
-      for (int j = 0; j < kWidth; ++j)                                         \
-        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                     \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {              \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {             \
-        src_uv[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) * 2 + j + 0 + OFF] =     \
-            (fastrand() & 0xff);                                               \
-        src_uv[(i * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) * 2 + j + 1 + OFF] =     \
-            (fastrand() & 0xff);                                               \
-      }                                                                        \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);                \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);                \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);              \
+    const int kPaddedWidth = (kWidth + (TILE_WIDTH - 1)) & ~(TILE_WIDTH - 1);  \
+    const int kPaddedHeight =                                                  \
+        (kHeight + (TILE_HEIGHT - 1)) & ~(TILE_HEIGHT - 1);                    \
+    const int kSrcHalfPaddedWidth = SUBSAMPLE(kPaddedWidth, SRC_SUBSAMP_X);    \
+    const int kSrcHalfPaddedHeight = SUBSAMPLE(kPaddedHeight, SRC_SUBSAMP_Y);  \
+    align_buffer_page_end(src_y, kPaddedWidth* kPaddedHeight* SRC_BPC + OFF);  \
+    align_buffer_page_end(                                                     \
+        src_uv,                                                                \
+        2 * kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * SRC_BPC + OFF);       \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                  \
+    align_buffer_page_end(dst_uv_c,                                            \
+                          2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);       \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);                \
+    align_buffer_page_end(dst_uv_opt,                                          \
+                          2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);       \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                    \
+    SRC_T* src_uv_p = reinterpret_cast<SRC_T*>(src_uv + OFF);                  \
+    for (int i = 0; i < kPaddedWidth * kPaddedHeight; ++i) {                   \
+      src_y_p[i] =                                                             \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
     }                                                                          \
-    memset(dst_y_c, 1, kWidth* kHeight);                                       \
-    memset(dst_uv_c, 2,                                                        \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));  \
-    memset(dst_y_opt, 101, kWidth* kHeight);                                   \
-    memset(dst_uv_opt, 102,                                                    \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * 2 * SUBSAMPLE(kHeight, SUBSAMP_Y));  \
+    for (int i = 0; i < kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * 2; ++i) { \
+      src_uv_p[i] =                                                            \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
+    }                                                                          \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                              \
+    memset(dst_uv_c, 2, 2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);         \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                          \
+    memset(dst_uv_opt, 102, 2 * kDstHalfWidth * kDstHalfHeight * DST_BPC);     \
     MaskCpuFlags(disable_cpu_flags_);                                          \
     SRC_FMT_PLANAR##To##FMT_PLANAR(                                            \
-        src_y + OFF, kWidth, src_uv + OFF,                                     \
-        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * 2, dst_y_c, kWidth, dst_uv_c,       \
-        SUBSAMPLE(kWidth, SUBSAMP_X) * 2, kWidth, NEG kHeight);                \
+        src_y_p, kWidth, src_uv_p, 2 * kSrcHalfWidth,                          \
+        DOY ? reinterpret_cast<DST_T*>(dst_y_c) : NULL, kWidth,                \
+        reinterpret_cast<DST_T*>(dst_uv_c), 2 * kDstHalfWidth, kWidth,         \
+        NEG kHeight);                                                          \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
     for (int i = 0; i < benchmark_iterations_; ++i) {                          \
       SRC_FMT_PLANAR##To##FMT_PLANAR(                                          \
-          src_y + OFF, kWidth, src_uv + OFF,                                   \
-          SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * 2, dst_y_opt, kWidth, dst_uv_opt, \
-          SUBSAMPLE(kWidth, SUBSAMP_X) * 2, kWidth, NEG kHeight);              \
+          src_y_p, kWidth, src_uv_p, 2 * kSrcHalfWidth,                        \
+          DOY ? reinterpret_cast<DST_T*>(dst_y_opt) : NULL, kWidth,            \
+          reinterpret_cast<DST_T*>(dst_uv_opt), 2 * kDstHalfWidth, kWidth,     \
+          NEG kHeight);                                                        \
     }                                                                          \
-    int max_diff = 0;                                                          \
-    for (int i = 0; i < kHeight; ++i) {                                        \
-      for (int j = 0; j < kWidth; ++j) {                                       \
-        int abs_diff = abs(static_cast<int>(dst_y_c[i * kWidth + j]) -         \
-                           static_cast<int>(dst_y_opt[i * kWidth + j]));       \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
+    if (DOY) {                                                                 \
+      for (int i = 0; i < kHeight; ++i) {                                      \
+        for (int j = 0; j < kWidth; ++j) {                                     \
+          EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);       \
         }                                                                      \
       }                                                                        \
     }                                                                          \
-    EXPECT_LE(max_diff, 1);                                                    \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X) * 2; ++j) {             \
-        int abs_diff =                                                         \
-            abs(static_cast<int>(                                              \
-                    dst_uv_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) * 2 + j]) -      \
-                static_cast<int>(                                              \
-                    dst_uv_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) * 2 + j]));    \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
-        }                                                                      \
+    for (int i = 0; i < kDstHalfHeight; ++i) {                                 \
+      for (int j = 0; j < 2 * kDstHalfWidth; ++j) {                            \
+        EXPECT_EQ(dst_uv_c[i * 2 * kDstHalfWidth + j],                         \
+                  dst_uv_opt[i * 2 * kDstHalfWidth + j]);                      \
       }                                                                        \
     }                                                                          \
-    EXPECT_LE(max_diff, 1);                                                    \
     free_aligned_buffer_page_end(dst_y_c);                                     \
     free_aligned_buffer_page_end(dst_uv_c);                                    \
     free_aligned_buffer_page_end(dst_y_opt);                                   \
@@ -472,112 +501,116 @@ TESTPLANARTOBP(I400, 2, 2, NV21, 2, 2)
     free_aligned_buffer_page_end(src_uv);                                      \
   }
 
-#define TESTBIPLANARTOBP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,        \
-                         FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                    \
-  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                    SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, 0)       \
-  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                    SUBSAMP_X, SUBSAMP_Y, benchmark_width_ - 4, _Any, +, 0)   \
-  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                    SUBSAMP_X, SUBSAMP_Y, benchmark_width, _Unaligned, +, 1)  \
-  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                    SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, -, 0)    \
-  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, \
-                    SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, 0)
-
-// TODO(fbarchard): Fix msan on this unittest
-// TESTBIPLANARTOBP(NV21, 2, 2, NV12, 2, 2)
-
-#define TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,         \
-                         FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, W1280, N, NEG, OFF, \
-                         DOY)                                                  \
+#define TESTBIPLANARTOBP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,        \
+                         SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,            \
+                         DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH, TILE_WIDTH,  \
+                         TILE_HEIGHT)                                          \
+  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,             \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,  \
+                    DST_SUBSAMP_Y, benchmark_width_ + 1, _Any, +, 0, 1,        \
+                    SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)                        \
+  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,             \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,  \
+                    DST_SUBSAMP_Y, benchmark_width_, _Unaligned, +, 2, 1,      \
+                    SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)                        \
+  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,             \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,  \
+                    DST_SUBSAMP_Y, benchmark_width_, _Invert, -, 0, 1,         \
+                    SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)                        \
+  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,             \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,  \
+                    DST_SUBSAMP_Y, benchmark_width_, _Opt, +, 0, 1, SRC_DEPTH, \
+                    TILE_WIDTH, TILE_HEIGHT)                                   \
+  TESTBIPLANARTOBPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,             \
+                    SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,  \
+                    DST_SUBSAMP_Y, benchmark_width_, _NullY, +, 0, 0,          \
+                    SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)
+
+TESTBIPLANARTOBP(NV21, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBIPLANARTOBP(NV12, uint8_t, 1, 2, 2, NV12Mirror, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBIPLANARTOBP(NV12, uint8_t, 1, 2, 2, NV24, uint8_t, 1, 1, 1, 8, 1, 1)
+TESTBIPLANARTOBP(NV16, uint8_t, 1, 2, 1, NV24, uint8_t, 1, 1, 1, 8, 1, 1)
+TESTBIPLANARTOBP(P010, uint16_t, 2, 2, 2, P410, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBIPLANARTOBP(P210, uint16_t, 2, 2, 1, P410, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBIPLANARTOBP(P012, uint16_t, 2, 2, 2, P412, uint16_t, 2, 1, 1, 10, 1, 1)
+TESTBIPLANARTOBP(P212, uint16_t, 2, 2, 1, P412, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBIPLANARTOBP(P016, uint16_t, 2, 2, 2, P416, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBIPLANARTOBP(P216, uint16_t, 2, 2, 1, P416, uint16_t, 2, 1, 1, 12, 1, 1)
+TESTBIPLANARTOBP(MM21, uint8_t, 1, 2, 2, NV12, uint8_t, 1, 2, 2, 8, 16, 32)
+
+#define TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,        \
+                         SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,            \
+                         DST_SUBSAMP_X, DST_SUBSAMP_Y, W1280, N, NEG, OFF,     \
+                         SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)                   \
   TEST_F(LibYUVConvertTest, SRC_FMT_PLANAR##To##FMT_PLANAR##N) {               \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    static_assert(SRC_BPC == 1 || SRC_BPC == 2, "SRC BPC unsupported");        \
+    static_assert(DST_BPC == 1 || DST_BPC == 2, "DST BPC unsupported");        \
+    static_assert(SRC_SUBSAMP_X == 1 || SRC_SUBSAMP_X == 2,                    \
+                  "SRC_SUBSAMP_X unsupported");                                \
+    static_assert(SRC_SUBSAMP_Y == 1 || SRC_SUBSAMP_Y == 2,                    \
+                  "SRC_SUBSAMP_Y unsupported");                                \
+    static_assert(DST_SUBSAMP_X == 1 || DST_SUBSAMP_X == 2,                    \
+                  "DST_SUBSAMP_X unsupported");                                \
+    static_assert(DST_SUBSAMP_Y == 1 || DST_SUBSAMP_Y == 2,                    \
+                  "DST_SUBSAMP_Y unsupported");                                \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
-    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                       \
-    align_buffer_page_end(src_uv, 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X) *       \
-                                          SUBSAMPLE(kHeight, SRC_SUBSAMP_Y) +  \
-                                      OFF);                                    \
-    align_buffer_page_end(dst_y_c, kWidth* kHeight);                           \
-    align_buffer_page_end(dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X) *              \
-                                       SUBSAMPLE(kHeight, SUBSAMP_Y));         \
-    align_buffer_page_end(dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X) *              \
-                                       SUBSAMPLE(kHeight, SUBSAMP_Y));         \
-    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                         \
-    align_buffer_page_end(dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *            \
-                                         SUBSAMPLE(kHeight, SUBSAMP_Y));       \
-    align_buffer_page_end(dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *            \
-                                         SUBSAMPLE(kHeight, SUBSAMP_Y));       \
-    for (int i = 0; i < kHeight; ++i)                                          \
-      for (int j = 0; j < kWidth; ++j)                                         \
-        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                     \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {              \
-      for (int j = 0; j < 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {         \
-        src_uv[(i * 2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X)) + j + OFF] =         \
-            (fastrand() & 0xff);                                               \
-      }                                                                        \
+    const int kSrcHalfWidth = SUBSAMPLE(kWidth, SRC_SUBSAMP_X);                \
+    const int kDstHalfWidth = SUBSAMPLE(kWidth, DST_SUBSAMP_X);                \
+    const int kDstHalfHeight = SUBSAMPLE(kHeight, DST_SUBSAMP_Y);              \
+    const int kPaddedWidth = (kWidth + (TILE_WIDTH - 1)) & ~(TILE_WIDTH - 1);  \
+    const int kPaddedHeight =                                                  \
+        (kHeight + (TILE_HEIGHT - 1)) & ~(TILE_HEIGHT - 1);                    \
+    const int kSrcHalfPaddedWidth = SUBSAMPLE(kPaddedWidth, SRC_SUBSAMP_X);    \
+    const int kSrcHalfPaddedHeight = SUBSAMPLE(kPaddedHeight, SRC_SUBSAMP_Y);  \
+    align_buffer_page_end(src_y, kPaddedWidth* kPaddedHeight* SRC_BPC + OFF);  \
+    align_buffer_page_end(                                                     \
+        src_uv, kSrcHalfPaddedWidth* kSrcHalfPaddedHeight* SRC_BPC * 2 + OFF); \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight* DST_BPC);                  \
+    align_buffer_page_end(dst_u_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);    \
+    align_buffer_page_end(dst_v_c, kDstHalfWidth* kDstHalfHeight* DST_BPC);    \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight* DST_BPC);                \
+    align_buffer_page_end(dst_u_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC);  \
+    align_buffer_page_end(dst_v_opt, kDstHalfWidth* kDstHalfHeight* DST_BPC);  \
+    SRC_T* src_y_p = reinterpret_cast<SRC_T*>(src_y + OFF);                    \
+    SRC_T* src_uv_p = reinterpret_cast<SRC_T*>(src_uv + OFF);                  \
+    for (int i = 0; i < kPaddedWidth * kPaddedHeight; ++i) {                   \
+      src_y_p[i] =                                                             \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
     }                                                                          \
-    memset(dst_y_c, 1, kWidth* kHeight);                                       \
-    memset(dst_u_c, 2,                                                         \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
-    memset(dst_v_c, 3,                                                         \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
-    memset(dst_y_opt, 101, kWidth* kHeight);                                   \
-    memset(dst_u_opt, 102,                                                     \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
-    memset(dst_v_opt, 103,                                                     \
-           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    for (int i = 0; i < kSrcHalfPaddedWidth * kSrcHalfPaddedHeight * 2; ++i) { \
+      src_uv_p[i] =                                                            \
+          (fastrand() & (((SRC_T)(-1)) << ((8 * SRC_BPC) - SRC_DEPTH)));       \
+    }                                                                          \
+    memset(dst_y_c, 1, kWidth* kHeight* DST_BPC);                              \
+    memset(dst_u_c, 2, kDstHalfWidth* kDstHalfHeight* DST_BPC);                \
+    memset(dst_v_c, 3, kDstHalfWidth* kDstHalfHeight* DST_BPC);                \
+    memset(dst_y_opt, 101, kWidth* kHeight* DST_BPC);                          \
+    memset(dst_u_opt, 102, kDstHalfWidth* kDstHalfHeight* DST_BPC);            \
+    memset(dst_v_opt, 103, kDstHalfWidth* kDstHalfHeight* DST_BPC);            \
     MaskCpuFlags(disable_cpu_flags_);                                          \
     SRC_FMT_PLANAR##To##FMT_PLANAR(                                            \
-        src_y + OFF, kWidth, src_uv + OFF,                                     \
-        2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X), DOY ? dst_y_c : NULL, kWidth,    \
-        dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X), dst_v_c,                        \
-        SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight);                    \
+        src_y_p, kWidth, src_uv_p, kSrcHalfWidth * 2,                          \
+        reinterpret_cast<DST_T*>(dst_y_c), kWidth,                             \
+        reinterpret_cast<DST_T*>(dst_u_c), kDstHalfWidth,                      \
+        reinterpret_cast<DST_T*>(dst_v_c), kDstHalfWidth, kWidth,              \
+        NEG kHeight);                                                          \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
     for (int i = 0; i < benchmark_iterations_; ++i) {                          \
       SRC_FMT_PLANAR##To##FMT_PLANAR(                                          \
-          src_y + OFF, kWidth, src_uv + OFF,                                   \
-          2 * SUBSAMPLE(kWidth, SRC_SUBSAMP_X), DOY ? dst_y_opt : NULL,        \
-          kWidth, dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X), dst_v_opt,          \
-          SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight);                  \
-    }                                                                          \
-    int max_diff = 0;                                                          \
-    if (DOY) {                                                                 \
-      for (int i = 0; i < kHeight; ++i) {                                      \
-        for (int j = 0; j < kWidth; ++j) {                                     \
-          int abs_diff = abs(static_cast<int>(dst_y_c[i * kWidth + j]) -       \
-                             static_cast<int>(dst_y_opt[i * kWidth + j]));     \
-          if (abs_diff > max_diff) {                                           \
-            max_diff = abs_diff;                                               \
-          }                                                                    \
-        }                                                                      \
-      }                                                                        \
-      EXPECT_LE(max_diff, 1);                                                  \
+          src_y_p, kWidth, src_uv_p, kSrcHalfWidth * 2,                        \
+          reinterpret_cast<DST_T*>(dst_y_opt), kWidth,                         \
+          reinterpret_cast<DST_T*>(dst_u_opt), kDstHalfWidth,                  \
+          reinterpret_cast<DST_T*>(dst_v_opt), kDstHalfWidth, kWidth,          \
+          NEG kHeight);                                                        \
     }                                                                          \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                 \
-        int abs_diff = abs(                                                    \
-            static_cast<int>(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -  \
-            static_cast<int>(                                                  \
-                dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]));             \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
-        }                                                                      \
-      }                                                                        \
+    for (int i = 0; i < kHeight * kWidth * DST_BPC; ++i) {                     \
+      EXPECT_EQ(dst_y_c[i], dst_y_opt[i]);                                     \
     }                                                                          \
-    EXPECT_LE(max_diff, 1);                                                    \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
-      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                 \
-        int abs_diff = abs(                                                    \
-            static_cast<int>(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]) -  \
-            static_cast<int>(                                                  \
-                dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]));             \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
-        }                                                                      \
-      }                                                                        \
+    for (int i = 0; i < kDstHalfWidth * kDstHalfHeight * DST_BPC; ++i) {       \
+      EXPECT_EQ(dst_u_c[i], dst_u_opt[i]);                                     \
+      EXPECT_EQ(dst_v_c[i], dst_v_opt[i]);                                     \
     }                                                                          \
-    EXPECT_LE(max_diff, 1);                                                    \
     free_aligned_buffer_page_end(dst_y_c);                                     \
     free_aligned_buffer_page_end(dst_u_c);                                     \
     free_aligned_buffer_page_end(dst_v_c);                                     \
@@ -588,29 +621,72 @@ TESTPLANARTOBP(I400, 2, 2, NV21, 2, 2)
     free_aligned_buffer_page_end(src_uv);                                      \
   }
 
-#define TESTBIPLANARTOP(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,         \
-                        FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y)                     \
-  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR,  \
-                   SUBSAMP_X, SUBSAMP_Y, benchmark_width_ - 4, _Any, +, 0, 1) \
-  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR,  \
-                   SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Unaligned, +, 1,  \
-                   1)                                                         \
-  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR,  \
-                   SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, -, 0, 1)  \
-  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR,  \
-                   SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, 0, 1)     \
-  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR,  \
-                   SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _NullY, +, 0, 0)
-
-TESTBIPLANARTOP(NV12, 2, 2, I420, 2, 2)
-TESTBIPLANARTOP(NV21, 2, 2, I420, 2, 2)
+#define TESTBIPLANARTOP(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,         \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC,             \
+                        DST_SUBSAMP_X, DST_SUBSAMP_Y, SRC_DEPTH, TILE_WIDTH,   \
+                        TILE_HEIGHT)                                           \
+  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                   DST_SUBSAMP_Y, benchmark_width_ + 1, _Any, +, 0, SRC_DEPTH, \
+                   TILE_WIDTH, TILE_HEIGHT)                                    \
+  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                   DST_SUBSAMP_Y, benchmark_width_, _Unaligned, +, 2,          \
+                   SRC_DEPTH, TILE_WIDTH, TILE_HEIGHT)                         \
+  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                   DST_SUBSAMP_Y, benchmark_width_, _Invert, -, 0, SRC_DEPTH,  \
+                   TILE_WIDTH, TILE_HEIGHT)                                    \
+  TESTBIPLANARTOPI(SRC_FMT_PLANAR, SRC_T, SRC_BPC, SRC_SUBSAMP_X,              \
+                   SRC_SUBSAMP_Y, FMT_PLANAR, DST_T, DST_BPC, DST_SUBSAMP_X,   \
+                   DST_SUBSAMP_Y, benchmark_width_, _Opt, +, 0, SRC_DEPTH,     \
+                   TILE_WIDTH, TILE_HEIGHT)
+
+TESTBIPLANARTOP(NV12, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBIPLANARTOP(NV21, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 1, 1)
+TESTBIPLANARTOP(MM21, uint8_t, 1, 2, 2, I420, uint8_t, 1, 2, 2, 8, 16, 32)
+
+// Provide matrix wrappers for full range bt.709
+#define F420ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F420ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+#define F422ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+#define F444ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuF709Constants, i, j)
+#define F444ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvF709Constants, i, j)
+
+// Provide matrix wrappers for full range bt.2020
+#define V420ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V420ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I420ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+#define V422ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V422ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I422ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+#define V444ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, e, f, c, d, g, h, &kYvuV2020Constants, i, j)
+#define V444ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I444ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvV2020Constants, i, j)
+
+#define I420ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I420ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I422ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I422ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
 
 #define ALIGNINT(V, ALIGN) (((V) + (ALIGN)-1) / (ALIGN) * (ALIGN))
 
 #define TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
                        YALIGN, W1280, N, NEG, OFF)                            \
   TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                       \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+    const int kWidth = W1280;                                                 \
     const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                  \
     const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                     \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
@@ -655,59 +731,123 @@ TESTBIPLANARTOP(NV21, 2, 2, I420, 2, 2)
     free_aligned_buffer_page_end(dst_argb_opt);                               \
   }
 
+#if defined(ENABLE_FULL_TESTS)
 #define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
                       YALIGN)                                                \
   TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                 YALIGN, benchmark_width_ - 4, _Any, +, 0)                   \
+                 YALIGN, benchmark_width_ + 1, _Any, +, 0)                   \
   TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                 YALIGN, benchmark_width_, _Unaligned, +, 1)                 \
+                 YALIGN, benchmark_width_, _Unaligned, +, 4)                 \
   TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
                  YALIGN, benchmark_width_, _Invert, -, 0)                    \
   TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
                  YALIGN, benchmark_width_, _Opt, +, 0)
+#else
+#define TESTPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                      YALIGN)                                                \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_ + 1, _Any, +, 0)                   \
+  TESTPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                 YALIGN, benchmark_width_, _Opt, +, 0)
+#endif
 
+#if defined(ENABLE_FULL_TESTS)
 TESTPLANARTOB(I420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1)
 TESTPLANARTOB(J420, 2, 2, ARGB, 4, 4, 1)
 TESTPLANARTOB(J420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(F420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(F420, 2, 2, ABGR, 4, 4, 1)
 TESTPLANARTOB(H420, 2, 2, ARGB, 4, 4, 1)
 TESTPLANARTOB(H420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(U420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(U420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(V420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(V420, 2, 2, ABGR, 4, 4, 1)
 TESTPLANARTOB(I420, 2, 2, BGRA, 4, 4, 1)
-TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1)
 TESTPLANARTOB(I420, 2, 2, RGBA, 4, 4, 1)
 TESTPLANARTOB(I420, 2, 2, RAW, 3, 3, 1)
 TESTPLANARTOB(I420, 2, 2, RGB24, 3, 3, 1)
+TESTPLANARTOB(J420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(J420, 2, 2, RGB24, 3, 3, 1)
 TESTPLANARTOB(H420, 2, 2, RAW, 3, 3, 1)
 TESTPLANARTOB(H420, 2, 2, RGB24, 3, 3, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
 TESTPLANARTOB(I420, 2, 2, RGB565, 2, 2, 1)
 TESTPLANARTOB(J420, 2, 2, RGB565, 2, 2, 1)
 TESTPLANARTOB(H420, 2, 2, RGB565, 2, 2, 1)
 TESTPLANARTOB(I420, 2, 2, ARGB1555, 2, 2, 1)
 TESTPLANARTOB(I420, 2, 2, ARGB4444, 2, 2, 1)
-TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1)
 TESTPLANARTOB(I422, 2, 1, RGB565, 2, 2, 1)
+#endif
+TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1)
 TESTPLANARTOB(J422, 2, 1, ARGB, 4, 4, 1)
 TESTPLANARTOB(J422, 2, 1, ABGR, 4, 4, 1)
 TESTPLANARTOB(H422, 2, 1, ARGB, 4, 4, 1)
 TESTPLANARTOB(H422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(U422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(U422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(V422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(V422, 2, 1, ABGR, 4, 4, 1)
 TESTPLANARTOB(I422, 2, 1, BGRA, 4, 4, 1)
-TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1)
 TESTPLANARTOB(I422, 2, 1, RGBA, 4, 4, 1)
 TESTPLANARTOB(I444, 1, 1, ARGB, 4, 4, 1)
-TESTPLANARTOB(J444, 1, 1, ARGB, 4, 4, 1)
 TESTPLANARTOB(I444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(J444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(J444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(H444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(H444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(U444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(U444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(V444, 1, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(V444, 1, 1, ABGR, 4, 4, 1)
 TESTPLANARTOB(I420, 2, 2, YUY2, 2, 4, 1)
 TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1)
 TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 1)
 TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 1)
 TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 1)
 TESTPLANARTOB(J420, 2, 2, J400, 1, 1, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
 TESTPLANARTOB(I420, 2, 2, AR30, 4, 4, 1)
 TESTPLANARTOB(H420, 2, 2, AR30, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, AB30, 4, 4, 1)
+TESTPLANARTOB(H420, 2, 2, AB30, 4, 4, 1)
+#endif
+TESTPLANARTOB(I420, 2, 2, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGBFilter, 4, 4, 1)
+#else
+TESTPLANARTOB(I420, 2, 2, ABGR, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, BGRA, 4, 4, 1)
+TESTPLANARTOB(I420, 2, 2, RAW, 3, 3, 1)
+TESTPLANARTOB(I420, 2, 2, RGB24, 3, 3, 1)
+TESTPLANARTOB(I420, 2, 2, RGBA, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOB(I420, 2, 2, RGB565, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB1555, 2, 2, 1)
+TESTPLANARTOB(I420, 2, 2, ARGB4444, 2, 2, 1)
+TESTPLANARTOB(I422, 2, 1, RGB565, 2, 2, 1)
+#endif
+TESTPLANARTOB(I420, 2, 2, I400, 1, 1, 1)
+TESTPLANARTOB(I420, 2, 2, UYVY, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, YUY2, 2, 4, 1)
+TESTPLANARTOB(I420, 2, 2, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGB, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, BGRA, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, RGBA, 4, 4, 1)
+TESTPLANARTOB(I422, 2, 1, UYVY, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, YUY2, 2, 4, 1)
+TESTPLANARTOB(I422, 2, 1, ARGBFilter, 4, 4, 1)
+TESTPLANARTOB(I444, 1, 1, ABGR, 4, 4, 1)
+TESTPLANARTOB(I444, 1, 1, ARGB, 4, 4, 1)
+#endif
 
 #define TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                        YALIGN, W1280, DIFF, N, NEG, OFF, ATTEN)               \
+                        YALIGN, W1280, N, NEG, OFF, ATTEN)                     \
   TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    const int kWidth = W1280;                                                  \
     const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
     const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
@@ -740,15 +880,9 @@ TESTPLANARTOB(H420, 2, 2, AR30, 4, 4, 1)
                             dst_argb_opt + OFF, kStrideB, kWidth, NEG kHeight, \
                             ATTEN);                                            \
     }                                                                          \
-    int max_diff = 0;                                                          \
     for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
-      int abs_diff = abs(static_cast<int>(dst_argb_c[i + OFF]) -               \
-                         static_cast<int>(dst_argb_opt[i + OFF]));             \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_opt[i + OFF]);                   \
     }                                                                          \
-    EXPECT_LE(max_diff, DIFF);                                                 \
     free_aligned_buffer_page_end(src_y);                                       \
     free_aligned_buffer_page_end(src_u);                                       \
     free_aligned_buffer_page_end(src_v);                                       \
@@ -757,26 +891,175 @@ TESTPLANARTOB(H420, 2, 2, AR30, 4, 4, 1)
     free_aligned_buffer_page_end(dst_argb_opt);                                \
   }
 
+#if defined(ENABLE_FULL_TESTS)
 #define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                       YALIGN, DIFF)                                          \
+                       YALIGN)                                                \
   TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                  YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, 0)          \
+                  YALIGN, benchmark_width_ + 1, _Any, +, 0, 0)                \
   TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                  YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, 0)        \
+                  YALIGN, benchmark_width_, _Unaligned, +, 2, 0)              \
   TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                  YALIGN, benchmark_width_, DIFF, _Invert, -, 0, 0)           \
+                  YALIGN, benchmark_width_, _Invert, -, 0, 0)                 \
   TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                  YALIGN, benchmark_width_, DIFF, _Opt, +, 0, 0)              \
+                  YALIGN, benchmark_width_, _Opt, +, 0, 0)                    \
+  TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Premult, +, 0, 1)
+#else
+#define TESTQPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN)                                                \
   TESTQPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                  YALIGN, benchmark_width_, DIFF, _Premult, +, 0, 1)
+                  YALIGN, benchmark_width_, _Opt, +, 0, 0)
+#endif
 
-TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1, 2)
-TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1, 2)
+#define J420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V420AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I420AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V420AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I420AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define J422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V422AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I422AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V422AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I422AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define J444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V444AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I444AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V444AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I444AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+
+#define I420AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I420AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+#define I422AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I422AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+
+#if defined(ENABLE_FULL_TESTS)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V420Alpha, 2, 2, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V422Alpha, 2, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(J444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(J444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(H444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(H444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(F444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(F444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(U444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(U444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(V444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(V444Alpha, 1, 1, ABGR, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGBFilter, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGBFilter, 4, 4, 1)
+#else
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I444Alpha, 1, 1, ARGB, 4, 4, 1)
+TESTQPLANARTOB(I420Alpha, 2, 2, ARGBFilter, 4, 4, 1)
+TESTQPLANARTOB(I422Alpha, 2, 1, ARGBFilter, 4, 4, 1)
+#endif
 
 #define TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C,       \
-                         BPP_B, W1280, DIFF, N, NEG, OFF)                      \
+                         BPP_B, W1280, N, NEG, OFF)                            \
   TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
     const int kStrideB = kWidth * BPP_B;                                       \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
@@ -813,18 +1096,12 @@ TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1, 2)
                   kHeight);                                                    \
     FMT_C##ToARGB(dst_argb_opt, kStrideB, dst_argb32_opt, kWidth * 4, kWidth,  \
                   kHeight);                                                    \
-    int max_diff = 0;                                                          \
     for (int i = 0; i < kHeight; ++i) {                                        \
       for (int j = 0; j < kWidth * 4; ++j) {                                   \
-        int abs_diff =                                                         \
-            abs(static_cast<int>(dst_argb32_c[i * kWidth * 4 + j]) -           \
-                static_cast<int>(dst_argb32_opt[i * kWidth * 4 + j]));         \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
-        }                                                                      \
+        EXPECT_EQ(dst_argb32_c[i * kWidth * 4 + j],                            \
+                  dst_argb32_opt[i * kWidth * 4 + j]);                         \
       }                                                                        \
     }                                                                          \
-    EXPECT_LE(max_diff, DIFF);                                                 \
     free_aligned_buffer_page_end(src_y);                                       \
     free_aligned_buffer_page_end(src_uv);                                      \
     free_aligned_buffer_page_end(dst_argb_c);                                  \
@@ -833,95 +1110,64 @@ TESTQPLANARTOB(I420Alpha, 2, 2, ABGR, 4, 4, 1, 2)
     free_aligned_buffer_page_end(dst_argb32_opt);                              \
   }
 
-#define TESTBIPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B, \
-                        DIFF)                                                  \
+#define TESTBIPLANARTOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B) \
   TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
-                   benchmark_width_ - 4, DIFF, _Any, +, 0)                     \
+                   benchmark_width_ + 1, _Any, +, 0)                           \
   TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
-                   benchmark_width_, DIFF, _Unaligned, +, 1)                   \
+                   benchmark_width_, _Unaligned, +, 2)                         \
   TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
-                   benchmark_width_, DIFF, _Invert, -, 0)                      \
+                   benchmark_width_, _Invert, -, 0)                            \
   TESTBIPLANARTOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, FMT_C, BPP_B,      \
-                   benchmark_width_, DIFF, _Opt, +, 0)
-
-TESTBIPLANARTOB(NV12, 2, 2, ARGB, ARGB, 4, 2)
-TESTBIPLANARTOB(NV21, 2, 2, ARGB, ARGB, 4, 2)
-TESTBIPLANARTOB(NV12, 2, 2, ABGR, ABGR, 4, 2)
-TESTBIPLANARTOB(NV21, 2, 2, ABGR, ABGR, 4, 2)
-TESTBIPLANARTOB(NV12, 2, 2, RGB24, RGB24, 3, 2)
-TESTBIPLANARTOB(NV21, 2, 2, RGB24, RGB24, 3, 2)
-TESTBIPLANARTOB(NV12, 2, 2, RAW, RAW, 3, 2)
-TESTBIPLANARTOB(NV21, 2, 2, RAW, RAW, 3, 2)
-TESTBIPLANARTOB(NV12, 2, 2, RGB565, RGB565, 2, 9)
-TESTBIPLANARTOB(NV21, 2, 2, YUV24, RAW, 3, 2)
-
-#ifdef DO_THREE_PLANES
-// Do 3 allocations for yuv.  conventional but slower.
-#define TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, \
-                       W1280, DIFF, N, NEG, OFF)                               \
-  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                        \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
-    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
-    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
-    const int kStride = (kStrideUV * SUBSAMP_X * 8 * BPP_A + 7) / 8;           \
-    align_buffer_page_end(src_argb, kStride* kHeight + OFF);                   \
-    align_buffer_page_end(dst_y_c, kWidth* kHeight);                           \
-    align_buffer_page_end(dst_u_c, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));  \
-    align_buffer_page_end(dst_v_c, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));  \
-    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                         \
-    align_buffer_page_end(dst_u_opt,                                           \
-                          kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));           \
-    align_buffer_page_end(dst_v_opt,                                           \
-                          kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));           \
-    memset(dst_y_c, 1, kWidth* kHeight);                                       \
-    memset(dst_u_c, 2, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));              \
-    memset(dst_v_c, 3, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));              \
-    memset(dst_y_opt, 101, kWidth* kHeight);                                   \
-    memset(dst_u_opt, 102, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));          \
-    memset(dst_v_opt, 103, kStrideUV* SUBSAMPLE(kHeight, SUBSAMP_Y));          \
-    for (int i = 0; i < kHeight; ++i)                                          \
-      for (int j = 0; j < kStride; ++j)                                        \
-        src_argb[(i * kStride) + j + OFF] = (fastrand() & 0xff);               \
-    MaskCpuFlags(disable_cpu_flags_);                                          \
-    FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_c, kWidth, dst_u_c,   \
-                          kStrideUV, dst_v_c, kStrideUV, kWidth, NEG kHeight); \
-    MaskCpuFlags(benchmark_cpu_info_);                                         \
-    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
-      FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_opt, kWidth,        \
-                            dst_u_opt, kStrideUV, dst_v_opt, kStrideUV,        \
-                            kWidth, NEG kHeight);                              \
-    }                                                                          \
-    for (int i = 0; i < kHeight; ++i) {                                        \
-      for (int j = 0; j < kWidth; ++j) {                                       \
-        EXPECT_NEAR(static_cast<int>(dst_y_c[i * kWidth + j]),                 \
-                    static_cast<int>(dst_y_opt[i * kWidth + j]), DIFF);        \
-      }                                                                        \
-    }                                                                          \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
-      for (int j = 0; j < kStrideUV; ++j) {                                    \
-        EXPECT_NEAR(static_cast<int>(dst_u_c[i * kStrideUV + j]),              \
-                    static_cast<int>(dst_u_opt[i * kStrideUV + j]), DIFF);     \
-      }                                                                        \
-    }                                                                          \
-    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                  \
-      for (int j = 0; j < kStrideUV; ++j) {                                    \
-        EXPECT_NEAR(static_cast<int>(dst_v_c[i * kStrideUV + j]),              \
-                    static_cast<int>(dst_v_opt[i * kStrideUV + j]), DIFF);     \
-      }                                                                        \
-    }                                                                          \
-    free_aligned_buffer_page_end(dst_y_c);                                     \
-    free_aligned_buffer_page_end(dst_u_c);                                     \
-    free_aligned_buffer_page_end(dst_v_c);                                     \
-    free_aligned_buffer_page_end(dst_y_opt);                                   \
-    free_aligned_buffer_page_end(dst_u_opt);                                   \
-    free_aligned_buffer_page_end(dst_v_opt);                                   \
-    free_aligned_buffer_page_end(src_argb);                                    \
-  }
-#else
+                   benchmark_width_, _Opt, +, 0)
+
+#define JNV12ToARGB(a, b, c, d, e, f, g, h) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV21ToARGB(a, b, c, d, e, f, g, h) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV12ToABGR(a, b, c, d, e, f, g, h) \
+  NV21ToARGBMatrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV21ToABGR(a, b, c, d, e, f, g, h) \
+  NV12ToARGBMatrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV12ToRGB24(a, b, c, d, e, f, g, h) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV21ToRGB24(a, b, c, d, e, f, g, h) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+#define JNV12ToRAW(a, b, c, d, e, f, g, h) \
+  NV21ToRGB24Matrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV21ToRAW(a, b, c, d, e, f, g, h) \
+  NV12ToRGB24Matrix(a, b, c, d, e, f, &kYvuJPEGConstants, g, h)
+#define JNV12ToRGB565(a, b, c, d, e, f, g, h) \
+  NV12ToRGB565Matrix(a, b, c, d, e, f, &kYuvJPEGConstants, g, h)
+
+TESTBIPLANARTOB(JNV12, 2, 2, ARGB, ARGB, 4)
+TESTBIPLANARTOB(JNV21, 2, 2, ARGB, ARGB, 4)
+TESTBIPLANARTOB(JNV12, 2, 2, ABGR, ABGR, 4)
+TESTBIPLANARTOB(JNV21, 2, 2, ABGR, ABGR, 4)
+TESTBIPLANARTOB(JNV12, 2, 2, RGB24, RGB24, 3)
+TESTBIPLANARTOB(JNV21, 2, 2, RGB24, RGB24, 3)
+TESTBIPLANARTOB(JNV12, 2, 2, RAW, RAW, 3)
+TESTBIPLANARTOB(JNV21, 2, 2, RAW, RAW, 3)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBIPLANARTOB(JNV12, 2, 2, RGB565, RGB565, 2)
+#endif
+
+TESTBIPLANARTOB(NV12, 2, 2, ARGB, ARGB, 4)
+TESTBIPLANARTOB(NV21, 2, 2, ARGB, ARGB, 4)
+TESTBIPLANARTOB(NV12, 2, 2, ABGR, ABGR, 4)
+TESTBIPLANARTOB(NV21, 2, 2, ABGR, ABGR, 4)
+TESTBIPLANARTOB(NV12, 2, 2, RGB24, RGB24, 3)
+TESTBIPLANARTOB(NV21, 2, 2, RGB24, RGB24, 3)
+TESTBIPLANARTOB(NV12, 2, 2, RAW, RAW, 3)
+TESTBIPLANARTOB(NV21, 2, 2, RAW, RAW, 3)
+TESTBIPLANARTOB(NV21, 2, 2, YUV24, RAW, 3)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBIPLANARTOB(NV12, 2, 2, RGB565, RGB565, 2)
+#endif
+
 #define TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, \
-                       W1280, DIFF, N, NEG, OFF)                               \
+                       W1280, N, NEG, OFF)                                     \
   TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                        \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    const int kWidth = W1280;                                                  \
     const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
     const int kStride = (kStrideUV * SUBSAMP_X * 8 * BPP_A + 7) / 8;           \
@@ -951,14 +1197,12 @@ TESTBIPLANARTOB(NV21, 2, 2, YUV24, RAW, 3, 2)
     }                                                                          \
     for (int i = 0; i < kHeight; ++i) {                                        \
       for (int j = 0; j < kWidth; ++j) {                                       \
-        EXPECT_NEAR(static_cast<int>(dst_y_c[i * kWidth + j]),                 \
-                    static_cast<int>(dst_y_opt[i * kWidth + j]), DIFF);        \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);         \
       }                                                                        \
     }                                                                          \
     for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y) * 2; ++i) {              \
       for (int j = 0; j < kStrideUV; ++j) {                                    \
-        EXPECT_NEAR(static_cast<int>(dst_uv_c[i * kStrideUV + j]),             \
-                    static_cast<int>(dst_uv_opt[i * kStrideUV + j]), DIFF);    \
+        EXPECT_EQ(dst_uv_c[i * kStrideUV + j], dst_uv_opt[i * kStrideUV + j]); \
       }                                                                        \
     }                                                                          \
     free_aligned_buffer_page_end(dst_y_c);                                     \
@@ -967,45 +1211,53 @@ TESTBIPLANARTOB(NV21, 2, 2, YUV24, RAW, 3, 2)
     free_aligned_buffer_page_end(dst_uv_opt);                                  \
     free_aligned_buffer_page_end(src_argb);                                    \
   }
-#endif
 
-#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, \
-                      DIFF)                                                   \
+#if defined(ENABLE_FULL_TESTS)
+#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
   TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                 benchmark_width_ - 4, DIFF, _Any, +, 0)                      \
+                 benchmark_width_ + 1, _Any, +, 0)                            \
   TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                 benchmark_width_, DIFF, _Unaligned, +, 1)                    \
+                 benchmark_width_, _Unaligned, +, 2)                          \
   TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                 benchmark_width_, DIFF, _Invert, -, 0)                       \
+                 benchmark_width_, _Invert, -, 0)                             \
   TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                 benchmark_width_, DIFF, _Opt, +, 0)
-
-TESTATOPLANAR(ABGR, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(ARGB, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(ARGB, 4, 1, I422, 2, 1, 2)
-TESTATOPLANAR(ARGB, 4, 1, I444, 1, 1, 2)
-TESTATOPLANAR(ARGB, 4, 1, J420, 2, 2, ARM_YUV_ERROR)
-TESTATOPLANAR(ARGB, 4, 1, J422, 2, 1, ARM_YUV_ERROR)
-TESTATOPLANAR(ARGB1555, 2, 1, I420, 2, 2, 15)
-TESTATOPLANAR(ARGB4444, 2, 1, I420, 2, 2, 17)
-TESTATOPLANAR(BGRA, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(I400, 1, 1, I420, 2, 2, 2)
-TESTATOPLANAR(J400, 1, 1, J420, 2, 2, 2)
-TESTATOPLANAR(RAW, 3, 1, I420, 2, 2, 4)
-TESTATOPLANAR(RGB24, 3, 1, I420, 2, 2, 4)
-// TODO(fbarchard): Investigate J420 error of 11 on Windows.
-TESTATOPLANAR(RGB24, 3, 1, J420, 2, 2, 11)
-TESTATOPLANAR(RGB565, 2, 1, I420, 2, 2, 5)
-TESTATOPLANAR(RGBA, 4, 1, I420, 2, 2, 4)
-TESTATOPLANAR(UYVY, 2, 1, I420, 2, 2, 2)
-TESTATOPLANAR(UYVY, 2, 1, I422, 2, 1, 2)
-TESTATOPLANAR(YUY2, 2, 1, I420, 2, 2, 2)
-TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1, 2)
+                 benchmark_width_, _Opt, +, 0)
+#else
+#define TESTATOPLANAR(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_ + 1, _Any, +, 0)                            \
+  TESTATOPLANARI(FMT_A, BPP_A, YALIGN, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                 benchmark_width_, _Opt, +, 0)
+#endif
+
+TESTATOPLANAR(ABGR, 4, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, I422, 2, 1)
+TESTATOPLANAR(ARGB, 4, 1, I444, 1, 1)
+TESTATOPLANAR(ARGB, 4, 1, J420, 2, 2)
+TESTATOPLANAR(ARGB, 4, 1, J422, 2, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOPLANAR(ARGB4444, 2, 1, I420, 2, 2)
+TESTATOPLANAR(RGB565, 2, 1, I420, 2, 2)
+TESTATOPLANAR(ARGB1555, 2, 1, I420, 2, 2)
+#endif
+TESTATOPLANAR(BGRA, 4, 1, I420, 2, 2)
+TESTATOPLANAR(I400, 1, 1, I420, 2, 2)
+TESTATOPLANAR(J400, 1, 1, J420, 2, 2)
+TESTATOPLANAR(RAW, 3, 1, I420, 2, 2)
+TESTATOPLANAR(RAW, 3, 1, J420, 2, 2)
+TESTATOPLANAR(RGB24, 3, 1, I420, 2, 2)
+TESTATOPLANAR(RGB24, 3, 1, J420, 2, 2)
+TESTATOPLANAR(RGBA, 4, 1, I420, 2, 2)
+TESTATOPLANAR(UYVY, 2, 1, I420, 2, 2)
+TESTATOPLANAR(UYVY, 2, 1, I422, 2, 1)
+TESTATOPLANAR(YUY2, 2, 1, I420, 2, 2)
+TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1)
 
 #define TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X,          \
                          SUBSAMP_Y, W1280, N, NEG, OFF)                       \
   TEST_F(LibYUVConvertTest, FMT_A##To##FMT_PLANAR##N) {                       \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
     const int kStride = SUBSAMPLE(kWidth, SUB_A) * BPP_A;                     \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
@@ -1031,28 +1283,17 @@ TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1, 2)
       FMT_A##To##FMT_PLANAR(src_argb + OFF, kStride, dst_y_opt, kWidth,       \
                             dst_uv_opt, kStrideUV * 2, kWidth, NEG kHeight);  \
     }                                                                         \
-    int max_diff = 0;                                                         \
     for (int i = 0; i < kHeight; ++i) {                                       \
       for (int j = 0; j < kWidth; ++j) {                                      \
-        int abs_diff = abs(static_cast<int>(dst_y_c[i * kWidth + j]) -        \
-                           static_cast<int>(dst_y_opt[i * kWidth + j]));      \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
       }                                                                       \
     }                                                                         \
-    EXPECT_LE(max_diff, 4);                                                   \
     for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
       for (int j = 0; j < kStrideUV * 2; ++j) {                               \
-        int abs_diff =                                                        \
-            abs(static_cast<int>(dst_uv_c[i * kStrideUV * 2 + j]) -           \
-                static_cast<int>(dst_uv_opt[i * kStrideUV * 2 + j]));         \
-        if (abs_diff > max_diff) {                                            \
-          max_diff = abs_diff;                                                \
-        }                                                                     \
+        EXPECT_EQ(dst_uv_c[i * kStrideUV * 2 + j],                            \
+                  dst_uv_opt[i * kStrideUV * 2 + j]);                         \
       }                                                                       \
     }                                                                         \
-    EXPECT_LE(max_diff, 4);                                                   \
     free_aligned_buffer_page_end(dst_y_c);                                    \
     free_aligned_buffer_page_end(dst_uv_c);                                   \
     free_aligned_buffer_page_end(dst_y_opt);                                  \
@@ -1062,9 +1303,9 @@ TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1, 2)
 
 #define TESTATOBIPLANAR(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y) \
   TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                   benchmark_width_ - 4, _Any, +, 0)                           \
+                   benchmark_width_ + 1, _Any, +, 0)                           \
   TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
-                   benchmark_width_, _Unaligned, +, 1)                         \
+                   benchmark_width_, _Unaligned, +, 2)                         \
   TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
                    benchmark_width_, _Invert, -, 0)                            \
   TESTATOBIPLANARI(FMT_A, SUB_A, BPP_A, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
@@ -1072,146 +1313,181 @@ TESTATOPLANAR(YUY2, 2, 1, I422, 2, 1, 2)
 
 TESTATOBIPLANAR(ARGB, 1, 4, NV12, 2, 2)
 TESTATOBIPLANAR(ARGB, 1, 4, NV21, 2, 2)
+TESTATOBIPLANAR(ABGR, 1, 4, NV12, 2, 2)
+TESTATOBIPLANAR(ABGR, 1, 4, NV21, 2, 2)
+TESTATOBIPLANAR(RAW, 1, 3, JNV21, 2, 2)
 TESTATOBIPLANAR(YUY2, 2, 4, NV12, 2, 2)
 TESTATOBIPLANAR(UYVY, 2, 4, NV12, 2, 2)
 TESTATOBIPLANAR(AYUV, 1, 4, NV12, 2, 2)
 TESTATOBIPLANAR(AYUV, 1, 4, NV21, 2, 2)
 
-#define TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,  \
-                  HEIGHT_B, W1280, DIFF, N, NEG, OFF)                        \
-  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) {                           \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                          \
-    const int kHeight = benchmark_height_;                                   \
-    const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
-    const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
-    const int kStrideA =                                                     \
-        (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;               \
-    const int kStrideB =                                                     \
-        (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;               \
-    align_buffer_page_end(src_argb, kStrideA* kHeightA + OFF);               \
-    align_buffer_page_end(dst_argb_c, kStrideB* kHeightB);                   \
-    align_buffer_page_end(dst_argb_opt, kStrideB* kHeightB);                 \
-    for (int i = 0; i < kStrideA * kHeightA; ++i) {                          \
-      src_argb[i + OFF] = (fastrand() & 0xff);                               \
-    }                                                                        \
-    memset(dst_argb_c, 1, kStrideB* kHeightB);                               \
-    memset(dst_argb_opt, 101, kStrideB* kHeightB);                           \
-    MaskCpuFlags(disable_cpu_flags_);                                        \
-    FMT_A##To##FMT_B(src_argb + OFF, kStrideA, dst_argb_c, kStrideB, kWidth, \
-                     NEG kHeight);                                           \
-    MaskCpuFlags(benchmark_cpu_info_);                                       \
-    for (int i = 0; i < benchmark_iterations_; ++i) {                        \
-      FMT_A##To##FMT_B(src_argb + OFF, kStrideA, dst_argb_opt, kStrideB,     \
-                       kWidth, NEG kHeight);                                 \
-    }                                                                        \
-    int max_diff = 0;                                                        \
-    for (int i = 0; i < kStrideB * kHeightB; ++i) {                          \
-      int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -                   \
-                         static_cast<int>(dst_argb_opt[i]));                 \
-      if (abs_diff > max_diff) {                                             \
-        max_diff = abs_diff;                                                 \
-      }                                                                      \
-    }                                                                        \
-    EXPECT_LE(max_diff, DIFF);                                               \
-    free_aligned_buffer_page_end(src_argb);                                  \
-    free_aligned_buffer_page_end(dst_argb_c);                                \
-    free_aligned_buffer_page_end(dst_argb_opt);                              \
+#define TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,     \
+                  EPP_B, STRIDE_B, HEIGHT_B, W1280, N, NEG, OFF)               \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##N) {                             \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = benchmark_height_;                                     \
+    const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;       \
+    const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;       \
+    const int kStrideA =                                                       \
+        (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                 \
+    const int kStrideB =                                                       \
+        (kWidth * EPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;                 \
+    align_buffer_page_end(src_argb,                                            \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeightB*(int)sizeof(TYPE_B)); \
+    align_buffer_page_end(dst_argb_opt,                                        \
+                          kStrideB* kHeightB*(int)sizeof(TYPE_B));             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
+      src_argb[i + OFF] = (fastrand() & 0xff);                                 \
+    }                                                                          \
+    memset(dst_argb_c, 1, kStrideB* kHeightB);                                 \
+    memset(dst_argb_opt, 101, kStrideB* kHeightB);                             \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_A##To##FMT_B((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_B*)dst_argb_c, \
+                     kStrideB, kWidth, NEG kHeight);                           \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_A##To##FMT_B((TYPE_A*)(src_argb + OFF), kStrideA,                    \
+                       (TYPE_B*)dst_argb_opt, kStrideB, kWidth, NEG kHeight);  \
+    }                                                                          \
+    for (int i = 0; i < kStrideB * kHeightB * (int)sizeof(TYPE_B); ++i) {      \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                               \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_argb);                                    \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
   }
 
-#define TESTATOBRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B,     \
-                       STRIDE_B, HEIGHT_B, DIFF)                           \
-  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##_Random) {                   \
-    for (int times = 0; times < benchmark_iterations_; ++times) {          \
-      const int kWidth = (fastrand() & 63) + 1;                            \
-      const int kHeight = (fastrand() & 31) + 1;                           \
-      const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A; \
-      const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B; \
-      const int kStrideA =                                                 \
-          (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;           \
-      const int kStrideB =                                                 \
-          (kWidth * BPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;           \
-      align_buffer_page_end(src_argb, kStrideA* kHeightA);                 \
-      align_buffer_page_end(dst_argb_c, kStrideB* kHeightB);               \
-      align_buffer_page_end(dst_argb_opt, kStrideB* kHeightB);             \
-      for (int i = 0; i < kStrideA * kHeightA; ++i) {                      \
-        src_argb[i] = (fastrand() & 0xff);                                 \
-      }                                                                    \
-      memset(dst_argb_c, 123, kStrideB* kHeightB);                         \
-      memset(dst_argb_opt, 123, kStrideB* kHeightB);                       \
-      MaskCpuFlags(disable_cpu_flags_);                                    \
-      FMT_A##To##FMT_B(src_argb, kStrideA, dst_argb_c, kStrideB, kWidth,   \
-                       kHeight);                                           \
-      MaskCpuFlags(benchmark_cpu_info_);                                   \
-      FMT_A##To##FMT_B(src_argb, kStrideA, dst_argb_opt, kStrideB, kWidth, \
-                       kHeight);                                           \
-      for (int i = 0; i < kStrideB * kHeightB; ++i) {                      \
-        EXPECT_NEAR(dst_argb_c[i], dst_argb_opt[i], DIFF);                 \
-      }                                                                    \
-      free_aligned_buffer_page_end(src_argb);                              \
-      free_aligned_buffer_page_end(dst_argb_c);                            \
-      free_aligned_buffer_page_end(dst_argb_opt);                          \
-    }                                                                      \
-  }
-
-#define TESTATOB(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
-                 HEIGHT_B, DIFF)                                           \
-  TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-            HEIGHT_B, benchmark_width_ - 4, DIFF, _Any, +, 0)              \
-  TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-            HEIGHT_B, benchmark_width_, DIFF, _Unaligned, +, 1)            \
-  TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-            HEIGHT_B, benchmark_width_, DIFF, _Invert, -, 0)               \
-  TESTATOBI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-            HEIGHT_B, benchmark_width_, DIFF, _Opt, +, 0)                  \
-  TESTATOBRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
-                 HEIGHT_B, DIFF)
-
-// TODO(fbarchard): make ARM version of C code that matches NEON.
-TESTATOB(AB30, 4, 4, 1, ABGR, 4, 4, 1, 0)
-TESTATOB(AB30, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ABGR, 4, 4, 1, AR30, 4, 4, 1, 0)
-TESTATOB(ABGR, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(AR30, 4, 4, 1, AB30, 4, 4, 1, 0)
-TESTATOB(AR30, 4, 4, 1, ABGR, 4, 4, 1, 0)
-TESTATOB(AR30, 4, 4, 1, AR30, 4, 4, 1, 0)
-TESTATOB(AR30, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ABGR, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, AR30, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGB1555, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGB4444, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, ARGBMirror, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, BGRA, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, I400, 1, 1, 1, 2)
-TESTATOB(ARGB, 4, 4, 1, J400, 1, 1, 1, 2)
-TESTATOB(ARGB, 4, 4, 1, RAW, 3, 3, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGB24, 3, 3, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGB565, 2, 2, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, RGBA, 4, 4, 1, 0)
-TESTATOB(ARGB, 4, 4, 1, UYVY, 2, 4, 1, 4)
-TESTATOB(ARGB, 4, 4, 1, YUY2, 2, 4, 1, 4)
-TESTATOB(ARGB1555, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(ARGB4444, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(BGRA, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(I400, 1, 1, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(I400, 1, 1, 1, I400, 1, 1, 1, 0)
-TESTATOB(I400, 1, 1, 1, I400Mirror, 1, 1, 1, 0)
-TESTATOB(J400, 1, 1, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(J400, 1, 1, 1, J400, 1, 1, 1, 0)
-TESTATOB(RAW, 3, 3, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RAW, 3, 3, 1, RGB24, 3, 3, 1, 0)
-TESTATOB(RGB24, 3, 3, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RGB24, 3, 3, 1, J400, 1, 1, 1, 0)
-TESTATOB(RGB565, 2, 2, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(RGBA, 4, 4, 1, ARGB, 4, 4, 1, 0)
-TESTATOB(UYVY, 2, 4, 1, ARGB, 4, 4, 1, ARM_YUV_ERROR)
-TESTATOB(YUY2, 2, 4, 1, ARGB, 4, 4, 1, ARM_YUV_ERROR)
-TESTATOB(YUY2, 2, 4, 1, Y, 1, 1, 1, 0)
+#define TESTATOBRANDOM(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B,        \
+                       TYPE_B, EPP_B, STRIDE_B, HEIGHT_B)                      \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##_Random) {                       \
+    for (int times = 0; times < benchmark_iterations_; ++times) {              \
+      const int kWidth = (fastrand() & 63) + 1;                                \
+      const int kHeight = (fastrand() & 31) + 1;                               \
+      const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
+      const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
+      const int kStrideA =                                                     \
+          (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;               \
+      const int kStrideB =                                                     \
+          (kWidth * EPP_B + STRIDE_B - 1) / STRIDE_B * STRIDE_B;               \
+      align_buffer_page_end(src_argb, kStrideA* kHeightA*(int)sizeof(TYPE_A)); \
+      align_buffer_page_end(dst_argb_c,                                        \
+                            kStrideB* kHeightB*(int)sizeof(TYPE_B));           \
+      align_buffer_page_end(dst_argb_opt,                                      \
+                            kStrideB* kHeightB*(int)sizeof(TYPE_B));           \
+      for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {    \
+        src_argb[i] = 0xfe;                                                    \
+      }                                                                        \
+      memset(dst_argb_c, 123, kStrideB* kHeightB);                             \
+      memset(dst_argb_opt, 123, kStrideB* kHeightB);                           \
+      MaskCpuFlags(disable_cpu_flags_);                                        \
+      FMT_A##To##FMT_B((TYPE_A*)src_argb, kStrideA, (TYPE_B*)dst_argb_c,       \
+                       kStrideB, kWidth, kHeight);                             \
+      MaskCpuFlags(benchmark_cpu_info_);                                       \
+      FMT_A##To##FMT_B((TYPE_A*)src_argb, kStrideA, (TYPE_B*)dst_argb_opt,     \
+                       kStrideB, kWidth, kHeight);                             \
+      for (int i = 0; i < kStrideB * kHeightB * (int)sizeof(TYPE_B); ++i) {    \
+        EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
+      }                                                                        \
+      free_aligned_buffer_page_end(src_argb);                                  \
+      free_aligned_buffer_page_end(dst_argb_c);                                \
+      free_aligned_buffer_page_end(dst_argb_opt);                              \
+    }                                                                          \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+#define TESTATOB(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)                                 \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_ + 1, _Any, +, 0)           \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Unaligned, +, 4)         \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Invert, -, 0)            \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Opt, +, 0)               \
+  TESTATOBRANDOM(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)
+#else
+#define TESTATOB(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B,   \
+                 EPP_B, STRIDE_B, HEIGHT_B)                                 \
+  TESTATOBI(FMT_A, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, FMT_B, TYPE_B, EPP_B, \
+            STRIDE_B, HEIGHT_B, benchmark_width_, _Opt, +, 0)
+#endif
+
+TESTATOB(AB30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AB30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ABGR, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(AR30, uint8_t, 4, 4, 1, AB30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(AR30, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(AR30, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+TESTATOB(AR30, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AR30, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB1555, uint8_t, 2, 2, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGB4444, uint8_t, 2, 2, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, ARGBMirror, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, BGRA, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, I400, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RGBA, uint8_t, 4, 4, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, RAW, uint8_t, 3, 3, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, RGB24, uint8_t, 3, 3, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGB565, uint8_t, 2, 2, 1)
+#endif
+TESTATOB(ARGB, uint8_t, 4, 4, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, UYVY, uint8_t, 2, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, YUY2, uint8_t, 2, 4, 1)  // 4
+TESTATOB(ARGB1555, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(ARGB4444, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(BGRA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, I400, uint8_t, 1, 1, 1)
+TESTATOB(I400, uint8_t, 1, 1, 1, I400Mirror, uint8_t, 1, 1, 1)
+TESTATOB(J400, uint8_t, 1, 1, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(J400, uint8_t, 1, 1, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, RGBA, uint8_t, 4, 4, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, RGB24, uint8_t, 3, 3, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+TESTATOB(RGB24, uint8_t, 3, 3, 1, RGB24Mirror, uint8_t, 3, 3, 1)
+TESTATOB(RAW, uint8_t, 3, 3, 1, J400, uint8_t, 1, 1, 1)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOB(RGB565, uint8_t, 2, 2, 1, ARGB, uint8_t, 4, 4, 1)
+#endif
+TESTATOB(RGBA, uint8_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(UYVY, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(YUY2, uint8_t, 2, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(YUY2, uint8_t, 2, 4, 1, Y, uint8_t, 1, 1, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+TESTATOB(ARGB, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
+TESTATOB(ABGR, uint8_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, ARGB, uint8_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, ABGR, uint8_t, 4, 4, 1)
+TESTATOB(AR64, uint16_t, 4, 4, 1, AB64, uint16_t, 4, 4, 1)
+TESTATOB(AB64, uint16_t, 4, 4, 1, AR64, uint16_t, 4, 4, 1)
 
 #define TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
-                   HEIGHT_B, W1280, DIFF, N, NEG, OFF)                       \
+                   HEIGHT_B, W1280, N, NEG, OFF)                             \
   TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither##N) {                   \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                          \
+    const int kWidth = W1280;                                                \
     const int kHeight = benchmark_height_;                                   \
     const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;     \
     const int kHeightB = (kHeight + HEIGHT_B - 1) / HEIGHT_B * HEIGHT_B;     \
@@ -1235,22 +1511,16 @@ TESTATOB(YUY2, 2, 4, 1, Y, 1, 1, 1, 0)
       FMT_A##To##FMT_B##Dither(src_argb + OFF, kStrideA, dst_argb_opt,       \
                                kStrideB, NULL, kWidth, NEG kHeight);         \
     }                                                                        \
-    int max_diff = 0;                                                        \
     for (int i = 0; i < kStrideB * kHeightB; ++i) {                          \
-      int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -                   \
-                         static_cast<int>(dst_argb_opt[i]));                 \
-      if (abs_diff > max_diff) {                                             \
-        max_diff = abs_diff;                                                 \
-      }                                                                      \
+      EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
     }                                                                        \
-    EXPECT_LE(max_diff, DIFF);                                               \
     free_aligned_buffer_page_end(src_argb);                                  \
     free_aligned_buffer_page_end(dst_argb_c);                                \
     free_aligned_buffer_page_end(dst_argb_opt);                              \
   }
 
 #define TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B,        \
-                        STRIDE_B, HEIGHT_B, DIFF)                              \
+                        STRIDE_B, HEIGHT_B)                                    \
   TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##Dither_Random) {                 \
     for (int times = 0; times < benchmark_iterations_; ++times) {              \
       const int kWidth = (fastrand() & 63) + 1;                                \
@@ -1275,15 +1545,9 @@ TESTATOB(YUY2, 2, 4, 1, Y, 1, 1, 1, 0)
       MaskCpuFlags(benchmark_cpu_info_);                                       \
       FMT_A##To##FMT_B##Dither(src_argb, kStrideA, dst_argb_opt, kStrideB,     \
                                NULL, kWidth, kHeight);                         \
-      int max_diff = 0;                                                        \
       for (int i = 0; i < kStrideB * kHeightB; ++i) {                          \
-        int abs_diff = abs(static_cast<int>(dst_argb_c[i]) -                   \
-                           static_cast<int>(dst_argb_opt[i]));                 \
-        if (abs_diff > max_diff) {                                             \
-          max_diff = abs_diff;                                                 \
-        }                                                                      \
+        EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                             \
       }                                                                        \
-      EXPECT_LE(max_diff, DIFF);                                               \
       free_aligned_buffer_page_end(src_argb);                                  \
       free_aligned_buffer_page_end(dst_argb_c);                                \
       free_aligned_buffer_page_end(dst_argb_opt);                              \
@@ -1291,49 +1555,57 @@ TESTATOB(YUY2, 2, 4, 1, Y, 1, 1, 1, 0)
   }
 
 #define TESTATOBD(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
-                  HEIGHT_B, DIFF)                                           \
+                  HEIGHT_B)                                                 \
   TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-             HEIGHT_B, benchmark_width_ - 4, DIFF, _Any, +, 0)              \
+             HEIGHT_B, benchmark_width_ + 1, _Any, +, 0)                    \
   TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-             HEIGHT_B, benchmark_width_, DIFF, _Unaligned, +, 1)            \
+             HEIGHT_B, benchmark_width_, _Unaligned, +, 2)                  \
   TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-             HEIGHT_B, benchmark_width_, DIFF, _Invert, -, 0)               \
+             HEIGHT_B, benchmark_width_, _Invert, -, 0)                     \
   TESTATOBDI(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B,      \
-             HEIGHT_B, benchmark_width_, DIFF, _Opt, +, 0)                  \
+             HEIGHT_B, benchmark_width_, _Opt, +, 0)                        \
   TESTATOBDRANDOM(FMT_A, BPP_A, STRIDE_A, HEIGHT_A, FMT_B, BPP_B, STRIDE_B, \
-                  HEIGHT_B, DIFF)
+                  HEIGHT_B)
 
-TESTATOBD(ARGB, 4, 4, 1, RGB565, 2, 2, 1, 0)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTATOBD(ARGB, 4, 4, 1, RGB565, 2, 2, 1)
+#endif
 
-#define TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A, W1280, N, NEG, OFF)      \
-  TEST_F(LibYUVConvertTest, FMT_ATOB##_Symetric##N) {                          \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+// These conversions called twice, produce the original result.
+// e.g. endian swap twice.
+#define TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, W1280, N, NEG,   \
+                 OFF)                                                          \
+  TEST_F(LibYUVConvertTest, FMT_ATOB##_Endswap##N) {                           \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
     const int kHeightA = (kHeight + HEIGHT_A - 1) / HEIGHT_A * HEIGHT_A;       \
     const int kStrideA =                                                       \
-        (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                 \
-    align_buffer_page_end(src_argb, kStrideA* kHeightA + OFF);                 \
-    align_buffer_page_end(dst_argb_c, kStrideA* kHeightA);                     \
-    align_buffer_page_end(dst_argb_opt, kStrideA* kHeightA);                   \
-    for (int i = 0; i < kStrideA * kHeightA; ++i) {                            \
+        (kWidth * EPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                 \
+    align_buffer_page_end(src_argb,                                            \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A) + OFF);       \
+    align_buffer_page_end(dst_argb_c, kStrideA* kHeightA*(int)sizeof(TYPE_A)); \
+    align_buffer_page_end(dst_argb_opt,                                        \
+                          kStrideA* kHeightA*(int)sizeof(TYPE_A));             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
       src_argb[i + OFF] = (fastrand() & 0xff);                                 \
     }                                                                          \
     memset(dst_argb_c, 1, kStrideA* kHeightA);                                 \
     memset(dst_argb_opt, 101, kStrideA* kHeightA);                             \
     MaskCpuFlags(disable_cpu_flags_);                                          \
-    FMT_ATOB(src_argb + OFF, kStrideA, dst_argb_c, kStrideA, kWidth,           \
-             NEG kHeight);                                                     \
+    FMT_ATOB((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_A*)dst_argb_c,         \
+             kStrideA, kWidth, NEG kHeight);                                   \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
     for (int i = 0; i < benchmark_iterations_; ++i) {                          \
-      FMT_ATOB(src_argb + OFF, kStrideA, dst_argb_opt, kStrideA, kWidth,       \
-               NEG kHeight);                                                   \
+      FMT_ATOB((TYPE_A*)(src_argb + OFF), kStrideA, (TYPE_A*)dst_argb_opt,     \
+               kStrideA, kWidth, NEG kHeight);                                 \
     }                                                                          \
     MaskCpuFlags(disable_cpu_flags_);                                          \
-    FMT_ATOB(dst_argb_c, kStrideA, dst_argb_c, kStrideA, kWidth, NEG kHeight); \
+    FMT_ATOB((TYPE_A*)dst_argb_c, kStrideA, (TYPE_A*)dst_argb_c, kStrideA,     \
+             kWidth, NEG kHeight);                                             \
     MaskCpuFlags(benchmark_cpu_info_);                                         \
-    FMT_ATOB(dst_argb_opt, kStrideA, dst_argb_opt, kStrideA, kWidth,           \
-             NEG kHeight);                                                     \
-    for (int i = 0; i < kStrideA * kHeightA; ++i) {                            \
+    FMT_ATOB((TYPE_A*)dst_argb_opt, kStrideA, (TYPE_A*)dst_argb_opt, kStrideA, \
+             kWidth, NEG kHeight);                                             \
+    for (int i = 0; i < kStrideA * kHeightA * (int)sizeof(TYPE_A); ++i) {      \
       EXPECT_EQ(src_argb[i + OFF], dst_argb_opt[i]);                           \
       EXPECT_EQ(dst_argb_c[i], dst_argb_opt[i]);                               \
     }                                                                          \
@@ -1342,32 +1614,25 @@ TESTATOBD(ARGB, 4, 4, 1, RGB565, 2, 2, 1, 0)
     free_aligned_buffer_page_end(dst_argb_opt);                                \
   }
 
-#define TESTSYM(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A)                           \
-  TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A, benchmark_width_ - 4, _Any, +, \
-           0)                                                                  \
-  TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A, benchmark_width_, _Unaligned,  \
-           +, 1)                                                               \
-  TESTSYMI(FMT_ATOB, BPP_A, STRIDE_A, HEIGHT_A, benchmark_width_, _Opt, +, 0)
-
-TESTSYM(ARGBToARGB, 4, 4, 1)
-TESTSYM(ARGBToBGRA, 4, 4, 1)
-TESTSYM(ARGBToABGR, 4, 4, 1)
-TESTSYM(BGRAToARGB, 4, 4, 1)
-TESTSYM(ABGRToARGB, 4, 4, 1)
+#if defined(ENABLE_FULL_TESTS)
+#define TESTEND(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A)                  \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_ + 1, \
+           _Any, +, 0)                                                        \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_,     \
+           _Unaligned, +, 2)                                                  \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_,     \
+           _Opt, +, 0)
+#else
+#define TESTEND(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A)              \
+  TESTENDI(FMT_ATOB, TYPE_A, EPP_A, STRIDE_A, HEIGHT_A, benchmark_width_, \
+           _Opt, +, 0)
+#endif
 
-TEST_F(LibYUVConvertTest, Test565) {
-  SIMD_ALIGNED(uint8_t orig_pixels[256][4]);
-  SIMD_ALIGNED(uint8_t pixels565[256][2]);
-
-  for (int i = 0; i < 256; ++i) {
-    for (int j = 0; j < 4; ++j) {
-      orig_pixels[i][j] = i;
-    }
-  }
-  ARGBToRGB565(&orig_pixels[0][0], 0, &pixels565[0][0], 0, 256, 1);
-  uint32_t checksum = HashDjb2(&pixels565[0][0], sizeof(pixels565), 5381);
-  EXPECT_EQ(610919429u, checksum);
-}
+TESTEND(ARGBToBGRA, uint8_t, 4, 4, 1)
+TESTEND(ARGBToABGR, uint8_t, 4, 4, 1)
+TESTEND(BGRAToARGB, uint8_t, 4, 4, 1)
+TESTEND(ABGRToARGB, uint8_t, 4, 4, 1)
+TESTEND(AB64ToAR64, uint16_t, 4, 4, 1)
 
 #ifdef HAVE_JPEG
 TEST_F(LibYUVConvertTest, ValidateJpeg) {
@@ -1883,6 +2148,65 @@ TEST_F(LibYUVConvertTest, TestMJPGToI420_NV21) {
   free_aligned_buffer_page_end(dst_vu);
 }
 
+TEST_F(LibYUVConvertTest, TestMJPGToI420_NV12) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  // Convert to NV12
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert to I420
+  align_buffer_page_end(dst2_y, width * height);
+  align_buffer_page_end(dst2_u, half_width * half_height);
+  align_buffer_page_end(dst2_v, half_width * half_height);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToI420(kTest2Jpg, kTest2JpgLen, dst2_y, width, dst2_u, half_width,
+                     dst2_v, half_width, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Convert I420 to NV12
+  align_buffer_page_end(dst3_y, width * height);
+  align_buffer_page_end(dst3_uv, half_width * half_height * 2);
+
+  I420ToNV12(dst2_y, width, dst2_u, half_width, dst2_v, half_width, dst3_y,
+             width, dst3_uv, half_width * 2, width, height);
+
+  for (int i = 0; i < width * height; ++i) {
+    EXPECT_EQ(dst_y[i], dst3_y[i]);
+  }
+  for (int i = 0; i < half_width * half_height * 2; ++i) {
+    EXPECT_EQ(dst_uv[i], dst3_uv[i]);
+    EXPECT_EQ(dst_uv[i], dst3_uv[i]);
+  }
+
+  free_aligned_buffer_page_end(dst3_y);
+  free_aligned_buffer_page_end(dst3_uv);
+
+  free_aligned_buffer_page_end(dst2_y);
+  free_aligned_buffer_page_end(dst2_u);
+  free_aligned_buffer_page_end(dst2_v);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
 TEST_F(LibYUVConvertTest, TestMJPGToNV21_420) {
   int width = 0;
   int height = 0;
@@ -1913,7 +2237,42 @@ TEST_F(LibYUVConvertTest, TestMJPGToNV21_420) {
   free_aligned_buffer_page_end(dst_uv);
 }
 
-TEST_F(LibYUVConvertTest, TestMJPGToNV21_422) {
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_420) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest2Jpg, kTest2JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest2Jpg, kTest2JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 1069662856u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
+// TODO(fbarchard): Improve test to compare against I422, not checksum
+TEST_F(LibYUVConvertTest, DISABLED_TestMJPGToNV21_422) {
   int width = 0;
   int height = 0;
   int ret = MJPGSize(kTest3Jpg, kTest3JpgLen, &width, &height);
@@ -1937,10 +2296,44 @@ TEST_F(LibYUVConvertTest, TestMJPGToNV21_422) {
   uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
   uint32_t dst_uv_hash = HashDjb2(dst_uv, half_width * half_height * 2, 5381);
   EXPECT_EQ(dst_y_hash, 2682851208u);
-  EXPECT_EQ(dst_uv_hash, 3543430771u);
+  EXPECT_EQ(dst_uv_hash, 493520167u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVConvertTest, DISABLED_TestMJPGToNV12_422) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest3Jpg, kTest3JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest3Jpg, kTest3JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 493520167u);
 
   free_aligned_buffer_page_end(dst_y);
   free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
 }
 
 TEST_F(LibYUVConvertTest, TestMJPGToNV21_400) {
@@ -1973,6 +2366,40 @@ TEST_F(LibYUVConvertTest, TestMJPGToNV21_400) {
   free_aligned_buffer_page_end(dst_uv);
 }
 
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_400) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest0Jpg, kTest0JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest0Jpg, kTest0JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 330644005u);
+  EXPECT_EQ(dst_vu_hash, 135214341u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
 TEST_F(LibYUVConvertTest, TestMJPGToNV21_444) {
   int width = 0;
   int height = 0;
@@ -2003,6 +2430,40 @@ TEST_F(LibYUVConvertTest, TestMJPGToNV21_444) {
   free_aligned_buffer_page_end(dst_uv);
 }
 
+TEST_F(LibYUVConvertTest, TestMJPGToNV12_444) {
+  int width = 0;
+  int height = 0;
+  int ret = MJPGSize(kTest1Jpg, kTest1JpgLen, &width, &height);
+  EXPECT_EQ(0, ret);
+
+  int half_width = (width + 1) / 2;
+  int half_height = (height + 1) / 2;
+  int benchmark_iterations = benchmark_iterations_ * benchmark_width_ *
+                             benchmark_height_ / (width * height);
+
+  align_buffer_page_end(dst_y, width * height);
+  align_buffer_page_end(dst_uv, half_width * half_height * 2);
+  for (int times = 0; times < benchmark_iterations; ++times) {
+    ret = MJPGToNV12(kTest1Jpg, kTest1JpgLen, dst_y, width, dst_uv,
+                     half_width * 2, width, height, width, height);
+  }
+  // Expect sucesss
+  EXPECT_EQ(0, ret);
+
+  // Test result matches known hash value. Hashes are for VU so flip the plane.
+  uint32_t dst_y_hash = HashDjb2(dst_y, width * height, 5381);
+  align_buffer_page_end(dst_vu, half_width * half_height * 2);
+  SwapUVPlane(dst_uv, half_width * 2, dst_vu, half_width * 2, half_width,
+              half_height);
+  uint32_t dst_vu_hash = HashDjb2(dst_vu, half_width * half_height * 2, 5381);
+  EXPECT_EQ(dst_y_hash, 2682851208u);
+  EXPECT_EQ(dst_vu_hash, 506143297u);
+
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+  free_aligned_buffer_page_end(dst_vu);
+}
+
 TEST_F(LibYUVConvertTest, TestMJPGToARGB) {
   int width = 0;
   int height = 0;
@@ -2022,7 +2483,11 @@ TEST_F(LibYUVConvertTest, TestMJPGToARGB) {
 
   // Test result matches known hash value.
   uint32_t dst_argb_hash = HashDjb2(dst_argb, width * height, 5381);
+#ifdef LIBYUV_UNLIMITED_DATA
+  EXPECT_EQ(dst_argb_hash, 3900633302u);
+#else
   EXPECT_EQ(dst_argb_hash, 2355976473u);
+#endif
 
   free_aligned_buffer_page_end(dst_argb);
 }
@@ -2178,7 +2643,7 @@ TEST_F(LibYUVConvertTest, I420CropOddY) {
   const int SUBSAMP_Y = 2;
   const int kWidth = benchmark_width_;
   const int kHeight = benchmark_height_;
-  const int crop_y = 1;
+  const int crop_y = benchmark_height_ > 1 ? 1 : 0;
   const int kDestWidth = benchmark_width_;
   const int kDestHeight = benchmark_height_ - crop_y * 2;
   const int kStrideU = SUBSAMPLE(kWidth, SUBSAMP_X);
@@ -2329,9 +2794,9 @@ TEST_F(LibYUVConvertTest, TestDither) {
 }
 
 #define TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                        YALIGN, W1280, DIFF, N, NEG, OFF, FMT_C, BPP_C)        \
+                        YALIGN, W1280, N, NEG, OFF, FMT_C, BPP_C)              \
   TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##Dither##N) {                \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+    const int kWidth = W1280;                                                  \
     const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
     const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
@@ -2360,7 +2825,6 @@ TEST_F(LibYUVConvertTest, TestDither) {
           src_y + OFF, kWidth, src_u + OFF, kStrideUV, src_v + OFF, kStrideUV, \
           dst_argb_opt + OFF, kStrideB, NULL, kWidth, NEG kHeight);            \
     }                                                                          \
-    int max_diff = 0;                                                          \
     /* Convert to ARGB so 565 is expanded to bytes that can be compared. */    \
     align_buffer_page_end(dst_argb32_c, kWidth* BPP_C* kHeight);               \
     align_buffer_page_end(dst_argb32_opt, kWidth* BPP_C* kHeight);             \
@@ -2371,13 +2835,8 @@ TEST_F(LibYUVConvertTest, TestDither) {
     FMT_B##To##FMT_C(dst_argb_opt + OFF, kStrideB, dst_argb32_opt,             \
                      kWidth * BPP_C, kWidth, kHeight);                         \
     for (int i = 0; i < kWidth * BPP_C * kHeight; ++i) {                       \
-      int abs_diff = abs(static_cast<int>(dst_argb32_c[i]) -                   \
-                         static_cast<int>(dst_argb32_opt[i]));                 \
-      if (abs_diff > max_diff) {                                               \
-        max_diff = abs_diff;                                                   \
-      }                                                                        \
+      EXPECT_EQ(dst_argb32_c[i], dst_argb32_opt[i]);                           \
     }                                                                          \
-    EXPECT_LE(max_diff, DIFF);                                                 \
     free_aligned_buffer_page_end(src_y);                                       \
     free_aligned_buffer_page_end(src_u);                                       \
     free_aligned_buffer_page_end(src_v);                                       \
@@ -2387,20 +2846,20 @@ TEST_F(LibYUVConvertTest, TestDither) {
     free_aligned_buffer_page_end(dst_argb32_opt);                              \
   }
 
-#define TESTPLANARTOBD(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
-                       YALIGN, DIFF, FMT_C, BPP_C)                             \
-  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,       \
-                  YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, FMT_C,       \
-                  BPP_C)                                                       \
-  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,       \
-                  YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, FMT_C,     \
-                  BPP_C)                                                       \
-  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,       \
-                  YALIGN, benchmark_width_, DIFF, _Invert, -, 0, FMT_C, BPP_C) \
-  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,       \
-                  YALIGN, benchmark_width_, DIFF, _Opt, +, 0, FMT_C, BPP_C)
-
-TESTPLANARTOBD(I420, 2, 2, RGB565, 2, 2, 1, 9, ARGB, 4)
+#define TESTPLANARTOBD(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                       YALIGN, FMT_C, BPP_C)                                  \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)     \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C)   \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Invert, -, 0, FMT_C, BPP_C)      \
+  TESTPLANARTOBID(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
+                  YALIGN, benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANARTOBD(I420, 2, 2, RGB565, 2, 2, 1, ARGB, 4)
+#endif
 
 #define TESTPTOB(NAME, UYVYTOI420, UYVYTONV12)                                \
   TEST_F(LibYUVConvertTest, NAME) {                                           \
@@ -2462,12 +2921,12 @@ TESTPLANARTOBD(I420, 2, 2, RGB565, 2, 2, 1, 9, ARGB, 4)
 TESTPTOB(TestYUY2ToNV12, YUY2ToI420, YUY2ToNV12)
 TESTPTOB(TestUYVYToNV12, UYVYToI420, UYVYToNV12)
 
-// Transitive tests.  A to B to C is same as A to C.
-
+// Transitive test.  A to B to C is same as A to C.
+// Benchmarks A To B to C for comparison to 1 step, benchmarked elsewhere.
 #define TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
                        W1280, N, NEG, OFF, FMT_C, BPP_C)                      \
-  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) {             \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##To##FMT_C##N) {            \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
     const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                    \
     const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
@@ -2484,23 +2943,23 @@ TESTPTOB(TestUYVYToNV12, UYVYToI420, UYVYToNV12)
       src_v[i + OFF] = (fastrand() & 0xff);                                   \
     }                                                                         \
     memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                          \
-    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
-      FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,      \
-                            src_v + OFF, kStrideUV, dst_argb_b + OFF,         \
-                            kStrideB, kWidth, NEG kHeight);                   \
-    }                                                                         \
+    FMT_PLANAR##To##FMT_B(src_y + OFF, kWidth, src_u + OFF, kStrideUV,        \
+                          src_v + OFF, kStrideUV, dst_argb_b + OFF, kStrideB, \
+                          kWidth, NEG kHeight);                               \
     /* Convert to a 3rd format in 1 step and 2 steps and compare  */          \
     const int kStrideC = kWidth * BPP_C;                                      \
     align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);               \
     align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);              \
     memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                          \
     memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                         \
-    FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth, src_u + OFF, kStrideUV,        \
-                          src_v + OFF, kStrideUV, dst_argb_c + OFF, kStrideC, \
-                          kWidth, NEG kHeight);                               \
-    /* Convert B to C */                                                      \
-    FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF, kStrideC, \
-                     kWidth, kHeight);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_PLANAR##To##FMT_C(src_y + OFF, kWidth, src_u + OFF, kStrideUV,      \
+                            src_v + OFF, kStrideUV, dst_argb_c + OFF,         \
+                            kStrideC, kWidth, NEG kHeight);                   \
+      /* Convert B to C */                                                    \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,         \
+                       kStrideC, kWidth, kHeight);                            \
+    }                                                                         \
     for (int i = 0; i < kStrideC * kHeight; ++i) {                            \
       EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                   \
     }                                                                         \
@@ -2512,58 +2971,109 @@ TESTPTOB(TestUYVYToNV12, UYVYToI420, UYVYToNV12)
     free_aligned_buffer_page_end(dst_argb_bc);                                \
   }
 
+#if defined(ENABLE_FULL_TESTS)
 #define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
                       FMT_C, BPP_C)                                          \
   TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-                 benchmark_width_ - 4, _Any, +, 0, FMT_C, BPP_C)             \
+                 benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)             \
   TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-                 benchmark_width_, _Unaligned, +, 1, FMT_C, BPP_C)           \
+                 benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C)           \
   TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
                  benchmark_width_, _Invert, -, 0, FMT_C, BPP_C)              \
   TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
                  benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+#else
+#define TESTPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                      FMT_C, BPP_C)                                          \
+  TESTPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                 benchmark_width_, _Opt, +, 0, FMT_C, BPP_C)
+#endif
 
+#if defined(ENABLE_FULL_TESTS)
+TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ABGR, 4)
-TESTPLANARTOE(J420, 2, 2, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(J420, 2, 2, ABGR, 1, 4, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB24, 3)
 TESTPLANARTOE(I420, 2, 2, BGRA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, RAW, 3)
-TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
-TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, ARGB, 4)
-TESTPLANARTOE(H420, 2, 2, RAW, 1, 3, RGB24, 3)
-TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, ABGR, 4)
 TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(H420, 2, 2, ARGB, 1, 4, RGB24, 3)
 TESTPLANARTOE(H420, 2, 2, RAW, 1, 3, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(H420, 2, 2, RGB24, 1, 3, RAW, 3)
+TESTPLANARTOE(J420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(J420, 2, 2, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U420, 2, 2, ARGB, 1, 4, ARGB, 4)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB565, 2)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB1555, 2)
 TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB4444, 2)
 TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, RGB565, 2)
+#endif
+TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(J422, 2, 1, ARGB, 1, 4, ARGB, 4)
 TESTPLANARTOE(J422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(H422, 2, 1, ARGB, 1, 4, ARGB, 4)
 TESTPLANARTOE(H422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U422, 2, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(V422, 2, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(V422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, BGRA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, RGBA, 1, 4, ARGB, 4)
-TESTPLANARTOE(I444, 1, 1, ARGB, 1, 4, ARGB, 4)
-TESTPLANARTOE(J444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(I444, 1, 1, ARGB, 1, 4, ABGR, 4)
 TESTPLANARTOE(I444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(J444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(J444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(H444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(H444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(U444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(U444, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(V444, 1, 1, ARGB, 1, 4, ARGB, 4)
+TESTPLANARTOE(V444, 1, 1, ABGR, 1, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, YUY2, 2, 4, ARGB, 4)
 TESTPLANARTOE(I420, 2, 2, UYVY, 2, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, YUY2, 2, 4, ARGB, 4)
 TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
+#else
+TESTPLANARTOE(I420, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB1555, 2)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, ARGB4444, 2)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, ARGB, 1, 4, RGB565, 2)
+TESTPLANARTOE(I420, 2, 2, BGRA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RAW, 1, 3, RGB24, 3)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, RGB24, 1, 3, RAW, 3)
+TESTPLANARTOE(I420, 2, 2, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, UYVY, 2, 4, ARGB, 4)
+TESTPLANARTOE(I420, 2, 2, YUY2, 2, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, ARGB, 1, 4, RGB565, 2)
+TESTPLANARTOE(I422, 2, 1, BGRA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, RGBA, 1, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
+TESTPLANARTOE(I422, 2, 1, YUY2, 2, 4, ARGB, 4)
+TESTPLANARTOE(I444, 1, 1, ABGR, 1, 4, ARGB, 4)
+#endif
 
+// Transitive test: Compare 1 step vs 2 step conversion for YUVA to ARGB.
+// Benchmark 2 step conversion for comparison to 1 step conversion.
 #define TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
                         W1280, N, NEG, OFF, FMT_C, BPP_C, ATTEN)               \
-  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##_##FMT_C##N) {              \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##To##FMT_C##N) {             \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
     const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                     \
     const int kSizeUV =                                                        \
@@ -2573,6 +3083,12 @@ TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
     align_buffer_page_end(src_v, kSizeUV + OFF);                               \
     align_buffer_page_end(src_a, kWidth* kHeight + OFF);                       \
     align_buffer_page_end(dst_argb_b, kStrideB* kHeight + OFF);                \
+    const int kStrideC = kWidth * BPP_C;                                       \
+    align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);               \
+    memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                           \
+    memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                          \
     for (int i = 0; i < kWidth * kHeight; ++i) {                               \
       src_y[i + OFF] = (fastrand() & 0xff);                                    \
       src_a[i + OFF] = (fastrand() & 0xff);                                    \
@@ -2581,26 +3097,21 @@ TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
       src_u[i + OFF] = (fastrand() & 0xff);                                    \
       src_v[i + OFF] = (fastrand() & 0xff);                                    \
     }                                                                          \
-    memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                           \
     for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      /* Convert A to B */                                                     \
       FMT_PLANAR##To##FMT_B(                                                   \
           src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),      \
           src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X), src_a + OFF, kWidth,      \
           dst_argb_b + OFF, kStrideB, kWidth, NEG kHeight, ATTEN);             \
+      /* Convert B to C */                                                     \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,          \
+                       kStrideC, kWidth, kHeight);                             \
     }                                                                          \
-    /* Convert to a 3rd format in 1 step and 2 steps and compare  */           \
-    const int kStrideC = kWidth * BPP_C;                                       \
-    align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);                \
-    align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);               \
-    memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                           \
-    memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                          \
+    /* Convert A to C */                                                       \
     FMT_PLANAR##To##FMT_C(                                                     \
         src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SUBSAMP_X),        \
         src_v + OFF, SUBSAMPLE(kWidth, SUBSAMP_X), src_a + OFF, kWidth,        \
         dst_argb_c + OFF, kStrideC, kWidth, NEG kHeight, ATTEN);               \
-    /* Convert B to C */                                                       \
-    FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF, kStrideC,  \
-                     kWidth, kHeight);                                         \
     for (int i = 0; i < kStrideC * kHeight; ++i) {                             \
       EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_bc[i + OFF]);                    \
     }                                                                          \
@@ -2613,26 +3124,71 @@ TESTPLANARTOE(I422, 2, 1, UYVY, 2, 4, ARGB, 4)
     free_aligned_buffer_page_end(dst_argb_bc);                                 \
   }
 
+#if defined(ENABLE_FULL_TESTS)
 #define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
                        FMT_C, BPP_C)                                          \
   TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-                  benchmark_width_ - 4, _Any, +, 0, FMT_C, BPP_C, 0)          \
+                  benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C, 0)          \
   TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
-                  benchmark_width_, _Unaligned, +, 1, FMT_C, BPP_C, 0)        \
+                  benchmark_width_, _Unaligned, +, 2, FMT_C, BPP_C, 0)        \
   TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
                   benchmark_width_, _Invert, -, 0, FMT_C, BPP_C, 0)           \
   TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
                   benchmark_width_, _Opt, +, 0, FMT_C, BPP_C, 0)              \
   TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
                   benchmark_width_, _Premult, +, 0, FMT_C, BPP_C, 1)
+#else
+#define TESTQPLANARTOE(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B, \
+                       FMT_C, BPP_C)                                          \
+  TESTQPLANARTOEI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, SUB_B, BPP_B,      \
+                  benchmark_width_, _Opt, +, 0, FMT_C, BPP_C, 0)
+#endif
 
+#if defined(ENABLE_FULL_TESTS)
 TESTQPLANARTOE(I420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
 TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(F420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(F420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V420Alpha, 2, 2, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(F422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(F422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V422Alpha, 2, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(J444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(J444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(H444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(H444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(U444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(U444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(V444Alpha, 1, 1, ARGB, 1, 4, ABGR, 4)
+TESTQPLANARTOE(V444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+#else
+TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I422Alpha, 2, 1, ABGR, 1, 4, ARGB, 4)
+TESTQPLANARTOE(I444Alpha, 1, 1, ABGR, 1, 4, ARGB, 4)
+#endif
 
 #define TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, W1280, N, NEG, \
                       OFF, FMT_C, BPP_C)                                       \
-  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##_##FMT_C##N) {                   \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                            \
+  TEST_F(LibYUVConvertTest, FMT_A##To##FMT_B##To##FMT_C##N) {                  \
+    const int kWidth = W1280;                                                  \
     const int kHeight = benchmark_height_;                                     \
     const int kStrideA = SUBSAMPLE(kWidth, SUB_A) * BPP_A;                     \
     const int kStrideB = SUBSAMPLE(kWidth, SUB_B) * BPP_B;                     \
@@ -2640,21 +3196,21 @@ TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
     align_buffer_page_end(dst_argb_b, kStrideB* kHeight + OFF);                \
     MemRandomize(src_argb_a + OFF, kStrideA * kHeight);                        \
     memset(dst_argb_b + OFF, 1, kStrideB * kHeight);                           \
-    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
-      FMT_A##To##FMT_B(src_argb_a + OFF, kStrideA, dst_argb_b + OFF, kStrideB, \
-                       kWidth, NEG kHeight);                                   \
-    }                                                                          \
+    FMT_A##To##FMT_B(src_argb_a + OFF, kStrideA, dst_argb_b + OFF, kStrideB,   \
+                     kWidth, NEG kHeight);                                     \
     /* Convert to a 3rd format in 1 step and 2 steps and compare  */           \
     const int kStrideC = kWidth * BPP_C;                                       \
     align_buffer_page_end(dst_argb_c, kStrideC* kHeight + OFF);                \
     align_buffer_page_end(dst_argb_bc, kStrideC* kHeight + OFF);               \
     memset(dst_argb_c + OFF, 2, kStrideC * kHeight);                           \
     memset(dst_argb_bc + OFF, 3, kStrideC * kHeight);                          \
-    FMT_A##To##FMT_C(src_argb_a + OFF, kStrideA, dst_argb_c + OFF, kStrideC,   \
-                     kWidth, NEG kHeight);                                     \
-    /* Convert B to C */                                                       \
-    FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF, kStrideC,  \
-                     kWidth, kHeight);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_A##To##FMT_C(src_argb_a + OFF, kStrideA, dst_argb_c + OFF, kStrideC, \
+                       kWidth, NEG kHeight);                                   \
+      /* Convert B to C */                                                     \
+      FMT_B##To##FMT_C(dst_argb_b + OFF, kStrideB, dst_argb_bc + OFF,          \
+                       kStrideC, kWidth, kHeight);                             \
+    }                                                                          \
     for (int i = 0; i < kStrideC * kHeight; i += 4) {                          \
       EXPECT_EQ(dst_argb_c[i + OFF + 0], dst_argb_bc[i + OFF + 0]);            \
       EXPECT_EQ(dst_argb_c[i + OFF + 1], dst_argb_bc[i + OFF + 1]);            \
@@ -2669,15 +3225,16 @@ TESTQPLANARTOE(I420Alpha, 2, 2, ABGR, 1, 4, ARGB, 4)
 
 #define TESTPLANETOE(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, FMT_C, BPP_C) \
   TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B,                    \
-                benchmark_width_ - 4, _Any, +, 0, FMT_C, BPP_C)              \
+                benchmark_width_ + 1, _Any, +, 0, FMT_C, BPP_C)              \
   TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
-                _Unaligned, +, 1, FMT_C, BPP_C)                              \
+                _Unaligned, +, 4, FMT_C, BPP_C)                              \
   TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
                 _Invert, -, 0, FMT_C, BPP_C)                                 \
   TESTPLANETOEI(FMT_A, SUB_A, BPP_A, FMT_B, SUB_B, BPP_B, benchmark_width_,  \
                 _Opt, +, 0, FMT_C, BPP_C)
 
 // Caveat: Destination needs to be 4 bytes
+#ifdef LITTLE_ENDIAN_ONLY_TEST
 TESTPLANETOE(ARGB, 1, 4, AR30, 1, 4, ARGB, 4)
 TESTPLANETOE(ABGR, 1, 4, AR30, 1, 4, ABGR, 4)
 TESTPLANETOE(AR30, 1, 4, ARGB, 1, 4, ABGR, 4)
@@ -2686,6 +3243,7 @@ TESTPLANETOE(ARGB, 1, 4, AB30, 1, 4, ARGB, 4)
 TESTPLANETOE(ABGR, 1, 4, AB30, 1, 4, ABGR, 4)
 TESTPLANETOE(AB30, 1, 4, ARGB, 1, 4, ABGR, 4)
 TESTPLANETOE(AB30, 1, 4, ABGR, 1, 4, ARGB, 4)
+#endif
 
 TEST_F(LibYUVConvertTest, RotateWithARGBSource) {
   // 2x2 frames
@@ -2790,79 +3348,506 @@ TEST_F(LibYUVConvertTest, ABGRToAR30Row_Opt) {
 }
 #endif  // HAS_ABGRTOAR30ROW_AVX2
 
+// Provide matrix wrappers for 12 bit YUV
+#define I012ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I012ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I012ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I012ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+
+#define I410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define H410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define H410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define U410ToARGB(a, b, c, d, e, f, g, h, i, j) \
+  I410ToARGBMatrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define U410ToABGR(a, b, c, d, e, f, g, h, i, j) \
+  I410ToABGRMatrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define I410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define I410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j)
+#define H410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define H410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuvH709Constants, i, j)
+#define U410ToAR30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAR30Matrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+#define U410ToAB30(a, b, c, d, e, f, g, h, i, j) \
+  I410ToAB30Matrix(a, b, c, d, e, f, g, h, &kYuv2020Constants, i, j)
+
+#define I010ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I010ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I010ToAR30Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I010ToAR30MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I210ToARGBFilter(a, b, c, d, e, f, g, h, i, j)                     \
+  I210ToARGBMatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+#define I210ToAR30Filter(a, b, c, d, e, f, g, h, i, j)                     \
+  I210ToAR30MatrixFilter(a, b, c, d, e, f, g, h, &kYuvI601Constants, i, j, \
+                         kFilterBilinear)
+
 // TODO(fbarchard): Fix clamping issue affected by U channel.
-#define TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,   \
-                         ALIGN, YALIGN, W1280, DIFF, N, NEG, SOFF, DOFF)   \
-  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                    \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                        \
-    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);               \
-    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                  \
-    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                    \
-    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);         \
-    const int kBpc = 2;                                                    \
-    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + SOFF);            \
-    align_buffer_page_end(src_u, kSizeUV* kBpc + SOFF);                    \
-    align_buffer_page_end(src_v, kSizeUV* kBpc + SOFF);                    \
-    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + DOFF);           \
-    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + DOFF);         \
-    for (int i = 0; i < kWidth * kHeight; ++i) {                           \
-      reinterpret_cast<uint16_t*>(src_y + SOFF)[i] = (fastrand() & 0x3ff); \
-    }                                                                      \
-    for (int i = 0; i < kSizeUV; ++i) {                                    \
-      reinterpret_cast<uint16_t*>(src_u + SOFF)[i] = (fastrand() & 0x3ff); \
-      reinterpret_cast<uint16_t*>(src_v + SOFF)[i] = (fastrand() & 0x3ff); \
-    }                                                                      \
-    memset(dst_argb_c + DOFF, 1, kStrideB * kHeight);                      \
-    memset(dst_argb_opt + DOFF, 101, kStrideB * kHeight);                  \
-    MaskCpuFlags(disable_cpu_flags_);                                      \
-    FMT_PLANAR##To##FMT_B(                                                 \
-        reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,                 \
-        reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,              \
-        reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,              \
-        dst_argb_c + DOFF, kStrideB, kWidth, NEG kHeight);                 \
-    MaskCpuFlags(benchmark_cpu_info_);                                     \
-    for (int i = 0; i < benchmark_iterations_; ++i) {                      \
-      FMT_PLANAR##To##FMT_B(                                               \
-          reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,               \
-          reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,            \
-          reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,            \
-          dst_argb_opt + DOFF, kStrideB, kWidth, NEG kHeight);             \
-    }                                                                      \
-    int max_diff = 0;                                                      \
-    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                   \
-      int abs_diff = abs(static_cast<int>(dst_argb_c[i + DOFF]) -          \
-                         static_cast<int>(dst_argb_opt[i + DOFF]));        \
-      if (abs_diff > max_diff) {                                           \
-        max_diff = abs_diff;                                               \
-      }                                                                    \
-    }                                                                      \
-    EXPECT_LE(max_diff, DIFF);                                             \
-    free_aligned_buffer_page_end(src_y);                                   \
-    free_aligned_buffer_page_end(src_u);                                   \
-    free_aligned_buffer_page_end(src_v);                                   \
-    free_aligned_buffer_page_end(dst_argb_c);                              \
-    free_aligned_buffer_page_end(dst_argb_opt);                            \
-  }
-
-#define TESTPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
-                        YALIGN, DIFF)                                          \
-  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                   YALIGN, benchmark_width_ - 4, DIFF, _Any, +, 0, 0)          \
-  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                   YALIGN, benchmark_width_, DIFF, _Unaligned, +, 1, 1)        \
-  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                   YALIGN, benchmark_width_, DIFF, _Invert, -, 0, 0)           \
-  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,      \
-                   YALIGN, benchmark_width_, DIFF, _Opt, +, 0, 0)
-
-TESTPLANAR16TOB(I010, 2, 2, ARGB, 4, 4, 1, 2)
-TESTPLANAR16TOB(I010, 2, 2, ABGR, 4, 4, 1, 2)
-TESTPLANAR16TOB(I010, 2, 2, AR30, 4, 4, 1, 2)
-TESTPLANAR16TOB(I010, 2, 2, AB30, 4, 4, 1, 2)
-TESTPLANAR16TOB(H010, 2, 2, ARGB, 4, 4, 1, 2)
-TESTPLANAR16TOB(H010, 2, 2, ABGR, 4, 4, 1, 2)
-TESTPLANAR16TOB(H010, 2, 2, AR30, 4, 4, 1, 2)
-TESTPLANAR16TOB(H010, 2, 2, AB30, 4, 4, 1, 2)
+#define TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B,   \
+                         BPP_B, ALIGN, YALIGN, W1280, N, NEG, SOFF, DOFF)     \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                       \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                  \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                     \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                       \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);            \
+    const int kBpc = 2;                                                       \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + SOFF);               \
+    align_buffer_page_end(src_u, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(src_v, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + DOFF);              \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + DOFF);            \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                              \
+      reinterpret_cast<uint16_t*>(src_y + SOFF)[i] = (fastrand() & FMT_MASK); \
+    }                                                                         \
+    for (int i = 0; i < kSizeUV; ++i) {                                       \
+      reinterpret_cast<uint16_t*>(src_u + SOFF)[i] = (fastrand() & FMT_MASK); \
+      reinterpret_cast<uint16_t*>(src_v + SOFF)[i] = (fastrand() & FMT_MASK); \
+    }                                                                         \
+    memset(dst_argb_c + DOFF, 1, kStrideB * kHeight);                         \
+    memset(dst_argb_opt + DOFF, 101, kStrideB * kHeight);                     \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    FMT_PLANAR##To##FMT_B(                                                    \
+        reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,                    \
+        reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,                 \
+        reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,                 \
+        dst_argb_c + DOFF, kStrideB, kWidth, NEG kHeight);                    \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      FMT_PLANAR##To##FMT_B(                                                  \
+          reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,                  \
+          reinterpret_cast<uint16_t*>(src_u + SOFF), kStrideUV,               \
+          reinterpret_cast<uint16_t*>(src_v + SOFF), kStrideUV,               \
+          dst_argb_opt + DOFF, kStrideB, kWidth, NEG kHeight);                \
+    }                                                                         \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                      \
+      EXPECT_EQ(dst_argb_c[i + DOFF], dst_argb_opt[i + DOFF]);                \
+    }                                                                         \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_u);                                      \
+    free_aligned_buffer_page_end(src_v);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                 \
+    free_aligned_buffer_page_end(dst_argb_opt);                               \
+  }
+
+#define TESTPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B,   \
+                        BPP_B, ALIGN, YALIGN)                                \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_ + 1, _Any, +, 0, 0)       \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Unaligned, +, 4, 4)     \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Invert, -, 0, 0)        \
+  TESTPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_MASK, FMT_B, BPP_B, \
+                   ALIGN, YALIGN, benchmark_width_, _Opt, +, 0, 0)
+
+// These conversions are only optimized for x86
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, ABGR, 4, 4, 1)
+TESTPLANAR16TOB(I012, 2, 2, 0xfff, ARGB, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, ARGBFilter, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, ARGBFilter, 4, 4, 1)
+
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U010, 2, 2, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U210, 2, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(H410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(U410, 1, 1, 0x3ff, AB30, 4, 4, 1)
+TESTPLANAR16TOB(I012, 2, 2, 0xfff, AR30, 4, 4, 1)
+TESTPLANAR16TOB(I010, 2, 2, 0x3ff, AR30Filter, 4, 4, 1)
+TESTPLANAR16TOB(I210, 2, 1, 0x3ff, AR30Filter, 4, 4, 1)
+#endif  // LITTLE_ENDIAN_ONLY_TEST
+#endif  // DISABLE_SLOW_TESTS
+
+#define TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,      \
+                          ALIGN, YALIGN, W1280, N, NEG, OFF, ATTEN, S_DEPTH)   \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X);                        \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y);             \
+    const int kBpc = 2;                                                        \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + OFF);                 \
+    align_buffer_page_end(src_u, kSizeUV* kBpc + OFF);                         \
+    align_buffer_page_end(src_v, kSizeUV* kBpc + OFF);                         \
+    align_buffer_page_end(src_a, kWidth* kHeight* kBpc + OFF);                 \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + OFF);                \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + OFF);              \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      reinterpret_cast<uint16_t*>(src_y + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+      reinterpret_cast<uint16_t*>(src_a + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+    }                                                                          \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      reinterpret_cast<uint16_t*>(src_u + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+      reinterpret_cast<uint16_t*>(src_v + OFF)[i] =                            \
+          (fastrand() & ((1 << S_DEPTH) - 1));                                 \
+    }                                                                          \
+    memset(dst_argb_c + OFF, 1, kStrideB * kHeight);                           \
+    memset(dst_argb_opt + OFF, 101, kStrideB * kHeight);                       \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + OFF), kWidth,    \
+                          reinterpret_cast<uint16_t*>(src_u + OFF), kStrideUV, \
+                          reinterpret_cast<uint16_t*>(src_v + OFF), kStrideUV, \
+                          reinterpret_cast<uint16_t*>(src_a + OFF), kWidth,    \
+                          dst_argb_c + OFF, kStrideB, kWidth, NEG kHeight,     \
+                          ATTEN);                                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(                                                   \
+          reinterpret_cast<uint16_t*>(src_y + OFF), kWidth,                    \
+          reinterpret_cast<uint16_t*>(src_u + OFF), kStrideUV,                 \
+          reinterpret_cast<uint16_t*>(src_v + OFF), kStrideUV,                 \
+          reinterpret_cast<uint16_t*>(src_a + OFF), kWidth,                    \
+          dst_argb_opt + OFF, kStrideB, kWidth, NEG kHeight, ATTEN);           \
+    }                                                                          \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb_c[i + OFF], dst_argb_opt[i + OFF]);                   \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_u);                                       \
+    free_aligned_buffer_page_end(src_v);                                       \
+    free_aligned_buffer_page_end(src_a);                                       \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+#define TESTQPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,    \
+                         ALIGN, YALIGN, S_DEPTH)                            \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_ + 1, _Any, +, 0, 0, S_DEPTH)   \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Unaligned, +, 2, 0, S_DEPTH) \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Invert, -, 0, 0, S_DEPTH)    \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Opt, +, 0, 0, S_DEPTH)       \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                    YALIGN, benchmark_width_, _Premult, +, 0, 1, S_DEPTH)
+#else
+#define TESTQPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,   \
+                         ALIGN, YALIGN, S_DEPTH)                           \
+  TESTQPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN, \
+                    YALIGN, benchmark_width_, _Opt, +, 0, 0, S_DEPTH)
+#endif
+
+#define I010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V010AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I010AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V010AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I010AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V210AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I210AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V210AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I210AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define I410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvI601Constants, k, \
+                        l, m)
+#define J410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define J410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvJPEGConstants, k, \
+                        l, m)
+#define F410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define F410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvF709Constants, k, \
+                        l, m)
+#define H410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define H410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvH709Constants, k, \
+                        l, m)
+#define U410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define U410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)               \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuv2020Constants, k, \
+                        l, m)
+#define V410AlphaToARGB(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I410AlphaToARGBMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define V410AlphaToABGR(a, b, c, d, e, f, g, h, i, j, k, l, m)                \
+  I410AlphaToABGRMatrix(a, b, c, d, e, f, g, h, i, j, &kYuvV2020Constants, k, \
+                        l, m)
+#define I010AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I010AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+#define I210AlphaToARGBFilter(a, b, c, d, e, f, g, h, i, j, k, l, m) \
+  I010AlphaToARGBMatrixFilter(a, b, c, d, e, f, g, h, i, j,          \
+                              &kYuvI601Constants, k, l, m, kFilterBilinear)
+
+// These conversions are only optimized for x86
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V010Alpha, 2, 2, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V010Alpha, 2, 2, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V210Alpha, 2, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V210Alpha, 2, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(J410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(H410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(F410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(U410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V410Alpha, 1, 1, ARGB, 4, 4, 1, 10)
+TESTQPLANAR16TOB(V410Alpha, 1, 1, ABGR, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I010Alpha, 2, 2, ARGBFilter, 4, 4, 1, 10)
+TESTQPLANAR16TOB(I210Alpha, 2, 1, ARGBFilter, 4, 4, 1, 10)
+#endif  // DISABLE_SLOW_TESTS
+
+#define TESTBIPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,     \
+                           ALIGN, YALIGN, W1280, N, NEG, SOFF, DOFF, S_DEPTH)  \
+  TEST_F(LibYUVConvertTest, FMT_PLANAR##To##FMT_B##N) {                        \
+    const int kWidth = W1280;                                                  \
+    const int kHeight = ALIGNINT(benchmark_height_, YALIGN);                   \
+    const int kStrideB = ALIGNINT(kWidth * BPP_B, ALIGN);                      \
+    const int kStrideUV = SUBSAMPLE(kWidth, SUBSAMP_X) * 2;                    \
+    const int kSizeUV = kStrideUV * SUBSAMPLE(kHeight, SUBSAMP_Y) * 2;         \
+    const int kBpc = 2;                                                        \
+    align_buffer_page_end(src_y, kWidth* kHeight* kBpc + SOFF);                \
+    align_buffer_page_end(src_uv, kSizeUV* kBpc + SOFF);                       \
+    align_buffer_page_end(dst_argb_c, kStrideB* kHeight + DOFF);               \
+    align_buffer_page_end(dst_argb_opt, kStrideB* kHeight + DOFF);             \
+    for (int i = 0; i < kWidth * kHeight; ++i) {                               \
+      reinterpret_cast<uint16_t*>(src_y + SOFF)[i] =                           \
+          (fastrand() & (((uint16_t)(-1)) << (16 - S_DEPTH)));                 \
+    }                                                                          \
+    for (int i = 0; i < kSizeUV; ++i) {                                        \
+      reinterpret_cast<uint16_t*>(src_uv + SOFF)[i] =                          \
+          (fastrand() & (((uint16_t)(-1)) << (16 - S_DEPTH)));                 \
+    }                                                                          \
+    memset(dst_argb_c + DOFF, 1, kStrideB * kHeight);                          \
+    memset(dst_argb_opt + DOFF, 101, kStrideB * kHeight);                      \
+    MaskCpuFlags(disable_cpu_flags_);                                          \
+    FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth,   \
+                          reinterpret_cast<uint16_t*>(src_uv + SOFF),          \
+                          kStrideUV, dst_argb_c + DOFF, kStrideB, kWidth,      \
+                          NEG kHeight);                                        \
+    MaskCpuFlags(benchmark_cpu_info_);                                         \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                          \
+      FMT_PLANAR##To##FMT_B(reinterpret_cast<uint16_t*>(src_y + SOFF), kWidth, \
+                            reinterpret_cast<uint16_t*>(src_uv + SOFF),        \
+                            kStrideUV, dst_argb_opt + DOFF, kStrideB, kWidth,  \
+                            NEG kHeight);                                      \
+    }                                                                          \
+    for (int i = 0; i < kWidth * BPP_B * kHeight; ++i) {                       \
+      EXPECT_EQ(dst_argb_c[i + DOFF], dst_argb_opt[i + DOFF]);                 \
+    }                                                                          \
+    free_aligned_buffer_page_end(src_y);                                       \
+    free_aligned_buffer_page_end(src_uv);                                      \
+    free_aligned_buffer_page_end(dst_argb_c);                                  \
+    free_aligned_buffer_page_end(dst_argb_opt);                                \
+  }
+
+#define TESTBIPLANAR16TOB(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B,    \
+                          ALIGN, YALIGN, S_DEPTH)                            \
+  TESTBIPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                     YALIGN, benchmark_width_ + 1, _Any, +, 0, 0, S_DEPTH)   \
+  TESTBIPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                     YALIGN, benchmark_width_, _Unaligned, +, 4, 4, S_DEPTH) \
+  TESTBIPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                     YALIGN, benchmark_width_, _Invert, -, 0, 0, S_DEPTH)    \
+  TESTBIPLANAR16TOBI(FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, FMT_B, BPP_B, ALIGN,  \
+                     YALIGN, benchmark_width_, _Opt, +, 0, 0, S_DEPTH)
+
+#define P010ToARGB(a, b, c, d, e, f, g, h) \
+  P010ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P210ToARGB(a, b, c, d, e, f, g, h) \
+  P210ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P010ToAR30(a, b, c, d, e, f, g, h) \
+  P010ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P210ToAR30(a, b, c, d, e, f, g, h) \
+  P210ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P012ToARGB(a, b, c, d, e, f, g, h) \
+  P012ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P212ToARGB(a, b, c, d, e, f, g, h) \
+  P212ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P012ToAR30(a, b, c, d, e, f, g, h) \
+  P012ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P212ToAR30(a, b, c, d, e, f, g, h) \
+  P212ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P016ToARGB(a, b, c, d, e, f, g, h) \
+  P016ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P216ToARGB(a, b, c, d, e, f, g, h) \
+  P216ToARGBMatrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P016ToAR30(a, b, c, d, e, f, g, h) \
+  P016ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+#define P216ToAR30(a, b, c, d, e, f, g, h) \
+  P216ToAR30Matrix(a, b, c, d, e, f, &kYuvH709Constants, g, h)
+
+#define P010ToARGBFilter(a, b, c, d, e, f, g, h)                     \
+  P010ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P210ToARGBFilter(a, b, c, d, e, f, g, h)                     \
+  P210ToARGBMatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P010ToAR30Filter(a, b, c, d, e, f, g, h)                     \
+  P010ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+#define P210ToAR30Filter(a, b, c, d, e, f, g, h)                     \
+  P210ToAR30MatrixFilter(a, b, c, d, e, f, &kYuvH709Constants, g, h, \
+                         kFilterBilinear)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+TESTBIPLANAR16TOB(P010, 2, 2, ARGB, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P210, 2, 1, ARGB, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P012, 2, 2, ARGB, 4, 4, 1, 12)
+TESTBIPLANAR16TOB(P212, 2, 1, ARGB, 4, 4, 1, 12)
+TESTBIPLANAR16TOB(P016, 2, 2, ARGB, 4, 4, 1, 16)
+TESTBIPLANAR16TOB(P216, 2, 1, ARGB, 4, 4, 1, 16)
+TESTBIPLANAR16TOB(P010, 2, 2, ARGBFilter, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P210, 2, 1, ARGBFilter, 4, 4, 1, 10)
+#ifdef LITTLE_ENDIAN_ONLY_TEST
+TESTBIPLANAR16TOB(P010, 2, 2, AR30, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P210, 2, 1, AR30, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P012, 2, 2, AR30, 4, 4, 1, 12)
+TESTBIPLANAR16TOB(P212, 2, 1, AR30, 4, 4, 1, 12)
+TESTBIPLANAR16TOB(P016, 2, 2, AR30, 4, 4, 1, 16)
+TESTBIPLANAR16TOB(P216, 2, 1, AR30, 4, 4, 1, 16)
+TESTBIPLANAR16TOB(P010, 2, 2, AR30Filter, 4, 4, 1, 10)
+TESTBIPLANAR16TOB(P210, 2, 1, AR30Filter, 4, 4, 1, 10)
+#endif  // LITTLE_ENDIAN_ONLY_TEST
+#endif  // DISABLE_SLOW_TESTS
 
 static int Clamp(int y) {
   if (y < 0) {
@@ -2918,10 +3903,11 @@ TEST_F(LibYUVConvertTest, TestH420ToARGB) {
     ++histogram_b[b];
     ++histogram_g[g];
     ++histogram_r[r];
-    int expected_y = Clamp(static_cast<int>((i - 16) * 1.164f));
-    EXPECT_NEAR(b, expected_y, 1);
-    EXPECT_NEAR(g, expected_y, 1);
-    EXPECT_NEAR(r, expected_y, 1);
+    // Reference formula for Y channel contribution in YUV to RGB conversions:
+    int expected_y = Clamp(static_cast<int>((i - 16) * 1.164f + 0.5f));
+    EXPECT_EQ(b, expected_y);
+    EXPECT_EQ(g, expected_y);
+    EXPECT_EQ(r, expected_y);
     EXPECT_EQ(a, 255);
   }
 
@@ -3043,7 +4029,7 @@ TEST_F(LibYUVConvertTest, TestH010ToAR30) {
     ++histogram_b[b10];
     ++histogram_g[g10];
     ++histogram_r[r10];
-    int expected_y = Clamp10(static_cast<int>((i - 64) * 1.164f));
+    int expected_y = Clamp10(static_cast<int>((i - 64) * 1.164f + 0.5));
     EXPECT_NEAR(b10, expected_y, 4);
     EXPECT_NEAR(g10, expected_y, 4);
     EXPECT_NEAR(r10, expected_y, 4);
@@ -3196,6 +4182,68 @@ TEST_F(LibYUVConvertTest, TestH420ToAR30) {
   free_aligned_buffer_page_end(ar30_pixels);
 }
 
+// Test I400 with jpeg matrix is same as J400
+TEST_F(LibYUVConvertTest, TestI400) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_i400, kSize);
+  align_buffer_page_end(argb_pixels_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_j400, kSize * 4);
+  align_buffer_page_end(argb_pixels_jpeg_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_h709_i400, kSize * 4);
+  align_buffer_page_end(argb_pixels_2020_i400, kSize * 4);
+
+  // Test grey scale
+  for (int i = 0; i < kSize; ++i) {
+    orig_i400[i] = i;
+  }
+
+  J400ToARGB(orig_i400, 0, argb_pixels_j400, 0, kSize, 1);
+  I400ToARGB(orig_i400, 0, argb_pixels_i400, 0, kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_jpeg_i400, 0, &kYuvJPEGConstants,
+                   kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_h709_i400, 0, &kYuvH709Constants,
+                   kSize, 1);
+  I400ToARGBMatrix(orig_i400, 0, argb_pixels_2020_i400, 0, &kYuv2020Constants,
+                   kSize, 1);
+
+  EXPECT_EQ(0, argb_pixels_i400[0]);
+  EXPECT_EQ(0, argb_pixels_j400[0]);
+  EXPECT_EQ(0, argb_pixels_jpeg_i400[0]);
+  EXPECT_EQ(0, argb_pixels_h709_i400[0]);
+  EXPECT_EQ(0, argb_pixels_2020_i400[0]);
+  EXPECT_EQ(0, argb_pixels_i400[16 * 4]);
+  EXPECT_EQ(16, argb_pixels_j400[16 * 4]);
+  EXPECT_EQ(16, argb_pixels_jpeg_i400[16 * 4]);
+  EXPECT_EQ(0, argb_pixels_h709_i400[16 * 4]);
+  EXPECT_EQ(0, argb_pixels_2020_i400[16 * 4]);
+  EXPECT_EQ(130, argb_pixels_i400[128 * 4]);
+  EXPECT_EQ(128, argb_pixels_j400[128 * 4]);
+  EXPECT_EQ(128, argb_pixels_jpeg_i400[128 * 4]);
+  EXPECT_EQ(130, argb_pixels_h709_i400[128 * 4]);
+  EXPECT_EQ(130, argb_pixels_2020_i400[128 * 4]);
+  EXPECT_EQ(255, argb_pixels_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_j400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_jpeg_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_h709_i400[255 * 4]);
+  EXPECT_EQ(255, argb_pixels_2020_i400[255 * 4]);
+
+  for (int i = 0; i < kSize * 4; ++i) {
+    if ((i & 3) == 3) {
+      EXPECT_EQ(255, argb_pixels_j400[i]);
+    } else {
+      EXPECT_EQ(i / 4, argb_pixels_j400[i]);
+    }
+    EXPECT_EQ(argb_pixels_jpeg_i400[i], argb_pixels_j400[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_i400);
+  free_aligned_buffer_page_end(argb_pixels_i400);
+  free_aligned_buffer_page_end(argb_pixels_j400);
+  free_aligned_buffer_page_end(argb_pixels_jpeg_i400);
+  free_aligned_buffer_page_end(argb_pixels_h709_i400);
+  free_aligned_buffer_page_end(argb_pixels_2020_i400);
+}
+
 // Test RGB24 to ARGB and back to RGB24
 TEST_F(LibYUVConvertTest, TestARGBToRGB24) {
   const int kSize = 256;
@@ -3220,4 +4268,74 @@ TEST_F(LibYUVConvertTest, TestARGBToRGB24) {
   free_aligned_buffer_page_end(dest_rgb24);
 }
 
+TEST_F(LibYUVConvertTest, Test565) {
+  SIMD_ALIGNED(uint8_t orig_pixels[256][4]);
+  SIMD_ALIGNED(uint8_t pixels565[256][2]);
+
+  for (int i = 0; i < 256; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      orig_pixels[i][j] = i;
+    }
+  }
+  ARGBToRGB565(&orig_pixels[0][0], 0, &pixels565[0][0], 0, 256, 1);
+  uint32_t checksum = HashDjb2(&pixels565[0][0], sizeof(pixels565), 5381);
+  EXPECT_EQ(610919429u, checksum);
+}
+
+// Test RGB24 to J420 is exact
+#if defined(LIBYUV_BIT_EXACT)
+TEST_F(LibYUVConvertTest, TestRGB24ToJ420) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_rgb24, kSize * 3 * 2);  // 2 rows of RGB24
+  align_buffer_page_end(dest_j420, kSize * 3 / 2 * 2);
+  int iterations256 = (benchmark_width_ * benchmark_height_ + (kSize * 2 - 1)) /
+                      (kSize * 2) * benchmark_iterations_;
+
+  for (int i = 0; i < kSize * 3 * 2; ++i) {
+    orig_rgb24[i] = i;
+  }
+
+  for (int i = 0; i < iterations256; ++i) {
+    RGB24ToJ420(orig_rgb24, kSize * 3, dest_j420, kSize,  // Y plane
+                dest_j420 + kSize * 2, kSize / 2,         // U plane
+                dest_j420 + kSize * 5 / 2, kSize / 2,     // V plane
+                kSize, 2);
+  }
+
+  uint32_t checksum = HashDjb2(dest_j420, kSize * 3 / 2 * 2, 5381);
+  EXPECT_EQ(2755440272u, checksum);
+
+  free_aligned_buffer_page_end(orig_rgb24);
+  free_aligned_buffer_page_end(dest_j420);
+}
+#endif
+
+// Test RGB24 to I420 is exact
+#if defined(LIBYUV_BIT_EXACT)
+TEST_F(LibYUVConvertTest, TestRGB24ToI420) {
+  const int kSize = 256;
+  align_buffer_page_end(orig_rgb24, kSize * 3 * 2);  // 2 rows of RGB24
+  align_buffer_page_end(dest_i420, kSize * 3 / 2 * 2);
+  int iterations256 = (benchmark_width_ * benchmark_height_ + (kSize * 2 - 1)) /
+                      (kSize * 2) * benchmark_iterations_;
+
+  for (int i = 0; i < kSize * 3 * 2; ++i) {
+    orig_rgb24[i] = i;
+  }
+
+  for (int i = 0; i < iterations256; ++i) {
+    RGB24ToI420(orig_rgb24, kSize * 3, dest_i420, kSize,  // Y plane
+                dest_i420 + kSize * 2, kSize / 2,         // U plane
+                dest_i420 + kSize * 5 / 2, kSize / 2,     // V plane
+                kSize, 2);
+  }
+
+  uint32_t checksum = HashDjb2(dest_i420, kSize * 3 / 2 * 2, 5381);
+  EXPECT_EQ(1526656597u, checksum);
+
+  free_aligned_buffer_page_end(orig_rgb24);
+  free_aligned_buffer_page_end(dest_i420);
+}
+#endif
+
 }  // namespace libyuv
diff --git a/files/unit_test/cpu_test.cc b/files/unit_test/cpu_test.cc
index a7991d2b..080778f5 100644
--- a/files/unit_test/cpu_test.cc
+++ b/files/unit_test/cpu_test.cc
@@ -40,6 +40,7 @@ TEST_F(LibYUVBaseTest, TestCpuHas) {
   int has_gfni = TestCpuFlag(kCpuHasGFNI);
   int has_avx512bw = TestCpuFlag(kCpuHasAVX512BW);
   int has_avx512vl = TestCpuFlag(kCpuHasAVX512VL);
+  int has_avx512vnni = TestCpuFlag(kCpuHasAVX512VNNI);
   int has_avx512vbmi = TestCpuFlag(kCpuHasAVX512VBMI);
   int has_avx512vbmi2 = TestCpuFlag(kCpuHasAVX512VBMI2);
   int has_avx512vbitalg = TestCpuFlag(kCpuHasAVX512VBITALG);
@@ -57,6 +58,7 @@ TEST_F(LibYUVBaseTest, TestCpuHas) {
   printf("Has GFNI %d\n", has_gfni);
   printf("Has AVX512BW %d\n", has_avx512bw);
   printf("Has AVX512VL %d\n", has_avx512vl);
+  printf("Has AVX512VNNI %d\n", has_avx512vnni);
   printf("Has AVX512VBMI %d\n", has_avx512vbmi);
   printf("Has AVX512VBMI2 %d\n", has_avx512vbmi2);
   printf("Has AVX512VBITALG %d\n", has_avx512vbitalg);
@@ -67,31 +69,113 @@ TEST_F(LibYUVBaseTest, TestCpuHas) {
   printf("Has MIPS %d\n", has_mips);
   int has_msa = TestCpuFlag(kCpuHasMSA);
   printf("Has MSA %d\n", has_msa);
-  int has_mmi = TestCpuFlag(kCpuHasMMI);
-  printf("Has MMI %d\n", has_mmi);
+#endif
+
+#if defined(__loongarch__)
+  int has_loongarch = TestCpuFlag(kCpuHasLOONGARCH);
+  printf("Has LOONGARCH %d\n", has_loongarch);
+  int has_lsx = TestCpuFlag(kCpuHasLSX);
+  printf("Has LSX %d\n", has_lsx);
+  int has_lasx = TestCpuFlag(kCpuHasLASX);
+  printf("Has LASX %d\n", has_lasx);
 #endif
 }
 
-TEST_F(LibYUVBaseTest, TestCpuCompilerEnabled) {
-#if defined(__aarch64__)
-  printf("Arm64 build\n");
+TEST_F(LibYUVBaseTest, TestCompilerMacros) {
+  // Tests all macros used in public headers.
+#ifdef __ATOMIC_RELAXED
+  printf("__ATOMIC_RELAXED %d\n", __ATOMIC_RELAXED);
+#endif
+#ifdef __cplusplus
+  printf("__cplusplus %ld\n", __cplusplus);
+#endif
+#ifdef __clang_major__
+  printf("__clang_major__ %d\n", __clang_major__);
+#endif
+#ifdef __clang_minor__
+  printf("__clang_minor__ %d\n", __clang_minor__);
 #endif
-#if defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON)
-  printf("Neon build enabled\n");
+#ifdef __GNUC__
+  printf("__GNUC__ %d\n", __GNUC__);
 #endif
-#if defined(__x86_64__) || defined(_M_X64)
-  printf("x64 build\n");
+#ifdef __GNUC_MINOR__
+  printf("__GNUC_MINOR__ %d\n", __GNUC_MINOR__);
+#endif
+#ifdef __i386__
+  printf("__i386__ %d\n", __i386__);
+#endif
+#ifdef __mips
+  printf("__mips %d\n", __mips);
+#endif
+#ifdef __mips_isa_rev
+  printf("__mips_isa_rev %d\n", __mips_isa_rev);
+#endif
+#ifdef __x86_64__
+  printf("__x86_64__ %d\n", __x86_64__);
 #endif
 #ifdef _MSC_VER
   printf("_MSC_VER %d\n", _MSC_VER);
 #endif
-#if !defined(LIBYUV_DISABLE_X86) &&                      \
-    (defined(GCC_HAS_AVX2) || defined(CLANG_HAS_AVX2) || \
-     defined(VISUALC_HAS_AVX2))
-  printf("Has AVX2 1\n");
-#else
-  printf("Has AVX2 0\n");
-// If compiler does not support AVX2, the following function not expected:
+#ifdef __aarch64__
+  printf("__aarch64__ %d\n", __aarch64__);
+#endif
+#ifdef __APPLE__
+  printf("__APPLE__ %d\n", __APPLE__);
+#endif
+#ifdef __arm__
+  printf("__arm__ %d\n", __arm__);
+#endif
+#ifdef __clang__
+  printf("__clang__ %d\n", __clang__);
+#endif
+#ifdef __CLR_VER
+  printf("__CLR_VER %d\n", __CLR_VER);
+#endif
+#ifdef __CYGWIN__
+  printf("__CYGWIN__ %d\n", __CYGWIN__);
+#endif
+#ifdef __llvm__
+  printf("__llvm__ %d\n", __llvm__);
+#endif
+#ifdef __mips_msa
+  printf("__mips_msa %d\n", __mips_msa);
+#endif
+#ifdef __native_client__
+  printf("__native_client__ %d\n", __native_client__);
+#endif
+#ifdef __pic__
+  printf("__pic__ %d\n", __pic__);
+#endif
+#ifdef __pnacl__
+  printf("__pnacl__ %d\n", __pnacl__);
+#endif
+#ifdef _M_IX86
+  printf("_M_IX86 %d\n", _M_IX86);
+#endif
+#ifdef _M_X64
+  printf("_M_X64 %d\n", _M_X64);
+#endif
+#ifdef _MIPS_ARCH_LOONGSON3A
+  printf("_MIPS_ARCH_LOONGSON3A %d\n", _MIPS_ARCH_LOONGSON3A);
+#endif
+#ifdef __loongarch__
+  printf("__loongarch__ %d\n", __loongarch__);
+#endif
+#ifdef _WIN32
+  printf("_WIN32 %d\n", _WIN32);
+#endif
+#ifdef GG_LONGLONG
+  printf("GG_LONGLONG %d\n", GG_LONGLONG);
+#endif
+#ifdef INT_TYPES_DEFINED
+  printf("INT_TYPES_DEFINED\n");
+#endif
+#ifdef __has_feature
+  printf("__has_feature\n");
+#if __has_feature(memory_sanitizer)
+  printf("__has_feature(memory_sanitizer) %d\n",
+         __has_feature(memory_sanitizer));
+#endif
 #endif
 }
 
@@ -156,11 +240,36 @@ TEST_F(LibYUVBaseTest, TestLinuxNeon) {
     printf("WARNING: unable to load \"../../unit_test/testdata/arm_v7.txt\"\n");
   }
 #if defined(__linux__) && defined(__ARM_NEON__)
-  EXPECT_EQ(kCpuHasNEON, ArmCpuCaps("/proc/cpuinfo"));
+  if (FileExists("/proc/cpuinfo")) {
+    if (kCpuHasNEON != ArmCpuCaps("/proc/cpuinfo")) {
+      // This can happen on ARM emulator but /proc/cpuinfo is from host.
+      printf("WARNING: Neon build enabled but CPU does not have NEON\n");
+    }
+  } else {
+    printf("WARNING: unable to load \"/proc/cpuinfo\"\n");
+  }
 #endif
 }
 
+TEST_F(LibYUVBaseTest, TestLinuxMipsMsa) {
+  if (FileExists("../../unit_test/testdata/mips.txt")) {
+    printf("Note: testing to load \"../../unit_test/testdata/mips.txt\"\n");
+
+    EXPECT_EQ(0, MipsCpuCaps("../../unit_test/testdata/mips.txt"));
+    EXPECT_EQ(kCpuHasMSA, MipsCpuCaps("../../unit_test/testdata/mips_msa.txt"));
+    EXPECT_EQ(kCpuHasMSA,
+              MipsCpuCaps("../../unit_test/testdata/mips_loongson2k.txt"));
+  } else {
+    printf("WARNING: unable to load \"../../unit_test/testdata/mips.txt\"\n");
+  }
+}
+
+// TODO(fbarchard): Fix clangcl test of cpuflags.
+#ifdef _MSC_VER
+TEST_F(LibYUVBaseTest, DISABLED_TestSetCpuFlags) {
+#else
 TEST_F(LibYUVBaseTest, TestSetCpuFlags) {
+#endif
   // Reset any masked flags that may have been set so auto init is enabled.
   MaskCpuFlags(0);
 
diff --git a/files/unit_test/cpu_thread_test.cc b/files/unit_test/cpu_thread_test.cc
index 59061b98..69aab74e 100644
--- a/files/unit_test/cpu_thread_test.cc
+++ b/files/unit_test/cpu_thread_test.cc
@@ -12,7 +12,7 @@
 
 #include "libyuv/cpu_id.h"
 
-#if defined(__clang__)
+#if defined(__clang__) && !defined(__wasm__)
 #if __has_include(<pthread.h>)
 #define LIBYUV_HAVE_PTHREAD 1
 #endif
@@ -30,7 +30,7 @@ namespace libyuv {
 void* ThreadMain(void* arg) {
   int* flags = static_cast<int*>(arg);
 
-  *flags = TestCpuFlag(kCpuHasSSSE3);
+  *flags = TestCpuFlag(kCpuInitialized);
   return nullptr;
 }
 #endif  // LIBYUV_HAVE_PTHREAD
diff --git a/files/unit_test/math_test.cc b/files/unit_test/math_test.cc
index 0abbad51..a1544c12 100644
--- a/files/unit_test/math_test.cc
+++ b/files/unit_test/math_test.cc
@@ -16,10 +16,14 @@
 #include "libyuv/basic_types.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/scale.h"
+
+#ifdef ENABLE_ROW_TESTS
 #include "libyuv/scale_row.h"
+#endif
 
 namespace libyuv {
 
+#ifdef ENABLE_ROW_TESTS
 TEST_F(LibYUVBaseTest, TestFixedDiv) {
   int num[1280];
   int div[1280];
@@ -151,5 +155,6 @@ TEST_F(LibYUVBaseTest, TestFixedDiv1_Opt) {
     EXPECT_NEAR(result_c[j], result_opt[j], 1);
   }
 }
+#endif  // ENABLE_ROW_TESTS
 
 }  // namespace libyuv
diff --git a/files/unit_test/planar_test.cc b/files/unit_test/planar_test.cc
index 70f8966e..3a8c470b 100644
--- a/files/unit_test/planar_test.cc
+++ b/files/unit_test/planar_test.cc
@@ -12,9 +12,6 @@
 #include <stdlib.h>
 #include <time.h>
 
-// row.h defines SIMD_ALIGNED, overriding unit_test.h
-#include "libyuv/row.h" /* For ScaleSumSamples_Neon */
-
 #include "../unit_test/unit_test.h"
 #include "libyuv/compare.h"
 #include "libyuv/convert.h"
@@ -24,6 +21,19 @@
 #include "libyuv/cpu_id.h"
 #include "libyuv/planar_functions.h"
 #include "libyuv/rotate.h"
+#include "libyuv/scale.h"
+
+#ifdef ENABLE_ROW_TESTS
+// row.h defines SIMD_ALIGNED, overriding unit_test.h
+// TODO(fbarchard): Remove row.h from unittests.  Test public functions.
+#include "libyuv/row.h" /* For ScaleSumSamples_Neon */
+#endif
+
+#if defined(LIBYUV_BIT_EXACT)
+#define EXPECTED_ATTENUATE_DIFF 0
+#else
+#define EXPECTED_ATTENUATE_DIFF 2
+#endif
 
 namespace libyuv {
 
@@ -96,9 +106,9 @@ TEST_F(LibYUVPlanarTest, TestAttenuate) {
   EXPECT_EQ(32, atten_pixels[128 * 4 + 1]);
   EXPECT_EQ(21, atten_pixels[128 * 4 + 2]);
   EXPECT_EQ(128, atten_pixels[128 * 4 + 3]);
-  EXPECT_NEAR(255, atten_pixels[255 * 4 + 0], 1);
-  EXPECT_NEAR(127, atten_pixels[255 * 4 + 1], 1);
-  EXPECT_NEAR(85, atten_pixels[255 * 4 + 2], 1);
+  EXPECT_NEAR(254, atten_pixels[255 * 4 + 0], EXPECTED_ATTENUATE_DIFF);
+  EXPECT_NEAR(127, atten_pixels[255 * 4 + 1], EXPECTED_ATTENUATE_DIFF);
+  EXPECT_NEAR(85, atten_pixels[255 * 4 + 2], EXPECTED_ATTENUATE_DIFF);
   EXPECT_EQ(255, atten_pixels[255 * 4 + 3]);
 
   free_aligned_buffer_page_end(atten2_pixels);
@@ -151,31 +161,32 @@ static int TestAttenuateI(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Any) {
-  int max_diff = TestAttenuateI(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestAttenuateI(benchmark_width_ + 1, benchmark_height_,
                                 benchmark_iterations_, disable_cpu_flags_,
                                 benchmark_cpu_info_, +1, 0);
-  EXPECT_LE(max_diff, 2);
+
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Unaligned) {
   int max_diff =
       TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
                      disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Invert) {
   int max_diff =
       TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
                      disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBAttenuate_Opt) {
   int max_diff =
       TestAttenuateI(benchmark_width_, benchmark_height_, benchmark_iterations_,
                      disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 static int TestUnattenuateI(int width,
@@ -224,31 +235,31 @@ static int TestUnattenuateI(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Any) {
-  int max_diff = TestUnattenuateI(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestUnattenuateI(benchmark_width_ + 1, benchmark_height_,
                                   benchmark_iterations_, disable_cpu_flags_,
                                   benchmark_cpu_info_, +1, 0);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Unaligned) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
                                   benchmark_iterations_, disable_cpu_flags_,
                                   benchmark_cpu_info_, +1, 1);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Invert) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
                                   benchmark_iterations_, disable_cpu_flags_,
                                   benchmark_cpu_info_, -1, 0);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBUnattenuate_Opt) {
   int max_diff = TestUnattenuateI(benchmark_width_, benchmark_height_,
                                   benchmark_iterations_, disable_cpu_flags_,
                                   benchmark_cpu_info_, +1, 0);
-  EXPECT_LE(max_diff, 2);
+  EXPECT_LE(max_diff, EXPECTED_ATTENUATE_DIFF);
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBComputeCumulativeSum) {
@@ -277,6 +288,7 @@ TEST_F(LibYUVPlanarTest, TestARGBComputeCumulativeSum) {
   }
 }
 
+// near is for legacy platforms.
 TEST_F(LibYUVPlanarTest, TestARGBGray) {
   SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   memset(orig_pixels, 0, sizeof(orig_pixels));
@@ -313,17 +325,17 @@ TEST_F(LibYUVPlanarTest, TestARGBGray) {
   orig_pixels[5][3] = 224u;
   // Do 16 to test asm version.
   ARGBGray(&orig_pixels[0][0], 0, 0, 0, 16, 1);
-  EXPECT_EQ(30u, orig_pixels[0][0]);
-  EXPECT_EQ(30u, orig_pixels[0][1]);
-  EXPECT_EQ(30u, orig_pixels[0][2]);
+  EXPECT_NEAR(29u, orig_pixels[0][0], 1);
+  EXPECT_NEAR(29u, orig_pixels[0][1], 1);
+  EXPECT_NEAR(29u, orig_pixels[0][2], 1);
   EXPECT_EQ(128u, orig_pixels[0][3]);
   EXPECT_EQ(149u, orig_pixels[1][0]);
   EXPECT_EQ(149u, orig_pixels[1][1]);
   EXPECT_EQ(149u, orig_pixels[1][2]);
   EXPECT_EQ(0u, orig_pixels[1][3]);
-  EXPECT_EQ(76u, orig_pixels[2][0]);
-  EXPECT_EQ(76u, orig_pixels[2][1]);
-  EXPECT_EQ(76u, orig_pixels[2][2]);
+  EXPECT_NEAR(77u, orig_pixels[2][0], 1);
+  EXPECT_NEAR(77u, orig_pixels[2][1], 1);
+  EXPECT_NEAR(77u, orig_pixels[2][2], 1);
   EXPECT_EQ(255u, orig_pixels[2][3]);
   EXPECT_EQ(0u, orig_pixels[3][0]);
   EXPECT_EQ(0u, orig_pixels[3][1]);
@@ -333,9 +345,9 @@ TEST_F(LibYUVPlanarTest, TestARGBGray) {
   EXPECT_EQ(255u, orig_pixels[4][1]);
   EXPECT_EQ(255u, orig_pixels[4][2]);
   EXPECT_EQ(255u, orig_pixels[4][3]);
-  EXPECT_EQ(96u, orig_pixels[5][0]);
-  EXPECT_EQ(96u, orig_pixels[5][1]);
-  EXPECT_EQ(96u, orig_pixels[5][2]);
+  EXPECT_NEAR(97u, orig_pixels[5][0], 1);
+  EXPECT_NEAR(97u, orig_pixels[5][1], 1);
+  EXPECT_NEAR(97u, orig_pixels[5][2], 1);
   EXPECT_EQ(224u, orig_pixels[5][3]);
   for (int i = 0; i < 1280; ++i) {
     orig_pixels[i][0] = i;
@@ -385,30 +397,30 @@ TEST_F(LibYUVPlanarTest, TestARGBGrayTo) {
   orig_pixels[5][3] = 224u;
   // Do 16 to test asm version.
   ARGBGrayTo(&orig_pixels[0][0], 0, &gray_pixels[0][0], 0, 16, 1);
-  EXPECT_EQ(30u, gray_pixels[0][0]);
-  EXPECT_EQ(30u, gray_pixels[0][1]);
-  EXPECT_EQ(30u, gray_pixels[0][2]);
-  EXPECT_EQ(128u, gray_pixels[0][3]);
-  EXPECT_EQ(149u, gray_pixels[1][0]);
-  EXPECT_EQ(149u, gray_pixels[1][1]);
-  EXPECT_EQ(149u, gray_pixels[1][2]);
-  EXPECT_EQ(0u, gray_pixels[1][3]);
-  EXPECT_EQ(76u, gray_pixels[2][0]);
-  EXPECT_EQ(76u, gray_pixels[2][1]);
-  EXPECT_EQ(76u, gray_pixels[2][2]);
-  EXPECT_EQ(255u, gray_pixels[2][3]);
-  EXPECT_EQ(0u, gray_pixels[3][0]);
-  EXPECT_EQ(0u, gray_pixels[3][1]);
-  EXPECT_EQ(0u, gray_pixels[3][2]);
-  EXPECT_EQ(255u, gray_pixels[3][3]);
-  EXPECT_EQ(255u, gray_pixels[4][0]);
-  EXPECT_EQ(255u, gray_pixels[4][1]);
-  EXPECT_EQ(255u, gray_pixels[4][2]);
-  EXPECT_EQ(255u, gray_pixels[4][3]);
-  EXPECT_EQ(96u, gray_pixels[5][0]);
-  EXPECT_EQ(96u, gray_pixels[5][1]);
-  EXPECT_EQ(96u, gray_pixels[5][2]);
-  EXPECT_EQ(224u, gray_pixels[5][3]);
+  EXPECT_NEAR(30u, gray_pixels[0][0], 1);
+  EXPECT_NEAR(30u, gray_pixels[0][1], 1);
+  EXPECT_NEAR(30u, gray_pixels[0][2], 1);
+  EXPECT_NEAR(128u, gray_pixels[0][3], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][0], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][1], 1);
+  EXPECT_NEAR(149u, gray_pixels[1][2], 1);
+  EXPECT_NEAR(0u, gray_pixels[1][3], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][0], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][1], 1);
+  EXPECT_NEAR(76u, gray_pixels[2][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[2][3], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][0], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][1], 1);
+  EXPECT_NEAR(0u, gray_pixels[3][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[3][3], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][0], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][1], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][2], 1);
+  EXPECT_NEAR(255u, gray_pixels[4][3], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][0], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][1], 1);
+  EXPECT_NEAR(96u, gray_pixels[5][2], 1);
+  EXPECT_NEAR(224u, gray_pixels[5][3], 1);
   for (int i = 0; i < 1280; ++i) {
     orig_pixels[i][0] = i;
     orig_pixels[i][1] = i / 2;
@@ -418,6 +430,20 @@ TEST_F(LibYUVPlanarTest, TestARGBGrayTo) {
   for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
     ARGBGrayTo(&orig_pixels[0][0], 0, &gray_pixels[0][0], 0, 1280, 1);
   }
+
+  for (int i = 0; i < 256; ++i) {
+    orig_pixels[i][0] = i;
+    orig_pixels[i][1] = i;
+    orig_pixels[i][2] = i;
+    orig_pixels[i][3] = i;
+  }
+  ARGBGray(&orig_pixels[0][0], 0, 0, 0, 256, 1);
+  for (int i = 0; i < 256; ++i) {
+    EXPECT_EQ(i, orig_pixels[i][0]);
+    EXPECT_EQ(i, orig_pixels[i][1]);
+    EXPECT_EQ(i, orig_pixels[i][2]);
+    EXPECT_EQ(i, orig_pixels[i][3]);
+  }
 }
 
 TEST_F(LibYUVPlanarTest, TestARGBSepia) {
@@ -763,27 +789,75 @@ TEST_F(LibYUVPlanarTest, TestARGBQuantize) {
   }
 }
 
-TEST_F(LibYUVPlanarTest, TestARGBMirror) {
-  SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
-  SIMD_ALIGNED(uint8_t dst_pixels[1280][4]);
+TEST_F(LibYUVPlanarTest, ARGBMirror_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_ * 4);
+  align_buffer_page_end(dst_pixels_opt,
+                        benchmark_width_ * benchmark_height_ * 4);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_ * 4);
 
-  for (int i = 0; i < 1280; ++i) {
-    orig_pixels[i][0] = i;
-    orig_pixels[i][1] = i / 2;
-    orig_pixels[i][2] = i / 3;
-    orig_pixels[i][3] = i / 4;
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_ * 4);
+  MaskCpuFlags(disable_cpu_flags_);
+  ARGBMirror(src_pixels, benchmark_width_ * 4, dst_pixels_c,
+             benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ARGBMirror(src_pixels, benchmark_width_ * 4, dst_pixels_opt,
+               benchmark_width_ * 4, benchmark_width_, benchmark_height_);
   }
-  ARGBMirror(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0, 1280, 1);
+  for (int i = 0; i < benchmark_width_ * benchmark_height_ * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
 
-  for (int i = 0; i < 1280; ++i) {
-    EXPECT_EQ(i & 255, dst_pixels[1280 - 1 - i][0]);
-    EXPECT_EQ((i / 2) & 255, dst_pixels[1280 - 1 - i][1]);
-    EXPECT_EQ((i / 3) & 255, dst_pixels[1280 - 1 - i][2]);
-    EXPECT_EQ((i / 4) & 255, dst_pixels[1280 - 1 - i][3]);
+TEST_F(LibYUVPlanarTest, MirrorPlane_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_pixels_opt, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_);
+
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_);
+  MaskCpuFlags(disable_cpu_flags_);
+  MirrorPlane(src_pixels, benchmark_width_, dst_pixels_c, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MirrorPlane(src_pixels, benchmark_width_, dst_pixels_opt, benchmark_width_,
+                benchmark_width_, benchmark_height_);
   }
-  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
-    ARGBMirror(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0, 1280, 1);
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MirrorUVPlane_Opt) {
+  align_buffer_page_end(src_pixels, benchmark_width_ * benchmark_height_ * 2);
+  align_buffer_page_end(dst_pixels_opt,
+                        benchmark_width_ * benchmark_height_ * 2);
+  align_buffer_page_end(dst_pixels_c, benchmark_width_ * benchmark_height_ * 2);
+
+  MemRandomize(src_pixels, benchmark_width_ * benchmark_height_ * 2);
+  MaskCpuFlags(disable_cpu_flags_);
+  MirrorUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_c,
+                benchmark_width_ * 2, benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MirrorUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_opt,
+                  benchmark_width_ * 2, benchmark_width_, benchmark_height_);
   }
+  for (int i = 0; i < benchmark_width_ * benchmark_height_ * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
 }
 
 TEST_F(LibYUVPlanarTest, TestShade) {
@@ -1006,10 +1080,91 @@ TEST_F(LibYUVPlanarTest, TestInterpolatePlane) {
   }
 }
 
+TEST_F(LibYUVPlanarTest, TestInterpolatePlane_16) {
+  SIMD_ALIGNED(uint16_t orig_pixels_0[1280]);
+  SIMD_ALIGNED(uint16_t orig_pixels_1[1280]);
+  SIMD_ALIGNED(uint16_t interpolate_pixels[1280]);
+  memset(orig_pixels_0, 0, sizeof(orig_pixels_0));
+  memset(orig_pixels_1, 0, sizeof(orig_pixels_1));
+
+  orig_pixels_0[0] = 16u;
+  orig_pixels_0[1] = 32u;
+  orig_pixels_0[2] = 64u;
+  orig_pixels_0[3] = 128u;
+  orig_pixels_0[4] = 0u;
+  orig_pixels_0[5] = 0u;
+  orig_pixels_0[6] = 0u;
+  orig_pixels_0[7] = 255u;
+  orig_pixels_0[8] = 0u;
+  orig_pixels_0[9] = 0u;
+  orig_pixels_0[10] = 0u;
+  orig_pixels_0[11] = 0u;
+  orig_pixels_0[12] = 0u;
+  orig_pixels_0[13] = 0u;
+  orig_pixels_0[14] = 0u;
+  orig_pixels_0[15] = 0u;
+
+  orig_pixels_1[0] = 0u;
+  orig_pixels_1[1] = 0u;
+  orig_pixels_1[2] = 0u;
+  orig_pixels_1[3] = 0u;
+  orig_pixels_1[4] = 0u;
+  orig_pixels_1[5] = 0u;
+  orig_pixels_1[6] = 0u;
+  orig_pixels_1[7] = 0u;
+  orig_pixels_1[8] = 0u;
+  orig_pixels_1[9] = 0u;
+  orig_pixels_1[10] = 0u;
+  orig_pixels_1[11] = 0u;
+  orig_pixels_1[12] = 255u;
+  orig_pixels_1[13] = 255u;
+  orig_pixels_1[14] = 255u;
+  orig_pixels_1[15] = 255u;
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 128);
+  EXPECT_EQ(8u, interpolate_pixels[0]);
+  EXPECT_EQ(16u, interpolate_pixels[1]);
+  EXPECT_EQ(32u, interpolate_pixels[2]);
+  EXPECT_EQ(64u, interpolate_pixels[3]);
+  EXPECT_EQ(0u, interpolate_pixels[4]);
+  EXPECT_EQ(0u, interpolate_pixels[5]);
+  EXPECT_EQ(0u, interpolate_pixels[6]);
+  EXPECT_EQ(128u, interpolate_pixels[7]);
+  EXPECT_EQ(0u, interpolate_pixels[8]);
+  EXPECT_EQ(0u, interpolate_pixels[9]);
+  EXPECT_EQ(0u, interpolate_pixels[10]);
+  EXPECT_EQ(0u, interpolate_pixels[11]);
+  EXPECT_EQ(128u, interpolate_pixels[12]);
+  EXPECT_EQ(128u, interpolate_pixels[13]);
+  EXPECT_EQ(128u, interpolate_pixels[14]);
+  EXPECT_EQ(128u, interpolate_pixels[15]);
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 0);
+  EXPECT_EQ(16u, interpolate_pixels[0]);
+  EXPECT_EQ(32u, interpolate_pixels[1]);
+  EXPECT_EQ(64u, interpolate_pixels[2]);
+  EXPECT_EQ(128u, interpolate_pixels[3]);
+
+  InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                      &interpolate_pixels[0], 0, 16, 1, 192);
+
+  EXPECT_EQ(4u, interpolate_pixels[0]);
+  EXPECT_EQ(8u, interpolate_pixels[1]);
+  EXPECT_EQ(16u, interpolate_pixels[2]);
+  EXPECT_EQ(32u, interpolate_pixels[3]);
+
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+    InterpolatePlane_16(&orig_pixels_0[0], 0, &orig_pixels_1[0], 0,
+                        &interpolate_pixels[0], 0, 1280, 1, 123);
+  }
+}
+
 #define TESTTERP(FMT_A, BPP_A, STRIDE_A, FMT_B, BPP_B, STRIDE_B, W1280, TERP, \
                  N, NEG, OFF)                                                 \
   TEST_F(LibYUVPlanarTest, ARGBInterpolate##TERP##N) {                        \
-    const int kWidth = ((W1280) > 0) ? (W1280) : 1;                           \
+    const int kWidth = W1280;                                                 \
     const int kHeight = benchmark_height_;                                    \
     const int kStrideA =                                                      \
         (kWidth * BPP_A + STRIDE_A - 1) / STRIDE_A * STRIDE_A;                \
@@ -1041,7 +1196,7 @@ TEST_F(LibYUVPlanarTest, TestInterpolatePlane) {
   }
 
 #define TESTINTERPOLATE(TERP)                                                \
-  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_ - 1, TERP, _Any, +, 0)   \
+  TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_ + 1, TERP, _Any, +, 0)   \
   TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Unaligned, +, 1) \
   TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Invert, -, 0)    \
   TESTTERP(ARGB, 4, 1, ARGB, 4, 1, benchmark_width_, TERP, _Opt, +, 0)
@@ -1058,7 +1213,8 @@ static int TestBlend(int width,
                      int disable_cpu_flags,
                      int benchmark_cpu_info,
                      int invert,
-                     int off) {
+                     int off,
+                     int attenuate) {
   if (width < 1) {
     width = 1;
   }
@@ -1072,10 +1228,12 @@ static int TestBlend(int width,
     src_argb_a[i + off] = (fastrand() & 0xff);
     src_argb_b[i + off] = (fastrand() & 0xff);
   }
-  ARGBAttenuate(src_argb_a + off, kStride, src_argb_a + off, kStride, width,
-                height);
-  ARGBAttenuate(src_argb_b + off, kStride, src_argb_b + off, kStride, width,
-                height);
+  MemRandomize(src_argb_a, kStride * height + off);
+  MemRandomize(src_argb_b, kStride * height + off);
+  if (attenuate) {
+    ARGBAttenuate(src_argb_a + off, kStride, src_argb_a + off, kStride, width,
+                  height);
+  }
   memset(dst_argb_c, 255, kStride * height);
   memset(dst_argb_opt, 255, kStride * height);
 
@@ -1104,29 +1262,36 @@ static int TestBlend(int width,
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Any) {
   int max_diff =
-      TestBlend(benchmark_width_ - 4, benchmark_height_, benchmark_iterations_,
-                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+      TestBlend(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Unaligned) {
   int max_diff =
       TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 1, 1);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Invert) {
   int max_diff =
       TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                disable_cpu_flags_, benchmark_cpu_info_, -1, 0);
+                disable_cpu_flags_, benchmark_cpu_info_, -1, 0, 1);
+  EXPECT_LE(max_diff, 1);
+}
+
+TEST_F(LibYUVPlanarTest, ARGBBlend_Unattenuated) {
+  int max_diff =
+      TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 0);
   EXPECT_LE(max_diff, 1);
 }
 
 TEST_F(LibYUVPlanarTest, ARGBBlend_Opt) {
   int max_diff =
       TestBlend(benchmark_width_, benchmark_height_, benchmark_iterations_,
-                disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
+                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, 1);
   EXPECT_LE(max_diff, 1);
 }
 
@@ -1203,7 +1368,7 @@ TEST_F(LibYUVPlanarTest, BlendPlane_Unaligned) {
                  disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
 }
 TEST_F(LibYUVPlanarTest, BlendPlane_Any) {
-  TestBlendPlane(benchmark_width_ - 4, benchmark_height_, benchmark_iterations_,
+  TestBlendPlane(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                  disable_cpu_flags_, benchmark_cpu_info_, +1, 1);
 }
 TEST_F(LibYUVPlanarTest, BlendPlane_Invert) {
@@ -1298,7 +1463,7 @@ TEST_F(LibYUVPlanarTest, I420Blend_Unaligned) {
 
 // TODO(fbarchard): DISABLED because _Any uses C.  Avoid C and re-enable.
 TEST_F(LibYUVPlanarTest, DISABLED_I420Blend_Any) {
-  TestI420Blend(benchmark_width_ - 4, benchmark_height_, benchmark_iterations_,
+  TestI420Blend(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                 disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
 }
 TEST_F(LibYUVPlanarTest, I420Blend_Invert) {
@@ -1400,6 +1565,212 @@ TEST_F(LibYUVPlanarTest, TestCopyPlane) {
   EXPECT_EQ(0, err);
 }
 
+TEST_F(LibYUVPlanarTest, CopyPlane_Opt) {
+  int i;
+  int y_plane_size = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_y, y_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+
+  MemRandomize(orig_y, y_plane_size);
+  memset(dst_c, 1, y_plane_size);
+  memset(dst_opt, 2, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  for (i = 0; i < benchmark_iterations_; i++) {
+    CopyPlane(orig_y, benchmark_width_, dst_c, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (i = 0; i < benchmark_iterations_; i++) {
+    CopyPlane(orig_y, benchmark_width_, dst_opt, benchmark_width_,
+              benchmark_width_, benchmark_height_);
+  }
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
+TEST_F(LibYUVPlanarTest, TestCopyPlaneZero) {
+  // Test to verify copying a rect with a zero height or width does
+  // not touch destination memory.
+  uint8_t src = 42;
+  uint8_t dst = 0;
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  CopyPlane(&src, 0, &dst, 0, 0, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 1, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 0, 1);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  CopyPlane(&src, 0, &dst, 0, 0, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 1, 0);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+
+  CopyPlane(&src, 1, &dst, 1, 0, 1);
+  EXPECT_EQ(src, 42);
+  EXPECT_EQ(dst, 0);
+}
+
+TEST_F(LibYUVPlanarTest, TestDetilePlane) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int orig_width = (benchmark_width_ + 15) & ~15;
+  int orig_height = (benchmark_height_ + 15) & ~15;
+  int orig_plane_size = orig_width * orig_height;
+  int y_plane_size = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_y, orig_plane_size);
+  align_buffer_page_end(dst_c, y_plane_size);
+  align_buffer_page_end(dst_opt, y_plane_size);
+
+  MemRandomize(orig_y, orig_plane_size);
+  memset(dst_c, 0, y_plane_size);
+  memset(dst_opt, 0, y_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(orig_y, orig_width, dst_c, benchmark_width_, benchmark_width_,
+                benchmark_height_, 16);
+  }
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(orig_y, orig_width, dst_opt, benchmark_width_, benchmark_width_,
+                benchmark_height_, 16);
+  }
+
+  for (i = 0; i < y_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_y);
+  free_aligned_buffer_page_end(dst_c);
+  free_aligned_buffer_page_end(dst_opt);
+}
+
+// Compares DetileSplitUV to 2 step Detile + SplitUV
+TEST_F(LibYUVPlanarTest, TestDetileSplitUVPlane_Correctness) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int orig_width = (benchmark_width_ + 15) & ~15;
+  int orig_height = (benchmark_height_ + 15) & ~15;
+  int orig_plane_size = orig_width * orig_height;
+  int uv_plane_size = ((benchmark_width_ + 1) / 2) * benchmark_height_;
+  align_buffer_page_end(orig_uv, orig_plane_size);
+  align_buffer_page_end(detiled_uv, orig_plane_size);
+  align_buffer_page_end(dst_u_two_stage, uv_plane_size);
+  align_buffer_page_end(dst_u_opt, uv_plane_size);
+  align_buffer_page_end(dst_v_two_stage, uv_plane_size);
+  align_buffer_page_end(dst_v_opt, uv_plane_size);
+
+  MemRandomize(orig_uv, orig_plane_size);
+  memset(detiled_uv, 0, orig_plane_size);
+  memset(dst_u_two_stage, 0, uv_plane_size);
+  memset(dst_u_opt, 0, uv_plane_size);
+  memset(dst_v_two_stage, 0, uv_plane_size);
+  memset(dst_v_opt, 0, uv_plane_size);
+
+  DetileSplitUVPlane(orig_uv, orig_width, dst_u_opt, (benchmark_width_ + 1) / 2,
+                     dst_v_opt, (benchmark_width_ + 1) / 2, benchmark_width_,
+                     benchmark_height_, 16);
+
+  // Benchmark 2 step conversion for comparison.
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetilePlane(orig_uv, orig_width, detiled_uv, benchmark_width_,
+                benchmark_width_, benchmark_height_, 16);
+    SplitUVPlane(detiled_uv, orig_width, dst_u_two_stage,
+                 (benchmark_width_ + 1) / 2, dst_v_two_stage,
+                 (benchmark_width_ + 1) / 2, (benchmark_width_ + 1) / 2,
+                 benchmark_height_);
+  }
+
+  for (i = 0; i < uv_plane_size; ++i) {
+    EXPECT_EQ(dst_u_two_stage[i], dst_u_opt[i]);
+    EXPECT_EQ(dst_v_two_stage[i], dst_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_uv);
+  free_aligned_buffer_page_end(detiled_uv);
+  free_aligned_buffer_page_end(dst_u_two_stage);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_two_stage);
+  free_aligned_buffer_page_end(dst_v_opt);
+}
+
+TEST_F(LibYUVPlanarTest, TestDetileSplitUVPlane_Benchmark) {
+  int i, j;
+
+  // orig is tiled.  Allocate enough memory for tiles.
+  int orig_width = (benchmark_width_ + 15) & ~15;
+  int orig_height = (benchmark_height_ + 15) & ~15;
+  int orig_plane_size = orig_width * orig_height;
+  int uv_plane_size = ((benchmark_width_ + 1) / 2) * benchmark_height_;
+  align_buffer_page_end(orig_uv, orig_plane_size);
+  align_buffer_page_end(dst_u_c, uv_plane_size);
+  align_buffer_page_end(dst_u_opt, uv_plane_size);
+  align_buffer_page_end(dst_v_c, uv_plane_size);
+  align_buffer_page_end(dst_v_opt, uv_plane_size);
+
+  MemRandomize(orig_uv, orig_plane_size);
+  memset(dst_u_c, 0, uv_plane_size);
+  memset(dst_u_opt, 0, uv_plane_size);
+  memset(dst_v_c, 0, uv_plane_size);
+  memset(dst_v_opt, 0, uv_plane_size);
+
+  // Disable all optimizations.
+  MaskCpuFlags(disable_cpu_flags_);
+
+  DetileSplitUVPlane(orig_uv, orig_width, dst_u_c, (benchmark_width_ + 1) / 2,
+                     dst_v_c, (benchmark_width_ + 1) / 2, benchmark_width_,
+                     benchmark_height_, 16);
+
+  // Enable optimizations.
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (j = 0; j < benchmark_iterations_; j++) {
+    DetileSplitUVPlane(
+        orig_uv, orig_width, dst_u_opt, (benchmark_width_ + 1) / 2, dst_v_opt,
+        (benchmark_width_ + 1) / 2, benchmark_width_, benchmark_height_, 16);
+  }
+
+  for (i = 0; i < uv_plane_size; ++i) {
+    EXPECT_EQ(dst_u_c[i], dst_u_opt[i]);
+    EXPECT_EQ(dst_v_c[i], dst_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(orig_uv);
+  free_aligned_buffer_page_end(dst_u_c);
+  free_aligned_buffer_page_end(dst_u_opt);
+  free_aligned_buffer_page_end(dst_v_c);
+  free_aligned_buffer_page_end(dst_v_opt);
+}
+
 static int TestMultiply(int width,
                         int height,
                         int benchmark_iterations,
@@ -1447,7 +1818,7 @@ static int TestMultiply(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBMultiply_Any) {
-  int max_diff = TestMultiply(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestMultiply(benchmark_width_ + 1, benchmark_height_,
                               benchmark_iterations_, disable_cpu_flags_,
                               benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
@@ -1522,7 +1893,7 @@ static int TestAdd(int width,
 
 TEST_F(LibYUVPlanarTest, ARGBAdd_Any) {
   int max_diff =
-      TestAdd(benchmark_width_ - 1, benchmark_height_, benchmark_iterations_,
+      TestAdd(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
               disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
 }
@@ -1595,7 +1966,7 @@ static int TestSubtract(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSubtract_Any) {
-  int max_diff = TestSubtract(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestSubtract(benchmark_width_ + 1, benchmark_height_,
                               benchmark_iterations_, disable_cpu_flags_,
                               benchmark_cpu_info_, +1, 0);
   EXPECT_LE(max_diff, 1);
@@ -1668,7 +2039,7 @@ static int TestSobel(int width,
 
 TEST_F(LibYUVPlanarTest, ARGBSobel_Any) {
   int max_diff =
-      TestSobel(benchmark_width_ - 1, benchmark_height_, benchmark_iterations_,
+      TestSobel(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                 disable_cpu_flags_, benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
 }
@@ -1741,7 +2112,7 @@ static int TestSobelToPlane(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelToPlane_Any) {
-  int max_diff = TestSobelToPlane(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestSobelToPlane(benchmark_width_ + 1, benchmark_height_,
                                   benchmark_iterations_, disable_cpu_flags_,
                                   benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
@@ -1813,7 +2184,7 @@ static int TestSobelXY(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBSobelXY_Any) {
-  int max_diff = TestSobelXY(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestSobelXY(benchmark_width_ + 1, benchmark_height_,
                              benchmark_iterations_, disable_cpu_flags_,
                              benchmark_cpu_info_, +1, 0);
   EXPECT_EQ(0, max_diff);
@@ -1889,29 +2260,35 @@ static int TestBlur(int width,
   return max_diff;
 }
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+#define DISABLED_ARM(name) name
+#else
+#define DISABLED_ARM(name) DISABLED_##name
+#endif
+
 static const int kBlurSize = 55;
-TEST_F(LibYUVPlanarTest, ARGBBlur_Any) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Any)) {
   int max_diff =
-      TestBlur(benchmark_width_ - 1, benchmark_height_, benchmark_iterations_,
+      TestBlur(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Unaligned) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Unaligned)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 1, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Invert) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Invert)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, -1, 0, kBlurSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlur_Opt) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlur_Opt)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSize);
@@ -1919,35 +2296,35 @@ TEST_F(LibYUVPlanarTest, ARGBBlur_Opt) {
 }
 
 static const int kBlurSmallSize = 5;
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Any) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Any)) {
   int max_diff =
-      TestBlur(benchmark_width_ - 1, benchmark_height_, benchmark_iterations_,
+      TestBlur(benchmark_width_ + 1, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Unaligned) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Unaligned)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 1, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Invert) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Invert)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, -1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, ARGBBlurSmall_Opt) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(ARGBBlurSmall_Opt)) {
   int max_diff =
       TestBlur(benchmark_width_, benchmark_height_, benchmark_iterations_,
                disable_cpu_flags_, benchmark_cpu_info_, +1, 0, kBlurSmallSize);
   EXPECT_LE(max_diff, 1);
 }
 
-TEST_F(LibYUVPlanarTest, TestARGBPolynomial) {
+TEST_F(LibYUVPlanarTest, DISABLED_ARM(TestARGBPolynomial)) {
   SIMD_ALIGNED(uint8_t orig_pixels[1280][4]);
   SIMD_ALIGNED(uint8_t dst_pixels_opt[1280][4]);
   SIMD_ALIGNED(uint8_t dst_pixels_c[1280][4]);
@@ -2426,7 +2803,7 @@ static int TestARGBRect(int width,
 }
 
 TEST_F(LibYUVPlanarTest, ARGBRect_Any) {
-  int max_diff = TestARGBRect(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestARGBRect(benchmark_width_ + 1, benchmark_height_,
                               benchmark_iterations_, disable_cpu_flags_,
                               benchmark_cpu_info_, +1, 0, 4);
   EXPECT_EQ(0, max_diff);
@@ -2454,7 +2831,7 @@ TEST_F(LibYUVPlanarTest, ARGBRect_Opt) {
 }
 
 TEST_F(LibYUVPlanarTest, SetPlane_Any) {
-  int max_diff = TestARGBRect(benchmark_width_ - 1, benchmark_height_,
+  int max_diff = TestARGBRect(benchmark_width_ + 1, benchmark_height_,
                               benchmark_iterations_, disable_cpu_flags_,
                               benchmark_cpu_info_, +1, 0, 1);
   EXPECT_EQ(0, max_diff);
@@ -2483,33 +2860,24 @@ TEST_F(LibYUVPlanarTest, SetPlane_Opt) {
 
 TEST_F(LibYUVPlanarTest, MergeUVPlane_Opt) {
   const int kPixels = benchmark_width_ * benchmark_height_;
-  align_buffer_page_end(src_pixels, kPixels * 2);
-  align_buffer_page_end(tmp_pixels_u, kPixels);
-  align_buffer_page_end(tmp_pixels_v, kPixels);
+  align_buffer_page_end(src_pixels_u, kPixels);
+  align_buffer_page_end(src_pixels_v, kPixels);
   align_buffer_page_end(dst_pixels_opt, kPixels * 2);
   align_buffer_page_end(dst_pixels_c, kPixels * 2);
 
-  MemRandomize(src_pixels, kPixels * 2);
-  MemRandomize(tmp_pixels_u, kPixels);
-  MemRandomize(tmp_pixels_v, kPixels);
+  MemRandomize(src_pixels_u, kPixels);
+  MemRandomize(src_pixels_v, kPixels);
   MemRandomize(dst_pixels_opt, kPixels * 2);
   MemRandomize(dst_pixels_c, kPixels * 2);
 
   MaskCpuFlags(disable_cpu_flags_);
-  SplitUVPlane(src_pixels, benchmark_width_ * 2, tmp_pixels_u, benchmark_width_,
-               tmp_pixels_v, benchmark_width_, benchmark_width_,
-               benchmark_height_);
-  MergeUVPlane(tmp_pixels_u, benchmark_width_, tmp_pixels_v, benchmark_width_,
+  MergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v, benchmark_width_,
                dst_pixels_c, benchmark_width_ * 2, benchmark_width_,
                benchmark_height_);
   MaskCpuFlags(benchmark_cpu_info_);
 
-  SplitUVPlane(src_pixels, benchmark_width_ * 2, tmp_pixels_u, benchmark_width_,
-               tmp_pixels_v, benchmark_width_, benchmark_width_,
-               benchmark_height_);
-
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    MergeUVPlane(tmp_pixels_u, benchmark_width_, tmp_pixels_v, benchmark_width_,
+    MergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v, benchmark_width_,
                  dst_pixels_opt, benchmark_width_ * 2, benchmark_width_,
                  benchmark_height_);
   }
@@ -2518,9 +2886,43 @@ TEST_F(LibYUVPlanarTest, MergeUVPlane_Opt) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
 
-  free_aligned_buffer_page_end(src_pixels);
-  free_aligned_buffer_page_end(tmp_pixels_u);
-  free_aligned_buffer_page_end(tmp_pixels_v);
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+// 16 bit channel split and merge
+TEST_F(LibYUVPlanarTest, MergeUVPlane_16_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_u, kPixels * 2);
+  align_buffer_page_end(src_pixels_v, kPixels * 2);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 2 * 2);
+  align_buffer_page_end(dst_pixels_c, kPixels * 2 * 2);
+  MemRandomize(src_pixels_u, kPixels * 2);
+  MemRandomize(src_pixels_v, kPixels * 2);
+  MemRandomize(dst_pixels_opt, kPixels * 2 * 2);
+  MemRandomize(dst_pixels_c, kPixels * 2 * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  MergeUVPlane_16((const uint16_t*)src_pixels_u, benchmark_width_,
+                  (const uint16_t*)src_pixels_v, benchmark_width_,
+                  (uint16_t*)dst_pixels_c, benchmark_width_ * 2,
+                  benchmark_width_, benchmark_height_, 12);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeUVPlane_16((const uint16_t*)src_pixels_u, benchmark_width_,
+                    (const uint16_t*)src_pixels_v, benchmark_width_,
+                    (uint16_t*)dst_pixels_opt, benchmark_width_ * 2,
+                    benchmark_width_, benchmark_height_, 12);
+  }
+
+  for (int i = 0; i < kPixels * 2 * 2; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
   free_aligned_buffer_page_end(dst_pixels_opt);
   free_aligned_buffer_page_end(dst_pixels_c);
 }
@@ -2528,47 +2930,112 @@ TEST_F(LibYUVPlanarTest, MergeUVPlane_Opt) {
 TEST_F(LibYUVPlanarTest, SplitUVPlane_Opt) {
   const int kPixels = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_pixels, kPixels * 2);
-  align_buffer_page_end(tmp_pixels_u, kPixels);
-  align_buffer_page_end(tmp_pixels_v, kPixels);
+  align_buffer_page_end(dst_pixels_u_c, kPixels);
+  align_buffer_page_end(dst_pixels_v_c, kPixels);
+  align_buffer_page_end(dst_pixels_u_opt, kPixels);
+  align_buffer_page_end(dst_pixels_v_opt, kPixels);
+
+  MemRandomize(src_pixels, kPixels * 2);
+  MemRandomize(dst_pixels_u_c, kPixels);
+  MemRandomize(dst_pixels_v_c, kPixels);
+  MemRandomize(dst_pixels_u_opt, kPixels);
+  MemRandomize(dst_pixels_v_opt, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_u_c,
+               benchmark_width_, dst_pixels_v_c, benchmark_width_,
+               benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_u_opt,
+                 benchmark_width_, dst_pixels_v_opt, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_u_c[i], dst_pixels_u_opt[i]);
+    EXPECT_EQ(dst_pixels_v_c[i], dst_pixels_v_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_u_c);
+  free_aligned_buffer_page_end(dst_pixels_v_c);
+  free_aligned_buffer_page_end(dst_pixels_u_opt);
+  free_aligned_buffer_page_end(dst_pixels_v_opt);
+}
+
+// 16 bit channel split
+TEST_F(LibYUVPlanarTest, SplitUVPlane_16_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 2 * 2);
+  align_buffer_page_end(dst_pixels_u_c, kPixels * 2);
+  align_buffer_page_end(dst_pixels_v_c, kPixels * 2);
+  align_buffer_page_end(dst_pixels_u_opt, kPixels * 2);
+  align_buffer_page_end(dst_pixels_v_opt, kPixels * 2);
+  MemRandomize(src_pixels, kPixels * 2 * 2);
+  MemRandomize(dst_pixels_u_c, kPixels * 2);
+  MemRandomize(dst_pixels_v_c, kPixels * 2);
+  MemRandomize(dst_pixels_u_opt, kPixels * 2);
+  MemRandomize(dst_pixels_v_opt, kPixels * 2);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitUVPlane_16((const uint16_t*)src_pixels, benchmark_width_ * 2,
+                  (uint16_t*)dst_pixels_u_c, benchmark_width_,
+                  (uint16_t*)dst_pixels_v_c, benchmark_width_, benchmark_width_,
+                  benchmark_height_, 10);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitUVPlane_16((const uint16_t*)src_pixels, benchmark_width_ * 2,
+                    (uint16_t*)dst_pixels_u_opt, benchmark_width_,
+                    (uint16_t*)dst_pixels_v_opt, benchmark_width_,
+                    benchmark_width_, benchmark_height_, 10);
+  }
+
+  for (int i = 0; i < kPixels * 2; ++i) {
+    EXPECT_EQ(dst_pixels_u_c[i], dst_pixels_u_opt[i]);
+    EXPECT_EQ(dst_pixels_v_c[i], dst_pixels_v_opt[i]);
+  }
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(dst_pixels_u_c);
+  free_aligned_buffer_page_end(dst_pixels_v_c);
+  free_aligned_buffer_page_end(dst_pixels_u_opt);
+  free_aligned_buffer_page_end(dst_pixels_v_opt);
+}
+
+TEST_F(LibYUVPlanarTest, SwapUVPlane_Opt) {
+  // Round count up to multiple of 16
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 2);
   align_buffer_page_end(dst_pixels_opt, kPixels * 2);
   align_buffer_page_end(dst_pixels_c, kPixels * 2);
 
   MemRandomize(src_pixels, kPixels * 2);
-  MemRandomize(tmp_pixels_u, kPixels);
-  MemRandomize(tmp_pixels_v, kPixels);
   MemRandomize(dst_pixels_opt, kPixels * 2);
   MemRandomize(dst_pixels_c, kPixels * 2);
 
   MaskCpuFlags(disable_cpu_flags_);
-  SplitUVPlane(src_pixels, benchmark_width_ * 2, tmp_pixels_u, benchmark_width_,
-               tmp_pixels_v, benchmark_width_, benchmark_width_,
-               benchmark_height_);
-  MergeUVPlane(tmp_pixels_u, benchmark_width_, tmp_pixels_v, benchmark_width_,
-               dst_pixels_c, benchmark_width_ * 2, benchmark_width_,
-               benchmark_height_);
+  SwapUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_c,
+              benchmark_width_ * 2, benchmark_width_, benchmark_height_);
   MaskCpuFlags(benchmark_cpu_info_);
 
   for (int i = 0; i < benchmark_iterations_; ++i) {
-    SplitUVPlane(src_pixels, benchmark_width_ * 2, tmp_pixels_u,
-                 benchmark_width_, tmp_pixels_v, benchmark_width_,
-                 benchmark_width_, benchmark_height_);
+    SwapUVPlane(src_pixels, benchmark_width_ * 2, dst_pixels_opt,
+                benchmark_width_ * 2, benchmark_width_, benchmark_height_);
   }
-  MergeUVPlane(tmp_pixels_u, benchmark_width_, tmp_pixels_v, benchmark_width_,
-               dst_pixels_opt, benchmark_width_ * 2, benchmark_width_,
-               benchmark_height_);
 
   for (int i = 0; i < kPixels * 2; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
 
   free_aligned_buffer_page_end(src_pixels);
-  free_aligned_buffer_page_end(tmp_pixels_u);
-  free_aligned_buffer_page_end(tmp_pixels_v);
   free_aligned_buffer_page_end(dst_pixels_opt);
   free_aligned_buffer_page_end(dst_pixels_c);
 }
 
 TEST_F(LibYUVPlanarTest, MergeRGBPlane_Opt) {
+  // Round count up to multiple of 16
   const int kPixels = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_pixels, kPixels * 3);
   align_buffer_page_end(tmp_pixels_r, kPixels);
@@ -2617,6 +3084,7 @@ TEST_F(LibYUVPlanarTest, MergeRGBPlane_Opt) {
 }
 
 TEST_F(LibYUVPlanarTest, SplitRGBPlane_Opt) {
+  // Round count up to multiple of 16
   const int kPixels = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_pixels, kPixels * 3);
   align_buffer_page_end(tmp_pixels_r, kPixels);
@@ -2663,10 +3131,373 @@ TEST_F(LibYUVPlanarTest, SplitRGBPlane_Opt) {
   free_aligned_buffer_page_end(dst_pixels_c);
 }
 
+TEST_F(LibYUVPlanarTest, MergeARGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(tmp_pixels_a, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(tmp_pixels_a, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_c, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g,
+                   benchmark_width_, tmp_pixels_b, benchmark_width_,
+                   tmp_pixels_a, benchmark_width_, dst_pixels_opt,
+                   benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(tmp_pixels_a);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitARGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(tmp_pixels_a, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(tmp_pixels_a, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 benchmark_width_, benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_c, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                   benchmark_width_, tmp_pixels_g, benchmark_width_,
+                   tmp_pixels_b, benchmark_width_, tmp_pixels_a,
+                   benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, tmp_pixels_a, benchmark_width_,
+                 dst_pixels_opt, benchmark_width_ * 4, benchmark_width_,
+                 benchmark_height_);
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(tmp_pixels_a);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, MergeXRGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_c,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g,
+                   benchmark_width_, tmp_pixels_b, benchmark_width_, NULL, 0,
+                   dst_pixels_opt, benchmark_width_ * 4, benchmark_width_,
+                   benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+TEST_F(LibYUVPlanarTest, SplitXRGBPlane_Opt) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels, kPixels * 4);
+  align_buffer_page_end(tmp_pixels_r, kPixels);
+  align_buffer_page_end(tmp_pixels_g, kPixels);
+  align_buffer_page_end(tmp_pixels_b, kPixels);
+  align_buffer_page_end(dst_pixels_opt, kPixels * 4);
+  align_buffer_page_end(dst_pixels_c, kPixels * 4);
+
+  MemRandomize(src_pixels, kPixels * 4);
+  MemRandomize(tmp_pixels_r, kPixels);
+  MemRandomize(tmp_pixels_g, kPixels);
+  MemRandomize(tmp_pixels_b, kPixels);
+  MemRandomize(dst_pixels_opt, kPixels * 4);
+  MemRandomize(dst_pixels_c, kPixels * 4);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                 benchmark_width_, tmp_pixels_g, benchmark_width_, tmp_pixels_b,
+                 benchmark_width_, NULL, 0, benchmark_width_,
+                 benchmark_height_);
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_c,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  MaskCpuFlags(benchmark_cpu_info_);
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    SplitARGBPlane(src_pixels, benchmark_width_ * 4, tmp_pixels_r,
+                   benchmark_width_, tmp_pixels_g, benchmark_width_,
+                   tmp_pixels_b, benchmark_width_, NULL, 0, benchmark_width_,
+                   benchmark_height_);
+  }
+
+  MergeARGBPlane(tmp_pixels_r, benchmark_width_, tmp_pixels_g, benchmark_width_,
+                 tmp_pixels_b, benchmark_width_, NULL, 0, dst_pixels_opt,
+                 benchmark_width_ * 4, benchmark_width_, benchmark_height_);
+
+  for (int i = 0; i < kPixels * 4; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels);
+  free_aligned_buffer_page_end(tmp_pixels_r);
+  free_aligned_buffer_page_end(tmp_pixels_g);
+  free_aligned_buffer_page_end(tmp_pixels_b);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(dst_pixels_c);
+}
+
+// Merge 4 channels
+#define TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)      \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_##DEPTH##N) {                        \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_a, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_a, kPixels * sizeof(STYPE) + OFF);              \
+    memset(dst_memory_c, 0, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_memory_opt, 0, kPixels * 4 * sizeof(DTYPE));                 \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    STYPE* src_pixels_a = reinterpret_cast<STYPE*>(src_memory_a + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, src_pixels_a, kWidth, dst_pixels_c, kWidth * 4,     \
+                kWidth, NEG benchmark_height_, DEPTH);                      \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, src_pixels_a, kWidth, dst_pixels_opt, kWidth * 4, \
+                  kWidth, NEG benchmark_height_, DEPTH);                    \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(src_memory_a);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+// Merge 3 channel RGB into 4 channel XRGB with opaque alpha
+#define TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)     \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_Opaque_##DEPTH##N) {                 \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    memset(dst_memory_c, 0, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_memory_opt, 0, kPixels * 4 * sizeof(DTYPE));                 \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, NULL, 0, dst_pixels_c, kWidth * 4, kWidth,          \
+                NEG benchmark_height_, DEPTH);                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, NULL, 0, dst_pixels_opt, kWidth * 4, kWidth,      \
+                  NEG benchmark_height_, DEPTH);                            \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+#define TESTQPLANARTOP(FUNC, STYPE, DTYPE, DEPTH)                              \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +, 0) \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +,  \
+                  2)                                                           \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0)  \
+  TESTQPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)     \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +,   \
+                   0)                                                          \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +, \
+                   2)                                                          \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0) \
+  TESTQPLANAROTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)
+
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 10)
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 12)
+TESTQPLANARTOP(MergeAR64, uint16_t, uint16_t, 16)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 10)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 12)
+TESTQPLANARTOP(MergeARGB16To8, uint16_t, uint8_t, 16)
+
+#define TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, W1280, N, NEG, OFF)      \
+  TEST_F(LibYUVPlanarTest, FUNC##Plane_##DEPTH##N) {                        \
+    const int kWidth = W1280;                                               \
+    const int kPixels = kWidth * benchmark_height_;                         \
+    align_buffer_page_end(src_memory_r, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_g, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(src_memory_b, kPixels * sizeof(STYPE) + OFF);     \
+    align_buffer_page_end(dst_memory_c, kPixels * 4 * sizeof(DTYPE));       \
+    align_buffer_page_end(dst_memory_opt, kPixels * 4 * sizeof(DTYPE));     \
+    MemRandomize(src_memory_r, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_g, kPixels * sizeof(STYPE) + OFF);              \
+    MemRandomize(src_memory_b, kPixels * sizeof(STYPE) + OFF);              \
+    STYPE* src_pixels_r = reinterpret_cast<STYPE*>(src_memory_r + OFF);     \
+    STYPE* src_pixels_g = reinterpret_cast<STYPE*>(src_memory_g + OFF);     \
+    STYPE* src_pixels_b = reinterpret_cast<STYPE*>(src_memory_b + OFF);     \
+    DTYPE* dst_pixels_c = reinterpret_cast<DTYPE*>(dst_memory_c);           \
+    DTYPE* dst_pixels_opt = reinterpret_cast<DTYPE*>(dst_memory_opt);       \
+    memset(dst_pixels_c, 1, kPixels * 4 * sizeof(DTYPE));                   \
+    memset(dst_pixels_opt, 2, kPixels * 4 * sizeof(DTYPE));                 \
+    MaskCpuFlags(disable_cpu_flags_);                                       \
+    FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b,   \
+                kWidth, dst_pixels_c, kWidth * 4, kWidth,                   \
+                NEG benchmark_height_, DEPTH);                              \
+    MaskCpuFlags(benchmark_cpu_info_);                                      \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                       \
+      FUNC##Plane(src_pixels_r, kWidth, src_pixels_g, kWidth, src_pixels_b, \
+                  kWidth, dst_pixels_opt, kWidth * 4, kWidth,               \
+                  NEG benchmark_height_, DEPTH);                            \
+    }                                                                       \
+    for (int i = 0; i < kPixels * 4; ++i) {                                 \
+      EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);                        \
+    }                                                                       \
+    free_aligned_buffer_page_end(src_memory_r);                             \
+    free_aligned_buffer_page_end(src_memory_g);                             \
+    free_aligned_buffer_page_end(src_memory_b);                             \
+    free_aligned_buffer_page_end(dst_memory_c);                             \
+    free_aligned_buffer_page_end(dst_memory_opt);                           \
+  }
+
+#define TESTTPLANARTOP(FUNC, STYPE, DTYPE, DEPTH)                              \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_ + 1, _Any, +, 0) \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Unaligned, +,  \
+                  2)                                                           \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Invert, -, 0)  \
+  TESTTPLANARTOPI(FUNC, STYPE, DTYPE, DEPTH, benchmark_width_, _Opt, +, 0)
+
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 10)
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 12)
+TESTTPLANARTOP(MergeXR30, uint16_t, uint8_t, 16)
+
 // TODO(fbarchard): improve test for platforms and cpu detect
 #ifdef HAS_MERGEUVROW_16_AVX2
 TEST_F(LibYUVPlanarTest, MergeUVRow_16_Opt) {
-  const int kPixels = benchmark_width_ * benchmark_height_;
+  // Round count up to multiple of 16
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 15) & ~15;
+
   align_buffer_page_end(src_pixels_u, kPixels * 2);
   align_buffer_page_end(src_pixels_v, kPixels * 2);
   align_buffer_page_end(dst_pixels_uv_opt, kPixels * 2 * 2);
@@ -2679,19 +3510,19 @@ TEST_F(LibYUVPlanarTest, MergeUVRow_16_Opt) {
 
   MergeUVRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_u),
                   reinterpret_cast<const uint16_t*>(src_pixels_v),
-                  reinterpret_cast<uint16_t*>(dst_pixels_uv_c), 64, kPixels);
+                  reinterpret_cast<uint16_t*>(dst_pixels_uv_c), 16, kPixels);
 
   int has_avx2 = TestCpuFlag(kCpuHasAVX2);
   for (int i = 0; i < benchmark_iterations_; ++i) {
     if (has_avx2) {
       MergeUVRow_16_AVX2(reinterpret_cast<const uint16_t*>(src_pixels_u),
                          reinterpret_cast<const uint16_t*>(src_pixels_v),
-                         reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 64,
+                         reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 16,
                          kPixels);
     } else {
       MergeUVRow_16_C(reinterpret_cast<const uint16_t*>(src_pixels_u),
                       reinterpret_cast<const uint16_t*>(src_pixels_v),
-                      reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 64,
+                      reinterpret_cast<uint16_t*>(dst_pixels_uv_opt), 16,
                       kPixels);
     }
   }
@@ -2710,7 +3541,9 @@ TEST_F(LibYUVPlanarTest, MergeUVRow_16_Opt) {
 // TODO(fbarchard): Improve test for more platforms.
 #ifdef HAS_MULTIPLYROW_16_AVX2
 TEST_F(LibYUVPlanarTest, MultiplyRow_16_Opt) {
-  const int kPixels = benchmark_width_ * benchmark_height_;
+  // Round count up to multiple of 32
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 31) & ~31;
+
   align_buffer_page_end(src_pixels_y, kPixels * 2);
   align_buffer_page_end(dst_pixels_y_opt, kPixels * 2);
   align_buffer_page_end(dst_pixels_y_c, kPixels * 2);
@@ -2776,6 +3609,65 @@ TEST_F(LibYUVPlanarTest, Convert16To8Plane) {
   free_aligned_buffer_page_end(dst_pixels_y_c);
 }
 
+TEST_F(LibYUVPlanarTest, YUY2ToY) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  YUY2ToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_c, benchmark_width_,
+          benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    YUY2ToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_opt,
+            benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+TEST_F(LibYUVPlanarTest, UYVYToY) {
+  const int kPixels = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  UYVYToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_c, benchmark_width_,
+          benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    UYVYToY(src_pixels_y, benchmark_width_ * 2, dst_pixels_y_opt,
+            benchmark_width_, benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+
+#ifdef ENABLE_ROW_TESTS
 // TODO(fbarchard): Improve test for more platforms.
 #ifdef HAS_CONVERT16TO8ROW_AVX2
 TEST_F(LibYUVPlanarTest, Convert16To8Row_Opt) {
@@ -2822,6 +3714,36 @@ TEST_F(LibYUVPlanarTest, Convert16To8Row_Opt) {
 }
 #endif  // HAS_CONVERT16TO8ROW_AVX2
 
+#ifdef HAS_UYVYTOYROW_NEON
+TEST_F(LibYUVPlanarTest, UYVYToYRow_Opt) {
+  // NEON does multiple of 16, so round count up
+  const int kPixels = (benchmark_width_ * benchmark_height_ + 15) & ~15;
+  align_buffer_page_end(src_pixels_y, kPixels * 2);
+  align_buffer_page_end(dst_pixels_y_opt, kPixels);
+  align_buffer_page_end(dst_pixels_y_c, kPixels);
+
+  MemRandomize(src_pixels_y, kPixels * 2);
+  memset(dst_pixels_y_opt, 0, kPixels);
+  memset(dst_pixels_y_c, 1, kPixels);
+
+  UYVYToYRow_C(src_pixels_y, dst_pixels_y_c, kPixels);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    UYVYToYRow_NEON(src_pixels_y, dst_pixels_y_opt, kPixels);
+  }
+
+  for (int i = 0; i < kPixels; ++i) {
+    EXPECT_EQ(dst_pixels_y_opt[i], dst_pixels_y_c[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_y);
+  free_aligned_buffer_page_end(dst_pixels_y_opt);
+  free_aligned_buffer_page_end(dst_pixels_y_c);
+}
+#endif  // HAS_UYVYTOYROW_NEON
+
+#endif  // ENABLE_ROW_TESTS
+
 TEST_F(LibYUVPlanarTest, Convert8To16Plane) {
   const int kPixels = benchmark_width_ * benchmark_height_;
   align_buffer_page_end(src_pixels_y, kPixels);
@@ -2855,6 +3777,7 @@ TEST_F(LibYUVPlanarTest, Convert8To16Plane) {
   free_aligned_buffer_page_end(dst_pixels_y_c);
 }
 
+#ifdef ENABLE_ROW_TESTS
 // TODO(fbarchard): Improve test for more platforms.
 #ifdef HAS_CONVERT8TO16ROW_AVX2
 TEST_F(LibYUVPlanarTest, Convert8To16Row_Opt) {
@@ -3173,33 +4096,33 @@ extern "C" void GaussRow_NEON(const uint32_t* src, uint16_t* dst, int width);
 extern "C" void GaussRow_C(const uint32_t* src, uint16_t* dst, int width);
 
 TEST_F(LibYUVPlanarTest, TestGaussRow_Opt) {
-  SIMD_ALIGNED(uint32_t orig_pixels[640 + 4]);
-  SIMD_ALIGNED(uint16_t dst_pixels_c[640]);
-  SIMD_ALIGNED(uint16_t dst_pixels_opt[640]);
+  SIMD_ALIGNED(uint32_t orig_pixels[1280 + 8]);
+  SIMD_ALIGNED(uint16_t dst_pixels_c[1280]);
+  SIMD_ALIGNED(uint16_t dst_pixels_opt[1280]);
 
   memset(orig_pixels, 0, sizeof(orig_pixels));
   memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
   memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
 
-  for (int i = 0; i < 640 + 4; ++i) {
+  for (int i = 0; i < 1280 + 8; ++i) {
     orig_pixels[i] = i * 256;
   }
-  GaussRow_C(&orig_pixels[0], &dst_pixels_c[0], 640);
-  for (int i = 0; i < benchmark_pixels_div1280_ * 2; ++i) {
+  GaussRow_C(&orig_pixels[0], &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
     int has_neon = TestCpuFlag(kCpuHasNEON);
     if (has_neon) {
-      GaussRow_NEON(&orig_pixels[0], &dst_pixels_opt[0], 640);
+      GaussRow_NEON(&orig_pixels[0], &dst_pixels_opt[0], 1280);
     } else {
-      GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 640);
+      GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
     }
 #else
-    GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 640);
+    GaussRow_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
 #endif
   }
 
-  for (int i = 0; i < 640; ++i) {
+  for (int i = 0; i < 1280; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
 
@@ -3225,141 +4148,139 @@ extern "C" void GaussCol_C(const uint16_t* src0,
                            int width);
 
 TEST_F(LibYUVPlanarTest, TestGaussCol_Opt) {
-  SIMD_ALIGNED(uint16_t orig_pixels[640 * 5]);
-  SIMD_ALIGNED(uint32_t dst_pixels_c[640]);
-  SIMD_ALIGNED(uint32_t dst_pixels_opt[640]);
+  SIMD_ALIGNED(uint16_t orig_pixels[1280 * 5]);
+  SIMD_ALIGNED(uint32_t dst_pixels_c[1280]);
+  SIMD_ALIGNED(uint32_t dst_pixels_opt[1280]);
 
   memset(orig_pixels, 0, sizeof(orig_pixels));
   memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
   memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
 
-  for (int i = 0; i < 640 * 5; ++i) {
-    orig_pixels[i] = i;
+  for (int i = 0; i < 1280 * 5; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
   }
-  GaussCol_C(&orig_pixels[0], &orig_pixels[640], &orig_pixels[640 * 2],
-             &orig_pixels[640 * 3], &orig_pixels[640 * 4], &dst_pixels_c[0],
-             640);
-  for (int i = 0; i < benchmark_pixels_div1280_ * 2; ++i) {
+  GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+             &orig_pixels[1280 * 3], &orig_pixels[1280 * 4], &dst_pixels_c[0],
+             1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
 #if !defined(LIBYUV_DISABLE_NEON) && \
     (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
     int has_neon = TestCpuFlag(kCpuHasNEON);
     if (has_neon) {
-      GaussCol_NEON(&orig_pixels[0], &orig_pixels[640], &orig_pixels[640 * 2],
-                    &orig_pixels[640 * 3], &orig_pixels[640 * 4],
-                    &dst_pixels_opt[0], 640);
+      GaussCol_NEON(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                    &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                    &dst_pixels_opt[0], 1280);
     } else {
-      GaussCol_C(&orig_pixels[0], &orig_pixels[640], &orig_pixels[640 * 2],
-                 &orig_pixels[640 * 3], &orig_pixels[640 * 4],
-                 &dst_pixels_opt[0], 640);
+      GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                 &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                 &dst_pixels_opt[0], 1280);
     }
 #else
-    GaussCol_C(&orig_pixels[0], &orig_pixels[640], &orig_pixels[640 * 2],
-               &orig_pixels[640 * 3], &orig_pixels[640 * 4], &dst_pixels_opt[0],
-               640);
+    GaussCol_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+               &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+               &dst_pixels_opt[0], 1280);
 #endif
   }
 
-  for (int i = 0; i < 640; ++i) {
+  for (int i = 0; i < 1280; ++i) {
     EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
-
-  EXPECT_EQ(dst_pixels_c[0],
-            static_cast<uint32_t>(0 * 1 + 640 * 4 + 640 * 2 * 6 + 640 * 3 * 4 +
-                                  640 * 4 * 1));
-  EXPECT_EQ(dst_pixels_c[639], static_cast<uint32_t>(30704));
 }
 
-float TestFloatDivToByte(int benchmark_width,
-                         int benchmark_height,
-                         int benchmark_iterations,
-                         float scale,
-                         bool opt) {
-  int i, j;
-  // NEON does multiple of 8, so round count up
-  const int kPixels = (benchmark_width * benchmark_height + 7) & ~7;
-  align_buffer_page_end(src_weights, kPixels * 4);
-  align_buffer_page_end(src_values, kPixels * 4);
-  align_buffer_page_end(dst_out_c, kPixels);
-  align_buffer_page_end(dst_out_opt, kPixels);
-  align_buffer_page_end(dst_mask_c, kPixels);
-  align_buffer_page_end(dst_mask_opt, kPixels);
-
-  // Randomize works but may contain some denormals affecting performance.
-  // MemRandomize(orig_y, kPixels * 4);
-  // large values are problematic.  audio is really -1 to 1.
-  for (i = 0; i < kPixels; ++i) {
-    (reinterpret_cast<float*>(src_weights))[i] = scale;
-    (reinterpret_cast<float*>(src_values))[i] =
-        sinf(static_cast<float>(i) * 0.1f);
-  }
-  memset(dst_out_c, 0, kPixels);
-  memset(dst_out_opt, 1, kPixels);
-  memset(dst_mask_c, 2, kPixels);
-  memset(dst_mask_opt, 3, kPixels);
+TEST_F(LibYUVPlanarTest, TestGaussRow_F32_Opt) {
+  SIMD_ALIGNED(float orig_pixels[1280 + 4]);
+  SIMD_ALIGNED(float dst_pixels_c[1280]);
+  SIMD_ALIGNED(float dst_pixels_opt[1280]);
 
-  FloatDivToByteRow_C(reinterpret_cast<float*>(src_weights),
-                      reinterpret_cast<float*>(src_values), dst_out_c,
-                      dst_mask_c, kPixels);
+  memset(orig_pixels, 0, sizeof(orig_pixels));
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
 
-  for (j = 0; j < benchmark_iterations; j++) {
-    if (opt) {
-#ifdef HAS_FLOATDIVTOBYTEROW_NEON
-      FloatDivToByteRow_NEON(reinterpret_cast<float*>(src_weights),
-                             reinterpret_cast<float*>(src_values), dst_out_opt,
-                             dst_mask_opt, kPixels);
-#else
-      FloatDivToByteRow_C(reinterpret_cast<float*>(src_weights),
-                          reinterpret_cast<float*>(src_values), dst_out_opt,
-                          dst_mask_opt, kPixels);
-#endif
+  for (int i = 0; i < 1280 + 4; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
+  }
+  GaussRow_F32_C(&orig_pixels[0], &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussRow_F32_NEON(&orig_pixels[0], &dst_pixels_opt[0], 1280);
     } else {
-      FloatDivToByteRow_C(reinterpret_cast<float*>(src_weights),
-                          reinterpret_cast<float*>(src_values), dst_out_opt,
-                          dst_mask_opt, kPixels);
+      GaussRow_F32_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
     }
+#else
+    GaussRow_F32_C(&orig_pixels[0], &dst_pixels_opt[0], 1280);
+#endif
   }
 
-  uint8_t max_diff = 0;
-  for (i = 0; i < kPixels; ++i) {
-    uint8_t abs_diff = abs(dst_out_c[i] - dst_out_opt[i]) +
-                       abs(dst_mask_c[i] - dst_mask_opt[i]);
-    if (abs_diff > max_diff) {
-      max_diff = abs_diff;
-    }
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
   }
+}
 
-  free_aligned_buffer_page_end(src_weights);
-  free_aligned_buffer_page_end(src_values);
-  free_aligned_buffer_page_end(dst_out_c);
-  free_aligned_buffer_page_end(dst_out_opt);
-  free_aligned_buffer_page_end(dst_mask_c);
-  free_aligned_buffer_page_end(dst_mask_opt);
+TEST_F(LibYUVPlanarTest, TestGaussCol_F32_Opt) {
+  SIMD_ALIGNED(float dst_pixels_c[1280]);
+  SIMD_ALIGNED(float dst_pixels_opt[1280]);
+  align_buffer_page_end(orig_pixels_buf, 1280 * 5 * 4);  // 5 rows
+  float* orig_pixels = reinterpret_cast<float*>(orig_pixels_buf);
 
-  return max_diff;
-}
+  memset(orig_pixels, 0, 1280 * 5 * 4);
+  memset(dst_pixels_c, 1, sizeof(dst_pixels_c));
+  memset(dst_pixels_opt, 2, sizeof(dst_pixels_opt));
 
-TEST_F(LibYUVPlanarTest, TestFloatDivToByte_C) {
-  float diff = TestFloatDivToByte(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_, 1.2f, false);
-  EXPECT_EQ(0, diff);
-}
+  for (int i = 0; i < 1280 * 5; ++i) {
+    orig_pixels[i] = static_cast<float>(i);
+  }
+  GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                 &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                 &dst_pixels_c[0], 1280);
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+    int has_neon = TestCpuFlag(kCpuHasNEON);
+    if (has_neon) {
+      GaussCol_F32_NEON(&orig_pixels[0], &orig_pixels[1280],
+                        &orig_pixels[1280 * 2], &orig_pixels[1280 * 3],
+                        &orig_pixels[1280 * 4], &dst_pixels_opt[0], 1280);
+    } else {
+      GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280],
+                     &orig_pixels[1280 * 2], &orig_pixels[1280 * 3],
+                     &orig_pixels[1280 * 4], &dst_pixels_opt[0], 1280);
+    }
+#else
+    GaussCol_F32_C(&orig_pixels[0], &orig_pixels[1280], &orig_pixels[1280 * 2],
+                   &orig_pixels[1280 * 3], &orig_pixels[1280 * 4],
+                   &dst_pixels_opt[0], 1280);
+#endif
+  }
 
-TEST_F(LibYUVPlanarTest, TestFloatDivToByte_Opt) {
-  float diff = TestFloatDivToByte(benchmark_width_, benchmark_height_,
-                                  benchmark_iterations_, 1.2f, true);
-  EXPECT_EQ(0, diff);
+  for (int i = 0; i < 1280; ++i) {
+    EXPECT_EQ(dst_pixels_c[i], dst_pixels_opt[i]);
+  }
+  free_aligned_buffer_page_end(orig_pixels_buf);
 }
 
-TEST_F(LibYUVPlanarTest, UVToVURow) {
+TEST_F(LibYUVPlanarTest, SwapUVRow) {
   const int kPixels = benchmark_width_ * benchmark_height_;
+  void (*SwapUVRow)(const uint8_t* src_uv, uint8_t* dst_vu, int width) =
+      SwapUVRow_C;
+
   align_buffer_page_end(src_pixels_vu, kPixels * 2);
   align_buffer_page_end(dst_pixels_uv, kPixels * 2);
-
   MemRandomize(src_pixels_vu, kPixels * 2);
   memset(dst_pixels_uv, 1, kPixels * 2);
 
-  UVToVURow_C(src_pixels_vu, dst_pixels_uv, kPixels);
+#if defined(HAS_SWAPUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SwapUVRow = SwapUVRow_Any_NEON;
+    if (IS_ALIGNED(kPixels, 16)) {
+      SwapUVRow = SwapUVRow_NEON;
+    }
+  }
+#endif
 
+  for (int j = 0; j < benchmark_iterations_; j++) {
+    SwapUVRow(src_pixels_vu, dst_pixels_uv, kPixels);
+  }
   for (int i = 0; i < kPixels; ++i) {
     EXPECT_EQ(dst_pixels_uv[i * 2 + 0], src_pixels_vu[i * 2 + 1]);
     EXPECT_EQ(dst_pixels_uv[i * 2 + 1], src_pixels_vu[i * 2 + 0]);
@@ -3368,5 +4289,144 @@ TEST_F(LibYUVPlanarTest, UVToVURow) {
   free_aligned_buffer_page_end(src_pixels_vu);
   free_aligned_buffer_page_end(dst_pixels_uv);
 }
+#endif  // ENABLE_ROW_TESTS
+
+TEST_F(LibYUVPlanarTest, TestGaussPlane_F32) {
+  const int kSize = benchmark_width_ * benchmark_height_ * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  align_buffer_page_end(dst_pixels_opt, kSize);
+  align_buffer_page_end(dst_pixels_c, kSize);
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    ((float*)(orig_pixels))[i] = (i & 1023) * 3.14f;
+  }
+  memset(dst_pixels_opt, 1, kSize);
+  memset(dst_pixels_c, 2, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  GaussPlane_F32((const float*)(orig_pixels), benchmark_width_,
+                 (float*)(dst_pixels_c), benchmark_width_, benchmark_width_,
+                 benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    GaussPlane_F32((const float*)(orig_pixels), benchmark_width_,
+                   (float*)(dst_pixels_opt), benchmark_width_, benchmark_width_,
+                   benchmark_height_);
+  }
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_NEAR(((float*)(dst_pixels_c))[i], ((float*)(dst_pixels_opt))[i], 1.f)
+        << i;
+  }
+
+  free_aligned_buffer_page_end(dst_pixels_c);
+  free_aligned_buffer_page_end(dst_pixels_opt);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVPlanarTest, HalfMergeUVPlane_Opt) {
+  int dst_width = (benchmark_width_ + 1) / 2;
+  int dst_height = (benchmark_height_ + 1) / 2;
+  align_buffer_page_end(src_pixels_u, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_pixels_v, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(tmp_pixels_u, dst_width * dst_height);
+  align_buffer_page_end(tmp_pixels_v, dst_width * dst_height);
+  align_buffer_page_end(dst_pixels_uv_opt, dst_width * 2 * dst_height);
+  align_buffer_page_end(dst_pixels_uv_c, dst_width * 2 * dst_height);
+
+  MemRandomize(src_pixels_u, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_pixels_v, benchmark_width_ * benchmark_height_);
+  MemRandomize(tmp_pixels_u, dst_width * dst_height);
+  MemRandomize(tmp_pixels_v, dst_width * dst_height);
+  MemRandomize(dst_pixels_uv_opt, dst_width * 2 * dst_height);
+  MemRandomize(dst_pixels_uv_c, dst_width * 2 * dst_height);
+
+  MaskCpuFlags(disable_cpu_flags_);
+  HalfMergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v,
+                   benchmark_width_, dst_pixels_uv_c, dst_width * 2,
+                   benchmark_width_, benchmark_height_);
+  MaskCpuFlags(benchmark_cpu_info_);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    HalfMergeUVPlane(src_pixels_u, benchmark_width_, src_pixels_v,
+                     benchmark_width_, dst_pixels_uv_opt, dst_width * 2,
+                     benchmark_width_, benchmark_height_);
+  }
+
+  for (int i = 0; i < dst_width * 2 * dst_height; ++i) {
+    EXPECT_EQ(dst_pixels_uv_c[i], dst_pixels_uv_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(src_pixels_u);
+  free_aligned_buffer_page_end(src_pixels_v);
+  free_aligned_buffer_page_end(tmp_pixels_u);
+  free_aligned_buffer_page_end(tmp_pixels_v);
+  free_aligned_buffer_page_end(dst_pixels_uv_opt);
+  free_aligned_buffer_page_end(dst_pixels_uv_c);
+}
+
+TEST_F(LibYUVPlanarTest, NV12Copy) {
+  const int halfwidth = (benchmark_width_ + 1) >> 1;
+  const int halfheight = (benchmark_height_ + 1) >> 1;
+  align_buffer_page_end(src_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_uv, halfwidth * 2 * halfheight);
+  align_buffer_page_end(dst_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_uv, halfwidth * 2 * halfheight);
+
+  MemRandomize(src_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_uv, halfwidth * 2 * halfheight);
+  MemRandomize(dst_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(dst_uv, halfwidth * 2 * halfheight);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    NV12Copy(src_y, benchmark_width_, src_uv, halfwidth * 2, dst_y,
+             benchmark_width_, dst_uv, halfwidth * 2, benchmark_width_,
+             benchmark_height_);
+  }
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(src_y[i], dst_y[i]);
+  }
+  for (int i = 0; i < halfwidth * 2 * halfheight; ++i) {
+    EXPECT_EQ(src_uv[i], dst_uv[i]);
+  }
+
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_uv);
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_uv);
+}
+
+TEST_F(LibYUVPlanarTest, NV21Copy) {
+  const int halfwidth = (benchmark_width_ + 1) >> 1;
+  const int halfheight = (benchmark_height_ + 1) >> 1;
+  align_buffer_page_end(src_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(src_vu, halfwidth * 2 * halfheight);
+  align_buffer_page_end(dst_y, benchmark_width_ * benchmark_height_);
+  align_buffer_page_end(dst_vu, halfwidth * 2 * halfheight);
+
+  MemRandomize(src_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(src_vu, halfwidth * 2 * halfheight);
+  MemRandomize(dst_y, benchmark_width_ * benchmark_height_);
+  MemRandomize(dst_vu, halfwidth * 2 * halfheight);
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    NV21Copy(src_y, benchmark_width_, src_vu, halfwidth * 2, dst_y,
+             benchmark_width_, dst_vu, halfwidth * 2, benchmark_width_,
+             benchmark_height_);
+  }
+
+  for (int i = 0; i < benchmark_width_ * benchmark_height_; ++i) {
+    EXPECT_EQ(src_y[i], dst_y[i]);
+  }
+  for (int i = 0; i < halfwidth * 2 * halfheight; ++i) {
+    EXPECT_EQ(src_vu[i], dst_vu[i]);
+  }
+
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_vu);
+  free_aligned_buffer_page_end(dst_y);
+  free_aligned_buffer_page_end(dst_vu);
+}
 
 }  // namespace libyuv
diff --git a/files/unit_test/rotate_argb_test.cc b/files/unit_test/rotate_argb_test.cc
index d2003895..01ed69ca 100644
--- a/files/unit_test/rotate_argb_test.cc
+++ b/files/unit_test/rotate_argb_test.cc
@@ -156,31 +156,73 @@ TEST_F(LibYUVRotateTest, RotatePlane270_Opt) {
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane0_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_width_ - 3, benchmark_height_ - 1, kRotate0,
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
                   benchmark_iterations_, disable_cpu_flags_,
                   benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane90_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_height_ - 1, benchmark_width_ - 3, kRotate90,
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
                   benchmark_iterations_, disable_cpu_flags_,
                   benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane180_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_width_ - 3, benchmark_height_ - 1, kRotate180,
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
                   benchmark_iterations_, disable_cpu_flags_,
                   benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_RotatePlane270_Odd) {
-  TestRotatePlane(benchmark_width_ - 3, benchmark_height_ - 1,
-                  benchmark_height_ - 1, benchmark_width_ - 3, kRotate270,
+  TestRotatePlane(benchmark_width_ + 1, benchmark_height_ + 1,
+                  benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
                   benchmark_iterations_, disable_cpu_flags_,
                   benchmark_cpu_info_);
 }
 
+TEST_F(LibYUVRotateTest, RotatePlane90_TestStride) {
+  int argb_plane_size = benchmark_width_ * 4 * abs(benchmark_height_);
+
+  align_buffer_page_end(src_argb, argb_plane_size);
+  align_buffer_page_end(dst_argb, argb_plane_size);
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          benchmark_width_ * 4, benchmark_width_,
+                          benchmark_height_, kRotate0));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                          benchmark_width_ * 4 - 1, benchmark_width_ - 1,
+                          benchmark_height_, kRotate0));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          benchmark_width_ * 4, benchmark_width_,
+                          benchmark_height_, kRotate180));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                          benchmark_width_ * 4 - 1, benchmark_width_ - 1,
+                          benchmark_height_, kRotate180));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          abs(benchmark_height_) * 4, benchmark_width_,
+                          benchmark_height_, kRotate90));
+
+  EXPECT_EQ(-1, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                           abs(benchmark_height_) * 4, benchmark_width_ - 1,
+                           benchmark_height_, kRotate90));
+
+  EXPECT_EQ(0, ARGBRotate(src_argb, benchmark_width_ * 4, dst_argb,
+                          abs(benchmark_height_) * 4, benchmark_width_,
+                          benchmark_height_, kRotate270));
+
+  EXPECT_EQ(-1, ARGBRotate(src_argb, benchmark_width_ * 4 - 1, dst_argb,
+                           abs(benchmark_height_) * 4, benchmark_width_ - 1,
+                           benchmark_height_, kRotate270));
+
+  free_aligned_buffer_page_end(dst_argb);
+  free_aligned_buffer_page_end(src_argb);
+}
+
 }  // namespace libyuv
diff --git a/files/unit_test/rotate_test.cc b/files/unit_test/rotate_test.cc
index 61941e63..d3887414 100644
--- a/files/unit_test/rotate_test.cc
+++ b/files/unit_test/rotate_test.cc
@@ -16,6 +16,8 @@
 
 namespace libyuv {
 
+#define SUBSAMPLE(v, a) ((((v) + (a)-1)) / (a))
+
 static void I420TestRotate(int src_width,
                            int src_height,
                            int dst_width,
@@ -108,33 +110,121 @@ TEST_F(LibYUVRotateTest, I420Rotate270_Opt) {
 // Odd width tests work but disabled because they use C code and can be
 // tested by passing an odd width command line or environment variable.
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate0_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate0,
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate90_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate90,
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate180_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate180,
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I420Rotate270_Odd) {
-  I420TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate270,
+  I420TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
+static void I422TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i422_y_size = src_width * Abs(src_height);
+  int src_i422_uv_size = ((src_width + 1) / 2) * Abs(src_height);
+  int src_i422_size = src_i422_y_size + src_i422_uv_size * 2;
+  align_buffer_page_end(src_i422, src_i422_size);
+  for (int i = 0; i < src_i422_size; ++i) {
+    src_i422[i] = fastrand() & 0xff;
+  }
+
+  int dst_i422_y_size = dst_width * dst_height;
+  int dst_i422_uv_size = ((dst_width + 1) / 2) * dst_height;
+  int dst_i422_size = dst_i422_y_size + dst_i422_uv_size * 2;
+  align_buffer_page_end(dst_i422_c, dst_i422_size);
+  align_buffer_page_end(dst_i422_opt, dst_i422_size);
+  memset(dst_i422_c, 2, dst_i422_size);
+  memset(dst_i422_opt, 3, dst_i422_size);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I422Rotate(src_i422, src_width, src_i422 + src_i422_y_size,
+             (src_width + 1) / 2, src_i422 + src_i422_y_size + src_i422_uv_size,
+             (src_width + 1) / 2, dst_i422_c, dst_width,
+             dst_i422_c + dst_i422_y_size, (dst_width + 1) / 2,
+             dst_i422_c + dst_i422_y_size + dst_i422_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I422Rotate(
+        src_i422, src_width, src_i422 + src_i422_y_size, (src_width + 1) / 2,
+        src_i422 + src_i422_y_size + src_i422_uv_size, (src_width + 1) / 2,
+        dst_i422_opt, dst_width, dst_i422_opt + dst_i422_y_size,
+        (dst_width + 1) / 2, dst_i422_opt + dst_i422_y_size + dst_i422_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i422_size; ++i) {
+    EXPECT_EQ(dst_i422_c[i], dst_i422_opt[i]);
+  }
+
+  free_aligned_buffer_page_end(dst_i422_c);
+  free_aligned_buffer_page_end(dst_i422_opt);
+  free_aligned_buffer_page_end(src_i422);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate0_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate90_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate180_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I422Rotate270_Opt) {
+  I422TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
 static void I444TestRotate(int src_width,
                            int src_height,
                            int dst_width,
@@ -225,29 +315,29 @@ TEST_F(LibYUVRotateTest, I444Rotate270_Opt) {
 // Odd width tests work but disabled because they use C code and can be
 // tested by passing an odd width command line or environment variable.
 TEST_F(LibYUVRotateTest, DISABLED_I444Rotate0_Odd) {
-  I444TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate0,
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I444Rotate90_Odd) {
-  I444TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate90,
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I444Rotate180_Odd) {
-  I444TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate180,
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_I444Rotate270_Odd) {
-  I444TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate270,
+  I444TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
@@ -340,29 +430,29 @@ TEST_F(LibYUVRotateTest, NV12Rotate270_Opt) {
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate0_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate0,
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate90_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate90,
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate180_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_width_ - 3, benchmark_height_ - 1, kRotate180,
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
 
 TEST_F(LibYUVRotateTest, DISABLED_NV12Rotate270_Odd) {
-  NV12TestRotate(benchmark_width_ - 3, benchmark_height_ - 1,
-                 benchmark_height_ - 1, benchmark_width_ - 3, kRotate270,
+  NV12TestRotate(benchmark_width_ + 1, benchmark_height_ + 1,
+                 benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
                  benchmark_iterations_, disable_cpu_flags_,
                  benchmark_cpu_info_);
 }
@@ -391,4 +481,119 @@ TEST_F(LibYUVRotateTest, NV12Rotate270_Invert) {
                  disable_cpu_flags_, benchmark_cpu_info_);
 }
 
+// Test Android 420 to I420 Rotate
+#define TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X,          \
+                        SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y,      \
+                        W1280, N, NEG, OFF, PN, OFF_U, OFF_V, ROT)            \
+  TEST_F(LibYUVRotateTest,                                                    \
+         SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate##ROT##To##PN##N) {            \
+    const int kWidth = W1280;                                                 \
+    const int kHeight = benchmark_height_;                                    \
+    const int kSizeUV =                                                       \
+        SUBSAMPLE(kWidth, SRC_SUBSAMP_X) * SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); \
+    align_buffer_page_end(src_y, kWidth* kHeight + OFF);                      \
+    align_buffer_page_end(src_uv,                                             \
+                          kSizeUV*((PIXEL_STRIDE == 3) ? 3 : 2) + OFF);       \
+    align_buffer_page_end(dst_y_c, kWidth* kHeight);                          \
+    align_buffer_page_end(dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_v_c, SUBSAMPLE(kWidth, SUBSAMP_X) *             \
+                                       SUBSAMPLE(kHeight, SUBSAMP_Y));        \
+    align_buffer_page_end(dst_y_opt, kWidth* kHeight);                        \
+    align_buffer_page_end(dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    align_buffer_page_end(dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X) *           \
+                                         SUBSAMPLE(kHeight, SUBSAMP_Y));      \
+    uint8_t* src_u = src_uv + OFF_U;                                          \
+    uint8_t* src_v = src_uv + (PIXEL_STRIDE == 1 ? kSizeUV : OFF_V);          \
+    int src_stride_uv = SUBSAMPLE(kWidth, SUBSAMP_X) * PIXEL_STRIDE;          \
+    for (int i = 0; i < kHeight; ++i)                                         \
+      for (int j = 0; j < kWidth; ++j)                                        \
+        src_y[i * kWidth + j + OFF] = (fastrand() & 0xff);                    \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SRC_SUBSAMP_Y); ++i) {             \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SRC_SUBSAMP_X); ++j) {            \
+        src_u[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+        src_v[(i * src_stride_uv) + j * PIXEL_STRIDE + OFF] =                 \
+            (fastrand() & 0xff);                                              \
+      }                                                                       \
+    }                                                                         \
+    memset(dst_y_c, 1, kWidth* kHeight);                                      \
+    memset(dst_u_c, 2,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_c, 3,                                                        \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_y_opt, 101, kWidth* kHeight);                                  \
+    memset(dst_u_opt, 102,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    memset(dst_v_opt, 103,                                                    \
+           SUBSAMPLE(kWidth, SUBSAMP_X) * SUBSAMPLE(kHeight, SUBSAMP_Y));     \
+    MaskCpuFlags(disable_cpu_flags_);                                         \
+    SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate(                                   \
+        src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X),   \
+        src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE, dst_y_c, \
+        kWidth, dst_u_c, SUBSAMPLE(kWidth, SUBSAMP_X), dst_v_c,               \
+        SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight,                    \
+        (libyuv::RotationMode)ROT);                                           \
+    MaskCpuFlags(benchmark_cpu_info_);                                        \
+    for (int i = 0; i < benchmark_iterations_; ++i) {                         \
+      SRC_FMT_PLANAR##To##FMT_PLANAR##Rotate(                                 \
+          src_y + OFF, kWidth, src_u + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), \
+          src_v + OFF, SUBSAMPLE(kWidth, SRC_SUBSAMP_X), PIXEL_STRIDE,        \
+          dst_y_opt, kWidth, dst_u_opt, SUBSAMPLE(kWidth, SUBSAMP_X),         \
+          dst_v_opt, SUBSAMPLE(kWidth, SUBSAMP_X), kWidth, NEG kHeight,       \
+          (libyuv::RotationMode)ROT);                                         \
+    }                                                                         \
+    for (int i = 0; i < kHeight; ++i) {                                       \
+      for (int j = 0; j < kWidth; ++j) {                                      \
+        EXPECT_EQ(dst_y_c[i * kWidth + j], dst_y_opt[i * kWidth + j]);        \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_u_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_u_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    for (int i = 0; i < SUBSAMPLE(kHeight, SUBSAMP_Y); ++i) {                 \
+      for (int j = 0; j < SUBSAMPLE(kWidth, SUBSAMP_X); ++j) {                \
+        EXPECT_EQ(dst_v_c[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j],              \
+                  dst_v_opt[i * SUBSAMPLE(kWidth, SUBSAMP_X) + j]);           \
+      }                                                                       \
+    }                                                                         \
+    free_aligned_buffer_page_end(dst_y_c);                                    \
+    free_aligned_buffer_page_end(dst_u_c);                                    \
+    free_aligned_buffer_page_end(dst_v_c);                                    \
+    free_aligned_buffer_page_end(dst_y_opt);                                  \
+    free_aligned_buffer_page_end(dst_u_opt);                                  \
+    free_aligned_buffer_page_end(dst_v_opt);                                  \
+    free_aligned_buffer_page_end(src_y);                                      \
+    free_aligned_buffer_page_end(src_uv);                                     \
+  }
+
+#define TESTAPLANARTOP(SRC_FMT_PLANAR, PN, PIXEL_STRIDE, OFF_U, OFF_V,         \
+                       SRC_SUBSAMP_X, SRC_SUBSAMP_Y, FMT_PLANAR, SUBSAMP_X,    \
+                       SUBSAMP_Y)                                              \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_ + 1,      \
+                  _Any, +, 0, PN, OFF_U, OFF_V, 0)                             \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_,          \
+                  _Unaligned, +, 2, PN, OFF_U, OFF_V, 0)                       \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Invert, \
+                  -, 0, PN, OFF_U, OFF_V, 0)                                   \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, \
+                  0, PN, OFF_U, OFF_V, 0)                                      \
+  TESTAPLANARTOPI(SRC_FMT_PLANAR, PIXEL_STRIDE, SRC_SUBSAMP_X, SRC_SUBSAMP_Y,  \
+                  FMT_PLANAR, SUBSAMP_X, SUBSAMP_Y, benchmark_width_, _Opt, +, \
+                  0, PN, OFF_U, OFF_V, 180)
+
+TESTAPLANARTOP(Android420, I420, 1, 0, 0, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV12, 2, 0, 1, 2, 2, I420, 2, 2)
+TESTAPLANARTOP(Android420, NV21, 2, 1, 0, 2, 2, I420, 2, 2)
+#undef TESTAPLANARTOP
+#undef TESTAPLANARTOPI
+
 }  // namespace libyuv
diff --git a/files/unit_test/scale_argb_test.cc b/files/unit_test/scale_argb_test.cc
index 94aef60e..f54a68f1 100644
--- a/files/unit_test/scale_argb_test.cc
+++ b/files/unit_test/scale_argb_test.cc
@@ -22,6 +22,12 @@ namespace libyuv {
 #define STRINGIZE(line) #line
 #define FILELINESTR(file, line) file ":" STRINGIZE(line)
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
 static int ARGBTestFilter(int src_width,
                           int src_height,
@@ -114,8 +120,8 @@ static int ARGBTestFilter(int src_width,
   return max_diff;
 }
 
-static const int kTileX = 8;
-static const int kTileY = 8;
+static const int kTileX = 64;
+static const int kTileY = 64;
 
 static int TileARGBScale(const uint8_t* src_argb,
                          int src_stride_argb,
@@ -232,7 +238,7 @@ static int ARGBClipTestFilter(int src_width,
 #define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
 #define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
 
-#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
+#define TEST_FACTOR1(DISABLED_, name, filter, nom, denom, max_diff)          \
   TEST_F(LibYUVScaleTest, ARGBScaleDownBy##name##_##filter) {                \
     int diff = ARGBTestFilter(                                               \
         SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
@@ -241,7 +247,7 @@ static int ARGBClipTestFilter(int src_width,
         benchmark_cpu_info_);                                                \
     EXPECT_LE(diff, max_diff);                                               \
   }                                                                          \
-  TEST_F(LibYUVScaleTest, ARGBScaleDownClipBy##name##_##filter) {            \
+  TEST_F(LibYUVScaleTest, DISABLED_##ARGBScaleDownClipBy##name##_##filter) { \
     int diff = ARGBClipTestFilter(                                           \
         SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
         DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
@@ -251,15 +257,30 @@ static int ARGBClipTestFilter(int src_width,
 
 // Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
 // filtering is different fixed point implementations for SSSE3, Neon and C.
-#define TEST_FACTOR(name, nom, denom)         \
-  TEST_FACTOR1(name, None, nom, denom, 0)     \
-  TEST_FACTOR1(name, Linear, nom, denom, 3)   \
-  TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
-  TEST_FACTOR1(name, Box, nom, denom, 3)
+#ifndef DISABLE_SLOW_TESTS
+#define TEST_FACTOR(name, nom, denom)           \
+  TEST_FACTOR1(, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(, name, Box, nom, denom, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_FACTOR(name, nom, denom)                    \
+  TEST_FACTOR1(DISABLED_, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(DISABLED_, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, 3)
+#else
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3)
+#endif
+#endif
 
 TEST_FACTOR(2, 1, 2)
 TEST_FACTOR(4, 1, 4)
+#ifndef DISABLE_SLOW_TESTS
 TEST_FACTOR(8, 1, 8)
+#endif
 TEST_FACTOR(3by4, 3, 4)
 TEST_FACTOR(3by8, 3, 8)
 TEST_FACTOR(3, 1, 3)
@@ -268,7 +289,7 @@ TEST_FACTOR(3, 1, 3)
 #undef SX
 #undef DX
 
-#define TEST_SCALETO1(name, width, height, filter, max_diff)                   \
+#define TEST_SCALETO1(DISABLED_, name, width, height, filter, max_diff)        \
   TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {             \
     int diff = ARGBTestFilter(benchmark_width_, benchmark_height_, width,      \
                               height, kFilter##filter, benchmark_iterations_,  \
@@ -282,34 +303,70 @@ TEST_FACTOR(3, 1, 3)
                               benchmark_cpu_info_);                            \
     EXPECT_LE(diff, max_diff);                                                 \
   }                                                                            \
-  TEST_F(LibYUVScaleTest, name##ClipTo##width##x##height##_##filter) {         \
+  TEST_F(LibYUVScaleTest,                                                      \
+         DISABLED_##name##ClipTo##width##x##height##_##filter) {               \
     int diff =                                                                 \
         ARGBClipTestFilter(benchmark_width_, benchmark_height_, width, height, \
                            kFilter##filter, benchmark_iterations_);            \
     EXPECT_LE(diff, max_diff);                                                 \
   }                                                                            \
-  TEST_F(LibYUVScaleTest, name##ClipFrom##width##x##height##_##filter) {       \
+  TEST_F(LibYUVScaleTest,                                                      \
+         DISABLED_##name##ClipFrom##width##x##height##_##filter) {             \
     int diff = ARGBClipTestFilter(width, height, Abs(benchmark_width_),        \
                                   Abs(benchmark_height_), kFilter##filter,     \
                                   benchmark_iterations_);                      \
     EXPECT_LE(diff, max_diff);                                                 \
   }
 
-/// Test scale to a specified size with all 4 filters.
-#define TEST_SCALETO(name, width, height)       \
-  TEST_SCALETO1(name, width, height, None, 0)   \
-  TEST_SCALETO1(name, width, height, Linear, 3) \
-  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#ifndef DISABLE_SLOW_TESTS
+// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)         \
+  TEST_SCALETO1(, name, width, height, None, 0)   \
+  TEST_SCALETO1(, name, width, height, Linear, 3) \
+  TEST_SCALETO1(, name, width, height, Bilinear, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_SCALETO(name, width, height)                  \
+  TEST_SCALETO1(DISABLED_, name, width, height, None, 0)   \
+  TEST_SCALETO1(DISABLED_, name, width, height, Linear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3)
+#endif
+#endif
 
 TEST_SCALETO(ARGBScale, 1, 1)
-TEST_SCALETO(ARGBScale, 320, 240)
 TEST_SCALETO(ARGBScale, 569, 480)
 TEST_SCALETO(ARGBScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(ARGBScale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(ARGBScale, 320, 240)
 TEST_SCALETO(ARGBScale, 1280, 720)
 TEST_SCALETO(ARGBScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
 #undef TEST_SCALETO1
 #undef TEST_SCALETO
 
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                       \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                      \
+    int diff = ARGBTestFilter(benchmark_width_, benchmark_height_,      \
+                              benchmark_height_, benchmark_width_,      \
+                              kFilter##filter, benchmark_iterations_,   \
+                              disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                          \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(ARGBScale, None, 0)
+TEST_SCALESWAPXY1(ARGBScale, Linear, 0)
+TEST_SCALESWAPXY1(ARGBScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(ARGBScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
 // Scale with YUV conversion to ARGB and clipping.
 // TODO(fbarchard): Add fourcc support.  All 4 ARGB formats is easy to support.
 LIBYUV_API
@@ -454,4 +511,78 @@ TEST_F(LibYUVScaleTest, YUVToRGBScaleDown) {
   EXPECT_LE(diff, 10);
 }
 
+TEST_F(LibYUVScaleTest, ARGBTest3x) {
+  const int kSrcStride = 480 * 4;
+  const int kDstStride = 160 * 4;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 4 + 0] = i;
+    orig_pixels[i * 4 + 1] = 255 - i;
+    orig_pixels[i * 4 + 2] = i + 1;
+    orig_pixels[i * 4 + 3] = i + 10;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ARGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+              kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+  EXPECT_EQ(226, dest_pixels[2]);
+  EXPECT_EQ(235, dest_pixels[3]);
+
+  ARGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+            kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+  EXPECT_EQ(226, dest_pixels[2]);
+  EXPECT_EQ(235, dest_pixels[3]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, ARGBTest4x) {
+  const int kSrcStride = 640 * 4;
+  const int kDstStride = 160 * 4;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 4 + 0] = i;
+    orig_pixels[i * 4 + 1] = 255 - i;
+    orig_pixels[i * 4 + 2] = i + 1;
+    orig_pixels[i * 4 + 3] = i + 10;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ARGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+              kFilterBilinear);
+  }
+
+  EXPECT_NEAR(66, dest_pixels[0], 4);
+  EXPECT_NEAR(255 - 66, dest_pixels[1], 4);
+  EXPECT_NEAR(67, dest_pixels[2], 4);
+  EXPECT_NEAR(76, dest_pixels[3], 4);
+
+  ARGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+            kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+  EXPECT_EQ(3, dest_pixels[2]);
+  EXPECT_EQ(12, dest_pixels[3]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
 }  // namespace libyuv
diff --git a/files/unit_test/scale_rgb_test.cc b/files/unit_test/scale_rgb_test.cc
new file mode 100644
index 00000000..8296abe3
--- /dev/null
+++ b/files/unit_test/scale_rgb_test.cc
@@ -0,0 +1,280 @@
+/*
+ *  Copyright 2022 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <time.h>
+
+#include "../unit_test/unit_test.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/scale_rgb.h"
+
+namespace libyuv {
+
+#define STRINGIZE(line) #line
+#define FILELINESTR(file, line) file ":" STRINGIZE(line)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int RGBTestFilter(int src_width,
+                         int src_height,
+                         int dst_width,
+                         int dst_height,
+                         FilterMode f,
+                         int benchmark_iterations,
+                         int disable_cpu_flags,
+                         int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  const int b = 0;  // 128 to test for padding/stride.
+  int64_t src_rgb_plane_size =
+      (Abs(src_width) + b * 3) * (Abs(src_height) + b * 3) * 3LL;
+  int src_stride_rgb = (b * 3 + Abs(src_width)) * 3;
+
+  align_buffer_page_end(src_rgb, src_rgb_plane_size);
+  if (!src_rgb) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_rgb, src_rgb_plane_size);
+
+  int64_t dst_rgb_plane_size = (dst_width + b * 3) * (dst_height + b * 3) * 3LL;
+  int dst_stride_rgb = (b * 3 + dst_width) * 3;
+
+  align_buffer_page_end(dst_rgb_c, dst_rgb_plane_size);
+  align_buffer_page_end(dst_rgb_opt, dst_rgb_plane_size);
+  if (!dst_rgb_c || !dst_rgb_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  memset(dst_rgb_c, 2, dst_rgb_plane_size);
+  memset(dst_rgb_opt, 3, dst_rgb_plane_size);
+
+  // Warm up both versions for consistent benchmarks.
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
+           dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
+           dst_stride_rgb, dst_width, dst_height, f);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+           src_height, dst_rgb_c + (dst_stride_rgb * b) + b * 3, dst_stride_rgb,
+           dst_width, dst_height, f);
+
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    RGBScale(src_rgb + (src_stride_rgb * b) + b * 3, src_stride_rgb, src_width,
+             src_height, dst_rgb_opt + (dst_stride_rgb * b) + b * 3,
+             dst_stride_rgb, dst_width, dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+
+  // Report performance of C vs OPT
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference isn't
+  //  over 2.
+  int max_diff = 0;
+  for (i = b; i < (dst_height + b); ++i) {
+    for (j = b * 3; j < (dst_width + b) * 3; ++j) {
+      int abs_diff = Abs(dst_rgb_c[(i * dst_stride_rgb) + j] -
+                         dst_rgb_opt[(i * dst_stride_rgb) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_rgb_c);
+  free_aligned_buffer_page_end(dst_rgb_opt);
+  free_aligned_buffer_page_end(src_rgb);
+  return max_diff;
+}
+
+// The following adjustments in dimensions ensure the scale factor will be
+// exactly achieved.
+#define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
+#define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
+
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
+  TEST_F(LibYUVScaleTest, RGBScaleDownBy##name##_##filter) {                 \
+    int diff = RGBTestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
+        benchmark_cpu_info_);                                                \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
+// filtering is different fixed point implementations for SSSE3, Neon and C.
+#define TEST_FACTOR(name, nom, denom)         \
+  TEST_FACTOR1(name, None, nom, denom, 0)     \
+  TEST_FACTOR1(name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(name, Box, nom, denom, 3)
+#else
+// Test a scale factor with Bilinear.
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3)
+#endif
+
+TEST_FACTOR(2, 1, 2)
+#ifndef DISABLE_SLOW_TESTS
+TEST_FACTOR(4, 1, 4)
+// TEST_FACTOR(8, 1, 8)  Disable for benchmark performance.
+TEST_FACTOR(3by4, 3, 4)
+TEST_FACTOR(3by8, 3, 8)
+TEST_FACTOR(3, 1, 3)
+#endif
+#undef TEST_FACTOR1
+#undef TEST_FACTOR
+#undef SX
+#undef DX
+
+#define TEST_SCALETO1(name, width, height, filter, max_diff)                 \
+  TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {           \
+    int diff = RGBTestFilter(benchmark_width_, benchmark_height_, width,     \
+                             height, kFilter##filter, benchmark_iterations_, \
+                             disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                               \
+  }                                                                          \
+  TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {         \
+    int diff = RGBTestFilter(width, height, Abs(benchmark_width_),           \
+                             Abs(benchmark_height_), kFilter##filter,        \
+                             benchmark_iterations_, disable_cpu_flags_,      \
+                             benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+/// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)       \
+  TEST_SCALETO1(name, width, height, None, 0)   \
+  TEST_SCALETO1(name, width, height, Linear, 3) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#endif
+
+TEST_SCALETO(RGBScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(RGBScale, 1, 1)
+TEST_SCALETO(RGBScale, 256, 144) /* 128x72 * 3 */
+TEST_SCALETO(RGBScale, 320, 240)
+TEST_SCALETO(RGBScale, 569, 480)
+TEST_SCALETO(RGBScale, 1280, 720)
+TEST_SCALETO(RGBScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
+#undef TEST_SCALETO1
+#undef TEST_SCALETO
+
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                      \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                     \
+    int diff = RGBTestFilter(benchmark_width_, benchmark_height_,      \
+                             benchmark_height_, benchmark_width_,      \
+                             kFilter##filter, benchmark_iterations_,   \
+                             disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                         \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(RGBScale, None, 0)
+TEST_SCALESWAPXY1(RGBScale, Linear, 0)
+TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(RGBScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
+TEST_F(LibYUVScaleTest, RGBTest3x) {
+  const int kSrcStride = 480 * 3;
+  const int kDstStride = 160 * 3;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 3 + 0] = i;
+    orig_pixels[i * 3 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+             kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  RGBScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+           kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, RGBTest4x) {
+  const int kSrcStride = 640 * 3;
+  const int kDstStride = 160 * 3;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 3 + 0] = i;
+    orig_pixels[i * 3 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    RGBScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+             kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+  EXPECT_EQ(190, dest_pixels[1]);
+
+  RGBScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
+           kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+}  // namespace libyuv
diff --git a/files/unit_test/scale_test.cc b/files/unit_test/scale_test.cc
index 811b2d04..a8c95268 100644
--- a/files/unit_test/scale_test.cc
+++ b/files/unit_test/scale_test.cc
@@ -14,11 +14,20 @@
 #include "../unit_test/unit_test.h"
 #include "libyuv/cpu_id.h"
 #include "libyuv/scale.h"
+
+#ifdef ENABLE_ROW_TESTS
 #include "libyuv/scale_row.h"  // For ScaleRowDown2Box_Odd_C
+#endif
 
 #define STRINGIZE(line) #line
 #define FILELINESTR(file, line) file ":" STRINGIZE(line)
 
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
 namespace libyuv {
 
 // Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
@@ -139,6 +148,123 @@ static int I420TestFilter(int src_width,
   return max_diff;
 }
 
+// Test scaling with 8 bit C vs 12 bit C and return maximum pixel difference.
+// 0 = exact.
+static int I420TestFilter_12(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int src_width_uv = (Abs(src_width) + 1) >> 1;
+  int src_height_uv = (Abs(src_height) + 1) >> 1;
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_12, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_12, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_12, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_12 || !src_u_12 || !src_v_12) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_12 = reinterpret_cast<uint16_t*>(src_y_12);
+  uint16_t* p_src_u_12 = reinterpret_cast<uint16_t*>(src_u_12);
+  uint16_t* p_src_v_12 = reinterpret_cast<uint16_t*>(src_v_12);
+
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_12[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_12[i] = src_u[i];
+    p_src_v_12[i] = src_v[i];
+  }
+
+  int dst_width_uv = (dst_width + 1) >> 1;
+  int dst_height_uv = (dst_height + 1) >> 1;
+
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_12, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_12, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_12, dst_uv_plane_size * 2);
+
+  uint16_t* p_dst_y_12 = reinterpret_cast<uint16_t*>(dst_y_12);
+  uint16_t* p_dst_u_12 = reinterpret_cast<uint16_t*>(dst_u_12);
+  uint16_t* p_dst_v_12 = reinterpret_cast<uint16_t*>(dst_v_12);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I420Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I420Scale_12(p_src_y_12, src_stride_y, p_src_u_12, src_stride_uv,
+                 p_src_v_12, src_stride_uv, src_width, src_height, p_dst_y_12,
+                 dst_stride_y, p_dst_u_12, dst_stride_uv, p_dst_v_12,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_12);
+  free_aligned_buffer_page_end(dst_u_12);
+  free_aligned_buffer_page_end(dst_v_12);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_12);
+  free_aligned_buffer_page_end(src_u_12);
+  free_aligned_buffer_page_end(src_v_12);
+
+  return max_diff;
+}
+
 // Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
 // 0 = exact.
 static int I420TestFilter_16(int src_width,
@@ -374,6 +500,123 @@ static int I444TestFilter(int src_width,
   return max_diff;
 }
 
+// Test scaling with 8 bit C vs 12 bit C and return maximum pixel difference.
+// 0 = exact.
+static int I444TestFilter_12(int src_width,
+                             int src_height,
+                             int dst_width,
+                             int dst_height,
+                             FilterMode f,
+                             int benchmark_iterations,
+                             int disable_cpu_flags,
+                             int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i;
+  int src_width_uv = Abs(src_width);
+  int src_height_uv = Abs(src_height);
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv);
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_u, src_uv_plane_size);
+  align_buffer_page_end(src_v, src_uv_plane_size);
+  align_buffer_page_end(src_y_12, src_y_plane_size * 2);
+  align_buffer_page_end(src_u_12, src_uv_plane_size * 2);
+  align_buffer_page_end(src_v_12, src_uv_plane_size * 2);
+  if (!src_y || !src_u || !src_v || !src_y_12 || !src_u_12 || !src_v_12) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  uint16_t* p_src_y_12 = reinterpret_cast<uint16_t*>(src_y_12);
+  uint16_t* p_src_u_12 = reinterpret_cast<uint16_t*>(src_u_12);
+  uint16_t* p_src_v_12 = reinterpret_cast<uint16_t*>(src_v_12);
+
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_u, src_uv_plane_size);
+  MemRandomize(src_v, src_uv_plane_size);
+
+  for (i = 0; i < src_y_plane_size; ++i) {
+    p_src_y_12[i] = src_y[i];
+  }
+  for (i = 0; i < src_uv_plane_size; ++i) {
+    p_src_u_12[i] = src_u[i];
+    p_src_v_12[i] = src_v[i];
+  }
+
+  int dst_width_uv = dst_width;
+  int dst_height_uv = dst_height;
+
+  int dst_y_plane_size = (dst_width) * (dst_height);
+  int dst_uv_plane_size = (dst_width_uv) * (dst_height_uv);
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv;
+
+  align_buffer_page_end(dst_y_8, dst_y_plane_size);
+  align_buffer_page_end(dst_u_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_v_8, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_12, dst_y_plane_size * 2);
+  align_buffer_page_end(dst_u_12, dst_uv_plane_size * 2);
+  align_buffer_page_end(dst_v_12, dst_uv_plane_size * 2);
+
+  uint16_t* p_dst_y_12 = reinterpret_cast<uint16_t*>(dst_y_12);
+  uint16_t* p_dst_u_12 = reinterpret_cast<uint16_t*>(dst_u_12);
+  uint16_t* p_dst_v_12 = reinterpret_cast<uint16_t*>(dst_v_12);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I444Scale(src_y, src_stride_y, src_u, src_stride_uv, src_v, src_stride_uv,
+            src_width, src_height, dst_y_8, dst_stride_y, dst_u_8,
+            dst_stride_uv, dst_v_8, dst_stride_uv, dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (i = 0; i < benchmark_iterations; ++i) {
+    I444Scale_12(p_src_y_12, src_stride_y, p_src_u_12, src_stride_uv,
+                 p_src_v_12, src_stride_uv, src_width, src_height, p_dst_y_12,
+                 dst_stride_y, p_dst_u_12, dst_stride_uv, p_dst_v_12,
+                 dst_stride_uv, dst_width, dst_height, f);
+  }
+
+  // Expect an exact match.
+  int max_diff = 0;
+  for (i = 0; i < dst_y_plane_size; ++i) {
+    int abs_diff = Abs(dst_y_8[i] - p_dst_y_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  for (i = 0; i < dst_uv_plane_size; ++i) {
+    int abs_diff = Abs(dst_u_8[i] - p_dst_u_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+    abs_diff = Abs(dst_v_8[i] - p_dst_v_12[i]);
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_8);
+  free_aligned_buffer_page_end(dst_u_8);
+  free_aligned_buffer_page_end(dst_v_8);
+  free_aligned_buffer_page_end(dst_y_12);
+  free_aligned_buffer_page_end(dst_u_12);
+  free_aligned_buffer_page_end(dst_v_12);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_u);
+  free_aligned_buffer_page_end(src_v);
+  free_aligned_buffer_page_end(src_y_12);
+  free_aligned_buffer_page_end(src_u_12);
+  free_aligned_buffer_page_end(src_v_12);
+
+  return max_diff;
+}
+
 // Test scaling with 8 bit C vs 16 bit C and return maximum pixel difference.
 // 0 = exact.
 static int I444TestFilter_16(int src_width,
@@ -491,57 +734,185 @@ static int I444TestFilter_16(int src_width,
   return max_diff;
 }
 
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int NV12TestFilter(int src_width,
+                          int src_height,
+                          int dst_width,
+                          int dst_height,
+                          FilterMode f,
+                          int benchmark_iterations,
+                          int disable_cpu_flags,
+                          int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  int src_width_uv = (Abs(src_width) + 1) >> 1;
+  int src_height_uv = (Abs(src_height) + 1) >> 1;
+
+  int64_t src_y_plane_size = (Abs(src_width)) * (Abs(src_height));
+  int64_t src_uv_plane_size = (src_width_uv) * (src_height_uv)*2;
+
+  int src_stride_y = Abs(src_width);
+  int src_stride_uv = src_width_uv * 2;
+
+  align_buffer_page_end(src_y, src_y_plane_size);
+  align_buffer_page_end(src_uv, src_uv_plane_size);
+  if (!src_y || !src_uv) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_y, src_y_plane_size);
+  MemRandomize(src_uv, src_uv_plane_size);
+
+  int dst_width_uv = (dst_width + 1) >> 1;
+  int dst_height_uv = (dst_height + 1) >> 1;
+
+  int64_t dst_y_plane_size = (dst_width) * (dst_height);
+  int64_t dst_uv_plane_size = (dst_width_uv) * (dst_height_uv)*2;
+
+  int dst_stride_y = dst_width;
+  int dst_stride_uv = dst_width_uv * 2;
+
+  align_buffer_page_end(dst_y_c, dst_y_plane_size);
+  align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_y_opt, dst_y_plane_size);
+  align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
+  if (!dst_y_c || !dst_uv_c || !dst_y_opt || !dst_uv_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  NV12Scale(src_y, src_stride_y, src_uv, src_stride_uv, src_width, src_height,
+            dst_y_c, dst_stride_y, dst_uv_c, dst_stride_uv, dst_width,
+            dst_height, f);
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    NV12Scale(src_y, src_stride_y, src_uv, src_stride_uv, src_width, src_height,
+              dst_y_opt, dst_stride_y, dst_uv_opt, dst_stride_uv, dst_width,
+              dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+  // Report performance of C vs OPT.
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference is not
+  //  over 3.
+  int max_diff = 0;
+  for (i = 0; i < (dst_height); ++i) {
+    for (j = 0; j < (dst_width); ++j) {
+      int abs_diff = Abs(dst_y_c[(i * dst_stride_y) + j] -
+                         dst_y_opt[(i * dst_stride_y) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  for (i = 0; i < (dst_height_uv); ++i) {
+    for (j = 0; j < (dst_width_uv * 2); ++j) {
+      int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
+                         dst_uv_opt[(i * dst_stride_uv) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_y_c);
+  free_aligned_buffer_page_end(dst_uv_c);
+  free_aligned_buffer_page_end(dst_y_opt);
+  free_aligned_buffer_page_end(dst_uv_opt);
+  free_aligned_buffer_page_end(src_y);
+  free_aligned_buffer_page_end(src_uv);
+
+  return max_diff;
+}
+
 // The following adjustments in dimensions ensure the scale factor will be
 // exactly achieved.
 // 2 is chroma subsample.
 #define DX(x, nom, denom) static_cast<int>(((Abs(x) / nom + 1) / 2) * nom * 2)
 #define SX(x, nom, denom) static_cast<int>(((x / nom + 1) / 2) * denom * 2)
 
-#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
-  TEST_F(LibYUVScaleTest, I420ScaleDownBy##name##_##filter) {                \
-    int diff = I420TestFilter(                                               \
-        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
-        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
-        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
-        benchmark_cpu_info_);                                                \
-    EXPECT_LE(diff, max_diff);                                               \
-  }                                                                          \
-  TEST_F(LibYUVScaleTest, I444ScaleDownBy##name##_##filter) {                \
-    int diff = I444TestFilter(                                               \
-        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
-        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
-        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
-        benchmark_cpu_info_);                                                \
-    EXPECT_LE(diff, max_diff);                                               \
-  }                                                                          \
-  TEST_F(LibYUVScaleTest, I420ScaleDownBy##name##_##filter##_16) {           \
-    int diff = I420TestFilter_16(                                            \
-        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
-        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
-        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
-        benchmark_cpu_info_);                                                \
-    EXPECT_LE(diff, max_diff);                                               \
-  }                                                                          \
-  TEST_F(LibYUVScaleTest, I444ScaleDownBy##name##_##filter##_16) {           \
-    int diff = I444TestFilter_16(                                            \
-        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
-        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
-        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
-        benchmark_cpu_info_);                                                \
-    EXPECT_LE(diff, max_diff);                                               \
+#define TEST_FACTOR1(DISABLED_, name, filter, nom, denom, max_diff)           \
+  TEST_F(LibYUVScaleTest, I420ScaleDownBy##name##_##filter) {                 \
+    int diff = I420TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, I444ScaleDownBy##name##_##filter) {                 \
+    int diff = I444TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420ScaleDownBy##name##_##filter##_12) { \
+    int diff = I420TestFilter_12(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444ScaleDownBy##name##_##filter##_12) { \
+    int diff = I444TestFilter_12(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12ScaleDownBy##name##_##filter) {                 \
+    int diff = NV12TestFilter(                                                \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),  \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),  \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,           \
+        benchmark_cpu_info_);                                                 \
+    EXPECT_LE(diff, max_diff);                                                \
   }
 
 // Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
 // filtering is different fixed point implementations for SSSE3, Neon and C.
-#define TEST_FACTOR(name, nom, denom, boxdiff) \
-  TEST_FACTOR1(name, None, nom, denom, 0)      \
-  TEST_FACTOR1(name, Linear, nom, denom, 3)    \
-  TEST_FACTOR1(name, Bilinear, nom, denom, 3)  \
-  TEST_FACTOR1(name, Box, nom, denom, boxdiff)
+#ifndef DISABLE_SLOW_TESTS
+#define TEST_FACTOR(name, nom, denom, boxdiff)  \
+  TEST_FACTOR1(, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(, name, Box, nom, denom, boxdiff)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_FACTOR(name, nom, denom, boxdiff)           \
+  TEST_FACTOR1(DISABLED_, name, None, nom, denom, 0)     \
+  TEST_FACTOR1(DISABLED_, name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, boxdiff)
+#else
+#define TEST_FACTOR(name, nom, denom, boxdiff)           \
+  TEST_FACTOR1(DISABLED_, name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(DISABLED_, name, Box, nom, denom, boxdiff)
+#endif
+#endif
 
 TEST_FACTOR(2, 1, 2, 0)
 TEST_FACTOR(4, 1, 4, 0)
+#ifndef DISABLE_SLOW_TESTS
 TEST_FACTOR(8, 1, 8, 0)
+#endif
 TEST_FACTOR(3by4, 3, 4, 1)
 TEST_FACTOR(3by8, 3, 8, 1)
 TEST_FACTOR(3, 1, 3, 0)
@@ -550,7 +921,7 @@ TEST_FACTOR(3, 1, 3, 0)
 #undef SX
 #undef DX
 
-#define TEST_SCALETO1(name, width, height, filter, max_diff)                  \
+#define TEST_SCALETO1(DISABLED_, name, width, height, filter, max_diff)       \
   TEST_F(LibYUVScaleTest, I420##name##To##width##x##height##_##filter) {      \
     int diff = I420TestFilter(benchmark_width_, benchmark_height_, width,     \
                               height, kFilter##filter, benchmark_iterations_, \
@@ -563,18 +934,40 @@ TEST_FACTOR(3, 1, 3, 0)
                               disable_cpu_flags_, benchmark_cpu_info_);       \
     EXPECT_LE(diff, max_diff);                                                \
   }                                                                           \
-  TEST_F(LibYUVScaleTest, I420##name##To##width##x##height##_##filter##_16) { \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##To##width##x##height##_##filter##_12) {       \
+    int diff = I420TestFilter_12(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##To##width##x##height##_##filter##_12) {       \
+    int diff = I444TestFilter_12(                                             \
+        benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
+        benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##To##width##x##height##_##filter##_16) {       \
     int diff = I420TestFilter_16(                                             \
         benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
         benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
     EXPECT_LE(diff, max_diff);                                                \
   }                                                                           \
-  TEST_F(LibYUVScaleTest, I444##name##To##width##x##height##_##filter##_16) { \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##To##width##x##height##_##filter##_16) {       \
     int diff = I444TestFilter_16(                                             \
         benchmark_width_, benchmark_height_, width, height, kFilter##filter,  \
         benchmark_iterations_, disable_cpu_flags_, benchmark_cpu_info_);      \
     EXPECT_LE(diff, max_diff);                                                \
   }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12##name##To##width##x##height##_##filter) {      \
+    int diff = NV12TestFilter(benchmark_width_, benchmark_height_, width,     \
+                              height, kFilter##filter, benchmark_iterations_, \
+                              disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
   TEST_F(LibYUVScaleTest, I420##name##From##width##x##height##_##filter) {    \
     int diff = I420TestFilter(width, height, Abs(benchmark_width_),           \
                               Abs(benchmark_height_), kFilter##filter,        \
@@ -590,7 +983,23 @@ TEST_FACTOR(3, 1, 3, 0)
     EXPECT_LE(diff, max_diff);                                                \
   }                                                                           \
   TEST_F(LibYUVScaleTest,                                                     \
-         I420##name##From##width##x##height##_##filter##_16) {                \
+         DISABLED_##I420##name##From##width##x##height##_##filter##_12) {     \
+    int diff = I420TestFilter_12(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I444##name##From##width##x##height##_##filter##_12) {     \
+    int diff = I444TestFilter_12(width, height, Abs(benchmark_width_),        \
+                                 Abs(benchmark_height_), kFilter##filter,     \
+                                 benchmark_iterations_, disable_cpu_flags_,   \
+                                 benchmark_cpu_info_);                        \
+    EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest,                                                     \
+         DISABLED_##I420##name##From##width##x##height##_##filter##_16) {     \
     int diff = I420TestFilter_16(width, height, Abs(benchmark_width_),        \
                                  Abs(benchmark_height_), kFilter##filter,     \
                                  benchmark_iterations_, disable_cpu_flags_,   \
@@ -598,30 +1007,126 @@ TEST_FACTOR(3, 1, 3, 0)
     EXPECT_LE(diff, max_diff);                                                \
   }                                                                           \
   TEST_F(LibYUVScaleTest,                                                     \
-         I444##name##From##width##x##height##_##filter##_16) {                \
+         DISABLED_##I444##name##From##width##x##height##_##filter##_16) {     \
     int diff = I444TestFilter_16(width, height, Abs(benchmark_width_),        \
                                  Abs(benchmark_height_), kFilter##filter,     \
                                  benchmark_iterations_, disable_cpu_flags_,   \
                                  benchmark_cpu_info_);                        \
     EXPECT_LE(diff, max_diff);                                                \
+  }                                                                           \
+  TEST_F(LibYUVScaleTest, NV12##name##From##width##x##height##_##filter) {    \
+    int diff = NV12TestFilter(width, height, Abs(benchmark_width_),           \
+                              Abs(benchmark_height_), kFilter##filter,        \
+                              benchmark_iterations_, disable_cpu_flags_,      \
+                              benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                                \
   }
 
+#ifndef DISABLE_SLOW_TESTS
 // Test scale to a specified size with all 4 filters.
-#define TEST_SCALETO(name, width, height)         \
-  TEST_SCALETO1(name, width, height, None, 0)     \
-  TEST_SCALETO1(name, width, height, Linear, 3)   \
-  TEST_SCALETO1(name, width, height, Bilinear, 3) \
-  TEST_SCALETO1(name, width, height, Box, 3)
+#define TEST_SCALETO(name, width, height)           \
+  TEST_SCALETO1(, name, width, height, None, 0)     \
+  TEST_SCALETO1(, name, width, height, Linear, 3)   \
+  TEST_SCALETO1(, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(, name, width, height, Box, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+#define TEST_SCALETO(name, width, height)                    \
+  TEST_SCALETO1(DISABLED_, name, width, height, None, 0)     \
+  TEST_SCALETO1(DISABLED_, name, width, height, Linear, 3)   \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Box, 3)
+#else
+#define TEST_SCALETO(name, width, height)                    \
+  TEST_SCALETO1(DISABLED_, name, width, height, Bilinear, 3) \
+  TEST_SCALETO1(DISABLED_, name, width, height, Box, 3)
+#endif
+#endif
 
 TEST_SCALETO(Scale, 1, 1)
-TEST_SCALETO(Scale, 320, 240)
 TEST_SCALETO(Scale, 569, 480)
 TEST_SCALETO(Scale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(Scale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(Scale, 320, 240)
 TEST_SCALETO(Scale, 1280, 720)
 TEST_SCALETO(Scale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
 #undef TEST_SCALETO1
 #undef TEST_SCALETO
 
+#define TEST_SCALESWAPXY1(DISABLED_, name, filter, max_diff)               \
+  TEST_F(LibYUVScaleTest, I420##name##SwapXY_##filter) {                   \
+    int diff = I420TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, I444##name##SwapXY_##filter) {                   \
+    int diff = I444TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420##name##SwapXY_##filter##_12) {   \
+    int diff = I420TestFilter_12(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444##name##SwapXY_##filter##_12) {   \
+    int diff = I444TestFilter_12(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I420##name##SwapXY_##filter##_16) {   \
+    int diff = I420TestFilter_16(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, DISABLED_##I444##name##SwapXY_##filter##_16) {   \
+    int diff = I444TestFilter_16(benchmark_width_, benchmark_height_,      \
+                                 benchmark_height_, benchmark_width_,      \
+                                 kFilter##filter, benchmark_iterations_,   \
+                                 disable_cpu_flags_, benchmark_cpu_info_); \
+    EXPECT_LE(diff, max_diff);                                             \
+  }                                                                        \
+  TEST_F(LibYUVScaleTest, NV12##name##SwapXY_##filter) {                   \
+    int diff = NV12TestFilter(benchmark_width_, benchmark_height_,         \
+                              benchmark_height_, benchmark_width_,         \
+                              kFilter##filter, benchmark_iterations_,      \
+                              disable_cpu_flags_, benchmark_cpu_info_);    \
+    EXPECT_LE(diff, max_diff);                                             \
+  }
+
+// Test scale to a specified size with all 4 filters.
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALESWAPXY1(, Scale, None, 0)
+TEST_SCALESWAPXY1(, Scale, Linear, 3)
+TEST_SCALESWAPXY1(, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(, Scale, Box, 3)
+#else
+#if defined(ENABLE_FULL_TESTS)
+TEST_SCALESWAPXY1(DISABLED_, Scale, None, 0)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Linear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Box, 3)
+#else
+TEST_SCALESWAPXY1(DISABLED_, Scale, Bilinear, 3)
+TEST_SCALESWAPXY1(DISABLED_, Scale, Box, 3)
+#endif
+#endif
+
+#undef TEST_SCALESWAPXY1
+
+#ifdef ENABLE_ROW_TESTS
 #ifdef HAS_SCALEROWDOWN2_SSSE3
 TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_Odd_SSSE3) {
   SIMD_ALIGNED(uint8_t orig_pixels[128 * 2]);
@@ -716,10 +1221,6 @@ extern "C" void ScaleRowUp2_16_NEON(const uint16_t* src_ptr,
                                     ptrdiff_t src_stride,
                                     uint16_t* dst,
                                     int dst_width);
-extern "C" void ScaleRowUp2_16_MMI(const uint16_t* src_ptr,
-                                   ptrdiff_t src_stride,
-                                   uint16_t* dst,
-                                   int dst_width);
 extern "C" void ScaleRowUp2_16_C(const uint16_t* src_ptr,
                                  ptrdiff_t src_stride,
                                  uint16_t* dst,
@@ -746,13 +1247,6 @@ TEST_F(LibYUVScaleTest, TestScaleRowUp2_16) {
     } else {
       ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
     }
-#elif !defined(LIBYUV_DISABLE_MMI) && defined(_MIPS_ARCH_LOONGSON3A)
-    int has_mmi = TestCpuFlag(kCpuHasMMI);
-    if (has_mmi) {
-      ScaleRowUp2_16_MMI(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
-    } else {
-      ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
-    }
 #else
     ScaleRowUp2_16_C(&orig_pixels[0], 640, &dst_pixels_opt[0], 1280);
 #endif
@@ -803,8 +1297,9 @@ TEST_F(LibYUVScaleTest, TestScaleRowDown2Box_16) {
   EXPECT_EQ(dst_pixels_c[0], (0 + 1 + 2560 + 2561 + 2) / 4);
   EXPECT_EQ(dst_pixels_c[1279], 3839);
 }
+#endif  // ENABLE_ROW_TESTS
 
-// Test scaling plane with 8 bit C vs 16 bit C and return maximum pixel
+// Test scaling plane with 8 bit C vs 12 bit C and return maximum pixel
 // difference.
 // 0 = exact.
 static int TestPlaneFilter_16(int src_width,
@@ -873,14 +1368,14 @@ static int TestPlaneFilter_16(int src_width,
 #define DX(x, nom, denom) static_cast<int>(((Abs(x) / nom + 1) / 2) * nom * 2)
 #define SX(x, nom, denom) static_cast<int>(((x / nom + 1) / 2) * denom * 2)
 
-#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
-  TEST_F(LibYUVScaleTest, ScalePlaneDownBy##name##_##filter##_16) {          \
-    int diff = TestPlaneFilter_16(                                           \
-        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
-        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
-        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
-        benchmark_cpu_info_);                                                \
-    EXPECT_LE(diff, max_diff);                                               \
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                       \
+  TEST_F(LibYUVScaleTest, DISABLED_##ScalePlaneDownBy##name##_##filter##_16) { \
+    int diff = TestPlaneFilter_16(                                             \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom),   \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom),   \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,            \
+        benchmark_cpu_info_);                                                  \
+    EXPECT_LE(diff, max_diff);                                                 \
   }
 
 // Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
@@ -893,7 +1388,7 @@ static int TestPlaneFilter_16(int src_width,
 
 TEST_FACTOR(2, 1, 2, 0)
 TEST_FACTOR(4, 1, 4, 0)
-TEST_FACTOR(8, 1, 8, 0)
+// TEST_FACTOR(8, 1, 8, 0) Disable for benchmark performance.  Takes 90 seconds.
 TEST_FACTOR(3by4, 3, 4, 1)
 TEST_FACTOR(3by8, 3, 8, 1)
 TEST_FACTOR(3, 1, 3, 0)
@@ -901,4 +1396,206 @@ TEST_FACTOR(3, 1, 3, 0)
 #undef TEST_FACTOR
 #undef SX
 #undef DX
+
+TEST_F(LibYUVScaleTest, PlaneTest3x) {
+  const int kSrcStride = 480;
+  const int kDstStride = 160;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ScalePlane(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+               kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+
+  ScalePlane(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+             kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest4x) {
+  const int kSrcStride = 640;
+  const int kDstStride = 160;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    ScalePlane(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+               kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+
+  ScalePlane(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+             kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+// Intent is to test 200x50 to 50x200 but width and height can be parameters.
+TEST_F(LibYUVScaleTest, PlaneTestRotate_None) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterNone);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterNone);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTestRotate_Bilinear) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterBilinear);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterBilinear);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+// Intent is to test 200x50 to 50x200 but width and height can be parameters.
+TEST_F(LibYUVScaleTest, PlaneTestRotate_Box) {
+  const int kSize = benchmark_width_ * benchmark_height_;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < kSize; ++i) {
+    orig_pixels[i] = i;
+  }
+  align_buffer_page_end(dest_opt_pixels, kSize);
+  align_buffer_page_end(dest_c_pixels, kSize);
+
+  MaskCpuFlags(disable_cpu_flags_);  // Disable all CPU optimization.
+  ScalePlane(orig_pixels, benchmark_width_, benchmark_width_, benchmark_height_,
+             dest_c_pixels, benchmark_height_, benchmark_height_,
+             benchmark_width_, kFilterBox);
+  MaskCpuFlags(benchmark_cpu_info_);  // Enable all CPU optimization.
+
+  for (int i = 0; i < benchmark_iterations_; ++i) {
+    ScalePlane(orig_pixels, benchmark_width_, benchmark_width_,
+               benchmark_height_, dest_opt_pixels, benchmark_height_,
+               benchmark_height_, benchmark_width_, kFilterBox);
+  }
+
+  for (int i = 0; i < kSize; ++i) {
+    EXPECT_EQ(dest_c_pixels[i], dest_opt_pixels[i]);
+  }
+
+  free_aligned_buffer_page_end(dest_c_pixels);
+  free_aligned_buffer_page_end(dest_opt_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest1_Box) {
+  align_buffer_page_end(orig_pixels, 3);
+  align_buffer_page_end(dst_pixels, 3);
+
+  // Pad the 1x1 byte image with invalid values before and after in case libyuv
+  // reads outside the memory boundaries.
+  orig_pixels[0] = 0;
+  orig_pixels[1] = 1;  // scale this pixel
+  orig_pixels[2] = 2;
+  dst_pixels[0] = 3;
+  dst_pixels[1] = 3;
+  dst_pixels[2] = 3;
+
+  libyuv::ScalePlane(orig_pixels + 1, /* src_stride= */ 1, /* src_width= */ 1,
+                     /* src_height= */ 1, dst_pixels, /* dst_stride= */ 1,
+                     /* dst_width= */ 1, /* dst_height= */ 2,
+                     libyuv::kFilterBox);
+
+  EXPECT_EQ(dst_pixels[0], 1);
+  EXPECT_EQ(dst_pixels[1], 1);
+  EXPECT_EQ(dst_pixels[2], 3);
+
+  free_aligned_buffer_page_end(dst_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, PlaneTest1_16_Box) {
+  align_buffer_page_end(orig_pixels_alloc, 3 * 2);
+  align_buffer_page_end(dst_pixels_alloc, 3 * 2);
+  uint16_t* orig_pixels = (uint16_t*)orig_pixels_alloc;
+  uint16_t* dst_pixels = (uint16_t*)dst_pixels_alloc;
+
+  // Pad the 1x1 byte image with invalid values before and after in case libyuv
+  // reads outside the memory boundaries.
+  orig_pixels[0] = 0;
+  orig_pixels[1] = 1;  // scale this pixel
+  orig_pixels[2] = 2;
+  dst_pixels[0] = 3;
+  dst_pixels[1] = 3;
+  dst_pixels[2] = 3;
+
+  libyuv::ScalePlane_16(
+      orig_pixels + 1, /* src_stride= */ 1, /* src_width= */ 1,
+      /* src_height= */ 1, dst_pixels, /* dst_stride= */ 1,
+      /* dst_width= */ 1, /* dst_height= */ 2, libyuv::kFilterNone);
+
+  EXPECT_EQ(dst_pixels[0], 1);
+  EXPECT_EQ(dst_pixels[1], 1);
+  EXPECT_EQ(dst_pixels[2], 3);
+
+  free_aligned_buffer_page_end(dst_pixels_alloc);
+  free_aligned_buffer_page_end(orig_pixels_alloc);
+}
 }  // namespace libyuv
diff --git a/files/unit_test/scale_uv_test.cc b/files/unit_test/scale_uv_test.cc
new file mode 100644
index 00000000..3d524bef
--- /dev/null
+++ b/files/unit_test/scale_uv_test.cc
@@ -0,0 +1,278 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+#include <time.h>
+
+#include "../unit_test/unit_test.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/scale_uv.h"
+
+namespace libyuv {
+
+#define STRINGIZE(line) #line
+#define FILELINESTR(file, line) file ":" STRINGIZE(line)
+
+#if !defined(DISABLE_SLOW_TESTS) || defined(__x86_64__) || defined(__i386__)
+// SLOW TESTS are those that are unoptimized C code.
+// FULL TESTS are optimized but test many variations of the same code.
+#define ENABLE_FULL_TESTS
+#endif
+
+// Test scaling with C vs Opt and return maximum pixel difference. 0 = exact.
+static int UVTestFilter(int src_width,
+                        int src_height,
+                        int dst_width,
+                        int dst_height,
+                        FilterMode f,
+                        int benchmark_iterations,
+                        int disable_cpu_flags,
+                        int benchmark_cpu_info) {
+  if (!SizeValid(src_width, src_height, dst_width, dst_height)) {
+    return 0;
+  }
+
+  int i, j;
+  const int b = 0;  // 128 to test for padding/stride.
+  int64_t src_uv_plane_size =
+      (Abs(src_width) + b * 2) * (Abs(src_height) + b * 2) * 2LL;
+  int src_stride_uv = (b * 2 + Abs(src_width)) * 2;
+
+  align_buffer_page_end(src_uv, src_uv_plane_size);
+  if (!src_uv) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  MemRandomize(src_uv, src_uv_plane_size);
+
+  int64_t dst_uv_plane_size = (dst_width + b * 2) * (dst_height + b * 2) * 2LL;
+  int dst_stride_uv = (b * 2 + dst_width) * 2;
+
+  align_buffer_page_end(dst_uv_c, dst_uv_plane_size);
+  align_buffer_page_end(dst_uv_opt, dst_uv_plane_size);
+  if (!dst_uv_c || !dst_uv_opt) {
+    printf("Skipped.  Alloc failed " FILELINESTR(__FILE__, __LINE__) "\n");
+    return 0;
+  }
+  memset(dst_uv_c, 2, dst_uv_plane_size);
+  memset(dst_uv_opt, 3, dst_uv_plane_size);
+
+  // Warm up both versions for consistent benchmarks.
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  double c_time = get_time();
+  UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+          src_height, dst_uv_c + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+          dst_width, dst_height, f);
+
+  c_time = (get_time() - c_time);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  double opt_time = get_time();
+  for (i = 0; i < benchmark_iterations; ++i) {
+    UVScale(src_uv + (src_stride_uv * b) + b * 2, src_stride_uv, src_width,
+            src_height, dst_uv_opt + (dst_stride_uv * b) + b * 2, dst_stride_uv,
+            dst_width, dst_height, f);
+  }
+  opt_time = (get_time() - opt_time) / benchmark_iterations;
+
+  // Report performance of C vs OPT
+  printf("filter %d - %8d us C - %8d us OPT\n", f,
+         static_cast<int>(c_time * 1e6), static_cast<int>(opt_time * 1e6));
+
+  // C version may be a little off from the optimized. Order of
+  //  operations may introduce rounding somewhere. So do a difference
+  //  of the buffers and look to see that the max difference isn't
+  //  over 2.
+  int max_diff = 0;
+  for (i = b; i < (dst_height + b); ++i) {
+    for (j = b * 2; j < (dst_width + b) * 2; ++j) {
+      int abs_diff = Abs(dst_uv_c[(i * dst_stride_uv) + j] -
+                         dst_uv_opt[(i * dst_stride_uv) + j]);
+      if (abs_diff > max_diff) {
+        max_diff = abs_diff;
+      }
+    }
+  }
+
+  free_aligned_buffer_page_end(dst_uv_c);
+  free_aligned_buffer_page_end(dst_uv_opt);
+  free_aligned_buffer_page_end(src_uv);
+  return max_diff;
+}
+
+// The following adjustments in dimensions ensure the scale factor will be
+// exactly achieved.
+#define DX(x, nom, denom) static_cast<int>((Abs(x) / nom) * nom)
+#define SX(x, nom, denom) static_cast<int>((x / nom) * denom)
+
+#define TEST_FACTOR1(name, filter, nom, denom, max_diff)                     \
+  TEST_F(LibYUVScaleTest, UVScaleDownBy##name##_##filter) {                  \
+    int diff = UVTestFilter(                                                 \
+        SX(benchmark_width_, nom, denom), SX(benchmark_height_, nom, denom), \
+        DX(benchmark_width_, nom, denom), DX(benchmark_height_, nom, denom), \
+        kFilter##filter, benchmark_iterations_, disable_cpu_flags_,          \
+        benchmark_cpu_info_);                                                \
+    EXPECT_LE(diff, max_diff);                                               \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test a scale factor with all 4 filters.  Expect unfiltered to be exact, but
+// filtering is different fixed point implementations for SSSE3, Neon and C.
+#define TEST_FACTOR(name, nom, denom)         \
+  TEST_FACTOR1(name, None, nom, denom, 0)     \
+  TEST_FACTOR1(name, Linear, nom, denom, 3)   \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3) \
+  TEST_FACTOR1(name, Box, nom, denom, 3)
+#else
+// Test a scale factor with Bilinear.
+#define TEST_FACTOR(name, nom, denom) \
+  TEST_FACTOR1(name, Bilinear, nom, denom, 3)
+#endif
+
+TEST_FACTOR(2, 1, 2)
+TEST_FACTOR(4, 1, 4)
+// TEST_FACTOR(8, 1, 8)  Disable for benchmark performance.
+TEST_FACTOR(3by4, 3, 4)
+TEST_FACTOR(3by8, 3, 8)
+TEST_FACTOR(3, 1, 3)
+#undef TEST_FACTOR1
+#undef TEST_FACTOR
+#undef SX
+#undef DX
+
+#define TEST_SCALETO1(name, width, height, filter, max_diff)                \
+  TEST_F(LibYUVScaleTest, name##To##width##x##height##_##filter) {          \
+    int diff = UVTestFilter(benchmark_width_, benchmark_height_, width,     \
+                            height, kFilter##filter, benchmark_iterations_, \
+                            disable_cpu_flags_, benchmark_cpu_info_);       \
+    EXPECT_LE(diff, max_diff);                                              \
+  }                                                                         \
+  TEST_F(LibYUVScaleTest, name##From##width##x##height##_##filter) {        \
+    int diff = UVTestFilter(width, height, Abs(benchmark_width_),           \
+                            Abs(benchmark_height_), kFilter##filter,        \
+                            benchmark_iterations_, disable_cpu_flags_,      \
+                            benchmark_cpu_info_);                           \
+    EXPECT_LE(diff, max_diff);                                              \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+/// Test scale to a specified size with all 4 filters.
+#define TEST_SCALETO(name, width, height)       \
+  TEST_SCALETO1(name, width, height, None, 0)   \
+  TEST_SCALETO1(name, width, height, Linear, 3) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#else
+#define TEST_SCALETO(name, width, height) \
+  TEST_SCALETO1(name, width, height, Bilinear, 3)
+#endif
+
+TEST_SCALETO(UVScale, 1, 1)
+TEST_SCALETO(UVScale, 569, 480)
+TEST_SCALETO(UVScale, 640, 360)
+#ifndef DISABLE_SLOW_TESTS
+TEST_SCALETO(UVScale, 256, 144) /* 128x72 * 2 */
+TEST_SCALETO(UVScale, 320, 240)
+TEST_SCALETO(UVScale, 1280, 720)
+TEST_SCALETO(UVScale, 1920, 1080)
+#endif  // DISABLE_SLOW_TESTS
+#undef TEST_SCALETO1
+#undef TEST_SCALETO
+
+#define TEST_SCALESWAPXY1(name, filter, max_diff)                              \
+  TEST_F(LibYUVScaleTest, name##SwapXY_##filter) {                             \
+    int diff =                                                                 \
+        UVTestFilter(benchmark_width_, benchmark_height_, benchmark_height_,   \
+                     benchmark_width_, kFilter##filter, benchmark_iterations_, \
+                     disable_cpu_flags_, benchmark_cpu_info_);                 \
+    EXPECT_LE(diff, max_diff);                                                 \
+  }
+
+#if defined(ENABLE_FULL_TESTS)
+// Test scale with swapped width and height with all 3 filters.
+TEST_SCALESWAPXY1(UVScale, None, 0)
+TEST_SCALESWAPXY1(UVScale, Linear, 0)
+TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
+#else
+TEST_SCALESWAPXY1(UVScale, Bilinear, 0)
+#endif
+#undef TEST_SCALESWAPXY1
+
+TEST_F(LibYUVScaleTest, UVTest3x) {
+  const int kSrcStride = 480 * 2;
+  const int kDstStride = 160 * 2;
+  const int kSize = kSrcStride * 3;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 480 * 3; ++i) {
+    orig_pixels[i * 2 + 0] = i;
+    orig_pixels[i * 2 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+            kFilterBilinear);
+  }
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  UVScale(orig_pixels, kSrcStride, 480, 3, dest_pixels, kDstStride, 160, 1,
+          kFilterNone);
+
+  EXPECT_EQ(225, dest_pixels[0]);
+  EXPECT_EQ(255 - 225, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+TEST_F(LibYUVScaleTest, UVTest4x) {
+  const int kSrcStride = 640 * 2;
+  const int kDstStride = 160 * 2;
+  const int kSize = kSrcStride * 4;
+  align_buffer_page_end(orig_pixels, kSize);
+  for (int i = 0; i < 640 * 4; ++i) {
+    orig_pixels[i * 2 + 0] = i;
+    orig_pixels[i * 2 + 1] = 255 - i;
+  }
+  align_buffer_page_end(dest_pixels, kDstStride);
+
+  int iterations160 = (benchmark_width_ * benchmark_height_ + (160 - 1)) / 160 *
+                      benchmark_iterations_;
+  for (int i = 0; i < iterations160; ++i) {
+    UVScale(orig_pixels, kSrcStride, 640, 4, dest_pixels, kDstStride, 160, 1,
+            kFilterBilinear);
+  }
+
+  EXPECT_EQ(66, dest_pixels[0]);
+  EXPECT_EQ(190, dest_pixels[1]);
+
+  UVScale(orig_pixels, kSrcStride, 64, 4, dest_pixels, kDstStride, 16, 1,
+          kFilterNone);
+
+  EXPECT_EQ(2, dest_pixels[0]);  // expect the 3rd pixel of the 3rd row
+  EXPECT_EQ(255 - 2, dest_pixels[1]);
+
+  free_aligned_buffer_page_end(dest_pixels);
+  free_aligned_buffer_page_end(orig_pixels);
+}
+
+}  // namespace libyuv
diff --git a/files/unit_test/testdata/mips.txt b/files/unit_test/testdata/mips.txt
new file mode 100644
index 00000000..d9f28cbf
--- /dev/null
+++ b/files/unit_test/testdata/mips.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz
+shadow register sets    : 1
diff --git a/files/unit_test/testdata/mips_loongson2k.txt b/files/unit_test/testdata/mips_loongson2k.txt
new file mode 100644
index 00000000..8a88d38f
--- /dev/null
+++ b/files/unit_test/testdata/mips_loongson2k.txt
@@ -0,0 +1,5 @@
+system type		: Loongson2K-SBC
+machine			: loongson,LS2k1000-EVP
+processor		: 0
+cpu model		: Loongson-2K V0.3  FPU V0.1
+BogoMIPS		: 1980.41
diff --git a/files/unit_test/testdata/mips_loongson3.txt b/files/unit_test/testdata/mips_loongson3.txt
new file mode 100644
index 00000000..1f540b12
--- /dev/null
+++ b/files/unit_test/testdata/mips_loongson3.txt
@@ -0,0 +1,10 @@
+system type		: generic-loongson-machine
+machine			: Unknown
+processor		: 0
+cpu model		: ICT Loongson-3 V0.9  FPU V0.1
+model name		: ICT Loongson-3A R3 (Loongson-3A3000) @ 1500MHz
+BogoMIPS		: 2990.15
+
+isa			: mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented	: dsp dsp2 vz
+shadow register sets	: 1
diff --git a/files/unit_test/testdata/mips_loongson_mmi.txt b/files/unit_test/testdata/mips_loongson_mmi.txt
new file mode 100644
index 00000000..0f10b8bb
--- /dev/null
+++ b/files/unit_test/testdata/mips_loongson_mmi.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz loongson-mmi loongson-ext
+shadow register sets    : 1
diff --git a/files/unit_test/testdata/mips_msa.txt b/files/unit_test/testdata/mips_msa.txt
new file mode 100644
index 00000000..ac930615
--- /dev/null
+++ b/files/unit_test/testdata/mips_msa.txt
@@ -0,0 +1,7 @@
+system type             : generic-loongson-machine
+machine                 : loongson,generic
+processor               : 0
+
+isa                     : mips1 mips2 mips3 mips4 mips5 mips32r1 mips32r2 mips64r1 mips64r2
+ASEs implemented        : vz msa
+shadow register sets    : 1
diff --git a/files/unit_test/unit_test.cc b/files/unit_test/unit_test.cc
index a1ae7ea3..61145a46 100644
--- a/files/unit_test/unit_test.cc
+++ b/files/unit_test/unit_test.cc
@@ -14,23 +14,28 @@
 
 #include <cstring>
 
-#ifdef LIBYUV_USE_GFLAGS
-#include "gflags/gflags.h"
+#ifdef LIBYUV_USE_ABSL_FLAGS
+#include "absl/flags/flag.h"
+#include "absl/flags/parse.h"
 #endif
 #include "libyuv/cpu_id.h"
 
 unsigned int fastrand_seed = 0xfb;
 
-#ifdef LIBYUV_USE_GFLAGS
-DEFINE_int32(libyuv_width, 0, "width of test image.");
-DEFINE_int32(libyuv_height, 0, "height of test image.");
-DEFINE_int32(libyuv_repeat, 0, "number of times to repeat test.");
-DEFINE_int32(libyuv_flags, 0, "cpu flags for reference code. 1 = C, -1 = SIMD");
-DEFINE_int32(libyuv_cpu_info,
-             0,
-             "cpu flags for benchmark code. 1 = C, -1 = SIMD");
+#ifdef LIBYUV_USE_ABSL_FLAGS
+ABSL_FLAG(int32_t, libyuv_width, 0, "width of test image.");
+ABSL_FLAG(int32_t, libyuv_height, 0, "height of test image.");
+ABSL_FLAG(int32_t, libyuv_repeat, 0, "number of times to repeat test.");
+ABSL_FLAG(int32_t,
+          libyuv_flags,
+          0,
+          "cpu flags for reference code. 1 = C, -1 = SIMD");
+ABSL_FLAG(int32_t,
+          libyuv_cpu_info,
+          0,
+          "cpu flags for benchmark code. 1 = C, -1 = SIMD");
 #else
-// Disable command line parameters if gflags disabled.
+// Disable command line parameters if absl/flags disabled.
 static const int32_t FLAGS_libyuv_width = 0;
 static const int32_t FLAGS_libyuv_height = 0;
 static const int32_t FLAGS_libyuv_repeat = 0;
@@ -38,6 +43,12 @@ static const int32_t FLAGS_libyuv_flags = 0;
 static const int32_t FLAGS_libyuv_cpu_info = 0;
 #endif
 
+#ifdef LIBYUV_USE_ABSL_FLAGS
+#define LIBYUV_GET_FLAG(f) absl::GetFlag(f)
+#else
+#define LIBYUV_GET_FLAG(f) f
+#endif
+
 // Test environment variable for disabling CPU features. Any non-zero value
 // to disable. Zero ignored to make it easy to set the variable on/off.
 #if !defined(__native_client__) && !defined(_M_ARM)
@@ -66,8 +77,15 @@ int TestCpuEnv(int cpu_info) {
   if (TestEnv("LIBYUV_DISABLE_MSA")) {
     cpu_info &= ~libyuv::kCpuHasMSA;
   }
-  if (TestEnv("LIBYUV_DISABLE_MMI")) {
-    cpu_info &= ~libyuv::kCpuHasMMI;
+#endif
+#if defined(__longarch__) && defined(__linux__)
+  if (TestEnv("LIBYUV_DISABLE_LSX")) {
+    cpu_info &= ~libyuv::kCpuHasLSX;
+  }
+#endif
+#if defined(__longarch__) && defined(__linux__)
+  if (TestEnv("LIBYUV_DISABLE_LASX")) {
+    cpu_info &= ~libyuv::kCpuHasLASX;
   }
 #endif
 #if !defined(__pnacl__) && !defined(__CLR_VER) &&                   \
@@ -109,6 +127,9 @@ int TestCpuEnv(int cpu_info) {
   if (TestEnv("LIBYUV_DISABLE_AVX512VL")) {
     cpu_info &= ~libyuv::kCpuHasAVX512VL;
   }
+  if (TestEnv("LIBYUV_DISABLE_AVX512VNNI")) {
+    cpu_info &= ~libyuv::kCpuHasAVX512VNNI;
+  }
   if (TestEnv("LIBYUV_DISABLE_AVX512VBMI")) {
     cpu_info &= ~libyuv::kCpuHasAVX512VBMI;
   }
@@ -145,8 +166,8 @@ LibYUVConvertTest::LibYUVConvertTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -156,29 +177,29 @@ LibYUVConvertTest::LibYUVConvertTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -201,8 +222,8 @@ LibYUVColorTest::LibYUVColorTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -212,29 +233,29 @@ LibYUVColorTest::LibYUVColorTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -257,8 +278,8 @@ LibYUVScaleTest::LibYUVScaleTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -268,29 +289,29 @@ LibYUVScaleTest::LibYUVScaleTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -313,8 +334,8 @@ LibYUVRotateTest::LibYUVRotateTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -324,29 +345,29 @@ LibYUVRotateTest::LibYUVRotateTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -369,8 +390,8 @@ LibYUVPlanarTest::LibYUVPlanarTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -380,29 +401,29 @@ LibYUVPlanarTest::LibYUVPlanarTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -425,8 +446,8 @@ LibYUVBaseTest::LibYUVBaseTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -436,29 +457,29 @@ LibYUVBaseTest::LibYUVBaseTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -481,8 +502,8 @@ LibYUVCompareTest::LibYUVCompareTest()
   if (repeat) {
     benchmark_iterations_ = atoi(repeat);  // NOLINT
   }
-  if (FLAGS_libyuv_repeat) {
-    benchmark_iterations_ = FLAGS_libyuv_repeat;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_repeat)) {
+    benchmark_iterations_ = LIBYUV_GET_FLAG(FLAGS_libyuv_repeat);
   }
   if (benchmark_iterations_ > 1) {
     benchmark_width_ = 1280;
@@ -492,29 +513,29 @@ LibYUVCompareTest::LibYUVCompareTest()
   if (width) {
     benchmark_width_ = atoi(width);  // NOLINT
   }
-  if (FLAGS_libyuv_width) {
-    benchmark_width_ = FLAGS_libyuv_width;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_width)) {
+    benchmark_width_ = LIBYUV_GET_FLAG(FLAGS_libyuv_width);
   }
   const char* height = getenv("LIBYUV_HEIGHT");
   if (height) {
     benchmark_height_ = atoi(height);  // NOLINT
   }
-  if (FLAGS_libyuv_height) {
-    benchmark_height_ = FLAGS_libyuv_height;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_height)) {
+    benchmark_height_ = LIBYUV_GET_FLAG(FLAGS_libyuv_height);
   }
   const char* cpu_flags = getenv("LIBYUV_FLAGS");
   if (cpu_flags) {
     disable_cpu_flags_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_flags) {
-    disable_cpu_flags_ = FLAGS_libyuv_flags;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_flags)) {
+    disable_cpu_flags_ = LIBYUV_GET_FLAG(FLAGS_libyuv_flags);
   }
   const char* cpu_info = getenv("LIBYUV_CPU_INFO");
   if (cpu_info) {
     benchmark_cpu_info_ = atoi(cpu_flags);  // NOLINT
   }
-  if (FLAGS_libyuv_cpu_info) {
-    benchmark_cpu_info_ = FLAGS_libyuv_cpu_info;
+  if (LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info)) {
+    benchmark_cpu_info_ = LIBYUV_GET_FLAG(FLAGS_libyuv_cpu_info);
   }
   disable_cpu_flags_ = TestCpuEnv(disable_cpu_flags_);
   benchmark_cpu_info_ = TestCpuEnv(benchmark_cpu_info_);
@@ -529,11 +550,8 @@ LibYUVCompareTest::LibYUVCompareTest()
 
 int main(int argc, char** argv) {
   ::testing::InitGoogleTest(&argc, argv);
-#ifdef LIBYUV_USE_GFLAGS
-  // AllowCommandLineParsing allows us to ignore flags passed on to us by
-  // Chromium build bots without having to explicitly disable them.
-  google::AllowCommandLineReparsing();
-  google::ParseCommandLineFlags(&argc, &argv, true);
+#ifdef LIBYUV_USE_ABSL_FLAGS
+  absl::ParseCommandLine(argc, argv);
 #endif
   return RUN_ALL_TESTS();
 }
diff --git a/files/unit_test/unit_test.h b/files/unit_test/unit_test.h
index 87907fa1..0a8df4d2 100644
--- a/files/unit_test/unit_test.h
+++ b/files/unit_test/unit_test.h
@@ -11,7 +11,7 @@
 #ifndef UNIT_TEST_UNIT_TEST_H_  // NOLINT
 #define UNIT_TEST_UNIT_TEST_H_
 
-#ifdef WIN32
+#ifdef _WIN32
 #include <windows.h>
 #else
 #include <sys/resource.h>
@@ -111,10 +111,13 @@ inline int fastrand() {
   return static_cast<int>((fastrand_seed >> 16) & 0xffff);
 }
 
+// ubsan fails if dst is unaligned unless we use uint8
 static inline void MemRandomize(uint8_t* dst, int64_t len) {
   int64_t i;
   for (i = 0; i < len - 1; i += 2) {
-    *reinterpret_cast<uint16_t*>(dst) = fastrand();
+    int r = fastrand();
+    dst[0] = static_cast<uint8_t>(r);
+    dst[1] = static_cast<uint8_t>(r >> 8);
     dst += 2;
   }
   for (; i < len; ++i) {
diff --git a/files/unit_test/video_common_test.cc b/files/unit_test/video_common_test.cc
index a84206a2..36728ea9 100644
--- a/files/unit_test/video_common_test.cc
+++ b/files/unit_test/video_common_test.cc
@@ -29,7 +29,7 @@ static bool TestValidFourCC(uint32_t fourcc, int bpp) {
       !TestValidChar((fourcc >> 24) & 0xff)) {
     return false;
   }
-  if (bpp < 0 || bpp > 32) {
+  if (bpp < 0 || bpp > 64) {
     return false;
   }
   return true;
@@ -65,13 +65,15 @@ TEST_F(LibYUVBaseTest, TestFourCC) {
   EXPECT_TRUE(TestValidFourCC(FOURCC_NV12, FOURCC_BPP_NV12));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YUY2, FOURCC_BPP_YUY2));
   EXPECT_TRUE(TestValidFourCC(FOURCC_UYVY, FOURCC_BPP_UYVY));
-  EXPECT_TRUE(TestValidFourCC(FOURCC_M420, FOURCC_BPP_M420));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_M420, FOURCC_BPP_M420));  // deprecated.
   EXPECT_TRUE(TestValidFourCC(FOURCC_Q420, FOURCC_BPP_Q420));  // deprecated.
   EXPECT_TRUE(TestValidFourCC(FOURCC_ARGB, FOURCC_BPP_ARGB));
   EXPECT_TRUE(TestValidFourCC(FOURCC_BGRA, FOURCC_BPP_BGRA));
   EXPECT_TRUE(TestValidFourCC(FOURCC_ABGR, FOURCC_BPP_ABGR));
   EXPECT_TRUE(TestValidFourCC(FOURCC_AR30, FOURCC_BPP_AR30));
   EXPECT_TRUE(TestValidFourCC(FOURCC_AB30, FOURCC_BPP_AB30));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AR64, FOURCC_BPP_AR64));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_AB64, FOURCC_BPP_AB64));
   EXPECT_TRUE(TestValidFourCC(FOURCC_24BG, FOURCC_BPP_24BG));
   EXPECT_TRUE(TestValidFourCC(FOURCC_RAW, FOURCC_BPP_RAW));
   EXPECT_TRUE(TestValidFourCC(FOURCC_RGBA, FOURCC_BPP_RGBA));
@@ -81,6 +83,11 @@ TEST_F(LibYUVBaseTest, TestFourCC) {
   EXPECT_TRUE(TestValidFourCC(FOURCC_H420, FOURCC_BPP_H420));
   EXPECT_TRUE(TestValidFourCC(FOURCC_H422, FOURCC_BPP_H422));
   EXPECT_TRUE(TestValidFourCC(FOURCC_H010, FOURCC_BPP_H010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_H210, FOURCC_BPP_H210));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_I010, FOURCC_BPP_I010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_I210, FOURCC_BPP_I210));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_P010, FOURCC_BPP_P010));
+  EXPECT_TRUE(TestValidFourCC(FOURCC_P210, FOURCC_BPP_P210));
   EXPECT_TRUE(TestValidFourCC(FOURCC_MJPG, FOURCC_BPP_MJPG));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YV12, FOURCC_BPP_YV12));
   EXPECT_TRUE(TestValidFourCC(FOURCC_YV16, FOURCC_BPP_YV16));
diff --git a/files/util/android/test_runner.py b/files/util/android/test_runner.py
deleted file mode 100755
index 8b06b7ea..00000000
--- a/files/util/android/test_runner.py
+++ /dev/null
@@ -1,37 +0,0 @@
-#!/usr/bin/env python
-# Copyright 2014 The LibYuv Project Authors. All rights reserved.
-#
-# Use of this source code is governed by a BSD-style license
-# that can be found in the LICENSE file in the root of the source
-# tree. An additional intellectual property rights grant can be found
-# in the file PATENTS. All contributing project authors may
-# be found in the AUTHORS file in the root of the source tree.
-
-"""
-Runs tests on Android devices.
-
-This script exists to avoid Libyuv being broken by changes in the Chrome Android
-test execution toolchain. It also conveniently sets the CHECKOUT_SOURCE_ROOT
-environment variable.
-"""
-
-import os
-import sys
-
-SCRIPT_DIR = os.path.dirname(__file__)
-ROOT_DIR = os.path.abspath(os.path.join(SCRIPT_DIR, os.pardir, os.pardir))
-CHROMIUM_BUILD_ANDROID_DIR = os.path.join(ROOT_DIR, 'build', 'android')
-sys.path.insert(0, CHROMIUM_BUILD_ANDROID_DIR)
-
-
-import test_runner  # pylint: disable=W0406
-
-def main():
-  # Override environment variable to make it possible for the scripts to find
-  # the root directory (our symlinking of the Chromium build toolchain would
-  # otherwise make them fail to do so).
-  os.environ['CHECKOUT_SOURCE_ROOT'] = ROOT_DIR
-  return test_runner.main()
-
-if __name__ == '__main__':
-  sys.exit(main())
diff --git a/files/util/color.cc b/files/util/color.cc
new file mode 100644
index 00000000..8c3bbefd
--- /dev/null
+++ b/files/util/color.cc
@@ -0,0 +1,120 @@
+/*
+ *  Copyright 2021 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+// This utility computes values needed to generate yuvconstants based on
+// white point values.
+// The yuv formulas are tuned for 8 bit YUV channels.
+
+// For those MCs that can be represented as kr and kb:
+// Full range
+// float M[3][3]
+// {{1,0,2*(1-kr)},{1,-((2*kb)/((2-kb)*(1-kb-kr))),-((2*kr)/((2-kr)*(1-kb-kr)))},{1,2*(1-kb),0}};
+// float B[3]
+// {1+(256*(1-kr))/255,1-(256*kb)/(255*(2-kb)*(1-kb-kr))-(256*kr)/(255*(2-kr)*(1-kb-kr)),1+(256*(1-kb))/255};
+// Limited range
+// float M[3][3]
+// {{85/73,0,255/112-(255*kr)/112},{85/73,-((255*kb)/(112*(2-kb)*(1-kb-kr))),-((255*kr)/(112*(2-kr)*(1-kb-kr)))},{85/73,255/112-(255*kb)/112,0}};
+// float B[3]
+// {77662/43435-(1537*kr)/1785,203/219-(1537*kb)/(1785*(2-kb)*(1-kb-kr))-(1537*kr)/(1785*(2-kr)*(1-kb-kr)),77662/43435-(1537*kb)/1785};
+
+// mc bt
+// 1 bt.709      KR = 0.2126; KB = 0.0722
+// 4 fcc         KR = 0.30;   KB = 0.11
+// 6 bt.601      KR = 0.299;  KB = 0.114
+// 7 SMPTE 240M  KR = 0.212;  KB = 0.087
+// 10 bt2020     KR = 0.2627; KB = 0.0593
+
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126
+//  KB = 0.0722
+
+// https://mymusing.co/bt601-yuv-to-rgb-conversion-color/
+
+// // Y contribution to R,G,B.  Scale and bias.
+// #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+// #define YB 32    /* 64 / 2 */
+//
+// // U and V contributions to R,G,B.
+// #define UB 113 /* round(1.77200 * 64) */
+// #define UG 22  /* round(0.34414 * 64) */
+// #define VG 46  /* round(0.71414 * 64) */
+// #define VR 90  /* round(1.40200 * 64) */
+//
+// // Bias values to round, and subtract 128 from U and V.
+// #define BB (-UB * 128 + YB)
+// #define BG (UG * 128 + VG * 128 + YB)
+// #define BR (-VR * 128 + YB)
+
+int round(float v) {
+  return (int)(v + 0.5);
+}
+
+int main(int argc, const char* argv[]) {
+  if (argc < 2) {
+    printf("color kr kb\n");
+    return -1;
+  }
+  float kr = atof(argv[1]);
+  float kb = atof(argv[2]);
+  float kg = 1 - kr - kb;
+
+  float vr = 2 * (1 - kr);
+  float ug = 2 * ((1 - kb) * kb / kg);
+  float vg = 2 * ((1 - kr) * kr / kg);
+  float ub = 2 * (1 - kb);
+
+  printf("Full range\n");
+  printf("R = Y                + V * %5f\n", vr);
+  printf("G = Y - U * %6f - V * %6f\n", ug, vg);
+  printf("B = Y + U * %5f\n", ub);
+
+  printf("KR = %4f; ", kr);
+  printf("KB = %4f\n", kb);
+  //  printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3d /* round(%f * 64) */\n", round(ub * 64), ub);
+  printf("UG %-3d /* round(%f * 64) */\n", round(ug * 64), ug);
+  printf("VG %-3d /* round(%f * 64) */\n", round(vg * 64), vg);
+  printf("VR %-3d /* round(%f * 64) */\n", round(vr * 64), vr);
+
+  vr = 255.f / 224.f * 2 * (1 - kr);
+  ug = 255.f / 224.f * 2 * ((1 - kb) * kb / kg);
+  vg = 255.f / 224.f * 2 * ((1 - kr) * kr / kg);
+  ub = 255.f / 224.f * 2 * (1 - kb);
+
+  printf("Limited range\n");
+  printf("R = (Y - 16) * 1.164                + V * %5f\n", vr);
+  printf("G = (Y - 16) * 1.164 - U * %6f - V * %6f\n", ug, vg);
+  printf("B = (Y - 16) * 1.164 + U * %5f\n", ub);
+
+  //  printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3d /* round(%f * 64) */\n", round(ub * 64), ub);
+  printf("UG %-3d /* round(%f * 64) */\n", round(ug * 64), ug);
+  printf("VG %-3d /* round(%f * 64) */\n", round(vg * 64), vg);
+  printf("VR %-3d /* round(%f * 64) */\n", round(vr * 64), vr);
+
+  return 0;
+}
diff --git a/files/util/cpuid.c b/files/util/cpuid.c
index 84c06022..b618bb10 100644
--- a/files/util/cpuid.c
+++ b/files/util/cpuid.c
@@ -12,16 +12,18 @@
 #include <stdlib.h>
 #include <string.h>
 
-#define INCLUDE_LIBYUV_COMPARE_H_
-#include "libyuv.h"
-#include "./psnr.h"
-#include "./ssim.h"
+#include "libyuv/cpu_id.h"
+
+#ifdef __cplusplus
+using namespace libyuv;
+#endif
 
 int main(int argc, const char* argv[]) {
   int cpu_flags = TestCpuFlag(-1);
   int has_arm = TestCpuFlag(kCpuHasARM);
   int has_mips = TestCpuFlag(kCpuHasMIPS);
   int has_x86 = TestCpuFlag(kCpuHasX86);
+  int has_loongarch = TestCpuFlag(kCpuHasLOONGARCH);
   (void)argc;
   (void)argv;
 
@@ -64,6 +66,7 @@ int main(int argc, const char* argv[]) {
   printf("Has ARM %x\n", has_arm);
   printf("Has MIPS %x\n", has_mips);
   printf("Has X86 %x\n", has_x86);
+  printf("Has LOONGARCH %x\n", has_loongarch);
   if (has_arm) {
     int has_neon = TestCpuFlag(kCpuHasNEON);
     printf("Has NEON %x\n", has_neon);
@@ -71,8 +74,12 @@ int main(int argc, const char* argv[]) {
   if (has_mips) {
     int has_msa = TestCpuFlag(kCpuHasMSA);
     printf("Has MSA %x\n", has_msa);
-    int has_mmi = TestCpuFlag(kCpuHasMMI);
-    printf("Has MMI %x\n", has_mmi);
+  }
+  if (has_loongarch) {
+    int has_lsx  = TestCpuFlag(kCpuHasLSX);
+    printf("Has LSX %x\n", has_lsx);
+    int has_lasx = TestCpuFlag(kCpuHasLASX);
+    printf("Has LASX %x\n", has_lasx);
   }
   if (has_x86) {
     int has_sse2 = TestCpuFlag(kCpuHasSSE2);
@@ -83,10 +90,11 @@ int main(int argc, const char* argv[]) {
     int has_avx2 = TestCpuFlag(kCpuHasAVX2);
     int has_erms = TestCpuFlag(kCpuHasERMS);
     int has_fma3 = TestCpuFlag(kCpuHasFMA3);
-    int has_f16c = TestCpuFlag(kCpuHasF16C); 
+    int has_f16c = TestCpuFlag(kCpuHasF16C);
     int has_gfni = TestCpuFlag(kCpuHasGFNI);
     int has_avx512bw = TestCpuFlag(kCpuHasAVX512BW);
     int has_avx512vl = TestCpuFlag(kCpuHasAVX512VL);
+    int has_avx512vnni = TestCpuFlag(kCpuHasAVX512VNNI);
     int has_avx512vbmi = TestCpuFlag(kCpuHasAVX512VBMI);
     int has_avx512vbmi2 = TestCpuFlag(kCpuHasAVX512VBMI2);
     int has_avx512vbitalg = TestCpuFlag(kCpuHasAVX512VBITALG);
@@ -103,6 +111,7 @@ int main(int argc, const char* argv[]) {
     printf("Has GFNI %x\n", has_gfni);
     printf("Has AVX512BW %x\n", has_avx512bw);
     printf("Has AVX512VL %x\n", has_avx512vl);
+    printf("Has AVX512VNNI %x\n", has_avx512vnni);
     printf("Has AVX512VBMI %x\n", has_avx512vbmi);
     printf("Has AVX512VBMI2 %x\n", has_avx512vbmi2);
     printf("Has AVX512VBITALG %x\n", has_avx512vbitalg);
diff --git a/files/util/i444tonv12_eg.cc b/files/util/i444tonv12_eg.cc
new file mode 100644
index 00000000..0fcb4095
--- /dev/null
+++ b/files/util/i444tonv12_eg.cc
@@ -0,0 +1,28 @@
+
+#include "libyuv/convert.h"
+
+#include <stdio.h>   // for printf
+#include <string.h>  // for memset
+
+int main(int, char**) {
+  unsigned char src_i444[640 * 400 * 3];
+  unsigned char dst_nv12[640 * 400 * 3 / 2];
+
+  for (size_t i = 0; i < sizeof(src_i444); ++i) {
+    src_i444[i] = i & 255;
+  }
+  memset(dst_nv12, 0, sizeof(dst_nv12));
+  libyuv::I444ToNV12(&src_i444[0], 640,              // source Y
+                     &src_i444[640 * 400], 640,      // source U
+                     &src_i444[640 * 400 * 2], 640,  // source V
+                     &dst_nv12[0], 640,              // dest Y
+                     &dst_nv12[640 * 400], 640,      // dest UV
+                     640, 400);                      // width and height
+
+  int checksum = 0;
+  for (size_t i = 0; i < sizeof(dst_nv12); ++i) {
+    checksum += dst_nv12[i];
+  }
+  printf("checksum %x %s\n", checksum, checksum == 0x2ec0c00 ? "PASS" : "FAIL");
+  return 0;
+}
\ No newline at end of file
diff --git a/files/util/psnr_main.cc b/files/util/psnr_main.cc
index a930b202..8b9fd972 100644
--- a/files/util/psnr_main.cc
+++ b/files/util/psnr_main.cc
@@ -248,13 +248,13 @@ bool UpdateMetrics(uint8_t* ch_org,
                    int number_of_frames,
                    metric* cur_distortion_psnr,
                    metric* distorted_frame,
-                   bool do_psnr) {
+                   bool compute_psnr) {
   const int uv_offset = (do_swap_uv ? uv_size : 0);
   const uint8_t* const u_org = ch_org + y_size + uv_offset;
   const uint8_t* const u_rec = ch_rec + y_size;
   const uint8_t* const v_org = ch_org + y_size + (uv_size - uv_offset);
   const uint8_t* const v_rec = ch_rec + y_size + uv_size;
-  if (do_psnr) {
+  if (compute_psnr) {
 #ifdef HAVE_JPEG
     double y_err = static_cast<double>(
         libyuv::ComputeSumSquareError(ch_org, ch_rec, y_size));
diff --git a/files/util/yuvconstants.c b/files/util/yuvconstants.c
new file mode 100644
index 00000000..037e0824
--- /dev/null
+++ b/files/util/yuvconstants.c
@@ -0,0 +1,105 @@
+/*
+ *  Copyright 2021 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+// This utility computes values needed to generate yuvconstants based on
+// white point values.
+// The yuv formulas are tuned for 8 bit YUV channels.
+
+// See Also
+// https://mymusing.co/bt601-yuv-to-rgb-conversion-color/
+
+// BT.709 full range YUV to RGB reference
+//  R = Y               + V * 1.5748
+//  G = Y - U * 0.18732 - V * 0.46812
+//  B = Y + U * 1.8556
+//  KR = 0.2126
+//  KB = 0.0722
+
+// // Y contribution to R,G,B.  Scale and bias.
+// #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+// #define YB 32    /* 64 / 2 */
+//
+// // U and V contributions to R,G,B.
+// #define UB 113 /* round(1.77200 * 64) */
+// #define UG 22  /* round(0.34414 * 64) */
+// #define VG 46  /* round(0.71414 * 64) */
+// #define VR 90  /* round(1.40200 * 64) */
+//
+// // Bias values to round, and subtract 128 from U and V.
+// #define BB (-UB * 128 + YB)
+// #define BG (UG * 128 + VG * 128 + YB)
+// #define BR (-VR * 128 + YB)
+
+int main(int argc, const char* argv[]) {
+  if (argc < 2) {
+    printf("yuvconstants Kr Kb\n");
+    printf("  MC BT          KR = 0.2126; KB = 0.0722\n");
+    printf("  1  BT.709      KR = 0.2126; KB = 0.0722\n");
+    printf("  4  FCC         KR = 0.30;   KB = 0.11\n");
+    printf("  6  BT.601      KR = 0.299;  KB = 0.114\n");
+    printf("  7  SMPTE 240M  KR = 0.212;  KB = 0.087\n");
+    printf("  9  BT.2020     KR = 0.2627; KB = 0.0593\n");
+    return -1;
+  }
+  float kr = atof(argv[1]);
+  float kb = atof(argv[2]);
+  float kg = 1 - kr - kb;
+
+  float vr = 2 * (1 - kr);
+  float ug = 2 * ((1 - kb) * kb / kg);
+  float vg = 2 * ((1 - kr) * kr / kg);
+  float ub = 2 * (1 - kb);
+
+  printf("Full range\n");
+  printf("R = Y                + V * %5f\n", vr);
+  printf("G = Y - U * %6f - V * %6f\n", ug, vg);
+  printf("B = Y + U * %5f\n", ub);
+
+  printf("KR = %4f; ", kr);
+  printf("KB = %4f\n", kb);
+  // printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ub * 64), ub, ub * 64);
+  printf("UG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ug * 64), ug, ug * 64);
+  printf("VG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vg * 64), vg, vg * 64);
+  printf("VR %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vr * 64), vr, vr * 64);
+
+  vr = 255.f / 224.f * 2 * (1 - kr);
+  ug = 255.f / 224.f * 2 * ((1 - kb) * kb / kg);
+  vg = 255.f / 224.f * 2 * ((1 - kr) * kr / kg);
+  ub = 255.f / 224.f * 2 * (1 - kb);
+
+  printf("\nLimited range\n");
+  printf("R = (Y - 16) * 1.164                + V * %5f\n", vr);
+  printf("G = (Y - 16) * 1.164 - U * %6f - V * %6f\n", ug, vg);
+  printf("B = (Y - 16) * 1.164 + U * %5f\n", ub);
+
+  // printf("KG = %4f\n", kg);
+  // #define YG 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+  // #define YB 32    /* 64 / 2 */
+  //
+  // // U and V contributions to R,G,B.
+
+  printf("UB %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ub * 64), ub, ub * 64);
+  printf("UG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(ug * 64), ug, ug * 64);
+  printf("VG %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vg * 64), vg, vg * 64);
+  printf("VR %-3.0f /* round(%f * 64 = %8.4f) */\n", round(vr * 64), vr, vr * 64);
+
+  return 0;
+}
diff --git a/files/util/yuvconvert.cc b/files/util/yuvconvert.cc
index 27cdfe9e..332699e3 100644
--- a/files/util/yuvconvert.cc
+++ b/files/util/yuvconvert.cc
@@ -165,23 +165,23 @@ static int TileARGBScale(const uint8_t* src_argb,
                          int src_height,
                          uint8_t* dst_argb,
                          int dst_stride_argb,
-                         int dst_width,
-                         int dst_height,
+                         int destination_width,
+                         int destination_height,
                          libyuv::FilterMode filtering) {
-  for (int y = 0; y < dst_height; y += kTileY) {
-    for (int x = 0; x < dst_width; x += kTileX) {
+  for (int y = 0; y < destination_height; y += kTileY) {
+    for (int x = 0; x < destination_width; x += kTileX) {
       int clip_width = kTileX;
-      if (x + clip_width > dst_width) {
-        clip_width = dst_width - x;
+      if (x + clip_width > destination_width) {
+        clip_width = destination_width - x;
       }
       int clip_height = kTileY;
-      if (y + clip_height > dst_height) {
-        clip_height = dst_height - y;
+      if (y + clip_height > destination_height) {
+        clip_height = destination_height - y;
       }
       int r = libyuv::ARGBScaleClip(src_argb, src_stride_argb, src_width,
                                     src_height, dst_argb, dst_stride_argb,
-                                    dst_width, dst_height, x, y, clip_width,
-                                    clip_height, filtering);
+                                    destination_width, destination_height, x, y,
+                                    clip_width, clip_height, filtering);
       if (r) {
         return r;
       }
diff --git a/files/winarm.mk b/files/winarm.mk
index c4307a43..b0a344ae 100644
--- a/files/winarm.mk
+++ b/files/winarm.mk
@@ -31,6 +31,7 @@ LOCAL_OBJ_FILES = \
 	source/scale_any.o\
 	source/scale_argb.o\
 	source/scale_common.o\
+	source/scale_uv.o\
 	source/video_common.o
 
 .cc.o: