Sync-patch with libwebp (ChangeId: Ia5475247)android-4.3_r3.1android-4.3_r3android-4.3_r2.3android-4.3_r2.2android-4.3_r2.1android-4.3_r2android-4.3_r1.1android-4.3_r1android-4.3_r0.9.1android-4.3_r0.9android-4.3.1_r1tools_r22.2jb-mr2.0.0-releasejb-mr2.0-releasejb-mr2-releasejb-mr2-dev

Added 16bit swapping of RGB565 / RGB4444 colorspace.
Added ARM/NEON code for decoder/encoder modules.
Speedup in WebP compression (method 3 and above).

Change-Id: I95a697338bef7c3ea08054eb5f850a97d1889eb9

63 files changed, 5489 insertions, 1566 deletions

diff --git a/ChangeLog b/ChangeLog
index 06c8316c..ef8b804b 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -4,3 +4,4 @@
 - 7/11: Updated WebP with head change#Ia53f845b
 - 8/12: release version 0.2.0-rc1 (head change#Ia5475247)
 -     : Updated WebP with head change#I3da2063b
+- 3/13: Updated WebP with head change#I9e5ae737
diff --git a/NEWS b/NEWS
index 9e3fa764..350327dd 100644
--- a/NEWS
+++ b/NEWS
@@ -1,2 +1,3 @@
 - 9/10: initial release version 0.1
 - 8/12: release version 0.2.0-rc1 (head change#Ia5475247).
+- 3/13: Updated WebP with head change#I9e5ae737
diff --git a/README b/README
index d3f40620..cd135850 100644
--- a/README
+++ b/README
@@ -4,7 +4,7 @@
           \__\__/\____/\_____/__/ ____  ___
                 / _/ /    \    \ /  _ \/ _/
                /  \_/   / /   \ \   __/  \__
-               \____/____/\_____/_____/____/v0.2.0
+               \____/____/\_____/_____/____/v0.2.1
 
 Description:
 ============
@@ -25,8 +25,8 @@ property rights grant can be found in the file PATENTS.
 Building:
 =========
 
-Windows build (without experimental features):
-----------------------------------------------
+Windows build:
+--------------
 
 By running:
 
@@ -38,49 +38,6 @@ contain the libwebp static library.
 The target architecture (x86/x64) is detected by Makefile.vc from the Visual
 Studio compiler (cl.exe) available in the system path.
 
-Windows build (with experimental features):
--------------------------------------------
-
-This release requires the zlib library. This library is not common under
-Windows nor can it be replaced with calls to the Windows API, so you will need
-to download it or build it yourself.
-
-You can either:
-(a) Link zlib dynamically (as a DLL).
-(b) Link zlib statically.
-
-Linking it dynamically is easier to do (as you can download a precompiled DLL),
-but a bit more cumbersome to use - you need to keep zlib1.dll in the same
-directory as the created EXEs.
-
-a. With dynamic zlib
-~~~~~~~~~~~~~~~~~~~~
-1. Download and unpack the archive from http://zlib.net/zlib125-dll.zip.
-2. Compile libwebp with the command:
-  nmake /f Makefile.vc CFG=release-static RTLIBCFG=static OBJDIR=output \
-      ZDLLDIR=C:\path\to\the\unpacked\archive EXPERIMENTAL=y
-3. Copy C:\path\to\the\unpacked\archive\zlib1.dll to
-   output\release-static\x86\bin for binaries there to work.
-Note: only a 32-bit DLL is currently available for download, so for a 64-bit
-build, you will need to build the DLL yourself.
-
-b. With static zlib
-~~~~~~~~~~~~~~~~~~~
-
-1. Download and unpack the source code from http://zlib.net/zlib125.zip.
-2. For a 32-bit build of zlib compatible with libwebp:
-  a. Edit zlib's win32\Makefile.msc, adding "-SAFESEH" to ASFLAGS.
-  b. Compile zlib with the command:
-    nmake /f win32/Makefile.msc LOC="-DASMV -DASMINF -MT" \
-        OBJA="inffas32.obj match686.obj"
-3. Compile libwebp with the command
-  nmake /f Makefile.vc CFG=release-static RTLIBCFG=static OBJDIR=output \
-      ZLIBDIR=C:\path\to\the\unpacked\archive experimental
-4. Use the binaries in output\release-static\x86\bin.
-Note: a 64-bit build of zlib requires different modifications of
-Makefile.msc.
-
-
 Unix build using makefile.unix:
 -------------------------------
 
@@ -194,12 +151,15 @@ options:
   -sns <int> ............. Spatial Noise Shaping (0:off, 100:max)
   -f <int> ............... filter strength (0=off..100)
   -sharpness <int> ....... filter sharpness (0:most .. 7:least sharp)
-  -strong ................ use strong filter instead of simple.
+  -strong ................ use strong filter instead of simple (default).
+  -nostrong .............. use simple filter instead of strong.
   -partition_limit <int> . limit quality to fit the 512k limit on
                            the first partition (0=no degradation ... 100=full)
   -pass <int> ............ analysis pass number (1..10)
   -crop <x> <y> <w> <h> .. crop picture with the given rectangle
   -resize <w> <h> ........ resize picture (after any cropping)
+  -mt .................... use multi-threading if available
+  -low_memory ............ reduce memory usage (slower encoding)
   -map <int> ............. print map of extra info.
   -print_psnr ............ prints averaged PSNR distortion.
   -print_ssim ............ prints averaged SSIM distortion.
@@ -214,6 +174,10 @@ options:
   -hint <string> ......... Specify image characteristics hint.
                            One of: photo, picture or graph
 
+  -metadata <string> ..... comma separated list of metadata to
+                           copy from the input to the output if present.
+                           Valid values: all, none (default), exif, iccp, xmp
+
   -short ................. condense printed message
   -quiet ................. don't print anything.
   -version ............... print version number and exit.
@@ -222,6 +186,7 @@ options:
   -progress .............. report encoding progress
 
 Experimental Options:
+  -jpeg_like ............. Roughly match expected JPEG size.
   -af .................... auto-adjust filter strength.
   -pre <int> ............. pre-processing filter
 
@@ -249,8 +214,8 @@ Namely:
      in-loop processing. The higher the value, the smoother the
      highly-compressed area will look. This is particularly useful when aiming
      at very small files. Typical values are around 20-30. Note that using the
-     option -strong will change the type of filtering. Use "-f 0" to turn
-     filtering off.
+     option -strong/-nostrong will change the type of filtering. Use "-f 0" to
+     turn filtering off.
   * 'm' controls the trade-off between encoding speed and quality. Default is 4.
      You can try -m 5 or -m 6 to explore more (time-consuming) encoding
      possibilities. A lower value will result in faster encoding at the expense
@@ -279,6 +244,8 @@ Use following options to convert into alternate image formats:
   -ppm ......... save the raw RGB samples as a color PPM
   -pgm ......... save the raw YUV samples as a grayscale PGM
                  file with IMC4 layout.
+  -yuv ......... save the raw YUV samples in flat layout.
+
  Other options are:
   -version  .... print version number and exit.
   -nofancy ..... don't use the fancy YUV420 upscaler.
diff --git a/README.android b/README.android
index 787a8a95..de94351e 100644
--- a/README.android
+++ b/README.android
@@ -25,6 +25,14 @@ Local modifications:
   color configs. Added the code under "ANDROID_WEBP_RGB" flag.
 - Sync-patch with libwebp ver 0.2.0-rc1 (head change#Ia5475247).
 - Updated WebP with head change#I3da2063b
+- Updated WebP with head change#I9e5ae737
+  - 16bit swapping of RGB565 / RGB4444 colorspace. Added compile-time flag
+    WEBP_SWAP_16BIT_CSP to support byte-swap (required for skia data type for
+    16bit values) output for these colorspaces.
+  - Added ARM/NEON code for decoder/encoder modules.
+  - Speedup in WebP compression (method 3 and above).
+  - Added multi-threaded alpha encoding for lossy compression.
+  - Updated WebP with head change#I9e5ae737
 
 The Android.mk file creates WebP Decoder and Encoder static libraries which
 can be added to any application by Adding to LOCAL_STATIC_LIBRARIES
diff --git a/include/webp/decode.h b/include/webp/decode.h
index 43b6c58f..d4eb710d 100644
--- a/include/webp/decode.h
+++ b/include/webp/decode.h
@@ -20,6 +20,18 @@ extern "C" {
 
 #define WEBP_DECODER_ABI_VERSION 0x0200    // MAJOR(8b) + MINOR(8b)
 
+typedef struct WebPRGBABuffer WebPRGBABuffer;
+typedef struct WebPYUVABuffer WebPYUVABuffer;
+typedef struct WebPDecBuffer WebPDecBuffer;
+#if !(defined(__cplusplus) || defined(c_plusplus))
+typedef enum VP8StatusCode VP8StatusCode;
+typedef enum WEBP_CSP_MODE WEBP_CSP_MODE;
+#endif
+typedef struct WebPIDecoder WebPIDecoder;
+typedef struct WebPBitstreamFeatures WebPBitstreamFeatures;
+typedef struct WebPDecoderOptions WebPDecoderOptions;
+typedef struct WebPDecoderConfig WebPDecoderConfig;
+
 // Return the decoder's version number, packed in hexadecimal using 8bits for
 // each of major/minor/revision. E.g: v2.5.7 is 0x020507.
 WEBP_EXTERN(int) WebPGetDecoderVersion(void);
@@ -118,20 +130,28 @@ WEBP_EXTERN(uint8_t*) WebPDecodeYUVInto(
 // Note: the naming describes the byte-ordering of packed samples in memory.
 // For instance, MODE_BGRA relates to samples ordered as B,G,R,A,B,G,R,A,...
 // Non-capital names (e.g.:MODE_Argb) relates to pre-multiplied RGB channels.
-// RGB-565 and RGBA-4444 are also endian-agnostic and byte-oriented.
-typedef enum { MODE_RGB = 0, MODE_RGBA = 1,
-               MODE_BGR = 2, MODE_BGRA = 3,
-               MODE_ARGB = 4, MODE_RGBA_4444 = 5,
-               MODE_RGB_565 = 6,
-               // RGB-premultiplied transparent modes (alpha value is preserved)
-               MODE_rgbA = 7,
-               MODE_bgrA = 8,
-               MODE_Argb = 9,
-               MODE_rgbA_4444 = 10,
-               // YUV modes must come after RGB ones.
-               MODE_YUV = 11, MODE_YUVA = 12,  // yuv 4:2:0
-               MODE_LAST = 13
-             } WEBP_CSP_MODE;
+// RGBA-4444 and RGB-565 colorspaces are represented by following byte-order:
+// RGBA-4444: [r3 r2 r1 r0 g3 g2 g1 g0], [b3 b2 b1 b0 a3 a2 a1 a0], ...
+// RGB-565: [r4 r3 r2 r1 r0 g5 g4 g3], [g2 g1 g0 b4 b3 b2 b1 b0], ...
+// In the case WEBP_SWAP_16BITS_CSP is defined, the bytes are swapped for
+// these two modes:
+// RGBA-4444: [b3 b2 b1 b0 a3 a2 a1 a0], [r3 r2 r1 r0 g3 g2 g1 g0], ...
+// RGB-565: [g2 g1 g0 b4 b3 b2 b1 b0], [r4 r3 r2 r1 r0 g5 g4 g3], ...
+
+enum WEBP_CSP_MODE {
+  MODE_RGB = 0, MODE_RGBA = 1,
+  MODE_BGR = 2, MODE_BGRA = 3,
+  MODE_ARGB = 4, MODE_RGBA_4444 = 5,
+  MODE_RGB_565 = 6,
+  // RGB-premultiplied transparent modes (alpha value is preserved)
+  MODE_rgbA = 7,
+  MODE_bgrA = 8,
+  MODE_Argb = 9,
+  MODE_rgbA_4444 = 10,
+  // YUV modes must come after RGB ones.
+  MODE_YUV = 11, MODE_YUVA = 12,  // yuv 4:2:0
+  MODE_LAST = 13
+};
 
 // Some useful macros:
 static WEBP_INLINE int WebPIsPremultipliedMode(WEBP_CSP_MODE mode) {
@@ -152,13 +172,13 @@ static WEBP_INLINE int WebPIsRGBMode(WEBP_CSP_MODE mode) {
 //------------------------------------------------------------------------------
 // WebPDecBuffer: Generic structure for describing the output sample buffer.
 
-typedef struct {    // view as RGBA
+struct WebPRGBABuffer {    // view as RGBA
   uint8_t* rgba;    // pointer to RGBA samples
   int stride;       // stride in bytes from one scanline to the next.
   size_t size;      // total size of the *rgba buffer.
-} WebPRGBABuffer;
+};
 
-typedef struct {              // view as YUVA
+struct WebPYUVABuffer {              // view as YUVA
   uint8_t* y, *u, *v, *a;     // pointer to luma, chroma U/V, alpha samples
   int y_stride;               // luma stride
   int u_stride, v_stride;     // chroma strides
@@ -166,10 +186,10 @@ typedef struct {              // view as YUVA
   size_t y_size;              // luma plane size
   size_t u_size, v_size;      // chroma planes size
   size_t a_size;              // alpha-plane size
-} WebPYUVABuffer;
+};
 
 // Output buffer
-typedef struct {
+struct WebPDecBuffer {
   WEBP_CSP_MODE colorspace;  // Colorspace.
   int width, height;         // Dimensions.
   int is_external_memory;    // If true, 'internal_memory' pointer is not used.
@@ -182,7 +202,7 @@ typedef struct {
   uint8_t* private_memory;   // Internally allocated memory (only when
                              // is_external_memory is false). Should not be used
                              // externally, but accessed via the buffer union.
-} WebPDecBuffer;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(int) WebPInitDecBufferInternal(WebPDecBuffer*, int);
@@ -200,7 +220,7 @@ WEBP_EXTERN(void) WebPFreeDecBuffer(WebPDecBuffer* buffer);
 //------------------------------------------------------------------------------
 // Enumeration of the status codes
 
-typedef enum {
+enum VP8StatusCode {
   VP8_STATUS_OK = 0,
   VP8_STATUS_OUT_OF_MEMORY,
   VP8_STATUS_INVALID_PARAM,
@@ -209,7 +229,7 @@ typedef enum {
   VP8_STATUS_SUSPENDED,
   VP8_STATUS_USER_ABORT,
   VP8_STATUS_NOT_ENOUGH_DATA
-} VP8StatusCode;
+};
 
 //------------------------------------------------------------------------------
 // Incremental decoding
@@ -237,8 +257,6 @@ typedef enum {
 //   }
 //   WebPIDelete(idec);
 
-typedef struct WebPIDecoder WebPIDecoder;
-
 // Creates a new incremental decoder with the supplied buffer parameter.
 // This output_buffer can be passed NULL, in which case a default output buffer
 // is used (with MODE_RGB). Otherwise, an internal reference to 'output_buffer'
@@ -251,19 +269,27 @@ WEBP_EXTERN(WebPIDecoder*) WebPINewDecoder(WebPDecBuffer* output_buffer);
 // will output the RGB/A samples specified by 'csp' into a preallocated
 // buffer 'output_buffer'. The size of this buffer is at least
 // 'output_buffer_size' and the stride (distance in bytes between two scanlines)
-// is specified by 'output_stride'. Returns NULL if the allocation failed.
+// is specified by 'output_stride'.
+// Additionally, output_buffer can be passed NULL in which case the output
+// buffer will be allocated automatically when the decoding starts. The
+// colorspace 'csp' is taken into account for allocating this buffer. All other
+// parameters are ignored.
+// Returns NULL if the allocation failed, or if some parameters are invalid.
 WEBP_EXTERN(WebPIDecoder*) WebPINewRGB(
     WEBP_CSP_MODE csp,
     uint8_t* output_buffer, size_t output_buffer_size, int output_stride);
 
 // This function allocates and initializes an incremental-decoder object, which
-// will output the raw luma/chroma samples into a preallocated planes. The luma
-// plane is specified by its pointer 'luma', its size 'luma_size' and its stride
-// 'luma_stride'. Similarly, the chroma-u plane is specified by the 'u',
-// 'u_size' and 'u_stride' parameters, and the chroma-v plane by 'v'
-// and 'v_size'. And same for the alpha-plane. The 'a' pointer can be pass
-// NULL in case one is not interested in the transparency plane.
-// Returns NULL if the allocation failed.
+// will output the raw luma/chroma samples into a preallocated planes if
+// supplied. The luma plane is specified by its pointer 'luma', its size
+// 'luma_size' and its stride 'luma_stride'. Similarly, the chroma-u plane
+// is specified by the 'u', 'u_size' and 'u_stride' parameters, and the chroma-v
+// plane by 'v' and 'v_size'. And same for the alpha-plane. The 'a' pointer
+// can be pass NULL in case one is not interested in the transparency plane.
+// Conversely, 'luma' can be passed NULL if no preallocated planes are supplied.
+// In this case, the output buffer will be automatically allocated (using
+// MODE_YUVA) when decoding starts. All parameters are then ignored.
+// Returns NULL if the allocation failed or if a parameter is invalid.
 WEBP_EXTERN(WebPIDecoder*) WebPINewYUVA(
     uint8_t* luma, size_t luma_size, int luma_stride,
     uint8_t* u, size_t u_size, int u_stride,
@@ -344,7 +370,7 @@ WEBP_EXTERN(const WebPDecBuffer*) WebPIDecodedArea(
      CHECK(WebPGetFeatures(data, data_size, &config.input) == VP8_STATUS_OK);
 
      // C) Adjust 'config', if needed
-     config.no_fancy = 1;
+     config.no_fancy_upsampling = 1;
      config.output.colorspace = MODE_BGRA;
      // etc.
 
@@ -365,7 +391,7 @@ WEBP_EXTERN(const WebPDecBuffer*) WebPIDecodedArea(
 */
 
 // Features gathered from the bitstream
-typedef struct {
+struct WebPBitstreamFeatures {
   int width;        // Width in pixels, as read from the bitstream.
   int height;       // Height in pixels, as read from the bitstream.
   int has_alpha;    // True if the bitstream contains an alpha channel.
@@ -377,7 +403,7 @@ typedef struct {
   int rotate;                   // TODO(later)
   int uv_sampling;              // should be 0 for now. TODO(later)
   uint32_t pad[3];              // padding for later use
-} WebPBitstreamFeatures;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(VP8StatusCode) WebPGetFeaturesInternal(
@@ -385,8 +411,9 @@ WEBP_EXTERN(VP8StatusCode) WebPGetFeaturesInternal(
 
 // Retrieve features from the bitstream. The *features structure is filled
 // with information gathered from the bitstream.
-// Returns false in case of error or version mismatch.
-// In case of error, features->bitstream_status will reflect the error code.
+// Returns VP8_STATUS_OK when the features are successfully retrieved. Returns
+// VP8_STATUS_NOT_ENOUGH_DATA when more data is needed to retrieve the
+// features from headers. Returns error in other cases.
 static WEBP_INLINE VP8StatusCode WebPGetFeatures(
     const uint8_t* data, size_t data_size,
     WebPBitstreamFeatures* features) {
@@ -395,7 +422,7 @@ static WEBP_INLINE VP8StatusCode WebPGetFeatures(
 }
 
 // Decoding options
-typedef struct {
+struct WebPDecoderOptions {
   int bypass_filtering;               // if true, skip the in-loop filtering
   int no_fancy_upsampling;            // if true, use faster pointwise upsampler
   int use_cropping;                   // if true, cropping is applied _first_
@@ -410,14 +437,14 @@ typedef struct {
   int force_rotation;                 // forced rotation (to be applied _last_)
   int no_enhancement;                 // if true, discard enhancement layer
   uint32_t pad[6];                    // padding for later use
-} WebPDecoderOptions;
+};
 
 // Main object storing the configuration for advanced decoding.
-typedef struct {
+struct WebPDecoderConfig {
   WebPBitstreamFeatures input;  // Immutable bitstream features (optional)
   WebPDecBuffer output;         // Output buffer (can point to external mem)
   WebPDecoderOptions options;   // Decoding options
-} WebPDecoderConfig;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(int) WebPInitDecoderConfigInternal(WebPDecoderConfig*, int);
diff --git a/include/webp/decode_vp8.h b/include/webp/decode_vp8.h
deleted file mode 100644
index cf1654fe..00000000
--- a/include/webp/decode_vp8.h
+++ /dev/null
@@ -1,14 +0,0 @@
-// Copyright 2010 Google Inc. All Rights Reserved.
-//
-// This code is licensed under the same terms as WebM:
-//  Software License Agreement:  http://www.webmproject.org/license/software/
-//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
-// -----------------------------------------------------------------------------
-//
-//  Dummy file retained to honor decode_vp8.h include in
-//  skia/src/images/SkImageDecoder_libwebp.cpp
-
-#ifndef WEBP_WEBP_DECODE_VP8_H_
-#define WEBP_WEBP_DECODE_VP8_H_
-
-#endif  /* WEBP_WEBP_DECODE_VP8_H_ */
diff --git a/include/webp/demux.h b/include/webp/demux.h
new file mode 100644
index 00000000..e227cc14
--- /dev/null
+++ b/include/webp/demux.h
@@ -0,0 +1,214 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// Demux API.
+// Enables extraction of image and extended format data from WebP files.
+
+// Code Example: Demuxing WebP data to extract all the frames, ICC profile
+// and EXIF/XMP metadata.
+//
+//   WebPDemuxer* demux = WebPDemux(&webp_data);
+//
+//   uint32_t width = WebPDemuxGetI(demux, WEBP_FF_CANVAS_WIDTH);
+//   uint32_t height = WebPDemuxGetI(demux, WEBP_FF_CANVAS_HEIGHT);
+//   // ... (Get information about the features present in the WebP file).
+//   uint32_t flags = WebPDemuxGetI(demux, WEBP_FF_FORMAT_FLAGS);
+//
+//   // ... (Iterate over all frames).
+//   WebPIterator iter;
+//   if (WebPDemuxGetFrame(demux, 1, &iter)) {
+//     do {
+//       // ... (Consume 'iter'; e.g. Decode 'iter.fragment' with WebPDecode(),
+//       // ... and get other frame properties like width, height, offsets etc.
+//       // ... see 'struct WebPIterator' below for more info).
+//     } while (WebPDemuxNextFrame(&iter));
+//     WebPDemuxReleaseIterator(&iter);
+//   }
+//
+//   // ... (Extract metadata).
+//   WebPChunkIterator chunk_iter;
+//   if (flags & ICCP_FLAG) WebPDemuxGetChunk(demux, "ICCP", 1, &chunk_iter);
+//   // ... (Consume the ICC profile in 'chunk_iter.chunk').
+//   WebPDemuxReleaseChunkIterator(&chunk_iter);
+//   if (flags & EXIF_FLAG) WebPDemuxGetChunk(demux, "EXIF", 1, &chunk_iter);
+//   // ... (Consume the EXIF metadata in 'chunk_iter.chunk').
+//   WebPDemuxReleaseChunkIterator(&chunk_iter);
+//   if (flags & XMP_FLAG) WebPDemuxGetChunk(demux, "XMP ", 1, &chunk_iter);
+//   // ... (Consume the XMP metadata in 'chunk_iter.chunk').
+//   WebPDemuxReleaseChunkIterator(&chunk_iter);
+//   WebPDemuxDelete(demux);
+
+#ifndef WEBP_WEBP_DEMUX_H_
+#define WEBP_WEBP_DEMUX_H_
+
+#include "./format_constants.h"
+#include "./mux_types.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+#define WEBP_DEMUX_ABI_VERSION 0x0100    // MAJOR(8b) + MINOR(8b)
+
+typedef struct WebPDemuxer WebPDemuxer;
+#if !(defined(__cplusplus) || defined(c_plusplus))
+typedef enum WebPDemuxState WebPDemuxState;
+typedef enum WebPFormatFeature WebPFormatFeature;
+#endif
+typedef struct WebPIterator WebPIterator;
+typedef struct WebPChunkIterator WebPChunkIterator;
+
+//------------------------------------------------------------------------------
+
+// Returns the version number of the demux library, packed in hexadecimal using
+// 8bits for each of major/minor/revision. E.g: v2.5.7 is 0x020507.
+WEBP_EXTERN(int) WebPGetDemuxVersion(void);
+
+//------------------------------------------------------------------------------
+// Life of a Demux object
+
+enum WebPDemuxState {
+  WEBP_DEMUX_PARSING_HEADER,  // Not enough data to parse full header.
+  WEBP_DEMUX_PARSED_HEADER,   // Header parsing complete, data may be available.
+  WEBP_DEMUX_DONE             // Entire file has been parsed.
+};
+
+// Internal, version-checked, entry point
+WEBP_EXTERN(WebPDemuxer*) WebPDemuxInternal(
+    const WebPData*, int, WebPDemuxState*, int);
+
+// Parses the WebP file given by 'data'.
+// A complete WebP file must be present in 'data' for the function to succeed.
+// Returns a WebPDemuxer object on successful parse, NULL otherwise.
+static WEBP_INLINE WebPDemuxer* WebPDemux(const WebPData* data) {
+  return WebPDemuxInternal(data, 0, NULL, WEBP_DEMUX_ABI_VERSION);
+}
+
+// Parses the WebP file given by 'data'.
+// If 'state' is non-NULL it will be set to indicate the status of the demuxer.
+// Returns a WebPDemuxer object on successful parse, NULL otherwise.
+static WEBP_INLINE WebPDemuxer* WebPDemuxPartial(
+    const WebPData* data, WebPDemuxState* state) {
+  return WebPDemuxInternal(data, 1, state, WEBP_DEMUX_ABI_VERSION);
+}
+
+// Frees memory associated with 'dmux'.
+WEBP_EXTERN(void) WebPDemuxDelete(WebPDemuxer* dmux);
+
+//------------------------------------------------------------------------------
+// Data/information extraction.
+
+enum WebPFormatFeature {
+  WEBP_FF_FORMAT_FLAGS,  // Extended format flags present in the 'VP8X' chunk.
+  WEBP_FF_CANVAS_WIDTH,
+  WEBP_FF_CANVAS_HEIGHT,
+  WEBP_FF_LOOP_COUNT,
+  WEBP_FF_BACKGROUND_COLOR,
+  WEBP_FF_FRAME_COUNT    // Number of frames present in the demux object.
+                         // In case of a partial demux, this is the number of
+                         // frames seen so far, with the last frame possibly
+                         // being partial.
+};
+
+// Get the 'feature' value from the 'dmux'.
+// NOTE: values are only valid if WebPDemux() was used or WebPDemuxPartial()
+// returned a state > WEBP_DEMUX_PARSING_HEADER.
+WEBP_EXTERN(uint32_t) WebPDemuxGetI(
+    const WebPDemuxer* dmux, WebPFormatFeature feature);
+
+//------------------------------------------------------------------------------
+// Frame iteration.
+
+struct WebPIterator {
+  int frame_num;
+  int num_frames;          // equivalent to WEBP_FF_FRAME_COUNT.
+  int fragment_num;
+  int num_fragments;
+  int x_offset, y_offset;  // offset relative to the canvas.
+  int width, height;       // dimensions of this frame or fragment.
+  int duration;            // display duration in milliseconds.
+  WebPMuxAnimDispose dispose_method;  // dispose method for the frame.
+  int complete;   // true if 'fragment' contains a full frame. partial images
+                  // may still be decoded with the WebP incremental decoder.
+  WebPData fragment;  // The frame or fragment given by 'frame_num' and
+                      // 'fragment_num'.
+
+  uint32_t pad[4];         // padding for later use.
+  void* private_;          // for internal use only.
+};
+
+// Retrieves frame 'frame_number' from 'dmux'.
+// 'iter->fragment' points to the first fragment on return from this function.
+// Individual fragments may be extracted using WebPDemuxSetFragment().
+// Setting 'frame_number' equal to 0 will return the last frame of the image.
+// Returns false if 'dmux' is NULL or frame 'frame_number' is not present.
+// Call WebPDemuxReleaseIterator() when use of the iterator is complete.
+// NOTE: 'dmux' must persist for the lifetime of 'iter'.
+WEBP_EXTERN(int) WebPDemuxGetFrame(
+    const WebPDemuxer* dmux, int frame_number, WebPIterator* iter);
+
+// Sets 'iter->fragment' to point to the next ('iter->frame_num' + 1) or
+// previous ('iter->frame_num' - 1) frame. These functions do not loop.
+// Returns true on success, false otherwise.
+WEBP_EXTERN(int) WebPDemuxNextFrame(WebPIterator* iter);
+WEBP_EXTERN(int) WebPDemuxPrevFrame(WebPIterator* iter);
+
+// Sets 'iter->fragment' to reflect fragment number 'fragment_num'.
+// Returns true if fragment 'fragment_num' is present, false otherwise.
+WEBP_EXTERN(int) WebPDemuxSelectFragment(WebPIterator* iter, int fragment_num);
+
+// Releases any memory associated with 'iter'.
+// Must be called before any subsequent calls to WebPDemuxGetChunk() on the same
+// iter. Also, must be called before destroying the associated WebPDemuxer with
+// WebPDemuxDelete().
+WEBP_EXTERN(void) WebPDemuxReleaseIterator(WebPIterator* iter);
+
+//------------------------------------------------------------------------------
+// Chunk iteration.
+
+struct WebPChunkIterator {
+  // The current and total number of chunks with the fourcc given to
+  // WebPDemuxGetChunk().
+  int chunk_num;
+  int num_chunks;
+  WebPData chunk;    // The payload of the chunk.
+
+  uint32_t pad[6];   // padding for later use
+  void* private_;
+};
+
+// Retrieves the 'chunk_number' instance of the chunk with id 'fourcc' from
+// 'dmux'.
+// 'fourcc' is a character array containing the fourcc of the chunk to return,
+// e.g., "ICCP", "XMP ", "EXIF", etc.
+// Setting 'chunk_number' equal to 0 will return the last chunk in a set.
+// Returns true if the chunk is found, false otherwise. Image related chunk
+// payloads are accessed through WebPDemuxGetFrame() and related functions.
+// Call WebPDemuxReleaseChunkIterator() when use of the iterator is complete.
+// NOTE: 'dmux' must persist for the lifetime of the iterator.
+WEBP_EXTERN(int) WebPDemuxGetChunk(const WebPDemuxer* dmux,
+                                   const char fourcc[4], int chunk_number,
+                                   WebPChunkIterator* iter);
+
+// Sets 'iter->chunk' to point to the next ('iter->chunk_num' + 1) or previous
+// ('iter->chunk_num' - 1) chunk. These functions do not loop.
+// Returns true on success, false otherwise.
+WEBP_EXTERN(int) WebPDemuxNextChunk(WebPChunkIterator* iter);
+WEBP_EXTERN(int) WebPDemuxPrevChunk(WebPChunkIterator* iter);
+
+// Releases any memory associated with 'iter'.
+// Must be called before destroying the associated WebPDemuxer with
+// WebPDemuxDelete().
+WEBP_EXTERN(void) WebPDemuxReleaseChunkIterator(WebPChunkIterator* iter);
+
+//------------------------------------------------------------------------------
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_DEMUX_H_ */
diff --git a/include/webp/encode.h b/include/webp/encode.h
index d87426bb..423622bd 100644
--- a/include/webp/encode.h
+++ b/include/webp/encode.h
@@ -18,7 +18,18 @@
 extern "C" {
 #endif
 
-#define WEBP_ENCODER_ABI_VERSION 0x0200    // MAJOR(8b) + MINOR(8b)
+#define WEBP_ENCODER_ABI_VERSION 0x0201    // MAJOR(8b) + MINOR(8b)
+
+#if !(defined(__cplusplus) || defined(c_plusplus))
+typedef enum WebPImageHint WebPImageHint;
+typedef enum WebPEncCSP WebPEncCSP;
+typedef enum WebPPreset WebPPreset;
+typedef enum WebPEncodingError WebPEncodingError;
+#endif
+typedef struct WebPConfig WebPConfig;
+typedef struct WebPPicture WebPPicture;   // main structure for I/O
+typedef struct WebPAuxStats WebPAuxStats;
+typedef struct WebPMemoryWriter WebPMemoryWriter;
 
 // Return the encoder's version number, packed in hexadecimal using 8bits for
 // each of major/minor/revision. E.g: v2.5.7 is 0x020507.
@@ -66,15 +77,16 @@ WEBP_EXTERN(size_t) WebPEncodeLosslessBGRA(const uint8_t* bgra,
 // Coding parameters
 
 // Image characteristics hint for the underlying encoder.
-typedef enum {
+enum WebPImageHint {
   WEBP_HINT_DEFAULT = 0,  // default preset.
   WEBP_HINT_PICTURE,      // digital picture, like portrait, inner shot
   WEBP_HINT_PHOTO,        // outdoor photograph, with natural lighting
   WEBP_HINT_GRAPH,        // Discrete tone image (graph, map-tile etc).
   WEBP_HINT_LAST
-} WebPImageHint;
+};
 
-typedef struct {
+// Compression parameters.
+struct WebPConfig {
   int lossless;           // Lossless encoding (0=lossy(default), 1=lossless).
   float quality;          // between 0 (smallest file) and 100 (biggest)
   int method;             // quality/speed trade-off (0=fast, 6=slower-better)
@@ -109,20 +121,26 @@ typedef struct {
   int partition_limit;    // quality degradation allowed to fit the 512k limit
                           // on prediction modes coding (0: no degradation,
                           // 100: maximum possible degradation).
-
-  uint32_t pad[8];        // padding for later use
-} WebPConfig;
+  int emulate_jpeg_size;  // If true, compression parameters will be remapped
+                          // to better match the expected output size from
+                          // JPEG compression. Generally, the output size will
+                          // be similar but the degradation will be lower.
+  int thread_level;       // If non-zero, try and use multi-threaded encoding.
+  int low_memory;         // If set, reduce memory usage (but increase CPU use).
+
+  uint32_t pad[5];        // padding for later use
+};
 
 // Enumerate some predefined settings for WebPConfig, depending on the type
 // of source picture. These presets are used when calling WebPConfigPreset().
-typedef enum {
+enum WebPPreset {
   WEBP_PRESET_DEFAULT = 0,  // default preset.
   WEBP_PRESET_PICTURE,      // digital picture, like portrait, inner shot
   WEBP_PRESET_PHOTO,        // outdoor photograph, with natural lighting
   WEBP_PRESET_DRAWING,      // hand or line drawing, with high-contrast details
   WEBP_PRESET_ICON,         // small-sized colorful images
   WEBP_PRESET_TEXT          // text-like
-} WebPPreset;
+};
 
 // Internal, version-checked, entry point
 WEBP_EXTERN(int) WebPConfigInitInternal(WebPConfig*, WebPPreset, float, int);
@@ -152,11 +170,9 @@ WEBP_EXTERN(int) WebPValidateConfig(const WebPConfig* config);
 
 //------------------------------------------------------------------------------
 // Input / Output
-
-typedef struct WebPPicture WebPPicture;   // main structure for I/O
-
 // Structure for storing auxiliary statistics (mostly for lossy encoding).
-typedef struct {
+
+struct WebPAuxStats {
   int coded_size;         // final size
 
   float PSNR[5];          // peak-signal-to-noise ratio for Y/U/V/All/Alpha
@@ -182,7 +198,7 @@ typedef struct {
   int lossless_size;           // final lossless size
 
   uint32_t pad[4];        // padding for later use
-} WebPAuxStats;
+};
 
 // Signature for output function. Should return true if writing was successful.
 // data/data_size is the segment of data to write, and 'picture' is for
@@ -192,18 +208,19 @@ typedef int (*WebPWriterFunction)(const uint8_t* data, size_t data_size,
 
 // WebPMemoryWrite: a special WebPWriterFunction that writes to memory using
 // the following WebPMemoryWriter object (to be set as a custom_ptr).
-typedef struct {
+struct WebPMemoryWriter {
   uint8_t* mem;       // final buffer (of size 'max_size', larger than 'size').
   size_t   size;      // final size
   size_t   max_size;  // total capacity
   uint32_t pad[1];    // padding for later use
-} WebPMemoryWriter;
+};
 
 // The following must be called first before any use.
 WEBP_EXTERN(void) WebPMemoryWriterInit(WebPMemoryWriter* writer);
 
 // The custom writer to be used with WebPMemoryWriter as custom_ptr. Upon
 // completion, writer.mem and writer.size will hold the coded data.
+// writer.mem must be freed using the call 'free(writer.mem)'.
 WEBP_EXTERN(int) WebPMemoryWrite(const uint8_t* data, size_t data_size,
                                  const WebPPicture* picture);
 
@@ -212,7 +229,8 @@ WEBP_EXTERN(int) WebPMemoryWrite(const uint8_t* data, size_t data_size,
 // everything is OK.
 typedef int (*WebPProgressHook)(int percent, const WebPPicture* picture);
 
-typedef enum {
+// Color spaces.
+enum WebPEncCSP {
   // chroma sampling
   WEBP_YUV420 = 0,   // 4:2:0
   WEBP_YUV422 = 1,   // 4:2:2
@@ -225,10 +243,10 @@ typedef enum {
   WEBP_YUV444A = 6,
   WEBP_YUV400A = 7,   // grayscale + alpha
   WEBP_CSP_ALPHA_BIT = 4   // bit that is set if alpha is present
-} WebPEncCSP;
+};
 
 // Encoding error conditions.
-typedef enum {
+enum WebPEncodingError {
   VP8_ENC_OK = 0,
   VP8_ENC_ERROR_OUT_OF_MEMORY,            // memory error allocating objects
   VP8_ENC_ERROR_BITSTREAM_OUT_OF_MEMORY,  // memory error while flushing bits
@@ -241,7 +259,7 @@ typedef enum {
   VP8_ENC_ERROR_FILE_TOO_BIG,             // file is bigger than 4G
   VP8_ENC_ERROR_USER_ABORT,               // abort request by user
   VP8_ENC_ERROR_LAST                      // list terminator. always last.
-} WebPEncodingError;
+};
 
 // maximum width/height allowed (inclusive), in pixels
 #define WEBP_MAX_DIMENSION 16383
@@ -350,10 +368,10 @@ WEBP_EXTERN(int) WebPPictureCopy(const WebPPicture* src, WebPPicture* dst);
 
 // Compute PSNR, SSIM or LSIM distortion metric between two pictures.
 // Result is in dB, stores in result[] in the Y/U/V/Alpha/All order.
-// Returns false in case of error (pic1 and pic2 don't have same dimension, ...)
+// Returns false in case of error (src and ref don't have same dimension, ...)
 // Warning: this function is rather CPU-intensive.
 WEBP_EXTERN(int) WebPPictureDistortion(
-    const WebPPicture* pic1, const WebPPicture* pic2,
+    const WebPPicture* src, const WebPPicture* ref,
     int metric_type,           // 0 = PSNR, 1 = SSIM, 2 = LSIM
     float result[5]);
 
diff --git a/include/webp/format_constants.h b/include/webp/format_constants.h
index 7ce498f6..2ac8fe67 100644
--- a/include/webp/format_constants.h
+++ b/include/webp/format_constants.h
@@ -65,23 +65,29 @@ typedef enum {
 #define CHUNK_SIZE_BYTES   4     // Size needed to store chunk's size.
 #define CHUNK_HEADER_SIZE  8     // Size of a chunk header.
 #define RIFF_HEADER_SIZE   12    // Size of the RIFF header ("RIFFnnnnWEBP").
-#define FRAME_CHUNK_SIZE   15    // Size of a FRM chunk.
-#define LOOP_CHUNK_SIZE    2     // Size of a LOOP chunk.
-#define TILE_CHUNK_SIZE    6     // Size of a TILE chunk.
+#define ANMF_CHUNK_SIZE    16    // Size of an ANMF chunk.
+#define ANIM_CHUNK_SIZE    6     // Size of an ANIM chunk.
+#define FRGM_CHUNK_SIZE    6     // Size of a FRGM chunk.
 #define VP8X_CHUNK_SIZE    10    // Size of a VP8X chunk.
 
-#define TILING_FLAG_BIT    0x01  // Set if tiles are possibly used.
-#define ANIMATION_FLAG_BIT 0x02  // Set if some animation is expected
-#define ICC_FLAG_BIT       0x04  // Whether ICC is present or not.
-#define METADATA_FLAG_BIT  0x08  // Set if some META chunk is possibly present.
-#define ALPHA_FLAG_BIT     0x10  // Should be same as the ALPHA_FLAG in mux.h
-#define ROTATION_FLAG_BITS 0xe0  // all 3 bits for rotation + symmetry
+// VP8X Feature Flags.
+#if !(defined(__cplusplus) || defined(c_plusplus))
+typedef enum WebPFeatureFlags WebPFeatureFlags;
+#endif
+enum WebPFeatureFlags {
+  FRAGMENTS_FLAG  = 0x00000001,
+  ANIMATION_FLAG  = 0x00000002,
+  XMP_FLAG        = 0x00000004,
+  EXIF_FLAG       = 0x00000008,
+  ALPHA_FLAG      = 0x00000010,
+  ICCP_FLAG       = 0x00000020
+};
 
 #define MAX_CANVAS_SIZE     (1 << 24)    // 24-bit max for VP8X width/height.
 #define MAX_IMAGE_AREA      (1ULL << 32) // 32-bit max for width x height.
 #define MAX_LOOP_COUNT      (1 << 16)    // maximum value for loop-count
 #define MAX_DURATION        (1 << 24)    // maximum duration
-#define MAX_POSITION_OFFSET (1 << 24)    // maximum frame/tile x/y offset
+#define MAX_POSITION_OFFSET (1 << 24)    // maximum frame/fragment x/y offset
 
 // Maximum chunk payload is such that adding the header and padding won't
 // overflow a uint32_t.
diff --git a/include/webp/mux_types.h b/include/webp/mux_types.h
new file mode 100644
index 00000000..84d94472
--- /dev/null
+++ b/include/webp/mux_types.h
@@ -0,0 +1,78 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// Internal header for some common data-types used by mux and demux libraries.
+//
+// Author: Urvang (urvang@google.com)
+
+#ifndef WEBP_WEBP_MUX_TYPES_H_
+#define WEBP_WEBP_MUX_TYPES_H_
+
+#include <stdlib.h>  // free()
+#include <string.h>  // memset()
+#include "./types.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+// Create fourcc of the chunk from the chunk tag characters.
+#define MKFOURCC(a, b, c, d) ((uint32_t)(a) | (b) << 8 | (c) << 16 | (d) << 24)
+
+// Dispose method (animation only). Indicates how the area used by the current
+// frame is to be treated before rendering the next frame on the canvas.
+#if !(defined(__cplusplus) || defined(c_plusplus))
+typedef enum WebPMuxAnimDispose WebPMuxAnimDispose;
+#endif
+enum WebPMuxAnimDispose {
+  WEBP_MUX_DISPOSE_NONE,       // Do not dispose.
+  WEBP_MUX_DISPOSE_BACKGROUND  // Dispose to background color.
+};
+
+// Data type used to describe 'raw' data, e.g., chunk data
+// (ICC profile, metadata) and WebP compressed image data.
+typedef struct WebPData WebPData;
+struct WebPData {
+  const uint8_t* bytes;
+  size_t size;
+};
+
+// Initializes the contents of the 'webp_data' object with default values.
+static WEBP_INLINE void WebPDataInit(WebPData* webp_data) {
+  if (webp_data != NULL) {
+    memset(webp_data, 0, sizeof(*webp_data));
+  }
+}
+
+// Clears the contents of the 'webp_data' object by calling free(). Does not
+// deallocate the object itself.
+static WEBP_INLINE void WebPDataClear(WebPData* webp_data) {
+  if (webp_data != NULL) {
+    free((void*)webp_data->bytes);
+    WebPDataInit(webp_data);
+  }
+}
+
+// Allocates necessary storage for 'dst' and copies the contents of 'src'.
+// Returns true on success.
+static WEBP_INLINE int WebPDataCopy(const WebPData* src, WebPData* dst) {
+  if (src == NULL || dst == NULL) return 0;
+  WebPDataInit(dst);
+  if (src->bytes != NULL && src->size != 0) {
+    dst->bytes = (uint8_t*)malloc(src->size);
+    if (dst->bytes == NULL) return 0;
+    memcpy((void*)dst->bytes, src->bytes, src->size);
+    dst->size = src->size;
+  }
+  return 1;
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_WEBP_MUX_TYPES_H_ */
diff --git a/src/dec/Android.mk b/src/dec/Android.mk
index ab795aee..34e188b0 100644
--- a/src/dec/Android.mk
+++ b/src/dec/Android.mk
@@ -28,27 +28,26 @@ LOCAL_SRC_FILES := \
         vp8l.c \
         webp.c \
         ../dsp/cpu.c \
+        ../dsp/cpu-features.c \
         ../dsp/dec.c \
         ../dsp/dec_neon.c \
         ../dsp/dec_sse2.c \
-        ../dsp/enc.c \
-        ../dsp/enc_sse2.c \
         ../dsp/lossless.c \
         ../dsp/upsampling.c \
+        ../dsp/upsampling_neon.c \
         ../dsp/upsampling_sse2.c \
         ../dsp/yuv.c \
+        ../demux/demux.c \
         ../utils/bit_reader.c \
-        ../utils/bit_writer.c \
         ../utils/color_cache.c \
         ../utils/filters.c \
         ../utils/huffman.c \
-        ../utils/huffman_encode.c \
-        ../utils/quant_levels.c \
+        ../utils/quant_levels_dec.c \
         ../utils/rescaler.c \
         ../utils/thread.c \
         ../utils/utils.c
 
-LOCAL_CFLAGS := -DANDROID
+LOCAL_CFLAGS := -DANDROID -DWEBP_SWAP_16BIT_CSP
 
 LOCAL_C_INCLUDES += \
         $(LOCAL_PATH) \
diff --git a/src/dec/alpha.c b/src/dec/alpha.c
index 3c19a892..f54b9f72 100644
--- a/src/dec/alpha.c
+++ b/src/dec/alpha.c
@@ -13,7 +13,7 @@
 #include "./vp8i.h"
 #include "./vp8li.h"
 #include "../utils/filters.h"
-#include "../utils/quant_levels.h"
+#include "../utils/quant_levels_dec.h"
 #include "webp/format_constants.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
diff --git a/src/dec/frame.c b/src/dec/frame.c
index 9c91a48e..911c7ffc 100644
--- a/src/dec/frame.c
+++ b/src/dec/frame.c
@@ -97,54 +97,51 @@ static void FilterRow(const VP8Decoder* const dec) {
 }
 
 //------------------------------------------------------------------------------
+// Precompute the filtering strength for each segment and each i4x4/i16x16 mode.
 
-void VP8StoreBlock(VP8Decoder* const dec) {
+static void PrecomputeFilterStrengths(VP8Decoder* const dec) {
   if (dec->filter_type_ > 0) {
-    VP8FInfo* const info = dec->f_info_ + dec->mb_x_;
-    const int skip = dec->mb_info_[dec->mb_x_].skip_;
-    int level = dec->filter_levels_[dec->segment_];
-    if (dec->filter_hdr_.use_lf_delta_) {
-      // TODO(skal): only CURRENT is handled for now.
-      level += dec->filter_hdr_.ref_lf_delta_[0];
-      if (dec->is_i4x4_) {
-        level += dec->filter_hdr_.mode_lf_delta_[0];
-      }
-    }
-    level = (level < 0) ? 0 : (level > 63) ? 63 : level;
-    info->f_level_ = level;
-
-    if (dec->filter_hdr_.sharpness_ > 0) {
-      if (dec->filter_hdr_.sharpness_ > 4) {
-        level >>= 2;
+    int s;
+    const VP8FilterHeader* const hdr = &dec->filter_hdr_;
+    for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+      int i4x4;
+      // First, compute the initial level
+      int base_level;
+      if (dec->segment_hdr_.use_segment_) {
+        base_level = dec->segment_hdr_.filter_strength_[s];
+        if (!dec->segment_hdr_.absolute_delta_) {
+          base_level += hdr->level_;
+        }
       } else {
-        level >>= 1;
+        base_level = hdr->level_;
       }
-      if (level > 9 - dec->filter_hdr_.sharpness_) {
-        level = 9 - dec->filter_hdr_.sharpness_;
+      for (i4x4 = 0; i4x4 <= 1; ++i4x4) {
+        VP8FInfo* const info = &dec->fstrengths_[s][i4x4];
+        int level = base_level;
+        if (hdr->use_lf_delta_) {
+          // TODO(skal): only CURRENT is handled for now.
+          level += hdr->ref_lf_delta_[0];
+          if (i4x4) {
+            level += hdr->mode_lf_delta_[0];
+          }
+        }
+        level = (level < 0) ? 0 : (level > 63) ? 63 : level;
+        info->f_level_ = level;
+
+        if (hdr->sharpness_ > 0) {
+          if (hdr->sharpness_ > 4) {
+            level >>= 2;
+          } else {
+            level >>= 1;
+          }
+          if (level > 9 - hdr->sharpness_) {
+            level = 9 - hdr->sharpness_;
+          }
+        }
+        info->f_ilevel_ = (level < 1) ? 1 : level;
+        info->f_inner_ = 0;
       }
     }
-
-    info->f_ilevel_ = (level < 1) ? 1 : level;
-    info->f_inner_ = (!skip || dec->is_i4x4_);
-  }
-  {
-    // Transfer samples to row cache
-    int y;
-    const int y_offset = dec->cache_id_ * 16 * dec->cache_y_stride_;
-    const int uv_offset = dec->cache_id_ * 8 * dec->cache_uv_stride_;
-    uint8_t* const ydst = dec->cache_y_ + dec->mb_x_ * 16 + y_offset;
-    uint8_t* const udst = dec->cache_u_ + dec->mb_x_ * 8 + uv_offset;
-    uint8_t* const vdst = dec->cache_v_ + dec->mb_x_ * 8 + uv_offset;
-    for (y = 0; y < 16; ++y) {
-      memcpy(ydst + y * dec->cache_y_stride_,
-             dec->yuv_b_ + Y_OFF + y * BPS, 16);
-    }
-    for (y = 0; y < 8; ++y) {
-      memcpy(udst + y * dec->cache_uv_stride_,
-           dec->yuv_b_ + U_OFF + y * BPS, 8);
-      memcpy(vdst + y * dec->cache_uv_stride_,
-           dec->yuv_b_ + V_OFF + y * BPS, 8);
-    }
   }
 }
 
@@ -339,6 +336,7 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io) {
       dec->br_mb_y_ = dec->mb_h_;
     }
   }
+  PrecomputeFilterStrengths(dec);
   return VP8_STATUS_OK;
 }
 
@@ -496,6 +494,7 @@ static int AllocateMemory(VP8Decoder* const dec) {
   // alpha plane
   dec->alpha_plane_ = alpha_size ? (uint8_t*)mem : NULL;
   mem += alpha_size;
+  assert(mem <= (uint8_t*)dec->mem_ + dec->mem_size_);
 
   // note: left-info is initialized once for all.
   memset(dec->mb_info_ - 1, 0, mb_info_size);
@@ -551,6 +550,7 @@ static WEBP_INLINE void Copy32b(uint8_t* dst, uint8_t* src) {
 }
 
 void VP8ReconstructBlock(VP8Decoder* const dec) {
+  int j;
   uint8_t* const y_dst = dec->yuv_b_ + Y_OFF;
   uint8_t* const u_dst = dec->yuv_b_ + U_OFF;
   uint8_t* const v_dst = dec->yuv_b_ + V_OFF;
@@ -558,7 +558,6 @@ void VP8ReconstructBlock(VP8Decoder* const dec) {
   // Rotate in the left samples from previously decoded block. We move four
   // pixels at a time for alignment reason, and because of in-loop filter.
   if (dec->mb_x_ > 0) {
-    int j;
     for (j = -1; j < 16; ++j) {
       Copy32b(&y_dst[j * BPS - 4], &y_dst[j * BPS + 12]);
     }
@@ -567,7 +566,6 @@ void VP8ReconstructBlock(VP8Decoder* const dec) {
       Copy32b(&v_dst[j * BPS - 4], &v_dst[j * BPS + 4]);
     }
   } else {
-    int j;
     for (j = 0; j < 16; ++j) {
       y_dst[j * BPS - 1] = 129;
     }
@@ -670,6 +668,21 @@ void VP8ReconstructBlock(VP8Decoder* const dec) {
       }
     }
   }
+  // Transfer reconstructed samples from yuv_b_ cache to final destination.
+  {
+    const int y_offset = dec->cache_id_ * 16 * dec->cache_y_stride_;
+    const int uv_offset = dec->cache_id_ * 8 * dec->cache_uv_stride_;
+    uint8_t* const y_out = dec->cache_y_ + dec->mb_x_ * 16 + y_offset;
+    uint8_t* const u_out = dec->cache_u_ + dec->mb_x_ * 8 + uv_offset;
+    uint8_t* const v_out = dec->cache_v_ + dec->mb_x_ * 8 + uv_offset;
+    for (j = 0; j < 16; ++j) {
+      memcpy(y_out + j * dec->cache_y_stride_, y_dst + j * BPS, 16);
+    }
+    for (j = 0; j < 8; ++j) {
+      memcpy(u_out + j * dec->cache_uv_stride_, u_dst + j * BPS, 8);
+      memcpy(v_out + j * dec->cache_uv_stride_, v_dst + j * BPS, 8);
+    }
+  }
 }
 
 //------------------------------------------------------------------------------
diff --git a/src/dec/idec.c b/src/dec/idec.c
index 7df790ce..17810c83 100644
--- a/src/dec/idec.c
+++ b/src/dec/idec.c
@@ -425,9 +425,8 @@ static VP8StatusCode DecodeRemaining(WebPIDecoder* const idec) {
         }
         return VP8_STATUS_SUSPENDED;
       }
+      // Reconstruct and emit samples.
       VP8ReconstructBlock(dec);
-      // Store data and save block's filtering params
-      VP8StoreBlock(dec);
 
       // Release buffer only if there is only one partition
       if (dec->num_parts_ == 1) {
@@ -596,12 +595,22 @@ void WebPIDelete(WebPIDecoder* idec) {
 
 WebPIDecoder* WebPINewRGB(WEBP_CSP_MODE mode, uint8_t* output_buffer,
                           size_t output_buffer_size, int output_stride) {
+  const int is_external_memory = (output_buffer != NULL);
   WebPIDecoder* idec;
+
   if (mode >= MODE_YUV) return NULL;
+  if (!is_external_memory) {    // Overwrite parameters to sane values.
+    output_buffer_size = 0;
+    output_stride = 0;
+  } else {  // A buffer was passed. Validate the other params.
+    if (output_stride == 0 || output_buffer_size == 0) {
+      return NULL;   // invalid parameter.
+    }
+  }
   idec = WebPINewDecoder(NULL);
   if (idec == NULL) return NULL;
   idec->output_.colorspace = mode;
-  idec->output_.is_external_memory = 1;
+  idec->output_.is_external_memory = is_external_memory;
   idec->output_.u.RGBA.rgba = output_buffer;
   idec->output_.u.RGBA.stride = output_stride;
   idec->output_.u.RGBA.size = output_buffer_size;
@@ -612,10 +621,30 @@ WebPIDecoder* WebPINewYUVA(uint8_t* luma, size_t luma_size, int luma_stride,
                            uint8_t* u, size_t u_size, int u_stride,
                            uint8_t* v, size_t v_size, int v_stride,
                            uint8_t* a, size_t a_size, int a_stride) {
-  WebPIDecoder* const idec = WebPINewDecoder(NULL);
+  const int is_external_memory = (luma != NULL);
+  WebPIDecoder* idec;
+  WEBP_CSP_MODE colorspace;
+
+  if (!is_external_memory) {    // Overwrite parameters to sane values.
+    luma_size = u_size = v_size = a_size = 0;
+    luma_stride = u_stride = v_stride = a_stride = 0;
+    u = v = a = NULL;
+    colorspace = MODE_YUVA;
+  } else {  // A luma buffer was passed. Validate the other parameters.
+    if (u == NULL || v == NULL) return NULL;
+    if (luma_size == 0 || u_size == 0 || v_size == 0) return NULL;
+    if (luma_stride == 0 || u_stride == 0 || v_stride == 0) return NULL;
+    if (a != NULL) {
+      if (a_size == 0 || a_stride == 0) return NULL;
+    }
+    colorspace = (a == NULL) ? MODE_YUV : MODE_YUVA;
+  }
+
+  idec = WebPINewDecoder(NULL);
   if (idec == NULL) return NULL;
-  idec->output_.colorspace = (a == NULL) ? MODE_YUV : MODE_YUVA;
-  idec->output_.is_external_memory = 1;
+
+  idec->output_.colorspace = colorspace;
+  idec->output_.is_external_memory = is_external_memory;
   idec->output_.u.YUVA.y = luma;
   idec->output_.u.YUVA.y_stride = luma_stride;
   idec->output_.u.YUVA.y_size = luma_size;
diff --git a/src/dec/io.c b/src/dec/io.c
index c5746f74..594804c2 100644
--- a/src/dec/io.c
+++ b/src/dec/io.c
@@ -111,7 +111,7 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
   const uint8_t* top_u = p->tmp_u;
   const uint8_t* top_v = p->tmp_v;
   int y = io->mb_y;
-  int y_end = io->mb_y + io->mb_h;
+  const int y_end = io->mb_y + io->mb_h;
   const int mb_w = io->mb_w;
   const int uv_w = (mb_w + 1) / 2;
 
@@ -150,7 +150,7 @@ static int EmitFancyRGB(const VP8Io* const io, WebPDecParams* const p) {
     // Process the very last row of even-sized picture
     if (!(y_end & 1)) {
       upsample(cur_y, NULL, cur_u, cur_v, cur_u, cur_v,
-              dst + buf->stride, NULL, mb_w);
+               dst + buf->stride, NULL, mb_w);
     }
   }
   return num_lines_out;
@@ -203,7 +203,7 @@ static int GetAlphaSourceRow(const VP8Io* const io,
       *alpha -= io->width;
     }
     if (io->crop_top + io->mb_y + io->mb_h == io->crop_bottom) {
-      // If it's the very last call, we process all the remaing rows!
+      // If it's the very last call, we process all the remaining rows!
       *num_rows = io->crop_bottom - io->crop_top - start_y;
     }
   }
@@ -214,32 +214,30 @@ static int EmitAlphaRGB(const VP8Io* const io, WebPDecParams* const p) {
   const uint8_t* alpha = io->a;
   if (alpha != NULL) {
     const int mb_w = io->mb_w;
-    int i, j;
     const WEBP_CSP_MODE colorspace = p->output->colorspace;
     const int alpha_first =
         (colorspace == MODE_ARGB || colorspace == MODE_Argb);
     const WebPRGBABuffer* const buf = &p->output->u.RGBA;
     int num_rows;
     const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
+    uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
+    uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
     uint32_t alpha_mask = 0xff;
+    int i, j;
 
-    {
-      uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
-      uint8_t* dst = base_rgba + (alpha_first ? 0 : 3);
-      for (j = 0; j < num_rows; ++j) {
-        for (i = 0; i < mb_w; ++i) {
-          const uint32_t alpha_value = alpha[i];
-          dst[4 * i] = alpha_value;
-          alpha_mask &= alpha_value;
-        }
-        alpha += io->width;
-        dst += buf->stride;
-      }
-      // alpha_mask is < 0xff if there's non-trivial alpha to premultiply with.
-      if (alpha_mask != 0xff && WebPIsPremultipliedMode(colorspace)) {
-        WebPApplyAlphaMultiply(base_rgba, alpha_first,
-                               mb_w, num_rows, buf->stride);
+    for (j = 0; j < num_rows; ++j) {
+      for (i = 0; i < mb_w; ++i) {
+        const uint32_t alpha_value = alpha[i];
+        dst[4 * i] = alpha_value;
+        alpha_mask &= alpha_value;
       }
+      alpha += io->width;
+      dst += buf->stride;
+    }
+    // alpha_mask is < 0xff if there's non-trivial alpha to premultiply with.
+    if (alpha_mask != 0xff && WebPIsPremultipliedMode(colorspace)) {
+      WebPApplyAlphaMultiply(base_rgba, alpha_first,
+                             mb_w, num_rows, buf->stride);
     }
   }
   return 0;
@@ -249,28 +247,27 @@ static int EmitAlphaRGBA4444(const VP8Io* const io, WebPDecParams* const p) {
   const uint8_t* alpha = io->a;
   if (alpha != NULL) {
     const int mb_w = io->mb_w;
-    int i, j;
+    const WEBP_CSP_MODE colorspace = p->output->colorspace;
     const WebPRGBABuffer* const buf = &p->output->u.RGBA;
     int num_rows;
     const int start_y = GetAlphaSourceRow(io, &alpha, &num_rows);
+    uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
+    uint8_t* alpha_dst = base_rgba + 1;
     uint32_t alpha_mask = 0x0f;
+    int i, j;
 
-    {
-      uint8_t* const base_rgba = buf->rgba + start_y * buf->stride;
-      uint8_t* alpha_dst = base_rgba + 1;
-      for (j = 0; j < num_rows; ++j) {
-        for (i = 0; i < mb_w; ++i) {
-          // Fill in the alpha value (converted to 4 bits).
-          const uint32_t alpha_value = alpha[i] >> 4;
-          alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_value;
-          alpha_mask &= alpha_value;
-        }
-        alpha += io->width;
-        alpha_dst += buf->stride;
-      }
-      if (alpha_mask != 0x0f && p->output->colorspace == MODE_rgbA_4444) {
-        WebPApplyAlphaMultiply4444(base_rgba, mb_w, num_rows, buf->stride);
+    for (j = 0; j < num_rows; ++j) {
+      for (i = 0; i < mb_w; ++i) {
+        // Fill in the alpha value (converted to 4 bits).
+        const uint32_t alpha_value = alpha[i] >> 4;
+        alpha_dst[2 * i] = (alpha_dst[2 * i] & 0xf0) | alpha_value;
+        alpha_mask &= alpha_value;
       }
+      alpha += io->width;
+      alpha_dst += buf->stride;
+    }
+    if (alpha_mask != 0x0f && WebPIsPremultipliedMode(colorspace)) {
+      WebPApplyAlphaMultiply4444(base_rgba, mb_w, num_rows, buf->stride);
     }
   }
   return 0;
@@ -497,8 +494,7 @@ static int InitRGBRescaler(const VP8Io* const io, WebPDecParams* const p) {
     tmp_size1 += work_size;
     tmp_size2 += out_width;
   }
-  p->memory =
-      calloc(1, tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp));
+  p->memory = calloc(1, tmp_size1 * sizeof(*work) + tmp_size2 * sizeof(*tmp));
   if (p->memory == NULL) {
     return 0;   // memory error
   }
@@ -595,7 +591,7 @@ static int CustomSetup(VP8Io* io) {
 //------------------------------------------------------------------------------
 
 static int CustomPut(const VP8Io* io) {
-  WebPDecParams* p = (WebPDecParams*)io->opaque;
+  WebPDecParams* const p = (WebPDecParams*)io->opaque;
   const int mb_w = io->mb_w;
   const int mb_h = io->mb_h;
   int num_lines_out;
diff --git a/src/dec/vp8.c b/src/dec/vp8.c
index b0ccfa2a..253cb6b6 100644
--- a/src/dec/vp8.c
+++ b/src/dec/vp8.c
@@ -236,20 +236,6 @@ static int ParseFilterHeader(VP8BitReader* br, VP8Decoder* const dec) {
     }
   }
   dec->filter_type_ = (hdr->level_ == 0) ? 0 : hdr->simple_ ? 1 : 2;
-  if (dec->filter_type_ > 0) {    // precompute filter levels per segment
-    if (dec->segment_hdr_.use_segment_) {
-      int s;
-      for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
-        int strength = dec->segment_hdr_.filter_strength_[s];
-        if (!dec->segment_hdr_.absolute_delta_) {
-          strength += hdr->level_;
-        }
-        dec->filter_levels_[s] = strength;
-      }
-    } else {
-      dec->filter_levels_[0] = hdr->level_;
-    }
-  }
   return !br->eof_;
 }
 
@@ -458,7 +444,7 @@ int VP8GetHeaders(VP8Decoder* const dec, VP8Io* const io) {
 //------------------------------------------------------------------------------
 // Residual decoding (Paragraph 13.2 / 13.3)
 
-static const uint8_t kBands[16 + 1] = {
+static const int kBands[16 + 1] = {
   0, 1, 2, 3, 6, 4, 5, 6, 6, 6, 6, 6, 6, 6, 6, 7,
   0  // extra entry as sentinel
 };
@@ -474,6 +460,39 @@ static const uint8_t kZigzag[16] = {
 };
 
 typedef const uint8_t (*ProbaArray)[NUM_CTX][NUM_PROBAS];  // for const-casting
+typedef const uint8_t (*ProbaCtxArray)[NUM_PROBAS];
+
+// See section 13-2: http://tools.ietf.org/html/rfc6386#section-13.2
+static int GetLargeValue(VP8BitReader* const br, const uint8_t* const p) {
+  int v;
+  if (!VP8GetBit(br, p[3])) {
+    if (!VP8GetBit(br, p[4])) {
+      v = 2;
+    } else {
+      v = 3 + VP8GetBit(br, p[5]);
+    }
+  } else {
+    if (!VP8GetBit(br, p[6])) {
+      if (!VP8GetBit(br, p[7])) {
+        v = 5 + VP8GetBit(br, 159);
+      } else {
+        v = 7 + 2 * VP8GetBit(br, 165);
+        v += VP8GetBit(br, 145);
+      }
+    } else {
+      const uint8_t* tab;
+      const int bit1 = VP8GetBit(br, p[8]);
+      const int bit0 = VP8GetBit(br, p[9 + bit1]);
+      const int cat = 2 * bit1 + bit0;
+      v = 0;
+      for (tab = kCat3456[cat]; *tab; ++tab) {
+        v += v + VP8GetBit(br, *tab);
+      }
+      v += 3 + (8 << cat);
+    }
+  }
+  return v;
+}
 
 // Returns the position of the last non-zero coeff plus one
 // (and 0 if there's no coeff at all)
@@ -484,54 +503,26 @@ static int GetCoeffs(VP8BitReader* const br, ProbaArray prob,
   if (!VP8GetBit(br, p[0])) {   // first EOB is more a 'CBP' bit.
     return 0;
   }
-  while (1) {
-    ++n;
+  for (; n < 16; ++n) {
+    const ProbaCtxArray p_ctx = prob[kBands[n + 1]];
     if (!VP8GetBit(br, p[1])) {
-      p = prob[kBands[n]][0];
+      p = p_ctx[0];
     } else {  // non zero coeff
-      int v, j;
+      int v;
       if (!VP8GetBit(br, p[2])) {
-        p = prob[kBands[n]][1];
         v = 1;
+        p = p_ctx[1];
       } else {
-        if (!VP8GetBit(br, p[3])) {
-          if (!VP8GetBit(br, p[4])) {
-            v = 2;
-          } else {
-            v = 3 + VP8GetBit(br, p[5]);
-          }
-        } else {
-          if (!VP8GetBit(br, p[6])) {
-            if (!VP8GetBit(br, p[7])) {
-              v = 5 + VP8GetBit(br, 159);
-            } else {
-              v = 7 + 2 * VP8GetBit(br, 165);
-              v += VP8GetBit(br, 145);
-            }
-          } else {
-            const uint8_t* tab;
-            const int bit1 = VP8GetBit(br, p[8]);
-            const int bit0 = VP8GetBit(br, p[9 + bit1]);
-            const int cat = 2 * bit1 + bit0;
-            v = 0;
-            for (tab = kCat3456[cat]; *tab; ++tab) {
-              v += v + VP8GetBit(br, *tab);
-            }
-            v += 3 + (8 << cat);
-          }
-        }
-        p = prob[kBands[n]][2];
+        v = GetLargeValue(br, p);
+        p = p_ctx[2];
       }
-      j = kZigzag[n - 1];
-      out[j] = VP8GetSigned(br, v) * dq[j > 0];
-      if (n == 16 || !VP8GetBit(br, p[0])) {   // EOB
-        return n;
+      out[kZigzag[n]] = VP8GetSigned(br, v) * dq[n > 0];
+      if (n < 15 && !VP8GetBit(br, p[0])) {   // EOB
+        return n + 1;
       }
     }
-    if (n == 16) {
-      return 16;
-    }
   }
+  return 16;
 }
 
 // Alias-safe way of converting 4bytes to 32bits.
@@ -670,6 +661,12 @@ int VP8DecodeMB(VP8Decoder* const dec, VP8BitReader* const token_br) {
     dec->non_zero_ac_ = 0;
   }
 
+  if (dec->filter_type_ > 0) {  // store filter info
+    VP8FInfo* const finfo = dec->f_info_ + dec->mb_x_;
+    *finfo = dec->fstrengths_[dec->segment_][dec->is_i4x4_];
+    finfo->f_inner_ = (!info->skip_ || dec->is_i4x4_);
+  }
+
   return (!token_br->eof_);
 }
 
@@ -693,10 +690,8 @@ static int ParseFrame(VP8Decoder* const dec, VP8Io* io) {
         return VP8SetError(dec, VP8_STATUS_NOT_ENOUGH_DATA,
                            "Premature end-of-file encountered.");
       }
+      // Reconstruct and emit samples.
       VP8ReconstructBlock(dec);
-
-      // Store data and save block's filtering params
-      VP8StoreBlock(dec);
     }
     if (!VP8ProcessRow(dec, io)) {
       return VP8SetError(dec, VP8_STATUS_USER_ABORT, "Output aborted.");
diff --git a/src/dec/vp8i.h b/src/dec/vp8i.h
index 4382edfd..4f5192e2 100644
--- a/src/dec/vp8i.h
+++ b/src/dec/vp8i.h
@@ -28,7 +28,7 @@ extern "C" {
 // version numbers
 #define DEC_MAJ_VERSION 0
 #define DEC_MIN_VERSION 2
-#define DEC_REV_VERSION 0
+#define DEC_REV_VERSION 1
 
 #define ONLY_KEYFRAME_CODE      // to remove any code related to P-Frames
 
@@ -157,7 +157,7 @@ typedef struct {  // filter specs
 } VP8FInfo;
 
 typedef struct {  // used for syntax-parsing
-  unsigned int nz_;          // non-zero AC/DC coeffs
+  unsigned int nz_:24;       // non-zero AC/DC coeffs (24bit)
   unsigned int dc_nz_:1;     // non-zero DC coeffs
   unsigned int skip_:1;      // block type
 } VP8MB;
@@ -269,9 +269,9 @@ struct VP8Decoder {
   uint32_t non_zero_ac_;
 
   // Filtering side-info
-  int filter_type_;                         // 0=off, 1=simple, 2=complex
-  int filter_row_;                          // per-row flag
-  uint8_t filter_levels_[NUM_MB_SEGMENTS];  // precalculated per-segment
+  int filter_type_;                          // 0=off, 1=simple, 2=complex
+  int filter_row_;                           // per-row flag
+  VP8FInfo fstrengths_[NUM_MB_SEGMENTS][2];  // precalculated per-segment/type
 
   // extensions
   const uint8_t* alpha_data_;   // compressed alpha data (if present)
@@ -312,8 +312,6 @@ VP8StatusCode VP8EnterCritical(VP8Decoder* const dec, VP8Io* const io);
 int VP8ExitCritical(VP8Decoder* const dec, VP8Io* const io);
 // Process the last decoded row (filtering + output)
 int VP8ProcessRow(VP8Decoder* const dec, VP8Io* const io);
-// Store a block, along with filtering params
-void VP8StoreBlock(VP8Decoder* const dec);
 // To be called at the start of a new scanline, to initialize predictors.
 void VP8InitScanline(VP8Decoder* const dec);
 // Decode one macroblock. Returns false if there is not enough data.
diff --git a/src/dec/vp8l.c b/src/dec/vp8l.c
index 897e4395..5659cb5d 100644
--- a/src/dec/vp8l.c
+++ b/src/dec/vp8l.c
@@ -149,31 +149,22 @@ static WEBP_INLINE int PlaneCodeToDistance(int xsize, int plane_code) {
 //------------------------------------------------------------------------------
 // Decodes the next Huffman code from bit-stream.
 // FillBitWindow(br) needs to be called at minimum every second call
-// to ReadSymbolUnsafe.
-static int ReadSymbolUnsafe(const HuffmanTree* tree, VP8LBitReader* const br) {
+// to ReadSymbol, in order to pre-fetch enough bits.
+static WEBP_INLINE int ReadSymbol(const HuffmanTree* tree,
+                                  VP8LBitReader* const br) {
   const HuffmanTreeNode* node = tree->root_;
+  int num_bits = 0;
+  uint32_t bits = VP8LPrefetchBits(br);
   assert(node != NULL);
   while (!HuffmanTreeNodeIsLeaf(node)) {
-    node = HuffmanTreeNextNode(node, VP8LReadOneBitUnsafe(br));
+    node = HuffmanTreeNextNode(node, bits & 1);
+    bits >>= 1;
+    ++num_bits;
   }
+  VP8LDiscardBits(br, num_bits);
   return node->symbol_;
 }
 
-static WEBP_INLINE int ReadSymbol(const HuffmanTree* tree,
-                                  VP8LBitReader* const br) {
-  const int read_safe = (br->pos_ + 8 > br->len_);
-  if (!read_safe) {
-    return ReadSymbolUnsafe(tree, br);
-  } else {
-    const HuffmanTreeNode* node = tree->root_;
-    assert(node != NULL);
-    while (!HuffmanTreeNodeIsLeaf(node)) {
-      node = HuffmanTreeNextNode(node, VP8LReadOneBit(br));
-    }
-    return node->symbol_;
-  }
-}
-
 static int ReadHuffmanCodeLengths(
     VP8LDecoder* const dec, const int* const code_length_code_lengths,
     int num_symbols, int* const code_lengths) {
@@ -327,10 +318,10 @@ static int ReadHuffmanCodes(VP8LDecoder* const dec, int xsize, int ysize,
     hdr->huffman_subsample_bits_ = huffman_precision;
     for (i = 0; i < huffman_pixs; ++i) {
       // The huffman data is stored in red and green bytes.
-      const int index = (huffman_image[i] >> 8) & 0xffff;
-      huffman_image[i] = index;
-      if (index >= num_htree_groups) {
-        num_htree_groups = index + 1;
+      const int group = (huffman_image[i] >> 8) & 0xffff;
+      huffman_image[i] = group;
+      if (group >= num_htree_groups) {
+        num_htree_groups = group + 1;
       }
     }
   }
diff --git a/src/dec/webp.c b/src/dec/webp.c
index 1edf6d97..0186fa18 100644
--- a/src/dec/webp.c
+++ b/src/dec/webp.c
@@ -40,8 +40,8 @@ extern "C" {
 //   20..23  VP8X flags bit-map corresponding to the chunk-types present.
 //   24..26  Width of the Canvas Image.
 //   27..29  Height of the Canvas Image.
-// There can be extra chunks after the "VP8X" chunk (ICCP, TILE, FRM, VP8,
-// META  ...)
+// There can be extra chunks after the "VP8X" chunk (ICCP, FRGM, ANMF, VP8,
+// VP8L, XMP, EXIF  ...)
 // All sizes are in little-endian order.
 // Note: chunk data size must be padded to multiple of 2 when written.
 
@@ -308,7 +308,7 @@ static VP8StatusCode ParseHeadersInternal(const uint8_t* data,
       // necessary to send VP8X chunk to the decoder.
       return VP8_STATUS_BITSTREAM_ERROR;
     }
-    if (has_alpha != NULL) *has_alpha = !!(flags & ALPHA_FLAG_BIT);
+    if (has_alpha != NULL) *has_alpha = !!(flags & ALPHA_FLAG);
     if (found_vp8x && headers == NULL) {
       return VP8_STATUS_OK;  // Return features from VP8X header.
     }
@@ -672,19 +672,13 @@ int WebPInitDecoderConfigInternal(WebPDecoderConfig* config,
 VP8StatusCode WebPGetFeaturesInternal(const uint8_t* data, size_t data_size,
                                       WebPBitstreamFeatures* features,
                                       int version) {
-  VP8StatusCode status;
   if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DECODER_ABI_VERSION)) {
     return VP8_STATUS_INVALID_PARAM;   // version mismatch
   }
   if (features == NULL) {
     return VP8_STATUS_INVALID_PARAM;
   }
-
-  status = GetFeatures(data, data_size, features);
-  if (status == VP8_STATUS_NOT_ENOUGH_DATA) {
-    return VP8_STATUS_BITSTREAM_ERROR;  // Not-enough-data treated as error.
-  }
-  return status;
+  return GetFeatures(data, data_size, features);
 }
 
 VP8StatusCode WebPDecode(const uint8_t* data, size_t data_size,
diff --git a/src/demux/demux.c b/src/demux/demux.c
new file mode 100644
index 00000000..60c216ac
--- /dev/null
+++ b/src/demux/demux.c
@@ -0,0 +1,935 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+//  WebP container demux.
+//
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "../utils/utils.h"
+#include "webp/decode.h"     // WebPGetFeatures
+#include "webp/demux.h"
+#include "webp/format_constants.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+#define DMUX_MAJ_VERSION 0
+#define DMUX_MIN_VERSION 1
+#define DMUX_REV_VERSION 0
+
+typedef struct {
+  size_t start_;        // start location of the data
+  size_t end_;          // end location
+  size_t riff_end_;     // riff chunk end location, can be > end_.
+  size_t buf_size_;     // size of the buffer
+  const uint8_t* buf_;
+} MemBuffer;
+
+typedef struct {
+  size_t offset_;
+  size_t size_;
+} ChunkData;
+
+typedef struct Frame {
+  int x_offset_, y_offset_;
+  int width_, height_;
+  int duration_;
+  WebPMuxAnimDispose dispose_method_;
+  int is_fragment_;  // this is a frame fragment (and not a full frame).
+  int frame_num_;  // the referent frame number for use in assembling fragments.
+  int complete_;   // img_components_ contains a full image.
+  ChunkData img_components_[2];  // 0=VP8{,L} 1=ALPH
+  struct Frame* next_;
+} Frame;
+
+typedef struct Chunk {
+  ChunkData data_;
+  struct Chunk* next_;
+} Chunk;
+
+struct WebPDemuxer {
+  MemBuffer mem_;
+  WebPDemuxState state_;
+  int is_ext_format_;
+  uint32_t feature_flags_;
+  int canvas_width_, canvas_height_;
+  int loop_count_;
+  uint32_t bgcolor_;
+  int num_frames_;
+  Frame* frames_;
+  Chunk* chunks_;  // non-image chunks
+};
+
+typedef enum {
+  PARSE_OK,
+  PARSE_NEED_MORE_DATA,
+  PARSE_ERROR
+} ParseStatus;
+
+typedef struct ChunkParser {
+  uint8_t id[4];
+  ParseStatus (*parse)(WebPDemuxer* const dmux);
+  int (*valid)(const WebPDemuxer* const dmux);
+} ChunkParser;
+
+static ParseStatus ParseSingleImage(WebPDemuxer* const dmux);
+static ParseStatus ParseVP8X(WebPDemuxer* const dmux);
+static int IsValidSimpleFormat(const WebPDemuxer* const dmux);
+static int IsValidExtendedFormat(const WebPDemuxer* const dmux);
+
+static const ChunkParser kMasterChunks[] = {
+  { { 'V', 'P', '8', ' ' }, ParseSingleImage, IsValidSimpleFormat },
+  { { 'V', 'P', '8', 'L' }, ParseSingleImage, IsValidSimpleFormat },
+  { { 'V', 'P', '8', 'X' }, ParseVP8X,        IsValidExtendedFormat },
+  { { '0', '0', '0', '0' }, NULL,             NULL },
+};
+
+//------------------------------------------------------------------------------
+
+int WebPGetDemuxVersion(void) {
+  return (DMUX_MAJ_VERSION << 16) | (DMUX_MIN_VERSION << 8) | DMUX_REV_VERSION;
+}
+
+// -----------------------------------------------------------------------------
+// MemBuffer
+
+static int RemapMemBuffer(MemBuffer* const mem,
+                          const uint8_t* data, size_t size) {
+  if (size < mem->buf_size_) return 0;  // can't remap to a shorter buffer!
+
+  mem->buf_ = data;
+  mem->end_ = mem->buf_size_ = size;
+  return 1;
+}
+
+static int InitMemBuffer(MemBuffer* const mem,
+                         const uint8_t* data, size_t size) {
+  memset(mem, 0, sizeof(*mem));
+  return RemapMemBuffer(mem, data, size);
+}
+
+// Return the remaining data size available in 'mem'.
+static WEBP_INLINE size_t MemDataSize(const MemBuffer* const mem) {
+  return (mem->end_ - mem->start_);
+}
+
+// Return true if 'size' exceeds the end of the RIFF chunk.
+static WEBP_INLINE int SizeIsInvalid(const MemBuffer* const mem, size_t size) {
+  return (size > mem->riff_end_ - mem->start_);
+}
+
+static WEBP_INLINE void Skip(MemBuffer* const mem, size_t size) {
+  mem->start_ += size;
+}
+
+static WEBP_INLINE void Rewind(MemBuffer* const mem, size_t size) {
+  mem->start_ -= size;
+}
+
+static WEBP_INLINE const uint8_t* GetBuffer(MemBuffer* const mem) {
+  return mem->buf_ + mem->start_;
+}
+
+// Read from 'mem' and skip the read bytes.
+static WEBP_INLINE uint8_t ReadByte(MemBuffer* const mem) {
+  const uint8_t byte = mem->buf_[mem->start_];
+  Skip(mem, 1);
+  return byte;
+}
+
+static WEBP_INLINE int ReadLE16s(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const int val = GetLE16(data);
+  Skip(mem, 2);
+  return val;
+}
+
+static WEBP_INLINE int ReadLE24s(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const int val = GetLE24(data);
+  Skip(mem, 3);
+  return val;
+}
+
+static WEBP_INLINE uint32_t ReadLE32(MemBuffer* const mem) {
+  const uint8_t* const data = mem->buf_ + mem->start_;
+  const uint32_t val = GetLE32(data);
+  Skip(mem, 4);
+  return val;
+}
+
+// -----------------------------------------------------------------------------
+// Secondary chunk parsing
+
+static void AddChunk(WebPDemuxer* const dmux, Chunk* const chunk) {
+  Chunk** c = &dmux->chunks_;
+  while (*c != NULL) c = &(*c)->next_;
+  *c = chunk;
+  chunk->next_ = NULL;
+}
+
+// Add a frame to the end of the list, ensuring the last frame is complete.
+// Returns true on success, false otherwise.
+static int AddFrame(WebPDemuxer* const dmux, Frame* const frame) {
+  const Frame* last_frame = NULL;
+  Frame** f = &dmux->frames_;
+  while (*f != NULL) {
+    last_frame = *f;
+    f = &(*f)->next_;
+  }
+  if (last_frame != NULL && !last_frame->complete_) return 0;
+  *f = frame;
+  frame->next_ = NULL;
+  return 1;
+}
+
+// Store image bearing chunks to 'frame'.
+// If 'has_vp8l_alpha' is not NULL, it will be set to true if the frame is a
+// lossless image with alpha.
+static ParseStatus StoreFrame(int frame_num, uint32_t min_size,
+                              MemBuffer* const mem, Frame* const frame,
+                              int* const has_vp8l_alpha) {
+  int alpha_chunks = 0;
+  int image_chunks = 0;
+  int done = (MemDataSize(mem) < min_size);
+  ParseStatus status = PARSE_OK;
+
+  if (has_vp8l_alpha != NULL) *has_vp8l_alpha = 0;  // Default.
+
+  if (done) return PARSE_NEED_MORE_DATA;
+
+  do {
+    const size_t chunk_start_offset = mem->start_;
+    const uint32_t fourcc = ReadLE32(mem);
+    const uint32_t payload_size = ReadLE32(mem);
+    const uint32_t payload_size_padded = payload_size + (payload_size & 1);
+    const size_t payload_available = (payload_size_padded > MemDataSize(mem))
+                                   ? MemDataSize(mem) : payload_size_padded;
+    const size_t chunk_size = CHUNK_HEADER_SIZE + payload_available;
+
+    if (payload_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+    if (SizeIsInvalid(mem, payload_size_padded)) return PARSE_ERROR;
+    if (payload_size_padded > MemDataSize(mem)) status = PARSE_NEED_MORE_DATA;
+
+    switch (fourcc) {
+      case MKFOURCC('A', 'L', 'P', 'H'):
+        if (alpha_chunks == 0) {
+          ++alpha_chunks;
+          frame->img_components_[1].offset_ = chunk_start_offset;
+          frame->img_components_[1].size_ = chunk_size;
+          frame->frame_num_ = frame_num;
+          Skip(mem, payload_available);
+        } else {
+          goto Done;
+        }
+        break;
+      case MKFOURCC('V', 'P', '8', ' '):
+      case MKFOURCC('V', 'P', '8', 'L'):
+        if (image_chunks == 0) {
+          // Extract the bitstream features, tolerating failures when the data
+          // is incomplete.
+          WebPBitstreamFeatures features;
+          const VP8StatusCode vp8_status =
+              WebPGetFeatures(mem->buf_ + chunk_start_offset, chunk_size,
+                              &features);
+          if (status == PARSE_NEED_MORE_DATA &&
+              vp8_status == VP8_STATUS_NOT_ENOUGH_DATA) {
+            return PARSE_NEED_MORE_DATA;
+          } else if (vp8_status != VP8_STATUS_OK) {
+            // We have enough data, and yet WebPGetFeatures() failed.
+            return PARSE_ERROR;
+          }
+          ++image_chunks;
+          frame->img_components_[0].offset_ = chunk_start_offset;
+          frame->img_components_[0].size_ = chunk_size;
+          frame->width_ = features.width;
+          frame->height_ = features.height;
+          if (has_vp8l_alpha != NULL) *has_vp8l_alpha = features.has_alpha;
+          frame->frame_num_ = frame_num;
+          frame->complete_ = (status == PARSE_OK);
+          Skip(mem, payload_available);
+        } else {
+          goto Done;
+        }
+        break;
+ Done:
+      default:
+        // Restore fourcc/size when moving up one level in parsing.
+        Rewind(mem, CHUNK_HEADER_SIZE);
+        done = 1;
+        break;
+    }
+
+    if (mem->start_ == mem->riff_end_) {
+      done = 1;
+    } else if (MemDataSize(mem) < CHUNK_HEADER_SIZE) {
+      status = PARSE_NEED_MORE_DATA;
+    }
+  } while (!done && status == PARSE_OK);
+
+  return status;
+}
+
+// Creates a new Frame if 'actual_size' is within bounds and 'mem' contains
+// enough data ('min_size') to parse the payload.
+// Returns PARSE_OK on success with *frame pointing to the new Frame.
+// Returns PARSE_NEED_MORE_DATA with insufficient data, PARSE_ERROR otherwise.
+static ParseStatus NewFrame(const MemBuffer* const mem,
+                            uint32_t min_size, uint32_t actual_size,
+                            Frame** frame) {
+  if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
+  if (actual_size < min_size) return PARSE_ERROR;
+  if (MemDataSize(mem) < min_size)  return PARSE_NEED_MORE_DATA;
+
+  *frame = (Frame*)calloc(1, sizeof(**frame));
+  return (*frame == NULL) ? PARSE_ERROR : PARSE_OK;
+}
+
+// Parse a 'ANMF' chunk and any image bearing chunks that immediately follow.
+// 'frame_chunk_size' is the previously validated, padded chunk size.
+static ParseStatus ParseFrame(
+    WebPDemuxer* const dmux, uint32_t frame_chunk_size) {
+  const int has_frames = !!(dmux->feature_flags_ & ANIMATION_FLAG);
+  const uint32_t anmf_payload_size = frame_chunk_size - ANMF_CHUNK_SIZE;
+  int added_frame = 0;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status =
+      NewFrame(mem, ANMF_CHUNK_SIZE, frame_chunk_size, &frame);
+  if (status != PARSE_OK) return status;
+
+  frame->x_offset_       = 2 * ReadLE24s(mem);
+  frame->y_offset_       = 2 * ReadLE24s(mem);
+  frame->width_          = 1 + ReadLE24s(mem);
+  frame->height_         = 1 + ReadLE24s(mem);
+  frame->duration_       = ReadLE24s(mem);
+  frame->dispose_method_ = (WebPMuxAnimDispose)(ReadByte(mem) & 1);
+  if (frame->width_ * (uint64_t)frame->height_ >= MAX_IMAGE_AREA) {
+    return PARSE_ERROR;
+  }
+
+  // Store a frame only if the animation flag is set and all data for this frame
+  // is available.
+  status = StoreFrame(dmux->num_frames_ + 1, anmf_payload_size, mem, frame,
+                      NULL);
+  if (status != PARSE_ERROR && has_frames && frame->frame_num_ > 0) {
+    added_frame = AddFrame(dmux, frame);
+    if (added_frame) {
+      ++dmux->num_frames_;
+    } else {
+      status = PARSE_ERROR;
+    }
+  }
+
+  if (!added_frame) free(frame);
+  return status;
+}
+
+// Parse a 'FRGM' chunk and any image bearing chunks that immediately follow.
+// 'fragment_chunk_size' is the previously validated, padded chunk size.
+static ParseStatus ParseFragment(WebPDemuxer* const dmux,
+                                 uint32_t fragment_chunk_size) {
+  const int frame_num = 1;  // All fragments belong to the 1st (and only) frame.
+  const int has_fragments = !!(dmux->feature_flags_ & FRAGMENTS_FLAG);
+  const uint32_t frgm_payload_size = fragment_chunk_size - FRGM_CHUNK_SIZE;
+  int added_fragment = 0;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status =
+      NewFrame(mem, FRGM_CHUNK_SIZE, fragment_chunk_size, &frame);
+  if (status != PARSE_OK) return status;
+
+  frame->is_fragment_  = 1;
+  frame->x_offset_ = 2 * ReadLE24s(mem);
+  frame->y_offset_ = 2 * ReadLE24s(mem);
+
+  // Store a fragment only if the fragments flag is set and all data for this
+  // fragment is available.
+  status = StoreFrame(frame_num, frgm_payload_size, mem, frame, NULL);
+  if (status != PARSE_ERROR && has_fragments && frame->frame_num_ > 0) {
+    added_fragment = AddFrame(dmux, frame);
+    if (!added_fragment) {
+      status = PARSE_ERROR;
+    } else {
+      dmux->num_frames_ = 1;
+    }
+  }
+
+  if (!added_fragment) free(frame);
+  return status;
+}
+
+// General chunk storage starting with the header at 'start_offset' allowing
+// the user to request the payload via a fourcc string. 'size' includes the
+// header and the unpadded payload size.
+// Returns true on success, false otherwise.
+static int StoreChunk(WebPDemuxer* const dmux,
+                      size_t start_offset, uint32_t size) {
+  Chunk* const chunk = (Chunk*)calloc(1, sizeof(*chunk));
+  if (chunk == NULL) return 0;
+
+  chunk->data_.offset_ = start_offset;
+  chunk->data_.size_ = size;
+  AddChunk(dmux, chunk);
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// Primary chunk parsing
+
+static int ReadHeader(MemBuffer* const mem) {
+  const size_t min_size = RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE;
+  uint32_t riff_size;
+
+  // Basic file level validation.
+  if (MemDataSize(mem) < min_size) return 0;
+  if (memcmp(GetBuffer(mem), "RIFF", CHUNK_SIZE_BYTES) ||
+      memcmp(GetBuffer(mem) + CHUNK_HEADER_SIZE, "WEBP", CHUNK_SIZE_BYTES)) {
+    return 0;
+  }
+
+  riff_size = GetLE32(GetBuffer(mem) + TAG_SIZE);
+  if (riff_size < CHUNK_HEADER_SIZE) return 0;
+  if (riff_size > MAX_CHUNK_PAYLOAD) return 0;
+
+  // There's no point in reading past the end of the RIFF chunk
+  mem->riff_end_ = riff_size + CHUNK_HEADER_SIZE;
+  if (mem->buf_size_ > mem->riff_end_) {
+    mem->buf_size_ = mem->end_ = mem->riff_end_;
+  }
+
+  Skip(mem, RIFF_HEADER_SIZE);
+  return 1;
+}
+
+static ParseStatus ParseSingleImage(WebPDemuxer* const dmux) {
+  const size_t min_size = CHUNK_HEADER_SIZE;
+  MemBuffer* const mem = &dmux->mem_;
+  Frame* frame;
+  ParseStatus status;
+  int has_vp8l_alpha = 0;  // Frame contains a lossless image with alpha.
+
+  if (dmux->frames_ != NULL) return PARSE_ERROR;
+  if (SizeIsInvalid(mem, min_size)) return PARSE_ERROR;
+  if (MemDataSize(mem) < min_size) return PARSE_NEED_MORE_DATA;
+
+  frame = (Frame*)calloc(1, sizeof(*frame));
+  if (frame == NULL) return PARSE_ERROR;
+
+  // For the single image case, we allow parsing of a partial frame. But we need
+  // at least CHUNK_HEADER_SIZE for parsing.
+  status = StoreFrame(1, CHUNK_HEADER_SIZE, &dmux->mem_, frame,
+                      &has_vp8l_alpha);
+  if (status != PARSE_ERROR) {
+    const int has_alpha = !!(dmux->feature_flags_ & ALPHA_FLAG);
+    // Clear any alpha when the alpha flag is missing.
+    if (!has_alpha && frame->img_components_[1].size_ > 0) {
+      frame->img_components_[1].offset_ = 0;
+      frame->img_components_[1].size_ = 0;
+    }
+
+    // Use the frame width/height as the canvas values for non-vp8x files.
+    // Also, set ALPHA_FLAG if this is a lossless image with alpha.
+    if (!dmux->is_ext_format_ && frame->width_ > 0 && frame->height_ > 0) {
+      dmux->state_ = WEBP_DEMUX_PARSED_HEADER;
+      dmux->canvas_width_ = frame->width_;
+      dmux->canvas_height_ = frame->height_;
+      dmux->feature_flags_ |= has_vp8l_alpha ? ALPHA_FLAG : 0;
+    }
+    AddFrame(dmux, frame);
+    dmux->num_frames_ = 1;
+  } else {
+    free(frame);
+  }
+
+  return status;
+}
+
+static ParseStatus ParseVP8X(WebPDemuxer* const dmux) {
+  MemBuffer* const mem = &dmux->mem_;
+  int anim_chunks = 0;
+  uint32_t vp8x_size;
+  ParseStatus status = PARSE_OK;
+
+  if (MemDataSize(mem) < CHUNK_HEADER_SIZE) return PARSE_NEED_MORE_DATA;
+
+  dmux->is_ext_format_ = 1;
+  Skip(mem, TAG_SIZE);  // VP8X
+  vp8x_size = ReadLE32(mem);
+  if (vp8x_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+  if (vp8x_size < VP8X_CHUNK_SIZE) return PARSE_ERROR;
+  vp8x_size += vp8x_size & 1;
+  if (SizeIsInvalid(mem, vp8x_size)) return PARSE_ERROR;
+  if (MemDataSize(mem) < vp8x_size) return PARSE_NEED_MORE_DATA;
+
+  dmux->feature_flags_ = ReadByte(mem);
+  Skip(mem, 3);  // Reserved.
+  dmux->canvas_width_  = 1 + ReadLE24s(mem);
+  dmux->canvas_height_ = 1 + ReadLE24s(mem);
+  if (dmux->canvas_width_ * (uint64_t)dmux->canvas_height_ >= MAX_IMAGE_AREA) {
+    return PARSE_ERROR;  // image final dimension is too large
+  }
+  Skip(mem, vp8x_size - VP8X_CHUNK_SIZE);  // skip any trailing data.
+  dmux->state_ = WEBP_DEMUX_PARSED_HEADER;
+
+  if (SizeIsInvalid(mem, CHUNK_HEADER_SIZE)) return PARSE_ERROR;
+  if (MemDataSize(mem) < CHUNK_HEADER_SIZE) return PARSE_NEED_MORE_DATA;
+
+  do {
+    int store_chunk = 1;
+    const size_t chunk_start_offset = mem->start_;
+    const uint32_t fourcc = ReadLE32(mem);
+    const uint32_t chunk_size = ReadLE32(mem);
+    const uint32_t chunk_size_padded = chunk_size + (chunk_size & 1);
+
+    if (chunk_size > MAX_CHUNK_PAYLOAD) return PARSE_ERROR;
+    if (SizeIsInvalid(mem, chunk_size_padded)) return PARSE_ERROR;
+
+    switch (fourcc) {
+      case MKFOURCC('V', 'P', '8', 'X'): {
+        return PARSE_ERROR;
+      }
+      case MKFOURCC('A', 'L', 'P', 'H'):
+      case MKFOURCC('V', 'P', '8', ' '):
+      case MKFOURCC('V', 'P', '8', 'L'): {
+        Rewind(mem, CHUNK_HEADER_SIZE);
+        status = ParseSingleImage(dmux);
+        break;
+      }
+      case MKFOURCC('A', 'N', 'I', 'M'): {
+        if (chunk_size_padded < ANIM_CHUNK_SIZE) return PARSE_ERROR;
+
+        if (MemDataSize(mem) < chunk_size_padded) {
+          status = PARSE_NEED_MORE_DATA;
+        } else if (anim_chunks == 0) {
+          ++anim_chunks;
+          dmux->bgcolor_ = ReadLE32(mem);
+          dmux->loop_count_ = ReadLE16s(mem);
+          Skip(mem, chunk_size_padded - ANIM_CHUNK_SIZE);
+        } else {
+          store_chunk = 0;
+          goto Skip;
+        }
+        break;
+      }
+      case MKFOURCC('A', 'N', 'M', 'F'): {
+        status = ParseFrame(dmux, chunk_size_padded);
+        break;
+      }
+      case MKFOURCC('F', 'R', 'G', 'M'): {
+        status = ParseFragment(dmux, chunk_size_padded);
+        break;
+      }
+      case MKFOURCC('I', 'C', 'C', 'P'): {
+        store_chunk = !!(dmux->feature_flags_ & ICCP_FLAG);
+        goto Skip;
+      }
+      case MKFOURCC('X', 'M', 'P', ' '): {
+        store_chunk = !!(dmux->feature_flags_ & XMP_FLAG);
+        goto Skip;
+      }
+      case MKFOURCC('E', 'X', 'I', 'F'): {
+        store_chunk = !!(dmux->feature_flags_ & EXIF_FLAG);
+        goto Skip;
+      }
+ Skip:
+      default: {
+        if (chunk_size_padded <= MemDataSize(mem)) {
+          if (store_chunk) {
+            // Store only the chunk header and unpadded size as only the payload
+            // will be returned to the user.
+            if (!StoreChunk(dmux, chunk_start_offset,
+                            CHUNK_HEADER_SIZE + chunk_size)) {
+              return PARSE_ERROR;
+            }
+          }
+          Skip(mem, chunk_size_padded);
+        } else {
+          status = PARSE_NEED_MORE_DATA;
+        }
+      }
+    }
+
+    if (mem->start_ == mem->riff_end_) {
+      break;
+    } else if (MemDataSize(mem) < CHUNK_HEADER_SIZE) {
+      status = PARSE_NEED_MORE_DATA;
+    }
+  } while (status == PARSE_OK);
+
+  return status;
+}
+
+// -----------------------------------------------------------------------------
+// Format validation
+
+static int IsValidSimpleFormat(const WebPDemuxer* const dmux) {
+  const Frame* const frame = dmux->frames_;
+  if (dmux->state_ == WEBP_DEMUX_PARSING_HEADER) return 1;
+
+  if (dmux->canvas_width_ <= 0 || dmux->canvas_height_ <= 0) return 0;
+  if (dmux->state_ == WEBP_DEMUX_DONE && frame == NULL) return 0;
+
+  if (frame->width_ <= 0 || frame->height_ <= 0) return 0;
+  return 1;
+}
+
+static int IsValidExtendedFormat(const WebPDemuxer* const dmux) {
+  const int has_fragments = !!(dmux->feature_flags_ & FRAGMENTS_FLAG);
+  const int has_frames = !!(dmux->feature_flags_ & ANIMATION_FLAG);
+  const Frame* f;
+
+  if (dmux->state_ == WEBP_DEMUX_PARSING_HEADER) return 1;
+
+  if (dmux->canvas_width_ <= 0 || dmux->canvas_height_ <= 0) return 0;
+  if (dmux->loop_count_ < 0) return 0;
+  if (dmux->state_ == WEBP_DEMUX_DONE && dmux->frames_ == NULL) return 0;
+
+  for (f = dmux->frames_; f != NULL; f = f->next_) {
+    const int cur_frame_set = f->frame_num_;
+    int frame_count = 0, fragment_count = 0;
+
+    // Check frame properties and if the image is composed of fragments that
+    // each fragment came from a fragment.
+    for (; f != NULL && f->frame_num_ == cur_frame_set; f = f->next_) {
+      const ChunkData* const image = f->img_components_;
+      const ChunkData* const alpha = f->img_components_ + 1;
+
+      if (!has_fragments && f->is_fragment_) return 0;
+      if (!has_frames && f->frame_num_ > 1) return 0;
+      if (f->x_offset_ < 0 || f->y_offset_ < 0) return 0;
+      if (f->complete_) {
+        if (alpha->size_ == 0 && image->size_ == 0) return 0;
+        // Ensure alpha precedes image bitstream.
+        if (alpha->size_ > 0 && alpha->offset_ > image->offset_) {
+          return 0;
+        }
+
+        if (f->width_ <= 0 || f->height_ <= 0) return 0;
+      } else {
+        // Ensure alpha precedes image bitstream.
+        if (alpha->size_ > 0 && image->size_ > 0 &&
+            alpha->offset_ > image->offset_) {
+          return 0;
+        }
+        // There shouldn't be any frames after an incomplete one.
+        if (f->next_ != NULL) return 0;
+      }
+
+      fragment_count += f->is_fragment_;
+      ++frame_count;
+    }
+    if (!has_fragments && frame_count > 1) return 0;
+    if (fragment_count > 0 && frame_count != fragment_count) return 0;
+    if (f == NULL) break;
+  }
+  return 1;
+}
+
+// -----------------------------------------------------------------------------
+// WebPDemuxer object
+
+static void InitDemux(WebPDemuxer* const dmux, const MemBuffer* const mem) {
+  dmux->state_ = WEBP_DEMUX_PARSING_HEADER;
+  dmux->loop_count_ = 1;
+  dmux->bgcolor_ = 0xFFFFFFFF;  // White background by default.
+  dmux->canvas_width_ = -1;
+  dmux->canvas_height_ = -1;
+  dmux->mem_ = *mem;
+}
+
+WebPDemuxer* WebPDemuxInternal(const WebPData* data, int allow_partial,
+                               WebPDemuxState* state, int version) {
+  const ChunkParser* parser;
+  int partial;
+  ParseStatus status = PARSE_ERROR;
+  MemBuffer mem;
+  WebPDemuxer* dmux;
+
+  if (WEBP_ABI_IS_INCOMPATIBLE(version, WEBP_DEMUX_ABI_VERSION)) return NULL;
+  if (data == NULL || data->bytes == NULL || data->size == 0) return NULL;
+
+  if (!InitMemBuffer(&mem, data->bytes, data->size)) return NULL;
+  if (!ReadHeader(&mem)) return NULL;
+
+  partial = (mem.buf_size_ < mem.riff_end_);
+  if (!allow_partial && partial) return NULL;
+
+  dmux = (WebPDemuxer*)calloc(1, sizeof(*dmux));
+  if (dmux == NULL) return NULL;
+  InitDemux(dmux, &mem);
+
+  for (parser = kMasterChunks; parser->parse != NULL; ++parser) {
+    if (!memcmp(parser->id, GetBuffer(&dmux->mem_), TAG_SIZE)) {
+      status = parser->parse(dmux);
+      if (status == PARSE_OK) dmux->state_ = WEBP_DEMUX_DONE;
+      if (status == PARSE_NEED_MORE_DATA && !partial) status = PARSE_ERROR;
+      if (status != PARSE_ERROR && !parser->valid(dmux)) status = PARSE_ERROR;
+      break;
+    }
+  }
+  if (state) *state = dmux->state_;
+
+  if (status == PARSE_ERROR) {
+    WebPDemuxDelete(dmux);
+    return NULL;
+  }
+  return dmux;
+}
+
+void WebPDemuxDelete(WebPDemuxer* dmux) {
+  Chunk* c;
+  Frame* f;
+  if (dmux == NULL) return;
+
+  for (f = dmux->frames_; f != NULL;) {
+    Frame* const cur_frame = f;
+    f = f->next_;
+    free(cur_frame);
+  }
+  for (c = dmux->chunks_; c != NULL;) {
+    Chunk* const cur_chunk = c;
+    c = c->next_;
+    free(cur_chunk);
+  }
+  free(dmux);
+}
+
+// -----------------------------------------------------------------------------
+
+uint32_t WebPDemuxGetI(const WebPDemuxer* dmux, WebPFormatFeature feature) {
+  if (dmux == NULL) return 0;
+
+  switch (feature) {
+    case WEBP_FF_FORMAT_FLAGS:     return dmux->feature_flags_;
+    case WEBP_FF_CANVAS_WIDTH:     return (uint32_t)dmux->canvas_width_;
+    case WEBP_FF_CANVAS_HEIGHT:    return (uint32_t)dmux->canvas_height_;
+    case WEBP_FF_LOOP_COUNT:       return (uint32_t)dmux->loop_count_;
+    case WEBP_FF_BACKGROUND_COLOR: return dmux->bgcolor_;
+    case WEBP_FF_FRAME_COUNT:      return (uint32_t)dmux->num_frames_;
+  }
+  return 0;
+}
+
+// -----------------------------------------------------------------------------
+// Frame iteration
+
+// Find the first 'frame_num' frame. There may be multiple such frames in a
+// fragmented frame.
+static const Frame* GetFrame(const WebPDemuxer* const dmux, int frame_num) {
+  const Frame* f;
+  for (f = dmux->frames_; f != NULL; f = f->next_) {
+    if (frame_num == f->frame_num_) break;
+  }
+  return f;
+}
+
+// Returns fragment 'fragment_num' and the total count.
+static const Frame* GetFragment(
+    const Frame* const frame_set, int fragment_num, int* const count) {
+  const int this_frame = frame_set->frame_num_;
+  const Frame* f = frame_set;
+  const Frame* fragment = NULL;
+  int total;
+
+  for (total = 0; f != NULL && f->frame_num_ == this_frame; f = f->next_) {
+    if (++total == fragment_num) fragment = f;
+  }
+  *count = total;
+  return fragment;
+}
+
+static const uint8_t* GetFramePayload(const uint8_t* const mem_buf,
+                                      const Frame* const frame,
+                                      size_t* const data_size) {
+  *data_size = 0;
+  if (frame != NULL) {
+    const ChunkData* const image = frame->img_components_;
+    const ChunkData* const alpha = frame->img_components_ + 1;
+    size_t start_offset = image->offset_;
+    *data_size = image->size_;
+
+    // if alpha exists it precedes image, update the size allowing for
+    // intervening chunks.
+    if (alpha->size_ > 0) {
+      const size_t inter_size = (image->offset_ > 0)
+                              ? image->offset_ - (alpha->offset_ + alpha->size_)
+                              : 0;
+      start_offset = alpha->offset_;
+      *data_size  += alpha->size_ + inter_size;
+    }
+    return mem_buf + start_offset;
+  }
+  return NULL;
+}
+
+// Create a whole 'frame' from VP8 (+ alpha) or lossless.
+static int SynthesizeFrame(const WebPDemuxer* const dmux,
+                           const Frame* const first_frame,
+                           int fragment_num, WebPIterator* const iter) {
+  const uint8_t* const mem_buf = dmux->mem_.buf_;
+  int num_fragments;
+  size_t payload_size = 0;
+  const Frame* const fragment =
+      GetFragment(first_frame, fragment_num, &num_fragments);
+  const uint8_t* const payload =
+      GetFramePayload(mem_buf, fragment, &payload_size);
+  if (payload == NULL) return 0;
+  assert(first_frame != NULL);
+
+  iter->frame_num      = first_frame->frame_num_;
+  iter->num_frames     = dmux->num_frames_;
+  iter->fragment_num   = fragment_num;
+  iter->num_fragments  = num_fragments;
+  iter->x_offset       = fragment->x_offset_;
+  iter->y_offset       = fragment->y_offset_;
+  iter->width          = fragment->width_;
+  iter->height         = fragment->height_;
+  iter->duration       = fragment->duration_;
+  iter->dispose_method = fragment->dispose_method_;
+  iter->complete       = fragment->complete_;
+  iter->fragment.bytes = payload;
+  iter->fragment.size  = payload_size;
+  // TODO(jzern): adjust offsets for 'FRGM's embedded in 'ANMF's
+  return 1;
+}
+
+static int SetFrame(int frame_num, WebPIterator* const iter) {
+  const Frame* frame;
+  const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
+  if (dmux == NULL || frame_num < 0) return 0;
+  if (frame_num > dmux->num_frames_) return 0;
+  if (frame_num == 0) frame_num = dmux->num_frames_;
+
+  frame = GetFrame(dmux, frame_num);
+  if (frame == NULL) return 0;
+
+  return SynthesizeFrame(dmux, frame, 1, iter);
+}
+
+int WebPDemuxGetFrame(const WebPDemuxer* dmux, int frame, WebPIterator* iter) {
+  if (iter == NULL) return 0;
+
+  memset(iter, 0, sizeof(*iter));
+  iter->private_ = (void*)dmux;
+  return SetFrame(frame, iter);
+}
+
+int WebPDemuxNextFrame(WebPIterator* iter) {
+  if (iter == NULL) return 0;
+  return SetFrame(iter->frame_num + 1, iter);
+}
+
+int WebPDemuxPrevFrame(WebPIterator* iter) {
+  if (iter == NULL) return 0;
+  if (iter->frame_num <= 1) return 0;
+  return SetFrame(iter->frame_num - 1, iter);
+}
+
+int WebPDemuxSelectFragment(WebPIterator* iter, int fragment_num) {
+  if (iter != NULL && iter->private_ != NULL && fragment_num > 0) {
+    const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
+    const Frame* const frame = GetFrame(dmux, iter->frame_num);
+    if (frame == NULL) return 0;
+
+    return SynthesizeFrame(dmux, frame, fragment_num, iter);
+  }
+  return 0;
+}
+
+void WebPDemuxReleaseIterator(WebPIterator* iter) {
+  (void)iter;
+}
+
+// -----------------------------------------------------------------------------
+// Chunk iteration
+
+static int ChunkCount(const WebPDemuxer* const dmux, const char fourcc[4]) {
+  const uint8_t* const mem_buf = dmux->mem_.buf_;
+  const Chunk* c;
+  int count = 0;
+  for (c = dmux->chunks_; c != NULL; c = c->next_) {
+    const uint8_t* const header = mem_buf + c->data_.offset_;
+    if (!memcmp(header, fourcc, TAG_SIZE)) ++count;
+  }
+  return count;
+}
+
+static const Chunk* GetChunk(const WebPDemuxer* const dmux,
+                             const char fourcc[4], int chunk_num) {
+  const uint8_t* const mem_buf = dmux->mem_.buf_;
+  const Chunk* c;
+  int count = 0;
+  for (c = dmux->chunks_; c != NULL; c = c->next_) {
+    const uint8_t* const header = mem_buf + c->data_.offset_;
+    if (!memcmp(header, fourcc, TAG_SIZE)) ++count;
+    if (count == chunk_num) break;
+  }
+  return c;
+}
+
+static int SetChunk(const char fourcc[4], int chunk_num,
+                    WebPChunkIterator* const iter) {
+  const WebPDemuxer* const dmux = (WebPDemuxer*)iter->private_;
+  int count;
+
+  if (dmux == NULL || fourcc == NULL || chunk_num < 0) return 0;
+  count = ChunkCount(dmux, fourcc);
+  if (count == 0) return 0;
+  if (chunk_num == 0) chunk_num = count;
+
+  if (chunk_num <= count) {
+    const uint8_t* const mem_buf = dmux->mem_.buf_;
+    const Chunk* const chunk = GetChunk(dmux, fourcc, chunk_num);
+    iter->chunk.bytes = mem_buf + chunk->data_.offset_ + CHUNK_HEADER_SIZE;
+    iter->chunk.size  = chunk->data_.size_ - CHUNK_HEADER_SIZE;
+    iter->num_chunks  = count;
+    iter->chunk_num   = chunk_num;
+    return 1;
+  }
+  return 0;
+}
+
+int WebPDemuxGetChunk(const WebPDemuxer* dmux,
+                      const char fourcc[4], int chunk_num,
+                      WebPChunkIterator* iter) {
+  if (iter == NULL) return 0;
+
+  memset(iter, 0, sizeof(*iter));
+  iter->private_ = (void*)dmux;
+  return SetChunk(fourcc, chunk_num, iter);
+}
+
+int WebPDemuxNextChunk(WebPChunkIterator* iter) {
+  if (iter != NULL) {
+    const char* const fourcc =
+        (const char*)iter->chunk.bytes - CHUNK_HEADER_SIZE;
+    return SetChunk(fourcc, iter->chunk_num + 1, iter);
+  }
+  return 0;
+}
+
+int WebPDemuxPrevChunk(WebPChunkIterator* iter) {
+  if (iter != NULL && iter->chunk_num > 1) {
+    const char* const fourcc =
+        (const char*)iter->chunk.bytes - CHUNK_HEADER_SIZE;
+    return SetChunk(fourcc, iter->chunk_num - 1, iter);
+  }
+  return 0;
+}
+
+void WebPDemuxReleaseChunkIterator(WebPChunkIterator* iter) {
+  (void)iter;
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}  // extern "C"
+#endif
diff --git a/src/dsp/cpu-features.c b/src/dsp/cpu-features.c
new file mode 100644
index 00000000..6a5cd8f1
--- /dev/null
+++ b/src/dsp/cpu-features.c
@@ -0,0 +1,396 @@
+/*
+ *  Copyright (c) 2012 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.
+ */
+
+#include <sys/system_properties.h>
+#ifdef __arm__
+#include <machine/cpu-features.h>
+#endif
+#include <pthread.h>
+#include "cpu-features.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <fcntl.h>
+#include <errno.h>
+
+static  pthread_once_t     g_once;
+static  AndroidCpuFamily   g_cpuFamily;
+static  uint64_t           g_cpuFeatures;
+static  int                g_cpuCount;
+
+static const int  android_cpufeatures_debug = 0;
+
+#ifdef __arm__
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_ARM
+#elif defined __i386__
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_X86
+#else
+#  define DEFAULT_CPU_FAMILY  ANDROID_CPU_FAMILY_UNKNOWN
+#endif
+
+#define  D(...) \
+    do { \
+        if (android_cpufeatures_debug) { \
+            printf(__VA_ARGS__); fflush(stdout); \
+        } \
+    } while (0)
+
+#ifdef __i386__
+static __inline__ void x86_cpuid(int func, int values[4])
+{
+    int a, b, c, d;
+    /* We need to preserve ebx since we're compiling PIC code */
+    /* this means we can't use "=b" for the second output register */
+    __asm__ __volatile__ ( \
+      "push %%ebx\n"
+      "cpuid\n" \
+      "mov %1, %%ebx\n"
+      "pop %%ebx\n"
+      : "=a" (a), "=r" (b), "=c" (c), "=d" (d) \
+      : "a" (func) \
+    );
+    values[0] = a;
+    values[1] = b;
+    values[2] = c;
+    values[3] = d;
+}
+#endif
+
+/* Read the content of /proc/cpuinfo into a user-provided buffer.
+ * Return the length of the data, or -1 on error. Does *not*
+ * zero-terminate the content. Will not read more
+ * than 'buffsize' bytes.
+ */
+static int
+read_file(const char*  pathname, char*  buffer, size_t  buffsize)
+{
+    int  fd, len;
+
+    fd = open(pathname, O_RDONLY);
+    if (fd < 0)
+        return -1;
+
+    do {
+        len = read(fd, buffer, buffsize);
+    } while (len < 0 && errno == EINTR);
+
+    close(fd);
+
+    return len;
+}
+
+/* Extract the content of a the first occurence of a given field in
+ * the content of /proc/cpuinfo and return it as a heap-allocated
+ * string that must be freed by the caller.
+ *
+ * Return NULL if not found
+ */
+static char*
+extract_cpuinfo_field(char* buffer, int buflen, const char* field)
+{
+    int  fieldlen = strlen(field);
+    char* bufend = buffer + buflen;
+    char* result = NULL;
+    int len, ignore;
+    const char *p, *q;
+
+    /* Look for first field occurence, and ensures it starts the line.
+     */
+    p = buffer;
+    bufend = buffer + buflen;
+    for (;;) {
+        p = memmem(p, bufend-p, field, fieldlen);
+        if (p == NULL)
+            goto EXIT;
+
+        if (p == buffer || p[-1] == '\n')
+            break;
+
+        p += fieldlen;
+    }
+
+    /* Skip to the first column followed by a space */
+    p += fieldlen;
+    p  = memchr(p, ':', bufend-p);
+    if (p == NULL || p[1] != ' ')
+        goto EXIT;
+
+    /* Find the end of the line */
+    p += 2;
+    q = memchr(p, '\n', bufend-p);
+    if (q == NULL)
+        q = bufend;
+
+    /* Copy the line into a heap-allocated buffer */
+    len = q-p;
+    result = malloc(len+1);
+    if (result == NULL)
+        goto EXIT;
+
+    memcpy(result, p, len);
+    result[len] = '\0';
+
+EXIT:
+    return result;
+}
+
+/* Count the number of occurences of a given field prefix in /proc/cpuinfo.
+ */
+static int
+count_cpuinfo_field(char* buffer, int buflen, const char* field)
+{
+    int fieldlen = strlen(field);
+    const char* p = buffer;
+    const char* bufend = buffer + buflen;
+    const char* q;
+    int count = 0;
+
+    for (;;) {
+        const char* q;
+
+        p = memmem(p, bufend-p, field, fieldlen);
+        if (p == NULL)
+            break;
+
+        /* Ensure that the field is at the start of a line */
+        if (p > buffer && p[-1] != '\n') {
+            p += fieldlen;
+            continue;
+        }
+
+
+        /* skip any whitespace */
+        q = p + fieldlen;
+        while (q < bufend && (*q == ' ' || *q == '\t'))
+            q++;
+
+        /* we must have a colon now */
+        if (q < bufend && *q == ':') {
+            count += 1;
+            q ++;
+        }
+        p = q;
+    }
+
+    return count;
+}
+
+/* Like strlen(), but for constant string literals */
+#define STRLEN_CONST(x)  ((sizeof(x)-1)
+
+
+/* Checks that a space-separated list of items contains one given 'item'.
+ * Returns 1 if found, 0 otherwise.
+ */
+static int
+has_list_item(const char* list, const char* item)
+{
+    const char*  p = list;
+    int itemlen = strlen(item);
+
+    if (list == NULL)
+        return 0;
+
+    while (*p) {
+        const char*  q;
+
+        /* skip spaces */
+        while (*p == ' ' || *p == '\t')
+            p++;
+
+        /* find end of current list item */
+        q = p;
+        while (*q && *q != ' ' && *q != '\t')
+            q++;
+
+        if (itemlen == q-p && !memcmp(p, item, itemlen))
+            return 1;
+
+        /* skip to next item */
+        p = q;
+    }
+    return 0;
+}
+
+
+static void
+android_cpuInit(void)
+{
+    char cpuinfo[4096];
+    int  cpuinfo_len;
+
+    g_cpuFamily   = DEFAULT_CPU_FAMILY;
+    g_cpuFeatures = 0;
+    g_cpuCount    = 1;
+
+    cpuinfo_len = read_file("/proc/cpuinfo", cpuinfo, sizeof cpuinfo);
+    D("cpuinfo_len is (%d):\n%.*s\n", cpuinfo_len,
+      cpuinfo_len >= 0 ? cpuinfo_len : 0, cpuinfo);
+
+    if (cpuinfo_len < 0)  /* should not happen */ {
+        return;
+    }
+
+    /* Count the CPU cores, the value may be 0 for single-core CPUs */
+    g_cpuCount = count_cpuinfo_field(cpuinfo, cpuinfo_len, "processor");
+    if (g_cpuCount == 0) {
+        g_cpuCount = count_cpuinfo_field(cpuinfo, cpuinfo_len, "Processor");
+        if (g_cpuCount == 0) {
+            g_cpuCount = 1;
+        }
+    }
+
+    D("found cpuCount = %d\n", g_cpuCount);
+
+#ifdef __ARM_ARCH__
+    {
+        char*  features = NULL;
+        char*  architecture = NULL;
+
+        /* Extract architecture from the "CPU Architecture" field.
+         * The list is well-known, unlike the the output of
+         * the 'Processor' field which can vary greatly.
+         *
+         * See the definition of the 'proc_arch' array in
+         * $KERNEL/arch/arm/kernel/setup.c and the 'c_show' function in
+         * same file.
+         */
+        char* cpuArch = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "CPU architecture");
+
+        if (cpuArch != NULL) {
+            char*  end;
+            long   archNumber;
+            int    hasARMv7 = 0;
+
+            D("found cpuArch = '%s'\n", cpuArch);
+
+            /* read the initial decimal number, ignore the rest */
+            archNumber = strtol(cpuArch, &end, 10);
+
+            /* Here we assume that ARMv8 will be upwards compatible with v7
+                * in the future. Unfortunately, there is no 'Features' field to
+                * indicate that Thumb-2 is supported.
+                */
+            if (end > cpuArch && archNumber >= 7) {
+                hasARMv7 = 1;
+            }
+
+            /* Unfortunately, it seems that certain ARMv6-based CPUs
+             * report an incorrect architecture number of 7!
+             *
+             * See http://code.google.com/p/android/issues/detail?id=10812
+             *
+             * We try to correct this by looking at the 'elf_format'
+             * field reported by the 'Processor' field, which is of the
+             * form of "(v7l)" for an ARMv7-based CPU, and "(v6l)" for
+             * an ARMv6-one.
+             */
+            if (hasARMv7) {
+                char* cpuProc = extract_cpuinfo_field(cpuinfo, cpuinfo_len,
+                                                      "Processor");
+                if (cpuProc != NULL) {
+                    D("found cpuProc = '%s'\n", cpuProc);
+                    if (has_list_item(cpuProc, "(v6l)")) {
+                        D("CPU processor and architecture mismatch!!\n");
+                        hasARMv7 = 0;
+                    }
+                    free(cpuProc);
+                }
+            }
+
+            if (hasARMv7) {
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_ARMv7;
+            }
+
+            /* The LDREX / STREX instructions are available from ARMv6 */
+            if (archNumber >= 6) {
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_LDREX_STREX;
+            }
+
+            free(cpuArch);
+        }
+
+        /* Extract the list of CPU features from 'Features' field */
+        char* cpuFeatures = extract_cpuinfo_field(cpuinfo, cpuinfo_len, "Features");
+
+        if (cpuFeatures != NULL) {
+
+            D("found cpuFeatures = '%s'\n", cpuFeatures);
+
+            if (has_list_item(cpuFeatures, "vfpv3"))
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3;
+
+            else if (has_list_item(cpuFeatures, "vfpv3d16"))
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_VFPv3;
+
+            if (has_list_item(cpuFeatures, "neon")) {
+                /* Note: Certain kernels only report neon but not vfpv3
+                    *       in their features list. However, ARM mandates
+                    *       that if Neon is implemented, so must be VFPv3
+                    *       so always set the flag.
+                    */
+                g_cpuFeatures |= ANDROID_CPU_ARM_FEATURE_NEON |
+                                 ANDROID_CPU_ARM_FEATURE_VFPv3;
+            }
+            free(cpuFeatures);
+        }
+    }
+#endif /* __ARM_ARCH__ */
+
+#ifdef __i386__
+    g_cpuFamily = ANDROID_CPU_FAMILY_X86;
+
+    int regs[4];
+
+/* According to http://en.wikipedia.org/wiki/CPUID */
+#define VENDOR_INTEL_b  0x756e6547
+#define VENDOR_INTEL_c  0x6c65746e
+#define VENDOR_INTEL_d  0x49656e69
+
+    x86_cpuid(0, regs);
+    int vendorIsIntel = (regs[1] == VENDOR_INTEL_b &&
+                         regs[2] == VENDOR_INTEL_c &&
+                         regs[3] == VENDOR_INTEL_d);
+
+    x86_cpuid(1, regs);
+    if ((regs[2] & (1 << 9)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_SSSE3;
+    }
+    if ((regs[2] & (1 << 23)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_POPCNT;
+    }
+    if (vendorIsIntel && (regs[2] & (1 << 22)) != 0) {
+        g_cpuFeatures |= ANDROID_CPU_X86_FEATURE_MOVBE;
+    }
+#endif
+}
+
+
+AndroidCpuFamily
+android_getCpuFamily(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuFamily;
+}
+
+
+uint64_t
+android_getCpuFeatures(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuFeatures;
+}
+
+
+int
+android_getCpuCount(void)
+{
+    pthread_once(&g_once, android_cpuInit);
+    return g_cpuCount;
+}
diff --git a/src/dsp/cpu-features.h b/src/dsp/cpu-features.h
new file mode 100644
index 00000000..f20c0bc4
--- /dev/null
+++ b/src/dsp/cpu-features.h
@@ -0,0 +1,56 @@
+/*
+ *  Copyright (c) 2012 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.
+ */
+
+// You can download Android source at
+// http://source.android.com/source/downloading.html
+// Original files are in ndk/sources/android/cpufeatures
+// Revision is Change-Id: I9a0629efba36a6023f05e5f092e7addcc1b7d2a9
+
+#ifndef CPU_FEATURES_H
+#define CPU_FEATURES_H
+
+#include <sys/cdefs.h>
+#include <stdint.h>
+
+__BEGIN_DECLS
+
+typedef enum {
+    ANDROID_CPU_FAMILY_UNKNOWN = 0,
+    ANDROID_CPU_FAMILY_ARM,
+    ANDROID_CPU_FAMILY_X86,
+
+    ANDROID_CPU_FAMILY_MAX  /* do not remove */
+
+} AndroidCpuFamily;
+
+/* Return family of the device's CPU */
+extern AndroidCpuFamily   android_getCpuFamily(void);
+
+enum {
+    ANDROID_CPU_ARM_FEATURE_ARMv7       = (1 << 0),
+    ANDROID_CPU_ARM_FEATURE_VFPv3       = (1 << 1),
+    ANDROID_CPU_ARM_FEATURE_NEON        = (1 << 2),
+    ANDROID_CPU_ARM_FEATURE_LDREX_STREX = (1 << 3),
+};
+
+enum {
+    ANDROID_CPU_X86_FEATURE_SSSE3  = (1 << 0),
+    ANDROID_CPU_X86_FEATURE_POPCNT = (1 << 1),
+    ANDROID_CPU_X86_FEATURE_MOVBE  = (1 << 2),
+};
+
+extern uint64_t    android_getCpuFeatures(void);
+
+/* Return the number of CPU cores detected on this device. */
+extern int         android_getCpuCount(void);
+
+__END_DECLS
+
+#endif /* CPU_FEATURES_H */
diff --git a/src/dsp/cpu.c b/src/dsp/cpu.c
index 2ee7812d..bf9ae0c7 100644
--- a/src/dsp/cpu.c
+++ b/src/dsp/cpu.c
@@ -11,9 +11,9 @@
 
 #include "./dsp.h"
 
-//#if defined(__ANDROID__)
-//#include <cpu-features.h>
-//#endif
+#if defined(__ANDROID__)
+#include "./cpu-features.h"
+#endif
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
@@ -57,17 +57,17 @@ static int x86CPUInfo(CPUFeature feature) {
   return 0;
 }
 VP8CPUInfo VP8GetCPUInfo = x86CPUInfo;
-//#elif defined(WEBP_ANDROID_NEON)
-//static int AndroidCPUInfo(CPUFeature feature) {
-//  const AndroidCpuFamily cpu_family = android_getCpuFamily();
-//  const uint64_t cpu_features = android_getCpuFeatures();
-//  if (feature == kNEON) {
-//    return (cpu_family == ANDROID_CPU_FAMILY_ARM &&
-//            0 != (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON));
-//  }
-//  return 0;
-//}
-//VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
+#elif defined(WEBP_ANDROID_NEON)
+static int AndroidCPUInfo(CPUFeature feature) {
+  const AndroidCpuFamily cpu_family = android_getCpuFamily();
+  const uint64_t cpu_features = android_getCpuFeatures();
+  if (feature == kNEON) {
+    return (cpu_family == ANDROID_CPU_FAMILY_ARM &&
+            0 != (cpu_features & ANDROID_CPU_ARM_FEATURE_NEON));
+  }
+  return 0;
+}
+VP8CPUInfo VP8GetCPUInfo = AndroidCPUInfo;
 #elif defined(__ARM_NEON__)
 // define a dummy function to enable turning off NEON at runtime by setting
 // VP8DecGetCPUInfo = NULL
diff --git a/src/dsp/dec.c b/src/dsp/dec.c
index 9ae7b6fa..758c6a57 100644
--- a/src/dsp/dec.c
+++ b/src/dsp/dec.c
@@ -426,11 +426,16 @@ static void HE8uv(uint8_t *dst) {    // horizontal
 }
 
 // helper for chroma-DC predictions
-static WEBP_INLINE void Put8x8uv(uint64_t v, uint8_t* dst) {
+static WEBP_INLINE void Put8x8uv(uint8_t value, uint8_t* dst) {
   int j;
+#ifndef WEBP_REFERENCE_IMPLEMENTATION
+  const uint64_t v = (uint64_t)value * 0x0101010101010101ULL;
   for (j = 0; j < 8; ++j) {
     *(uint64_t*)(dst + j * BPS) = v;
   }
+#else
+  for (j = 0; j < 8; ++j) memset(dst + j * BPS, value, 8);
+#endif
 }
 
 static void DC8uv(uint8_t *dst) {     // DC
@@ -439,7 +444,7 @@ static void DC8uv(uint8_t *dst) {     // DC
   for (i = 0; i < 8; ++i) {
     dc0 += dst[i - BPS] + dst[-1 + i * BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 4) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 4, dst);
 }
 
 static void DC8uvNoLeft(uint8_t *dst) {   // DC with no left samples
@@ -448,7 +453,7 @@ static void DC8uvNoLeft(uint8_t *dst) {   // DC with no left samples
   for (i = 0; i < 8; ++i) {
     dc0 += dst[i - BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 3, dst);
 }
 
 static void DC8uvNoTop(uint8_t *dst) {  // DC with no top samples
@@ -457,11 +462,11 @@ static void DC8uvNoTop(uint8_t *dst) {  // DC with no top samples
   for (i = 0; i < 8; ++i) {
     dc0 += dst[-1 + i * BPS];
   }
-  Put8x8uv((uint64_t)((dc0 >> 3) * 0x0101010101010101ULL), dst);
+  Put8x8uv(dc0 >> 3, dst);
 }
 
 static void DC8uvNoTopLeft(uint8_t *dst) {    // DC with nothing
-  Put8x8uv(0x8080808080808080ULL, dst);
+  Put8x8uv(0x80, dst);
 }
 
 //------------------------------------------------------------------------------
diff --git a/src/dsp/dec_neon.c b/src/dsp/dec_neon.c
index ec824b79..5d7cff15 100644
--- a/src/dsp/dec_neon.c
+++ b/src/dsp/dec_neon.c
@@ -12,14 +12,14 @@
 
 #include "./dsp.h"
 
-#if defined(WEBP_USE_NEON)
-
-#include "../dec/vp8i.h"
-
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
 #endif
 
+#if defined(WEBP_USE_NEON)
+
+#include "../dec/vp8i.h"
+
 #define QRegs "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",                  \
               "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
 
@@ -155,6 +155,9 @@ static void SimpleHFilter16iNEON(uint8_t* p, int stride, int thresh) {
   }
 }
 
+//-----------------------------------------------------------------------------
+// Inverse transforms (Paragraph 14.4)
+
 static void TransformOneNEON(const int16_t *in, uint8_t *dst) {
   const int kBPS = BPS;
   const int16_t constants[] = {20091, 17734, 0, 0};
@@ -311,19 +314,92 @@ static void TransformTwoNEON(const int16_t* in, uint8_t* dst, int do_two) {
   }
 }
 
+static void TransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32; // The store is only incrementing the pointer as if we
+                        // had stored a single byte.
+  __asm__ volatile (
+    // part 1
+    // load data into q0, q1
+    "vld1.16         {q0, q1}, [%[in]]           \n"
+
+    "vaddl.s16       q2, d0, d3                  \n" // a0 = in[0] + in[12]
+    "vaddl.s16       q3, d1, d2                  \n" // a1 = in[4] + in[8]
+    "vsubl.s16       q4, d1, d2                  \n" // a2 = in[4] - in[8]
+    "vsubl.s16       q5, d0, d3                  \n" // a3 = in[0] - in[12]
+
+    "vadd.s32        q0, q2, q3                  \n" // tmp[0] = a0 + a1
+    "vsub.s32        q2, q2, q3                  \n" // tmp[8] = a0 - a1
+    "vadd.s32        q1, q5, q4                  \n" // tmp[4] = a3 + a2
+    "vsub.s32        q3, q5, q4                  \n" // tmp[12] = a3 - a2
+
+    // Transpose
+    // q0 = tmp[0, 4, 8, 12], q1 = tmp[2, 6, 10, 14]
+    // q2 = tmp[1, 5, 9, 13], q3 = tmp[3, 7, 11, 15]
+    "vswp            d1, d4                      \n" // vtrn.64 q0, q2
+    "vswp            d3, d6                      \n" // vtrn.64 q1, q3
+    "vtrn.32         q0, q1                      \n"
+    "vtrn.32         q2, q3                      \n"
+
+    "vmov.s32        q4, #3                      \n" // dc = 3
+    "vadd.s32        q0, q0, q4                  \n" // dc = tmp[0] + 3
+    "vadd.s32        q6, q0, q3                  \n" // a0 = dc + tmp[3]
+    "vadd.s32        q7, q1, q2                  \n" // a1 = tmp[1] + tmp[2]
+    "vsub.s32        q8, q1, q2                  \n" // a2 = tmp[1] - tmp[2]
+    "vsub.s32        q9, q0, q3                  \n" // a3 = dc - tmp[3]
+
+    "vadd.s32        q0, q6, q7                  \n"
+    "vshrn.s32       d0, q0, #3                  \n" // (a0 + a1) >> 3
+    "vadd.s32        q1, q9, q8                  \n"
+    "vshrn.s32       d1, q1, #3                  \n" // (a3 + a2) >> 3
+    "vsub.s32        q2, q6, q7                  \n"
+    "vshrn.s32       d2, q2, #3                  \n" // (a0 - a1) >> 3
+    "vsub.s32        q3, q9, q8                  \n"
+    "vshrn.s32       d3, q3, #3                  \n" // (a3 - a2) >> 3
+
+    // set the results to output
+    "vst1.16         d0[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[0], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[1], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[2], [%[out]], %[kStep]   \n"
+    "vst1.16         d0[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d1[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d2[3], [%[out]], %[kStep]   \n"
+    "vst1.16         d3[3], [%[out]], %[kStep]   \n"
+
+    : [out] "+r"(out)  // modified registers
+    : [in] "r"(in), [kStep] "r"(kStep)  // constants
+    : "memory", "q0", "q1", "q2", "q3", "q4",
+      "q5", "q6", "q7", "q8", "q9"  // clobbered
+  );
+}
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+// Entry point
+
 extern void VP8DspInitNEON(void);
 
 void VP8DspInitNEON(void) {
+#if defined(WEBP_USE_NEON)
   VP8Transform = TransformTwoNEON;
+  VP8TransformWHT = TransformWHT;
 
   VP8SimpleVFilter16 = SimpleVFilter16NEON;
   VP8SimpleHFilter16 = SimpleHFilter16NEON;
   VP8SimpleVFilter16i = SimpleVFilter16iNEON;
   VP8SimpleHFilter16i = SimpleHFilter16iNEON;
+#endif   // WEBP_USE_NEON
 }
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
-
-#endif   // WEBP_USE_NEON
diff --git a/src/dsp/dec_sse2.c b/src/dsp/dec_sse2.c
index 472b68ec..1cac1b84 100644
--- a/src/dsp/dec_sse2.c
+++ b/src/dsp/dec_sse2.c
@@ -12,15 +12,15 @@
 
 #include "./dsp.h"
 
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
 #if defined(WEBP_USE_SSE2)
 
 #include <emmintrin.h>
 #include "../dec/vp8i.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
 
@@ -194,7 +194,7 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
   // Add inverse transform to 'dst' and store.
   {
-    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i zero = _mm_setzero_si128();
     // Load the reference(s).
     __m128i dst0, dst1, dst2, dst3;
     if (do_two) {
@@ -278,14 +278,14 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
 #define GET_NOTHEV(p1, p0, q0, q1, hev_thresh, not_hev) {                      \
   const __m128i zero = _mm_setzero_si128();                                    \
-  const __m128i t1 = MM_ABS(p1, p0);                                           \
-  const __m128i t2 = MM_ABS(q1, q0);                                           \
+  const __m128i t_1 = MM_ABS(p1, p0);                                          \
+  const __m128i t_2 = MM_ABS(q1, q0);                                          \
                                                                                \
   const __m128i h = _mm_set1_epi8(hev_thresh);                                 \
-  const __m128i t3 = _mm_subs_epu8(t1, h);  /* abs(p1 - p0) - hev_tresh */     \
-  const __m128i t4 = _mm_subs_epu8(t2, h);  /* abs(q1 - q0) - hev_tresh */     \
+  const __m128i t_3 = _mm_subs_epu8(t_1, h);  /* abs(p1 - p0) - hev_tresh */   \
+  const __m128i t_4 = _mm_subs_epu8(t_2, h);  /* abs(q1 - q0) - hev_tresh */   \
                                                                                \
-  not_hev = _mm_or_si128(t3, t4);                                              \
+  not_hev = _mm_or_si128(t_3, t_4);                                            \
   not_hev = _mm_cmpeq_epi8(not_hev, zero); /* not_hev <= t1 && not_hev <= t2 */\
 }
 
@@ -314,13 +314,13 @@ static void TransformSSE2(const int16_t* in, uint8_t* dst, int do_two) {
 
 // Updates values of 2 pixels at MB edge during complex filtering.
 // Update operations:
-// q = q - a and p = p + a; where a = [(a_hi >> 7), (a_lo >> 7)]
+// q = q - delta and p = p + delta; where delta = [(a_hi >> 7), (a_lo >> 7)]
 #define UPDATE_2PIXELS(pi, qi, a_lo, a_hi) {                                   \
   const __m128i a_lo7 = _mm_srai_epi16(a_lo, 7);                               \
   const __m128i a_hi7 = _mm_srai_epi16(a_hi, 7);                               \
-  const __m128i a = _mm_packs_epi16(a_lo7, a_hi7);                             \
-  pi = _mm_adds_epi8(pi, a);                                                   \
-  qi = _mm_subs_epi8(qi, a);                                                   \
+  const __m128i delta = _mm_packs_epi16(a_lo7, a_hi7);                         \
+  pi = _mm_adds_epi8(pi, delta);                                               \
+  qi = _mm_subs_epi8(qi, delta);                                               \
 }
 
 static void NeedsFilter(const __m128i* p1, const __m128i* p0, const __m128i* q0,
@@ -876,9 +876,15 @@ static void HFilter8iSSE2(uint8_t* u, uint8_t* v, int stride,
   Store16x4(u, v, stride, &p1, &p0, &q0, &q1);
 }
 
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Entry point
+
 extern void VP8DspInitSSE2(void);
 
 void VP8DspInitSSE2(void) {
+#if defined(WEBP_USE_SSE2)
   VP8Transform = TransformSSE2;
 
   VP8VFilter16 = VFilter16SSE2;
@@ -894,10 +900,9 @@ void VP8DspInitSSE2(void) {
   VP8SimpleHFilter16 = SimpleHFilter16SSE2;
   VP8SimpleVFilter16i = SimpleVFilter16iSSE2;
   VP8SimpleHFilter16i = SimpleHFilter16iSSE2;
+#endif   // WEBP_USE_SSE2
 }
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
-
-#endif   // WEBP_USE_SSE2
diff --git a/src/dsp/dsp.h b/src/dsp/dsp.h
index 3aad3095..fdf53e7f 100644
--- a/src/dsp/dsp.h
+++ b/src/dsp/dsp.h
@@ -29,11 +29,11 @@ extern "C" {
 #define WEBP_USE_SSE2
 #endif
 
-//#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
-//#define WEBP_ANDROID_NEON  // Android targets that might support NEON
-//#endif
+#if defined(__ANDROID__) && defined(__ARM_ARCH_7A__)
+#define WEBP_ANDROID_NEON  // Android targets that might support NEON
+#endif
 
-#if defined(__ARM_NEON__) || defined(WEBP_ANDROID_NEON)
+#if defined(__ARM_NEON__)
 #define WEBP_USE_NEON
 #endif
 
@@ -49,8 +49,6 @@ extern VP8CPUInfo VP8GetCPUInfo;
 //------------------------------------------------------------------------------
 // Encoding
 
-int VP8GetAlpha(const int histo[]);
-
 // Transforms
 // VP8Idct: Does one of two inverse transforms. If do_two is set, the transforms
 //          will be done for (ref, in, dst) and (ref + 4, in + 16, dst + 4).
@@ -85,10 +83,11 @@ typedef int (*VP8QuantizeBlock)(int16_t in[16], int16_t out[16],
                                 int n, const struct VP8Matrix* const mtx);
 extern VP8QuantizeBlock VP8EncQuantizeBlock;
 
-// Compute susceptibility based on DCT-coeff histograms:
-// the higher, the "easier" the macroblock is to compress.
-typedef int (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
-                         int start_block, int end_block);
+// Collect histogram for susceptibility calculation and accumulate in histo[].
+struct VP8Histogram;
+typedef void (*VP8CHisto)(const uint8_t* ref, const uint8_t* pred,
+                          int start_block, int end_block,
+                          struct VP8Histogram* const histo);
 extern const int VP8DspScan[16 + 4 + 4];
 extern VP8CHisto VP8CollectHistogram;
 
@@ -104,7 +103,7 @@ extern VP8DecIdct2 VP8Transform;
 extern VP8DecIdct VP8TransformUV;
 extern VP8DecIdct VP8TransformDC;
 extern VP8DecIdct VP8TransformDCUV;
-extern void (*VP8TransformWHT)(const int16_t* in, int16_t* out);
+extern VP8WHT VP8TransformWHT;
 
 // *dst is the destination block, with stride BPS. Boundary samples are
 // assumed accessible when needed.
@@ -159,6 +158,9 @@ extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 // Initializes SSE2 version of the fancy upsamplers.
 void WebPInitUpsamplersSSE2(void);
 
+// NEON version
+void WebPInitUpsamplersNEON(void);
+
 #endif    // FANCY_UPSAMPLING
 
 // Point-sampling methods.
@@ -200,6 +202,7 @@ extern void (*WebPApplyAlphaMultiply4444)(
 void WebPInitPremultiply(void);
 
 void WebPInitPremultiplySSE2(void);   // should not be called directly.
+void WebPInitPremultiplyNEON(void);
 
 //------------------------------------------------------------------------------
 
diff --git a/src/dsp/enc.c b/src/dsp/enc.c
index 02234564..ae2c830a 100644
--- a/src/dsp/enc.c
+++ b/src/dsp/enc.c
@@ -17,31 +17,18 @@
 extern "C" {
 #endif
 
-//------------------------------------------------------------------------------
-// Compute susceptibility based on DCT-coeff histograms:
-// the higher, the "easier" the macroblock is to compress.
-
-static int ClipAlpha(int alpha) {
-  return alpha < 0 ? 0 : alpha > 255 ? 255 : alpha;
+static WEBP_INLINE uint8_t clip_8b(int v) {
+  return (!(v & ~0xff)) ? v : (v < 0) ? 0 : 255;
 }
 
-int VP8GetAlpha(const int histo[MAX_COEFF_THRESH + 1]) {
-  int num = 0, den = 0, val = 0;
-  int k;
-  int alpha;
-  // note: changing this loop to avoid the numerous "k + 1" slows things down.
-  for (k = 0; k < MAX_COEFF_THRESH; ++k) {
-    if (histo[k + 1]) {
-      val += histo[k + 1];
-      num += val * (k + 1);
-      den += (k + 1) * (k + 1);
-    }
-  }
-  // we scale the value to a usable [0..255] range
-  alpha = den ? 10 * num / den - 5 : 0;
-  return ClipAlpha(alpha);
+static WEBP_INLINE int clip_max(int v, int max) {
+  return (v > max) ? max : v;
 }
 
+//------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
 const int VP8DspScan[16 + 4 + 4] = {
   // Luma
   0 +  0 * BPS,  4 +  0 * BPS, 8 +  0 * BPS, 12 +  0 * BPS,
@@ -53,27 +40,23 @@ const int VP8DspScan[16 + 4 + 4] = {
   8 + 0 * BPS,  12 + 0 * BPS, 8 + 4 * BPS, 12 + 4 * BPS     // V
 };
 
-static int CollectHistogram(const uint8_t* ref, const uint8_t* pred,
-                            int start_block, int end_block) {
-  int histo[MAX_COEFF_THRESH + 1] = { 0 };
-  int16_t out[16];
-  int j, k;
+static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
+                             int start_block, int end_block,
+                             VP8Histogram* const histo) {
+  int j;
   for (j = start_block; j < end_block; ++j) {
-    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
+    int k;
+    int16_t out[16];
 
-    // Convert coefficients to bin (within out[]).
-    for (k = 0; k < 16; ++k) {
-      const int v = abs(out[k]) >> 2;
-      out[k] = (v > MAX_COEFF_THRESH) ? MAX_COEFF_THRESH : v;
-    }
+    VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
 
-    // Use bin to update histogram.
+    // Convert coefficients to bin.
     for (k = 0; k < 16; ++k) {
-      histo[out[k]]++;
+      const int v = abs(out[k]) >> 3;  // TODO(skal): add rounding?
+      const int clipped_value = clip_max(v, MAX_COEFF_THRESH);
+      histo->distribution[clipped_value]++;
     }
   }
-
-  return VP8GetAlpha(histo);
 }
 
 //------------------------------------------------------------------------------
@@ -89,15 +72,12 @@ static void InitTables(void) {
   if (!tables_ok) {
     int i;
     for (i = -255; i <= 255 + 255; ++i) {
-      clip1[255 + i] = (i < 0) ? 0 : (i > 255) ? 255 : i;
+      clip1[255 + i] = clip_8b(i);
     }
     tables_ok = 1;
   }
 }
 
-static WEBP_INLINE uint8_t clip_8b(int v) {
-  return (!(v & ~0xff)) ? v : v < 0 ? 0 : 255;
-}
 
 //------------------------------------------------------------------------------
 // Transforms (Paragraph 14.4)
@@ -154,25 +134,25 @@ static void FTransform(const uint8_t* src, const uint8_t* ref, int16_t* out) {
   int i;
   int tmp[16];
   for (i = 0; i < 4; ++i, src += BPS, ref += BPS) {
-    const int d0 = src[0] - ref[0];
+    const int d0 = src[0] - ref[0];   // 9bit dynamic range ([-255,255])
     const int d1 = src[1] - ref[1];
     const int d2 = src[2] - ref[2];
     const int d3 = src[3] - ref[3];
-    const int a0 = (d0 + d3) << 3;
-    const int a1 = (d1 + d2) << 3;
-    const int a2 = (d1 - d2) << 3;
-    const int a3 = (d0 - d3) << 3;
-    tmp[0 + i * 4] = (a0 + a1);
-    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 14500) >> 12;
-    tmp[2 + i * 4] = (a0 - a1);
-    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  7500) >> 12;
+    const int a0 = (d0 + d3);         // 10b                      [-510,510]
+    const int a1 = (d1 + d2);
+    const int a2 = (d1 - d2);
+    const int a3 = (d0 - d3);
+    tmp[0 + i * 4] = (a0 + a1) << 3;  // 14b                      [-8160,8160]
+    tmp[1 + i * 4] = (a2 * 2217 + a3 * 5352 + 1812) >> 9;      // [-7536,7542]
+    tmp[2 + i * 4] = (a0 - a1) << 3;
+    tmp[3 + i * 4] = (a3 * 2217 - a2 * 5352 +  937) >> 9;
   }
   for (i = 0; i < 4; ++i) {
-    const int a0 = (tmp[0 + i] + tmp[12 + i]);
+    const int a0 = (tmp[0 + i] + tmp[12 + i]);  // 15b
     const int a1 = (tmp[4 + i] + tmp[ 8 + i]);
     const int a2 = (tmp[4 + i] - tmp[ 8 + i]);
     const int a3 = (tmp[0 + i] - tmp[12 + i]);
-    out[0 + i] = (a0 + a1 + 7) >> 4;
+    out[0 + i] = (a0 + a1 + 7) >> 4;            // 12b
     out[4 + i] = ((a2 * 2217 + a3 * 5352 + 12000) >> 16) + (a3 != 0);
     out[8 + i] = (a0 - a1 + 7) >> 4;
     out[12+ i] = ((a3 * 2217 - a2 * 5352 + 51000) >> 16);
@@ -589,30 +569,30 @@ static int TTransform(const uint8_t* in, const uint16_t* w) {
   int i;
   // horizontal pass
   for (i = 0; i < 4; ++i, in += BPS) {
-    const int a0 = (in[0] + in[2]) << 2;
-    const int a1 = (in[1] + in[3]) << 2;
-    const int a2 = (in[1] - in[3]) << 2;
-    const int a3 = (in[0] - in[2]) << 2;
-    tmp[0 + i * 4] = a0 + a1 + (a0 != 0);
+    const int a0 = in[0] + in[2];
+    const int a1 = in[1] + in[3];
+    const int a2 = in[1] - in[3];
+    const int a3 = in[0] - in[2];
+    tmp[0 + i * 4] = a0 + a1;
     tmp[1 + i * 4] = a3 + a2;
     tmp[2 + i * 4] = a3 - a2;
     tmp[3 + i * 4] = a0 - a1;
   }
   // vertical pass
   for (i = 0; i < 4; ++i, ++w) {
-    const int a0 = (tmp[0 + i] + tmp[8 + i]);
-    const int a1 = (tmp[4 + i] + tmp[12+ i]);
-    const int a2 = (tmp[4 + i] - tmp[12+ i]);
-    const int a3 = (tmp[0 + i] - tmp[8 + i]);
+    const int a0 = tmp[0 + i] + tmp[8 + i];
+    const int a1 = tmp[4 + i] + tmp[12+ i];
+    const int a2 = tmp[4 + i] - tmp[12+ i];
+    const int a3 = tmp[0 + i] - tmp[8 + i];
     const int b0 = a0 + a1;
     const int b1 = a3 + a2;
     const int b2 = a3 - a2;
     const int b3 = a0 - a1;
-    // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3
-    sum += w[ 0] * ((abs(b0) + 3) >> 3);
-    sum += w[ 4] * ((abs(b1) + 3) >> 3);
-    sum += w[ 8] * ((abs(b2) + 3) >> 3);
-    sum += w[12] * ((abs(b3) + 3) >> 3);
+
+    sum += w[ 0] * abs(b0);
+    sum += w[ 4] * abs(b1);
+    sum += w[ 8] * abs(b2);
+    sum += w[12] * abs(b3);
   }
   return sum;
 }
@@ -621,7 +601,7 @@ static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
                     const uint16_t* const w) {
   const int sum1 = TTransform(a, w);
   const int sum2 = TTransform(b, w);
-  return (abs(sum2 - sum1) + 8) >> 4;
+  return abs(sum2 - sum1) >> 5;
 }
 
 static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
@@ -706,6 +686,7 @@ VP8QuantizeBlock VP8EncQuantizeBlock;
 VP8BlockCopy VP8Copy4x4;
 
 extern void VP8EncDspInitSSE2(void);
+extern void VP8EncDspInitNEON(void);
 
 void VP8EncDspInit(void) {
   InitTables();
@@ -734,6 +715,10 @@ void VP8EncDspInit(void) {
     if (VP8GetCPUInfo(kSSE2)) {
       VP8EncDspInitSSE2();
     }
+#elif defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      VP8EncDspInitNEON();
+    }
 #endif
   }
 }
diff --git a/src/dsp/enc_neon.c b/src/dsp/enc_neon.c
new file mode 100644
index 00000000..b5a1fbaf
--- /dev/null
+++ b/src/dsp/enc_neon.c
@@ -0,0 +1,661 @@
+// Copyright 2012 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// ARM NEON version of speed-critical encoding functions.
+//
+// adapted from libvpx (http://www.webmproject.org/code/)
+
+#include "./dsp.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+#if defined(WEBP_USE_NEON)
+
+#include "../enc/vp8enci.h"
+
+//------------------------------------------------------------------------------
+// Transforms (Paragraph 14.4)
+
+// Inverse transform.
+// This code is pretty much the same as TransformOneNEON in the decoder, except
+// for subtraction to *ref. See the comments there for algorithmic explanations.
+static void ITransformOne(const uint8_t* ref,
+                          const int16_t* in, uint8_t* dst) {
+  const int kBPS = BPS;
+  const int16_t kC1C2[] = { 20091, 17734, 0, 0 };  // kC1 / (kC2 >> 1) / 0 / 0
+
+  __asm__ volatile (
+    "vld1.16         {q1, q2}, [%[in]]           \n"
+    "vld1.16         {d0}, [%[kC1C2]]            \n"
+
+    // d2: in[0]
+    // d3: in[8]
+    // d4: in[4]
+    // d5: in[12]
+    "vswp            d3, d4                      \n"
+
+    // q8 = {in[4], in[12]} * kC1 * 2 >> 16
+    // q9 = {in[4], in[12]} * kC2 >> 16
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    // d22 = a = in[0] + in[8]
+    // d23 = b = in[0] - in[8]
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    //  q8 = in[4]/[12] * kC1 >> 16
+    "vshr.s16        q8, q8, #1                  \n"
+
+    // Add {in[4], in[12]} back after the multiplication.
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    // d20 = c = in[4]*kC2 - in[12]*kC1
+    // d21 = d = in[4]*kC1 + in[12]*kC2
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    // d2 = tmp[0] = a + d
+    // d3 = tmp[1] = b + c
+    // d4 = tmp[2] = b - c
+    // d5 = tmp[3] = a - d
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    "vswp            d3, d4                      \n"
+
+    // q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16
+    // q9 = {tmp[4], tmp[12]} * kC2 >> 16
+    "vqdmulh.s16     q8, q2, d0[0]               \n"
+    "vqdmulh.s16     q9, q2, d0[1]               \n"
+
+    // d22 = a = tmp[0] + tmp[8]
+    // d23 = b = tmp[0] - tmp[8]
+    "vqadd.s16       d22, d2, d3                 \n"
+    "vqsub.s16       d23, d2, d3                 \n"
+
+    "vshr.s16        q8, q8, #1                  \n"
+    "vqadd.s16       q8, q2, q8                  \n"
+
+    // d20 = c = in[4]*kC2 - in[12]*kC1
+    // d21 = d = in[4]*kC1 + in[12]*kC2
+    "vqsub.s16       d20, d18, d17               \n"
+    "vqadd.s16       d21, d19, d16               \n"
+
+    // d2 = tmp[0] = a + d
+    // d3 = tmp[1] = b + c
+    // d4 = tmp[2] = b - c
+    // d5 = tmp[3] = a - d
+    "vqadd.s16       d2, d22, d21                \n"
+    "vqadd.s16       d3, d23, d20                \n"
+    "vqsub.s16       d4, d23, d20                \n"
+    "vqsub.s16       d5, d22, d21                \n"
+
+    "vld1.32         d6[0], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d6[1], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d7[0], [%[ref]], %[kBPS]    \n"
+    "vld1.32         d7[1], [%[ref]], %[kBPS]    \n"
+
+    "sub         %[ref], %[ref], %[kBPS], lsl #2 \n"
+
+    // (val) + 4 >> 3
+    "vrshr.s16       d2, d2, #3                  \n"
+    "vrshr.s16       d3, d3, #3                  \n"
+    "vrshr.s16       d4, d4, #3                  \n"
+    "vrshr.s16       d5, d5, #3                  \n"
+
+    "vzip.16         q1, q2                      \n"
+    "vzip.16         q1, q2                      \n"
+
+    // Must accumulate before saturating
+    "vmovl.u8        q8, d6                      \n"
+    "vmovl.u8        q9, d7                      \n"
+
+    "vqadd.s16       q1, q1, q8                  \n"
+    "vqadd.s16       q2, q2, q9                  \n"
+
+    "vqmovun.s16     d0, q1                      \n"
+    "vqmovun.s16     d1, q2                      \n"
+
+    "vst1.32         d0[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d0[1], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[0], [%[dst]], %[kBPS]    \n"
+    "vst1.32         d1[1], [%[dst]]             \n"
+
+    : [in] "+r"(in), [dst] "+r"(dst)               // modified registers
+    : [kBPS] "r"(kBPS), [kC1C2] "r"(kC1C2), [ref] "r"(ref)  // constants
+    : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11"  // clobbered
+  );
+}
+
+static void ITransform(const uint8_t* ref,
+                       const int16_t* in, uint8_t* dst, int do_two) {
+  ITransformOne(ref, in, dst);
+  if (do_two) {
+    ITransformOne(ref + 4, in + 16, dst + 4);
+  }
+}
+
+// Same code as dec_neon.c
+static void ITransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32; // The store is only incrementing the pointer as if we
+                        // had stored a single byte.
+  __asm__ volatile (
+    // part 1
+    // load data into q0, q1
+    "vld1.16         {q0, q1}, [%[in]]           \n"
+
+    "vaddl.s16       q2, d0, d3                  \n" // a0 = in[0] + in[12]
+    "vaddl.s16       q3, d1, d2                  \n" // a1 = in[4] + in[8]
+    "vsubl.s16       q4, d1, d2                  \n" // a2 = in[4] - in[8]
+    "vsubl.s16       q5, d0, d3                  \n" // a3 = in[0] - in[12]
+
+    "vadd.s32        q0, q2, q3                  \n" // tmp[0] = a0 + a1
+    "vsub.s32        q2, q2, q3                  \n" // tmp[8] = a0 - a1
+    "vadd.s32        q1, q5, q4                  \n" // tmp[4] = a3 + a2
+    "vsub.s32        q3, q5, q4                  \n" // tmp[12] = a3 - a2
+
+    // Transpose
+    // q0 = tmp[0, 4, 8, 12], q1 = tmp[2, 6, 10, 14]
+    // q2 = tmp[1, 5, 9, 13], q3 = tmp[3, 7, 11, 15]
+    "vswp            d1, d4                      \n" // vtrn.64 q0, q2
+    "vswp            d3, d6                      \n" // vtrn.64 q1, q3
+    "vtrn.32         q0, q1                      \n"
+    "vtrn.32         q2, q3                      \n"
+
+    "vmov.s32        q4, #3                      \n" // dc = 3
+    "vadd.s32        q0, q0, q4                  \n" // dc = tmp[0] + 3
+    "vadd.s32        q6, q0, q3                  \n" // a0 = dc + tmp[3]
+    "vadd.s32        q7, q1, q2                  \n" // a1 = tmp[1] + tmp[2]
+    "vsub.s32        q8, q1, q2                  \n" // a2 = tmp[1] - tmp[2]
+    "vsub.s32        q9, q0, q3                  \n" // a3 = dc - tmp[3]
+
+    "vadd.s32        q0, q6, q7                  \n"
+    "vshrn.s32       d0, q0, #3                  \n" // (a0 + a1) >> 3
+    "vadd.s32        q1, q9, q8                  \n"
+    "vshrn.s32       d1, q1, #3                  \n" // (a3 + a2) >> 3
+    "vsub.s32        q2, q6, q7                  \n"
+    "vshrn.s32       d2, q2, #3                  \n" // (a0 - a1) >> 3
+    "vsub.s32        q3, q9, q8                  \n"
+    "vshrn.s32       d3, q3, #3                  \n" // (a3 - a2) >> 3
+
+    // set the results to output
+    "vst1.16         d0[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[0], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[1], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[2], [%[out]], %[kStep]      \n"
+    "vst1.16         d0[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d1[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d2[3], [%[out]], %[kStep]      \n"
+    "vst1.16         d3[3], [%[out]], %[kStep]      \n"
+
+    : [out] "+r"(out)  // modified registers
+    : [in] "r"(in), [kStep] "r"(kStep)  // constants
+    : "memory", "q0", "q1", "q2", "q3", "q4",
+      "q5", "q6", "q7", "q8", "q9" // clobbered
+  );
+}
+
+// Forward transform.
+
+// adapted from vp8/encoder/arm/neon/shortfdct_neon.asm
+static const int16_t kCoeff16[] = {
+  5352,  5352,  5352, 5352, 2217,  2217,  2217, 2217
+};
+static const int32_t kCoeff32[] = {
+   1812,  1812,  1812,  1812,
+    937,   937,   937,   937,
+  12000, 12000, 12000, 12000,
+  51000, 51000, 51000, 51000
+};
+
+static void FTransform(const uint8_t* src, const uint8_t* ref,
+                       int16_t* out) {
+  const int kBPS = BPS;
+  const uint8_t* src_ptr = src;
+  const uint8_t* ref_ptr = ref;
+  const int16_t* coeff16 = kCoeff16;
+  const int32_t* coeff32 = kCoeff32;
+
+  __asm__ volatile (
+    // load src into q4, q5 in high half
+    "vld1.8 {d8},  [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d10}, [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d9},  [%[src_ptr]], %[kBPS]      \n"
+    "vld1.8 {d11}, [%[src_ptr]]               \n"
+
+    // load ref into q6, q7 in high half
+    "vld1.8 {d12}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d14}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d13}, [%[ref_ptr]], %[kBPS]      \n"
+    "vld1.8 {d15}, [%[ref_ptr]]               \n"
+
+    // Pack the high values in to q4 and q6
+    "vtrn.32     q4, q5                       \n"
+    "vtrn.32     q6, q7                       \n"
+
+    // d[0-3] = src - ref
+    "vsubl.u8    q0, d8, d12                  \n"
+    "vsubl.u8    q1, d9, d13                  \n"
+
+    // load coeff16 into q8(d16=5352, d17=2217)
+    "vld1.16     {q8}, [%[coeff16]]           \n"
+
+    // load coeff32 high half into q9 = 1812, q10 = 937
+    "vld1.32     {q9, q10}, [%[coeff32]]!     \n"
+
+    // load coeff32 low half into q11=12000, q12=51000
+    "vld1.32     {q11,q12}, [%[coeff32]]      \n"
+
+    // part 1
+    // Transpose. Register dN is the same as dN in C
+    "vtrn.32         d0, d2                   \n"
+    "vtrn.32         d1, d3                   \n"
+    "vtrn.16         d0, d1                   \n"
+    "vtrn.16         d2, d3                   \n"
+
+    "vadd.s16        d4, d0, d3               \n" // a0 = d0 + d3
+    "vadd.s16        d5, d1, d2               \n" // a1 = d1 + d2
+    "vsub.s16        d6, d1, d2               \n" // a2 = d1 - d2
+    "vsub.s16        d7, d0, d3               \n" // a3 = d0 - d3
+
+    "vadd.s16        d0, d4, d5               \n" // a0 + a1
+    "vshl.s16        d0, d0, #3               \n" // temp[0+i*4] = (a0+a1) << 3
+    "vsub.s16        d2, d4, d5               \n" // a0 - a1
+    "vshl.s16        d2, d2, #3               \n" // (temp[2+i*4] = (a0-a1) << 3
+
+    "vmlal.s16       q9, d7, d16              \n" // a3*5352 + 1812
+    "vmlal.s16       q10, d7, d17             \n" // a3*2217 + 937
+    "vmlal.s16       q9, d6, d17              \n" // a2*2217 + a3*5352 + 1812
+    "vmlsl.s16       q10, d6, d16             \n" // a3*2217 + 937 - a2*5352
+
+    // temp[1+i*4] = (d2*2217 + d3*5352 + 1812) >> 9
+    // temp[3+i*4] = (d3*2217 + 937 - d2*5352) >> 9
+    "vshrn.s32       d1, q9, #9               \n"
+    "vshrn.s32       d3, q10, #9              \n"
+
+    // part 2
+    // transpose d0=ip[0], d1=ip[4], d2=ip[8], d3=ip[12]
+    "vtrn.32         d0, d2                   \n"
+    "vtrn.32         d1, d3                   \n"
+    "vtrn.16         d0, d1                   \n"
+    "vtrn.16         d2, d3                   \n"
+
+    "vmov.s16        d26, #7                  \n"
+
+    "vadd.s16        d4, d0, d3               \n" // a1 = ip[0] + ip[12]
+    "vadd.s16        d5, d1, d2               \n" // b1 = ip[4] + ip[8]
+    "vsub.s16        d6, d1, d2               \n" // c1 = ip[4] - ip[8]
+    "vadd.s16        d4, d4, d26              \n" // a1 + 7
+    "vsub.s16        d7, d0, d3               \n" // d1 = ip[0] - ip[12]
+
+    "vadd.s16        d0, d4, d5               \n" // op[0] = a1 + b1 + 7
+    "vsub.s16        d2, d4, d5               \n" // op[8] = a1 - b1 + 7
+
+    "vmlal.s16       q11, d7, d16             \n" // d1*5352 + 12000
+    "vmlal.s16       q12, d7, d17             \n" // d1*2217 + 51000
+
+    "vceq.s16        d4, d7, #0               \n"
+
+    "vshr.s16        d0, d0, #4               \n"
+    "vshr.s16        d2, d2, #4               \n"
+
+    "vmlal.s16       q11, d6, d17             \n" // c1*2217 + d1*5352 + 12000
+    "vmlsl.s16       q12, d6, d16             \n" // d1*2217 - c1*5352 + 51000
+
+    "vmvn.s16        d4, d4                   \n"
+    // op[4] = (c1*2217 + d1*5352 + 12000)>>16
+    "vshrn.s32       d1, q11, #16             \n"
+    // op[4] += (d1!=0)
+    "vsub.s16        d1, d1, d4               \n"
+    // op[12]= (d1*2217 - c1*5352 + 51000)>>16
+    "vshrn.s32       d3, q12, #16             \n"
+
+    // set result to out array
+    "vst1.16         {q0, q1}, [%[out]]   \n"
+    : [src_ptr] "+r"(src_ptr), [ref_ptr] "+r"(ref_ptr),
+      [coeff32] "+r"(coeff32)          // modified registers
+    : [kBPS] "r"(kBPS), [coeff16] "r"(coeff16),
+      [out] "r"(out)                   // constants
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+      "q10", "q11", "q12", "q13"       // clobbered
+  );
+}
+
+static void FTransformWHT(const int16_t* in, int16_t* out) {
+  const int kStep = 32;
+  __asm__ volatile (
+    // d0 = in[0 * 16] , d1 = in[1 * 16]
+    // d2 = in[2 * 16] , d3 = in[3 * 16]
+    "vld1.16         d0[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[0], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[1], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[2], [%[in]], %[kStep]   \n"
+    "vld1.16         d0[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d1[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d2[3], [%[in]], %[kStep]   \n"
+    "vld1.16         d3[3], [%[in]], %[kStep]   \n"
+
+    "vaddl.s16       q2, d0, d2                 \n"
+    "vshl.s32        q2, q2, #2                 \n" // a0=(in[0*16]+in[2*16])<<2
+    "vaddl.s16       q3, d1, d3                 \n"
+    "vshl.s32        q3, q3, #2                 \n" // a1=(in[1*16]+in[3*16])<<2
+    "vsubl.s16       q4, d1, d3                 \n"
+    "vshl.s32        q4, q4, #2                 \n" // a2=(in[1*16]-in[3*16])<<2
+    "vsubl.s16       q5, d0, d2                 \n"
+    "vshl.s32        q5, q5, #2                 \n" // a3=(in[0*16]-in[2*16])<<2
+
+    "vceq.s32        q10, q2, #0                \n"
+    "vmvn.s32        q10, q10                   \n" // (a0 != 0)
+    "vqadd.s32       q6, q2, q3                 \n" // (a0 + a1)
+    "vqsub.s32       q6, q6, q10                \n" // (a0 + a1) + (a0 != 0)
+    "vqadd.s32       q7, q5, q4                 \n" // a3 + a2
+    "vqsub.s32       q8, q5, q4                 \n" // a3 - a2
+    "vqsub.s32       q9, q2, q3                 \n" // a0 - a1
+
+    // Transpose
+    // q6 = tmp[0, 1,  2,  3] ; q7 = tmp[ 4,  5,  6,  7]
+    // q8 = tmp[8, 9, 10, 11] ; q9 = tmp[12, 13, 14, 15]
+    "vswp            d13, d16                   \n" // vtrn.64 q0, q2
+    "vswp            d15, d18                   \n" // vtrn.64 q1, q3
+    "vtrn.32         q6, q7                     \n"
+    "vtrn.32         q8, q9                     \n"
+
+    "vqadd.s32       q0, q6, q8                 \n" // a0 = tmp[0] + tmp[8]
+    "vqadd.s32       q1, q7, q9                 \n" // a1 = tmp[4] + tmp[12]
+    "vqsub.s32       q2, q7, q9                 \n" // a2 = tmp[4] - tmp[12]
+    "vqsub.s32       q3, q6, q8                 \n" // a3 = tmp[0] - tmp[8]
+
+    "vqadd.s32       q4, q0, q1                 \n" // b0 = a0 + a1
+    "vqadd.s32       q5, q3, q2                 \n" // b1 = a3 + a2
+    "vqsub.s32       q6, q3, q2                 \n" // b2 = a3 - a2
+    "vqsub.s32       q7, q0, q1                 \n" // b3 = a0 - a1
+
+    "vmov.s32         q0, #3                    \n" // q0 = 3
+
+    "vcgt.s32        q1, q4, #0                 \n" // (b0>0)
+    "vqsub.s32       q2, q4, q1                 \n" // (b0+(b0>0))
+    "vqadd.s32       q3, q2, q0                 \n" // (b0+(b0>0)+3)
+    "vshrn.s32       d18, q3, #3                \n" // (b0+(b0>0)+3) >> 3
+
+    "vcgt.s32        q1, q5, #0                 \n" // (b1>0)
+    "vqsub.s32       q2, q5, q1                 \n" // (b1+(b1>0))
+    "vqadd.s32       q3, q2, q0                 \n" // (b1+(b1>0)+3)
+    "vshrn.s32       d19, q3, #3                \n" // (b1+(b1>0)+3) >> 3
+
+    "vcgt.s32        q1, q6, #0                 \n" // (b2>0)
+    "vqsub.s32       q2, q6, q1                 \n" // (b2+(b2>0))
+    "vqadd.s32       q3, q2, q0                 \n" // (b2+(b2>0)+3)
+    "vshrn.s32       d20, q3, #3                \n" // (b2+(b2>0)+3) >> 3
+
+    "vcgt.s32        q1, q7, #0                 \n" // (b3>0)
+    "vqsub.s32       q2, q7, q1                 \n" // (b3+(b3>0))
+    "vqadd.s32       q3, q2, q0                 \n" // (b3+(b3>0)+3)
+    "vshrn.s32       d21, q3, #3                \n" // (b3+(b3>0)+3) >> 3
+
+    "vst1.16         {q9, q10}, [%[out]]        \n"
+
+    : [in] "+r"(in)
+    : [kStep] "r"(kStep), [out] "r"(out)
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5",
+      "q6", "q7", "q8", "q9", "q10"       // clobbered
+  ) ;
+}
+
+//------------------------------------------------------------------------------
+// Texture distortion
+//
+// We try to match the spectral content (weighted) between source and
+// reconstructed samples.
+
+// Hadamard transform
+// Returns the weighted sum of the absolute value of transformed coefficients.
+// This uses a TTransform helper function in C
+static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
+                    const uint16_t* const w) {
+  const int kBPS = BPS;
+  const uint8_t* A = a;
+  const uint8_t* B = b;
+  const uint16_t* W = w;
+  int sum;
+  __asm__ volatile (
+    "vld1.32         d0[0], [%[a]], %[kBPS]   \n"
+    "vld1.32         d0[1], [%[a]], %[kBPS]   \n"
+    "vld1.32         d2[0], [%[a]], %[kBPS]   \n"
+    "vld1.32         d2[1], [%[a]]            \n"
+
+    "vld1.32         d1[0], [%[b]], %[kBPS]   \n"
+    "vld1.32         d1[1], [%[b]], %[kBPS]   \n"
+    "vld1.32         d3[0], [%[b]], %[kBPS]   \n"
+    "vld1.32         d3[1], [%[b]]            \n"
+
+    // a d0/d2, b d1/d3
+    // d0/d1: 01 01 01 01
+    // d2/d3: 23 23 23 23
+    // But: it goes 01 45 23 67
+    // Notice the middle values are transposed
+    "vtrn.16         q0, q1                   \n"
+
+    // {a0, a1} = {in[0] + in[2], in[1] + in[3]}
+    "vaddl.u8        q2, d0, d2               \n"
+    "vaddl.u8        q10, d1, d3              \n"
+    // {a3, a2} = {in[0] - in[2], in[1] - in[3]}
+    "vsubl.u8        q3, d0, d2               \n"
+    "vsubl.u8        q11, d1, d3              \n"
+
+    // tmp[0] = a0 + a1
+    "vpaddl.s16      q0, q2                   \n"
+    "vpaddl.s16      q8, q10                  \n"
+
+    // tmp[1] = a3 + a2
+    "vpaddl.s16      q1, q3                   \n"
+    "vpaddl.s16      q9, q11                  \n"
+
+    // No pair subtract
+    // q2 = {a0, a3}
+    // q3 = {a1, a2}
+    "vtrn.16         q2, q3                   \n"
+    "vtrn.16         q10, q11                 \n"
+
+    // {tmp[3], tmp[2]} = {a0 - a1, a3 - a2}
+    "vsubl.s16       q12, d4, d6              \n"
+    "vsubl.s16       q13, d5, d7              \n"
+    "vsubl.s16       q14, d20, d22            \n"
+    "vsubl.s16       q15, d21, d23            \n"
+
+    // separate tmp[3] and tmp[2]
+    // q12 = tmp[3]
+    // q13 = tmp[2]
+    "vtrn.32         q12, q13                 \n"
+    "vtrn.32         q14, q15                 \n"
+
+    // Transpose tmp for a
+    "vswp            d1, d26                  \n" // vtrn.64
+    "vswp            d3, d24                  \n" // vtrn.64
+    "vtrn.32         q0, q1                   \n"
+    "vtrn.32         q13, q12                 \n"
+
+    // Transpose tmp for b
+    "vswp            d17, d30                 \n" // vtrn.64
+    "vswp            d19, d28                 \n" // vtrn.64
+    "vtrn.32         q8, q9                   \n"
+    "vtrn.32         q15, q14                 \n"
+
+    // The first Q register is a, the second b.
+    // q0/8 tmp[0-3]
+    // q13/15 tmp[4-7]
+    // q1/9 tmp[8-11]
+    // q12/14 tmp[12-15]
+
+    // These are still in 01 45 23 67 order. We fix it easily in the addition
+    // case but the subtraction propegates them.
+    "vswp            d3, d27                  \n"
+    "vswp            d19, d31                 \n"
+
+    // a0 = tmp[0] + tmp[8]
+    "vadd.s32        q2, q0, q1               \n"
+    "vadd.s32        q3, q8, q9               \n"
+
+    // a1 = tmp[4] + tmp[12]
+    "vadd.s32        q10, q13, q12            \n"
+    "vadd.s32        q11, q15, q14            \n"
+
+    // a2 = tmp[4] - tmp[12]
+    "vsub.s32        q13, q13, q12            \n"
+    "vsub.s32        q15, q15, q14            \n"
+
+    // a3 = tmp[0] - tmp[8]
+    "vsub.s32        q0, q0, q1               \n"
+    "vsub.s32        q8, q8, q9               \n"
+
+    // b0 = a0 + a1
+    "vadd.s32        q1, q2, q10              \n"
+    "vadd.s32        q9, q3, q11              \n"
+
+    // b1 = a3 + a2
+    "vadd.s32        q12, q0, q13             \n"
+    "vadd.s32        q14, q8, q15             \n"
+
+    // b2 = a3 - a2
+    "vsub.s32        q0, q0, q13              \n"
+    "vsub.s32        q8, q8, q15              \n"
+
+    // b3 = a0 - a1
+    "vsub.s32        q2, q2, q10              \n"
+    "vsub.s32        q3, q3, q11              \n"
+
+    "vld1.64         {q10, q11}, [%[w]]       \n"
+
+    // abs(b0)
+    "vabs.s32        q1, q1                   \n"
+    "vabs.s32        q9, q9                   \n"
+    // abs(b1)
+    "vabs.s32        q12, q12                 \n"
+    "vabs.s32        q14, q14                 \n"
+    // abs(b2)
+    "vabs.s32        q0, q0                   \n"
+    "vabs.s32        q8, q8                   \n"
+    // abs(b3)
+    "vabs.s32        q2, q2                   \n"
+    "vabs.s32        q3, q3                   \n"
+
+    // expand w before using.
+    "vmovl.u16       q13, d20                 \n"
+    "vmovl.u16       q15, d21                 \n"
+
+    // w[0] * abs(b0)
+    "vmul.u32        q1, q1, q13              \n"
+    "vmul.u32        q9, q9, q13              \n"
+
+    // w[4] * abs(b1)
+    "vmla.u32        q1, q12, q15             \n"
+    "vmla.u32        q9, q14, q15             \n"
+
+    // expand w before using.
+    "vmovl.u16       q13, d22                 \n"
+    "vmovl.u16       q15, d23                 \n"
+
+    // w[8] * abs(b1)
+    "vmla.u32        q1, q0, q13              \n"
+    "vmla.u32        q9, q8, q13              \n"
+
+    // w[12] * abs(b1)
+    "vmla.u32        q1, q2, q15              \n"
+    "vmla.u32        q9, q3, q15              \n"
+
+    // Sum the arrays
+    "vpaddl.u32      q1, q1                   \n"
+    "vpaddl.u32      q9, q9                   \n"
+    "vadd.u64        d2, d3                   \n"
+    "vadd.u64        d18, d19                 \n"
+
+    // Hadamard transform needs 4 bits of extra precision (2 bits in each
+    // direction) for dynamic raw. Weights w[] are 16bits at max, so the maximum
+    // precision for coeff is 8bit of input + 4bits of Hadamard transform +
+    // 16bits for w[] + 2 bits of abs() summation.
+    //
+    // This uses a maximum of 31 bits (signed). Discarding the top 32 bits is
+    // A-OK.
+
+    // sum2 - sum1
+    "vsub.u32        d0, d2, d18              \n"
+    // abs(sum2 - sum1)
+    "vabs.s32        d0, d0                   \n"
+    // abs(sum2 - sum1) >> 5
+    "vshr.u32        d0, #5                   \n"
+
+    // It would be better to move the value straight into r0 but I'm not
+    // entirely sure how this works with inline assembly.
+    "vmov.32         %[sum], d0[0]            \n"
+
+    : [sum] "=r"(sum), [a] "+r"(A), [b] "+r"(B), [w] "+r"(W)
+    : [kBPS] "r"(kBPS)
+    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
+      "q10", "q11", "q12", "q13", "q14", "q15"  // clobbered
+  ) ;
+
+  return sum;
+}
+
+static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
+                      const uint16_t* const w) {
+  int D = 0;
+  int x, y;
+  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
+    for (x = 0; x < 16; x += 4) {
+      D += Disto4x4(a + x + y, b + x + y, w);
+    }
+  }
+  return D;
+}
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+// Entry point
+
+extern void VP8EncDspInitNEON(void);
+
+void VP8EncDspInitNEON(void) {
+#if defined(WEBP_USE_NEON)
+  VP8ITransform = ITransform;
+  VP8FTransform = FTransform;
+
+  VP8ITransformWHT = ITransformWHT;
+  VP8FTransformWHT = FTransformWHT;
+
+  VP8TDisto4x4 = Disto4x4;
+  VP8TDisto16x16 = Disto16x16;
+#endif   // WEBP_USE_NEON
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
diff --git a/src/dsp/enc_sse2.c b/src/dsp/enc_sse2.c
index b046761d..c4148b56 100644
--- a/src/dsp/enc_sse2.c
+++ b/src/dsp/enc_sse2.c
@@ -11,27 +11,58 @@
 
 #include "./dsp.h"
 
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
 #if defined(WEBP_USE_SSE2)
 #include <stdlib.h>  // for abs()
 #include <emmintrin.h>
 
 #include "../enc/vp8enci.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
+//------------------------------------------------------------------------------
+// Quite useful macro for debugging. Left here for convenience.
+
+#if 0
+#include <stdio.h>
+static void PrintReg(const __m128i r, const char* const name, int size) {
+  int n;
+  union {
+    __m128i r;
+    uint8_t i8[16];
+    uint16_t i16[8];
+    uint32_t i32[4];
+    uint64_t i64[2];
+  } tmp;
+  tmp.r = r;
+  printf("%s\t: ", name);
+  if (size == 8) {
+    for (n = 0; n < 16; ++n) printf("%.2x ", tmp.i8[n]);
+  } else if (size == 16) {
+    for (n = 0; n < 8; ++n) printf("%.4x ", tmp.i16[n]);
+  } else if (size == 32) {
+    for (n = 0; n < 4; ++n) printf("%.8x ", tmp.i32[n]);
+  } else {
+    for (n = 0; n < 2; ++n) printf("%.16lx ", tmp.i64[n]);
+  }
+  printf("\n");
+}
 #endif
 
 //------------------------------------------------------------------------------
 // Compute susceptibility based on DCT-coeff histograms:
 // the higher, the "easier" the macroblock is to compress.
 
-static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
-                                int start_block, int end_block) {
-  int histo[MAX_COEFF_THRESH + 1] = { 0 };
-  int16_t out[16];
-  int j, k;
+static void CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
+                                 int start_block, int end_block,
+                                 VP8Histogram* const histo) {
   const __m128i max_coeff_thresh = _mm_set1_epi16(MAX_COEFF_THRESH);
+  int j;
   for (j = start_block; j < end_block; ++j) {
+    int16_t out[16];
+    int k;
+
     VP8FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
 
     // Convert coefficients to bin (within out[]).
@@ -47,9 +78,9 @@ static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
       const __m128i xor1 = _mm_xor_si128(out1, sign1);
       const __m128i abs0 = _mm_sub_epi16(xor0, sign0);
       const __m128i abs1 = _mm_sub_epi16(xor1, sign1);
-      // v = abs(out) >> 2
-      const __m128i v0 = _mm_srai_epi16(abs0, 2);
-      const __m128i v1 = _mm_srai_epi16(abs1, 2);
+      // v = abs(out) >> 3
+      const __m128i v0 = _mm_srai_epi16(abs0, 3);
+      const __m128i v1 = _mm_srai_epi16(abs1, 3);
       // bin = min(v, MAX_COEFF_THRESH)
       const __m128i bin0 = _mm_min_epi16(v0, max_coeff_thresh);
       const __m128i bin1 = _mm_min_epi16(v1, max_coeff_thresh);
@@ -58,13 +89,11 @@ static int CollectHistogramSSE2(const uint8_t* ref, const uint8_t* pred,
       _mm_storeu_si128((__m128i*)&out[8], bin1);
     }
 
-    // Use bin to update histogram.
+    // Convert coefficients to bin.
     for (k = 0; k < 16; ++k) {
-      histo[out[k]]++;
+      histo->distribution[out[k]]++;
     }
   }
-
-  return VP8GetAlpha(histo);
 }
 
 //------------------------------------------------------------------------------
@@ -243,7 +272,7 @@ static void ITransformSSE2(const uint8_t* ref, const int16_t* in, uint8_t* dst,
 
   // Add inverse transform to 'ref' and store.
   {
-    const __m128i zero = _mm_set1_epi16(0);
+    const __m128i zero = _mm_setzero_si128();
     // Load the reference(s).
     __m128i ref0, ref1, ref2, ref3;
     if (do_two) {
@@ -295,17 +324,23 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
                            int16_t* out) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i seven = _mm_set1_epi16(7);
-  const __m128i k7500 = _mm_set1_epi32(7500);
-  const __m128i k14500 = _mm_set1_epi32(14500);
+  const __m128i k937 = _mm_set1_epi32(937);
+  const __m128i k1812 = _mm_set1_epi32(1812);
   const __m128i k51000 = _mm_set1_epi32(51000);
   const __m128i k12000_plus_one = _mm_set1_epi32(12000 + (1 << 16));
   const __m128i k5352_2217 = _mm_set_epi16(5352,  2217, 5352,  2217,
                                            5352,  2217, 5352,  2217);
   const __m128i k2217_5352 = _mm_set_epi16(2217, -5352, 2217, -5352,
                                            2217, -5352, 2217, -5352);
-
+  const __m128i k88p = _mm_set_epi16(8, 8, 8, 8, 8, 8, 8, 8);
+  const __m128i k88m = _mm_set_epi16(-8, 8, -8, 8, -8, 8, -8, 8);
+  const __m128i k5352_2217p = _mm_set_epi16(2217, 5352, 2217, 5352,
+                                            2217, 5352, 2217, 5352);
+  const __m128i k5352_2217m = _mm_set_epi16(-5352, 2217, -5352, 2217,
+                                            -5352, 2217, -5352, 2217);
   __m128i v01, v32;
 
+
   // Difference between src and ref and initial transpose.
   {
     // Load src and convert to 16b.
@@ -326,73 +361,52 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
     const __m128i ref_1 = _mm_unpacklo_epi8(ref1, zero);
     const __m128i ref_2 = _mm_unpacklo_epi8(ref2, zero);
     const __m128i ref_3 = _mm_unpacklo_epi8(ref3, zero);
-    // Compute difference.
+    // Compute difference. -> 00 01 02 03 00 00 00 00
     const __m128i diff0 = _mm_sub_epi16(src_0, ref_0);
     const __m128i diff1 = _mm_sub_epi16(src_1, ref_1);
     const __m128i diff2 = _mm_sub_epi16(src_2, ref_2);
     const __m128i diff3 = _mm_sub_epi16(src_3, ref_3);
 
-    // Transpose.
+
+    // Unpack and shuffle
     // 00 01 02 03   0 0 0 0
     // 10 11 12 13   0 0 0 0
     // 20 21 22 23   0 0 0 0
     // 30 31 32 33   0 0 0 0
-    const __m128i transpose0_0 = _mm_unpacklo_epi16(diff0, diff1);
-    const __m128i transpose0_1 = _mm_unpacklo_epi16(diff2, diff3);
-    // 00 10 01 11   02 12 03 13
-    // 20 30 21 31   22 32 23 33
-    const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
-    v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
-    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
-    // a02 a12 a22 a32   a03 a13 a23 a33
-    // a00 a10 a20 a30   a01 a11 a21 a31
-    // a03 a13 a23 a33   a02 a12 a22 a32
-  }
-
-  // First pass and subsequent transpose.
-  {
-    // Same operations are done on the (0,3) and (1,2) pairs.
-    // b0 = (a0 + a3) << 3
-    // b1 = (a1 + a2) << 3
-    // b3 = (a0 - a3) << 3
-    // b2 = (a1 - a2) << 3
-    const __m128i a01 = _mm_add_epi16(v01, v32);
-    const __m128i a32 = _mm_sub_epi16(v01, v32);
-    const __m128i b01 = _mm_slli_epi16(a01, 3);
-    const __m128i b32 = _mm_slli_epi16(a32, 3);
-    const __m128i b11 = _mm_unpackhi_epi64(b01, b01);
-    const __m128i b22 = _mm_unpackhi_epi64(b32, b32);
-
-    // e0 = b0 + b1
-    // e2 = b0 - b1
-    const __m128i e0 = _mm_add_epi16(b01, b11);
-    const __m128i e2 = _mm_sub_epi16(b01, b11);
-    const __m128i e02 = _mm_unpacklo_epi64(e0, e2);
-
-    // e1 = (b3 * 5352 + b2 * 2217 + 14500) >> 12
-    // e3 = (b3 * 2217 - b2 * 5352 +  7500) >> 12
-    const __m128i b23 = _mm_unpacklo_epi16(b22, b32);
-    const __m128i c1 = _mm_madd_epi16(b23, k5352_2217);
-    const __m128i c3 = _mm_madd_epi16(b23, k2217_5352);
-    const __m128i d1 = _mm_add_epi32(c1, k14500);
-    const __m128i d3 = _mm_add_epi32(c3, k7500);
-    const __m128i e1 = _mm_srai_epi32(d1, 12);
-    const __m128i e3 = _mm_srai_epi32(d3, 12);
-    const __m128i e13 = _mm_packs_epi32(e1, e3);
-
-    // Transpose.
-    // 00 01 02 03  20 21 22 23
-    // 10 11 12 13  30 31 32 33
-    const __m128i transpose0_0 = _mm_unpacklo_epi16(e02, e13);
-    const __m128i transpose0_1 = _mm_unpackhi_epi16(e02, e13);
-    // 00 10 01 11   02 12 03 13
-    // 20 30 21 31   22 32 23 33
-    const __m128i v23 = _mm_unpackhi_epi32(transpose0_0, transpose0_1);
-    v01 = _mm_unpacklo_epi32(transpose0_0, transpose0_1);
-    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));
-    // 02 12 22 32   03 13 23 33
-    // 00 10 20 30   01 11 21 31
-    // 03 13 23 33   02 12 22 32
+    const __m128i shuf01 = _mm_unpacklo_epi32(diff0, diff1);
+    const __m128i shuf23 = _mm_unpacklo_epi32(diff2, diff3);
+    // 00 01 10 11 02 03 12 13
+    // 20 21 30 31 22 23 32 33
+    const __m128i shuf01_p =
+        _mm_shufflehi_epi16(shuf01, _MM_SHUFFLE(2, 3, 0, 1));
+    const __m128i shuf23_p =
+        _mm_shufflehi_epi16(shuf23, _MM_SHUFFLE(2, 3, 0, 1));
+    // 00 01 10 11 03 02 13 12
+    // 20 21 30 31 23 22 33 32
+    const __m128i s01 = _mm_unpacklo_epi64(shuf01_p, shuf23_p);
+    const __m128i s32 = _mm_unpackhi_epi64(shuf01_p, shuf23_p);
+    // 00 01 10 11 20 21 30 31
+    // 03 02 13 12 23 22 33 32
+    const __m128i a01 = _mm_add_epi16(s01, s32);
+    const __m128i a32 = _mm_sub_epi16(s01, s32);
+    // [d0 + d3 | d1 + d2 | ...] = [a0 a1 | a0' a1' | ... ]
+    // [d0 - d3 | d1 - d2 | ...] = [a3 a2 | a3' a2' | ... ]
+
+    const __m128i tmp0 = _mm_madd_epi16(a01, k88p);  // [ (a0 + a1) << 3, ... ]
+    const __m128i tmp2 = _mm_madd_epi16(a01, k88m);  // [ (a0 - a1) << 3, ... ]
+    const __m128i tmp1_1 = _mm_madd_epi16(a32, k5352_2217p);
+    const __m128i tmp3_1 = _mm_madd_epi16(a32, k5352_2217m);
+    const __m128i tmp1_2 = _mm_add_epi32(tmp1_1, k1812);
+    const __m128i tmp3_2 = _mm_add_epi32(tmp3_1, k937);
+    const __m128i tmp1   = _mm_srai_epi32(tmp1_2, 9);
+    const __m128i tmp3   = _mm_srai_epi32(tmp3_2, 9);
+    const __m128i s03 = _mm_packs_epi32(tmp0, tmp2);
+    const __m128i s12 = _mm_packs_epi32(tmp1, tmp3);
+    const __m128i s_lo = _mm_unpacklo_epi16(s03, s12);   // 0 1 0 1 0 1...
+    const __m128i s_hi = _mm_unpackhi_epi16(s03, s12);   // 2 3 2 3 2 3
+    const __m128i v23 = _mm_unpackhi_epi32(s_lo, s_hi);
+    v01 = _mm_unpacklo_epi32(s_lo, s_hi);
+    v32 = _mm_shuffle_epi32(v23, _MM_SHUFFLE(1, 0, 3, 2));  // 3 2 3 2 3 2..
   }
 
   // Second pass
@@ -406,13 +420,12 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
     const __m128i a32 = _mm_sub_epi16(v01, v32);
     const __m128i a11 = _mm_unpackhi_epi64(a01, a01);
     const __m128i a22 = _mm_unpackhi_epi64(a32, a32);
+    const __m128i a01_plus_7 = _mm_add_epi16(a01, seven);
 
     // d0 = (a0 + a1 + 7) >> 4;
     // d2 = (a0 - a1 + 7) >> 4;
-    const __m128i b0 = _mm_add_epi16(a01, a11);
-    const __m128i b2 = _mm_sub_epi16(a01, a11);
-    const __m128i c0 = _mm_add_epi16(b0, seven);
-    const __m128i c2 = _mm_add_epi16(b2, seven);
+    const __m128i c0 = _mm_add_epi16(a01_plus_7, a11);
+    const __m128i c2 = _mm_sub_epi16(a01_plus_7, a11);
     const __m128i d0 = _mm_srai_epi16(c0, 4);
     const __m128i d2 = _mm_srai_epi16(c2, 4);
 
@@ -430,6 +443,7 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
     // f1 = f1 + (a3 != 0);
     // The compare will return (0xffff, 0) for (==0, !=0). To turn that into the
     // desired (0, 1), we add one earlier through k12000_plus_one.
+    // -> f1 = f1 + 1 - (a3 == 0)
     const __m128i g1 = _mm_add_epi16(f1, _mm_cmpeq_epi16(a32, zero));
 
     _mm_storel_epi64((__m128i*)&out[ 0], d0);
@@ -442,10 +456,101 @@ static void FTransformSSE2(const uint8_t* src, const uint8_t* ref,
 //------------------------------------------------------------------------------
 // Metric
 
+static int SSE_Nx4SSE2(const uint8_t* a, const uint8_t* b,
+                       int num_quads, int do_16) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum1 = zero;
+  __m128i sum2 = zero;
+
+  while (num_quads-- > 0) {
+    // Note: for the !do_16 case, we read 16 pixels instead of 8 but that's ok,
+    // thanks to buffer over-allocation to that effect.
+    const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]);
+    const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]);
+    const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]);
+    const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]);
+    const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]);
+    const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]);
+    const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]);
+    const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]);
+
+    // compute clip0(a-b) and clip0(b-a)
+    const __m128i a0p = _mm_subs_epu8(a0, b0);
+    const __m128i a0m = _mm_subs_epu8(b0, a0);
+    const __m128i a1p = _mm_subs_epu8(a1, b1);
+    const __m128i a1m = _mm_subs_epu8(b1, a1);
+    const __m128i a2p = _mm_subs_epu8(a2, b2);
+    const __m128i a2m = _mm_subs_epu8(b2, a2);
+    const __m128i a3p = _mm_subs_epu8(a3, b3);
+    const __m128i a3m = _mm_subs_epu8(b3, a3);
+
+    // compute |a-b| with 8b arithmetic as clip0(a-b) | clip0(b-a)
+    const __m128i diff0 = _mm_or_si128(a0p, a0m);
+    const __m128i diff1 = _mm_or_si128(a1p, a1m);
+    const __m128i diff2 = _mm_or_si128(a2p, a2m);
+    const __m128i diff3 = _mm_or_si128(a3p, a3m);
+
+    // unpack (only four operations, instead of eight)
+    const __m128i low0 = _mm_unpacklo_epi8(diff0, zero);
+    const __m128i low1 = _mm_unpacklo_epi8(diff1, zero);
+    const __m128i low2 = _mm_unpacklo_epi8(diff2, zero);
+    const __m128i low3 = _mm_unpacklo_epi8(diff3, zero);
+
+    // multiply with self
+    const __m128i low_madd0 = _mm_madd_epi16(low0, low0);
+    const __m128i low_madd1 = _mm_madd_epi16(low1, low1);
+    const __m128i low_madd2 = _mm_madd_epi16(low2, low2);
+    const __m128i low_madd3 = _mm_madd_epi16(low3, low3);
+
+    // collect in a cascading way
+    const __m128i low_sum0 = _mm_add_epi32(low_madd0, low_madd1);
+    const __m128i low_sum1 = _mm_add_epi32(low_madd2, low_madd3);
+    sum1 = _mm_add_epi32(sum1, low_sum0);
+    sum2 = _mm_add_epi32(sum2, low_sum1);
+
+    if (do_16) {  // if necessary, process the higher 8 bytes similarly
+      const __m128i hi0 = _mm_unpackhi_epi8(diff0, zero);
+      const __m128i hi1 = _mm_unpackhi_epi8(diff1, zero);
+      const __m128i hi2 = _mm_unpackhi_epi8(diff2, zero);
+      const __m128i hi3 = _mm_unpackhi_epi8(diff3, zero);
+
+      const __m128i hi_madd0 = _mm_madd_epi16(hi0, hi0);
+      const __m128i hi_madd1 = _mm_madd_epi16(hi1, hi1);
+      const __m128i hi_madd2 = _mm_madd_epi16(hi2, hi2);
+      const __m128i hi_madd3 = _mm_madd_epi16(hi3, hi3);
+      const __m128i hi_sum0 = _mm_add_epi32(hi_madd0, hi_madd1);
+      const __m128i hi_sum1 = _mm_add_epi32(hi_madd2, hi_madd3);
+      sum1 = _mm_add_epi32(sum1, hi_sum0);
+      sum2 = _mm_add_epi32(sum2, hi_sum1);
+    }
+    a += 4 * BPS;
+    b += 4 * BPS;
+  }
+  {
+    int32_t tmp[4];
+    const __m128i sum = _mm_add_epi32(sum1, sum2);
+    _mm_storeu_si128((__m128i*)tmp, sum);
+    return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
+  }
+}
+
+static int SSE16x16SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 4, 1);
+}
+
+static int SSE16x8SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 2, 1);
+}
+
+static int SSE8x8SSE2(const uint8_t* a, const uint8_t* b) {
+  return SSE_Nx4SSE2(a, b, 2, 0);
+}
+
 static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
-  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i zero = _mm_setzero_si128();
 
-  // Load values.
+  // Load values. Note that we read 8 pixels instead of 4,
+  // but the a/b buffers are over-allocated to that effect.
   const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]);
   const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]);
   const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]);
@@ -483,6 +588,7 @@ static int SSE4x4SSE2(const uint8_t* a, const uint8_t* b) {
   const __m128i sum0 = _mm_add_epi32(madd0, madd1);
   const __m128i sum1 = _mm_add_epi32(madd2, madd3);
   const __m128i sum2 = _mm_add_epi32(sum0, sum1);
+
   int32_t tmp[4];
   _mm_storeu_si128((__m128i*)tmp, sum2);
   return (tmp[3] + tmp[2] + tmp[1] + tmp[0]);
@@ -502,8 +608,6 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
   int32_t sum[4];
   __m128i tmp_0, tmp_1, tmp_2, tmp_3;
   const __m128i zero = _mm_setzero_si128();
-  const __m128i one = _mm_set1_epi16(1);
-  const __m128i three = _mm_set1_epi16(3);
 
   // Load, combine and tranpose inputs.
   {
@@ -550,17 +654,14 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
   // Horizontal pass and subsequent transpose.
   {
     // Calculate a and b (two 4x4 at once).
-    const __m128i a0 = _mm_slli_epi16(_mm_add_epi16(tmp_0, tmp_2), 2);
-    const __m128i a1 = _mm_slli_epi16(_mm_add_epi16(tmp_1, tmp_3), 2);
-    const __m128i a2 = _mm_slli_epi16(_mm_sub_epi16(tmp_1, tmp_3), 2);
-    const __m128i a3 = _mm_slli_epi16(_mm_sub_epi16(tmp_0, tmp_2), 2);
-    // b0_extra = (a0 != 0);
-    const __m128i b0_extra = _mm_andnot_si128(_mm_cmpeq_epi16 (a0, zero), one);
-    const __m128i b0_base = _mm_add_epi16(a0, a1);
+    const __m128i a0 = _mm_add_epi16(tmp_0, tmp_2);
+    const __m128i a1 = _mm_add_epi16(tmp_1, tmp_3);
+    const __m128i a2 = _mm_sub_epi16(tmp_1, tmp_3);
+    const __m128i a3 = _mm_sub_epi16(tmp_0, tmp_2);
+    const __m128i b0 = _mm_add_epi16(a0, a1);
     const __m128i b1 = _mm_add_epi16(a3, a2);
     const __m128i b2 = _mm_sub_epi16(a3, a2);
     const __m128i b3 = _mm_sub_epi16(a0, a1);
-    const __m128i b0 = _mm_add_epi16(b0_base, b0_extra);
     // a00 a01 a02 a03   b00 b01 b02 b03
     // a10 a11 a12 a13   b10 b11 b12 b13
     // a20 a21 a22 a23   b20 b21 b22 b23
@@ -635,19 +736,6 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
       B_b2 = _mm_sub_epi16(B_b2, sign_B_b2);
     }
 
-    // b = abs(b) + 3
-    A_b0 = _mm_add_epi16(A_b0, three);
-    A_b2 = _mm_add_epi16(A_b2, three);
-    B_b0 = _mm_add_epi16(B_b0, three);
-    B_b2 = _mm_add_epi16(B_b2, three);
-
-    // abs((b + (b<0) + 3) >> 3) = (abs(b) + 3) >> 3
-    // b = (abs(b) + 3) >> 3
-    A_b0 = _mm_srai_epi16(A_b0, 3);
-    A_b2 = _mm_srai_epi16(A_b2, 3);
-    B_b0 = _mm_srai_epi16(B_b0, 3);
-    B_b2 = _mm_srai_epi16(B_b2, 3);
-
     // weighted sums
     A_b0 = _mm_madd_epi16(A_b0, w_0);
     A_b2 = _mm_madd_epi16(A_b2, w_8);
@@ -666,7 +754,7 @@ static int TTransformSSE2(const uint8_t* inA, const uint8_t* inB,
 static int Disto4x4SSE2(const uint8_t* const a, const uint8_t* const b,
                         const uint16_t* const w) {
   const int diff_sum = TTransformSSE2(a, b, w);
-  return (abs(diff_sum) + 8) >> 4;
+  return abs(diff_sum) >> 5;
 }
 
 static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
@@ -681,7 +769,6 @@ static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
   return D;
 }
 
-
 //------------------------------------------------------------------------------
 // Quantization
 //
@@ -690,8 +777,7 @@ static int Disto16x16SSE2(const uint8_t* const a, const uint8_t* const b,
 static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
                              int n, const VP8Matrix* const mtx) {
   const __m128i max_coeff_2047 = _mm_set1_epi16(2047);
-  const __m128i zero = _mm_set1_epi16(0);
-  __m128i sign0, sign8;
+  const __m128i zero = _mm_setzero_si128();
   __m128i coeff0, coeff8;
   __m128i out0, out8;
   __m128i packed_out;
@@ -713,8 +799,8 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   const __m128i zthresh8 = _mm_loadu_si128((__m128i*)&mtx->zthresh_[8]);
 
   // sign(in) = in >> 15  (0x0000 if positive, 0xffff if negative)
-  sign0 = _mm_srai_epi16(in0, 15);
-  sign8 = _mm_srai_epi16(in8, 15);
+  const __m128i sign0 = _mm_srai_epi16(in0, 15);
+  const __m128i sign8 = _mm_srai_epi16(in8, 15);
 
   // coeff = abs(in) = (in ^ sign) - sign
   coeff0 = _mm_xor_si128(in0, sign0);
@@ -819,19 +905,28 @@ static int QuantizeBlockSSE2(int16_t in[16], int16_t out[16],
   }
 }
 
+#endif   // WEBP_USE_SSE2
+
+//------------------------------------------------------------------------------
+// Entry point
+
 extern void VP8EncDspInitSSE2(void);
+
 void VP8EncDspInitSSE2(void) {
+#if defined(WEBP_USE_SSE2)
   VP8CollectHistogram = CollectHistogramSSE2;
   VP8EncQuantizeBlock = QuantizeBlockSSE2;
   VP8ITransform = ITransformSSE2;
   VP8FTransform = FTransformSSE2;
+  VP8SSE16x16 = SSE16x16SSE2;
+  VP8SSE16x8 = SSE16x8SSE2;
+  VP8SSE8x8 = SSE8x8SSE2;
   VP8SSE4x4 = SSE4x4SSE2;
   VP8TDisto4x4 = Disto4x4SSE2;
   VP8TDisto16x16 = Disto16x16SSE2;
+#endif   // WEBP_USE_SSE2
 }
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
-
-#endif   // WEBP_USE_SSE2
diff --git a/src/dsp/lossless.c b/src/dsp/lossless.c
index 6d3094fd..080b3e63 100644
--- a/src/dsp/lossless.c
+++ b/src/dsp/lossless.c
@@ -11,27 +11,31 @@
 //          Jyrki Alakuijala (jyrki@google.com)
 //          Urvang Joshi (urvang@google.com)
 
-#define ANDROID_WEBP_RGB
+#include "./dsp.h"
+
+// Define the following if target arch is sure to have SSE2
+// #define WEBP_TARGET_HAS_SSE2
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
 #endif
 
+#if defined(WEBP_TARGET_HAS_SSE2)
+#include <emmintrin.h>
+#endif
+
 #include <math.h>
 #include <stdlib.h>
 #include "./lossless.h"
 #include "../dec/vp8li.h"
-#include "../dsp/yuv.h"
-#include "../dsp/dsp.h"
-#include "../enc/histogram.h"
+#include "./yuv.h"
 
 #define MAX_DIFF_COST (1e30f)
 
 // lookup table for small values of log2(int)
 #define APPROX_LOG_MAX  4096
 #define LOG_2_RECIPROCAL 1.44269504088896338700465094007086
-#define LOG_LOOKUP_IDX_MAX 256
-static const float kLog2Table[LOG_LOOKUP_IDX_MAX] = {
+const float kLog2Table[LOG_LOOKUP_IDX_MAX] = {
   0.0000000000000000f, 0.0000000000000000f,
   1.0000000000000000f, 1.5849625007211560f,
   2.0000000000000000f, 2.3219280948873621f,
@@ -162,16 +166,97 @@ static const float kLog2Table[LOG_LOOKUP_IDX_MAX] = {
   7.9886846867721654f, 7.9943534368588577f
 };
 
-float VP8LFastLog2(int v) {
-  if (v < LOG_LOOKUP_IDX_MAX) {
-    return kLog2Table[v];
-  } else if (v < APPROX_LOG_MAX) {
+const float kSLog2Table[LOG_LOOKUP_IDX_MAX] = {
+  0.00000000f,    0.00000000f,  2.00000000f,   4.75488750f,
+  8.00000000f,   11.60964047f,  15.50977500f,  19.65148445f,
+  24.00000000f,  28.52932501f,  33.21928095f,  38.05374781f,
+  43.01955001f,  48.10571634f,  53.30296891f,  58.60335893f,
+  64.00000000f,  69.48686830f,  75.05865003f,  80.71062276f,
+  86.43856190f,  92.23866588f,  98.10749561f,  104.04192499f,
+  110.03910002f, 116.09640474f, 122.21143267f, 128.38196256f,
+  134.60593782f, 140.88144886f, 147.20671787f, 153.58008562f,
+  160.00000000f, 166.46500594f, 172.97373660f, 179.52490559f,
+  186.11730005f, 192.74977453f, 199.42124551f, 206.13068654f,
+  212.87712380f, 219.65963219f, 226.47733176f, 233.32938445f,
+  240.21499122f, 247.13338933f, 254.08384998f, 261.06567603f,
+  268.07820003f, 275.12078236f, 282.19280949f, 289.29369244f,
+  296.42286534f, 303.57978409f, 310.76392512f, 317.97478424f,
+  325.21187564f, 332.47473081f, 339.76289772f, 347.07593991f,
+  354.41343574f, 361.77497759f, 369.16017124f, 376.56863518f,
+  384.00000000f, 391.45390785f, 398.93001188f, 406.42797576f,
+  413.94747321f, 421.48818752f, 429.04981119f, 436.63204548f,
+  444.23460010f, 451.85719280f, 459.49954906f, 467.16140179f,
+  474.84249102f, 482.54256363f, 490.26137307f, 497.99867911f,
+  505.75424759f, 513.52785023f, 521.31926438f, 529.12827280f,
+  536.95466351f, 544.79822957f, 552.65876890f, 560.53608414f,
+  568.42998244f, 576.34027536f, 584.26677867f, 592.20931226f,
+  600.16769996f, 608.14176943f, 616.13135206f, 624.13628279f,
+  632.15640007f, 640.19154569f, 648.24156472f, 656.30630539f,
+  664.38561898f, 672.47935976f, 680.58738488f, 688.70955430f,
+  696.84573069f, 704.99577935f, 713.15956818f, 721.33696754f,
+  729.52785023f, 737.73209140f, 745.94956849f, 754.18016116f,
+  762.42375127f, 770.68022275f, 778.94946161f, 787.23135586f,
+  795.52579543f, 803.83267219f, 812.15187982f, 820.48331383f,
+  828.82687147f, 837.18245171f, 845.54995518f, 853.92928416f,
+  862.32034249f, 870.72303558f, 879.13727036f, 887.56295522f,
+  896.00000000f, 904.44831595f, 912.90781569f, 921.37841320f,
+  929.86002376f, 938.35256392f, 946.85595152f, 955.37010560f,
+  963.89494641f, 972.43039537f, 980.97637504f, 989.53280911f,
+  998.09962237f, 1006.67674069f, 1015.26409097f, 1023.86160116f,
+  1032.46920021f, 1041.08681805f, 1049.71438560f, 1058.35183469f,
+  1066.99909811f, 1075.65610955f, 1084.32280357f, 1092.99911564f,
+  1101.68498204f, 1110.38033993f, 1119.08512727f, 1127.79928282f,
+  1136.52274614f, 1145.25545758f, 1153.99735821f, 1162.74838989f,
+  1171.50849518f, 1180.27761738f, 1189.05570047f, 1197.84268914f,
+  1206.63852876f, 1215.44316535f, 1224.25654560f, 1233.07861684f,
+  1241.90932703f, 1250.74862473f, 1259.59645914f, 1268.45278005f,
+  1277.31753781f, 1286.19068338f, 1295.07216828f, 1303.96194457f,
+  1312.85996488f, 1321.76618236f, 1330.68055071f, 1339.60302413f,
+  1348.53355734f, 1357.47210556f, 1366.41862452f, 1375.37307041f,
+  1384.33539991f, 1393.30557020f, 1402.28353887f, 1411.26926400f,
+  1420.26270412f, 1429.26381818f, 1438.27256558f, 1447.28890615f,
+  1456.31280014f, 1465.34420819f, 1474.38309138f, 1483.42941118f,
+  1492.48312945f, 1501.54420843f, 1510.61261078f, 1519.68829949f,
+  1528.77123795f, 1537.86138993f, 1546.95871952f, 1556.06319119f,
+  1565.17476976f, 1574.29342040f, 1583.41910860f, 1592.55180020f,
+  1601.69146137f, 1610.83805860f, 1619.99155871f, 1629.15192882f,
+  1638.31913637f, 1647.49314911f, 1656.67393509f, 1665.86146266f,
+  1675.05570047f, 1684.25661744f, 1693.46418280f, 1702.67836605f,
+  1711.89913698f, 1721.12646563f, 1730.36032233f, 1739.60067768f,
+  1748.84750254f, 1758.10076802f, 1767.36044551f, 1776.62650662f,
+  1785.89892323f, 1795.17766747f, 1804.46271172f, 1813.75402857f,
+  1823.05159087f, 1832.35537170f, 1841.66534438f, 1850.98148244f,
+  1860.30375965f, 1869.63214999f, 1878.96662767f, 1888.30716711f,
+  1897.65374295f, 1907.00633003f, 1916.36490342f, 1925.72943838f,
+  1935.09991037f, 1944.47629506f, 1953.85856831f, 1963.24670620f,
+  1972.64068498f, 1982.04048108f, 1991.44607117f, 2000.85743204f,
+  2010.27454072f, 2019.69737440f, 2029.12591044f, 2038.56012640f
+};
+
+float VP8LFastSLog2Slow(int v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_MAX) {
     int log_cnt = 0;
+    const float v_f = (float)v;
     while (v >= LOG_LOOKUP_IDX_MAX) {
       ++log_cnt;
       v = v >> 1;
     }
-    return kLog2Table[v] + (float)log_cnt;
+    return v_f * (kLog2Table[v] + log_cnt);
+  } else {
+    return (float)(LOG_2_RECIPROCAL * v * log((double)v));
+  }
+}
+
+float VP8LFastLog2Slow(int v) {
+  assert(v >= LOG_LOOKUP_IDX_MAX);
+  if (v < APPROX_LOG_MAX) {
+    int log_cnt = 0;
+    while (v >= LOG_LOOKUP_IDX_MAX) {
+      ++log_cnt;
+      v = v >> 1;
+    }
+    return kLog2Table[v] + log_cnt;
   } else {
     return (float)(LOG_2_RECIPROCAL * log((double)v));
   }
@@ -200,6 +285,61 @@ static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1,
   return Average2(Average2(a0, a1), Average2(a2, a3));
 }
 
+#if defined(WEBP_TARGET_HAS_SSE2)
+static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero);
+  const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero);
+  const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i V1 = _mm_add_epi16(C0, C1);
+  const __m128i V2 = _mm_sub_epi16(V1, C2);
+  const __m128i b = _mm_packus_epi16(V2, V2);
+  const uint32_t output = _mm_cvtsi128_si32(b);
+  return output;
+}
+
+static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
+                                                   uint32_t c2) {
+  const uint32_t ave = Average2(c0, c1);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(ave), zero);
+  const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero);
+  const __m128i A1 = _mm_sub_epi16(A0, B0);
+  const __m128i BgtA = _mm_cmpgt_epi16(B0, A0);
+  const __m128i A2 = _mm_sub_epi16(A1, BgtA);
+  const __m128i A3 = _mm_srai_epi16(A2, 1);
+  const __m128i A4 = _mm_add_epi16(A0, A3);
+  const __m128i A5 = _mm_packus_epi16(A4, A4);
+  const uint32_t output = _mm_cvtsi128_si32(A5);
+  return output;
+}
+
+static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
+  int pa_minus_pb;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i A0 = _mm_cvtsi32_si128(a);
+  const __m128i B0 = _mm_cvtsi32_si128(b);
+  const __m128i C0 = _mm_cvtsi32_si128(c);
+  const __m128i AC0 = _mm_subs_epu8(A0, C0);
+  const __m128i CA0 = _mm_subs_epu8(C0, A0);
+  const __m128i BC0 = _mm_subs_epu8(B0, C0);
+  const __m128i CB0 = _mm_subs_epu8(C0, B0);
+  const __m128i AC = _mm_or_si128(AC0, CA0);
+  const __m128i BC = _mm_or_si128(BC0, CB0);
+  const __m128i pa = _mm_unpacklo_epi8(AC, zero);  // |a - c|
+  const __m128i pb = _mm_unpacklo_epi8(BC, zero);  // |b - c|
+  const __m128i diff = _mm_sub_epi16(pb, pa);
+  {
+    int16_t out[8];
+    _mm_storeu_si128((__m128i*)out, diff);
+    pa_minus_pb = out[0] + out[1] + out[2] + out[3];
+  }
+  return (pa_minus_pb <= 0) ? a : b;
+}
+
+#else
+
 static WEBP_INLINE uint32_t Clip255(uint32_t a) {
   if (a < 256) {
     return a;
@@ -241,9 +381,9 @@ static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1,
 }
 
 static WEBP_INLINE int Sub3(int a, int b, int c) {
-  const int pa = b - c;
-  const int pb = a - c;
-  return abs(pa) - abs(pb);
+  const int pb = b - c;
+  const int pa = a - c;
+  return abs(pb) - abs(pa);
 }
 
 static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
@@ -252,9 +392,9 @@ static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) {
       Sub3((a >> 16) & 0xff, (b >> 16) & 0xff, (c >> 16) & 0xff) +
       Sub3((a >>  8) & 0xff, (b >>  8) & 0xff, (c >>  8) & 0xff) +
       Sub3((a      ) & 0xff, (b      ) & 0xff, (c      ) & 0xff);
-
   return (pa_minus_pb <= 0) ? a : b;
 }
+#endif
 
 //------------------------------------------------------------------------------
 // Predictors
@@ -342,35 +482,36 @@ static float PredictionCostSpatial(const int* counts,
   return (float)(-0.1 * bits);
 }
 
-// Compute the Shanon's entropy: Sum(p*log2(p))
-static float ShannonEntropy(const int* const array, int n) {
+// Compute the combined Shanon's entropy for distribution {X} and {X+Y}
+static float CombinedShannonEntropy(const int* const X,
+                                    const int* const Y, int n) {
   int i;
-  float retval = 0.f;
-  int sum = 0;
+  double retval = 0.;
+  int sumX = 0, sumXY = 0;
   for (i = 0; i < n; ++i) {
-    if (array[i] != 0) {
-      sum += array[i];
-      retval -= VP8LFastSLog2(array[i]);
+    const int x = X[i];
+    const int xy = X[i] + Y[i];
+    if (x != 0) {
+      sumX += x;
+      retval -= VP8LFastSLog2(x);
+    }
+    if (xy != 0) {
+      sumXY += xy;
+      retval -= VP8LFastSLog2(xy);
     }
   }
-  retval += VP8LFastSLog2(sum);
-  return retval;
+  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
+  return (float)retval;
 }
 
 static float PredictionCostSpatialHistogram(int accumulated[4][256],
                                             int tile[4][256]) {
   int i;
-  int k;
-  int combo[256];
   double retval = 0;
   for (i = 0; i < 4; ++i) {
-    const double exp_val = 0.94;
-    retval += PredictionCostSpatial(&tile[i][0], 1, exp_val);
-    retval += ShannonEntropy(&tile[i][0], 256);
-    for (k = 0; k < 256; ++k) {
-      combo[k] = accumulated[i][k] + tile[i][k];
-    }
-    retval += ShannonEntropy(&combo[0], 256);
+    const double kExpValue = 0.94;
+    retval += PredictionCostSpatial(tile[i], 1, kExpValue);
+    retval += CombinedShannonEntropy(tile[i], accumulated[i], 256);
   }
   return (float)retval;
 }
@@ -574,8 +715,21 @@ static void PredictorInverseTransform(const VP8LTransform* const transform,
 }
 
 void VP8LSubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixs) {
-  int i;
-  for (i = 0; i < num_pixs; ++i) {
+  int i = 0;
+#if defined(WEBP_TARGET_HAS_SSE2)
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  for (; i + 4 < num_pixs; i += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_sub_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)&argb_data[i], out);
+  }
+  // fallthrough and finish off with plain-C
+#endif
+  for (; i < num_pixs; ++i) {
     const uint32_t argb = argb_data[i];
     const uint32_t green = (argb >> 8) & 0xff;
     const uint32_t new_r = (((argb >> 16) & 0xff) - green) & 0xff;
@@ -590,9 +744,21 @@ static void AddGreenToBlueAndRed(const VP8LTransform* const transform,
                                  int y_start, int y_end, uint32_t* data) {
   const int width = transform->xsize_;
   const uint32_t* const data_end = data + (y_end - y_start) * width;
+#if defined(WEBP_TARGET_HAS_SSE2)
+  const __m128i mask = _mm_set1_epi32(0x0000ff00);
+  for (; data + 4 < data_end; data += 4) {
+    const __m128i in = _mm_loadu_si128((__m128i*)data);
+    const __m128i in_00g0 = _mm_and_si128(in, mask);     // 00g0|00g0|...
+    const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8);  // 0g00|0g00|...
+    const __m128i in_000g = _mm_srli_epi32(in_00g0, 8);  // 000g|000g|...
+    const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g);
+    const __m128i out = _mm_add_epi8(in, in_0g0g);
+    _mm_storeu_si128((__m128i*)data, out);
+  }
+  // fallthrough and finish off with plain-C
+#endif
   while (data < data_end) {
     const uint32_t argb = *data;
-    // "* 0001001u" is equivalent to "(green << 16) + green)"
     const uint32_t green = ((argb >> 8) & 0xff);
     uint32_t red_blue = (argb & 0x00ff00ffu);
     red_blue += (green << 16) | green;
@@ -657,6 +823,25 @@ static WEBP_INLINE uint32_t TransformColor(const Multipliers* const m,
   return (argb & 0xff00ff00u) | (new_red << 16) | (new_blue);
 }
 
+static WEBP_INLINE uint8_t TransformColorRed(uint8_t green_to_red,
+                                             uint32_t argb) {
+  const uint32_t green = argb >> 8;
+  uint32_t new_red = argb >> 16;
+  new_red -= ColorTransformDelta(green_to_red, green);
+  return (new_red & 0xff);
+}
+
+static WEBP_INLINE uint8_t TransformColorBlue(uint8_t green_to_blue,
+                                              uint8_t red_to_blue,
+                                              uint32_t argb) {
+  const uint32_t green = argb >> 8;
+  const uint32_t red = argb >> 16;
+  uint8_t new_blue = argb;
+  new_blue -= ColorTransformDelta(green_to_blue, green);
+  new_blue -= ColorTransformDelta(red_to_blue, red);
+  return (new_blue & 0xff);
+}
+
 static WEBP_INLINE int SkipRepeatedPixels(const uint32_t* const argb,
                                           int ix, int xsize) {
   const uint32_t v = argb[ix];
@@ -677,14 +862,10 @@ static WEBP_INLINE int SkipRepeatedPixels(const uint32_t* const argb,
 static float PredictionCostCrossColor(const int accumulated[256],
                                       const int counts[256]) {
   // Favor low entropy, locally and globally.
-  int i;
-  int combo[256];
-  for (i = 0; i < 256; ++i) {
-    combo[i] = accumulated[i] + counts[i];
-  }
-  return ShannonEntropy(combo, 256) +
-         ShannonEntropy(counts, 256) +
-         PredictionCostSpatial(counts, 3, 2.4);  // Favor small absolute values.
+  // Favor small absolute values for PredictionCostSpatial
+  static const double kExpValue = 2.4;
+  return CombinedShannonEntropy(counts, accumulated, 256) +
+         PredictionCostSpatial(counts, 3, kExpValue);
 }
 
 static Multipliers GetBestColorTransformForTile(
@@ -714,85 +895,75 @@ static Multipliers GetBestColorTransformForTile(
   if (all_y_max > ysize) {
     all_y_max = ysize;
   }
+
   for (green_to_red = -64; green_to_red <= 64; green_to_red += halfstep) {
     int histo[256] = { 0 };
     int all_y;
-    Multipliers tx;
-    MultipliersClear(&tx);
-    tx.green_to_red_ = green_to_red & 0xff;
 
     for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
-      uint32_t predict;
       int ix = all_y * xsize + tile_x_offset;
       int all_x;
       for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
         if (SkipRepeatedPixels(argb, ix, xsize)) {
           continue;
         }
-        predict = TransformColor(&tx, argb[ix], 0);
-        ++histo[(predict >> 16) & 0xff];  // red.
+        ++histo[TransformColorRed(green_to_red, argb[ix])];  // red.
       }
     }
     cur_diff = PredictionCostCrossColor(&accumulated_red_histo[0], &histo[0]);
-    if (tx.green_to_red_ == prevX.green_to_red_) {
+    if ((uint8_t)green_to_red == prevX.green_to_red_) {
       cur_diff -= 3;  // favor keeping the areas locally similar
     }
-    if (tx.green_to_red_ == prevY.green_to_red_) {
+    if ((uint8_t)green_to_red == prevY.green_to_red_) {
       cur_diff -= 3;  // favor keeping the areas locally similar
     }
-    if (tx.green_to_red_ == 0) {
+    if (green_to_red == 0) {
       cur_diff -= 3;
     }
     if (cur_diff < best_diff) {
       best_diff = cur_diff;
-      best_tx = tx;
+      best_tx.green_to_red_ = green_to_red;
     }
   }
   best_diff = MAX_DIFF_COST;
-  green_to_red = best_tx.green_to_red_;
   for (green_to_blue = -32; green_to_blue <= 32; green_to_blue += step) {
     for (red_to_blue = -32; red_to_blue <= 32; red_to_blue += step) {
       int all_y;
       int histo[256] = { 0 };
-      Multipliers tx;
-      tx.green_to_red_ = green_to_red;
-      tx.green_to_blue_ = green_to_blue;
-      tx.red_to_blue_ = red_to_blue;
       for (all_y = tile_y_offset; all_y < all_y_max; ++all_y) {
-        uint32_t predict;
         int all_x;
         int ix = all_y * xsize + tile_x_offset;
         for (all_x = tile_x_offset; all_x < all_x_max; ++all_x, ++ix) {
           if (SkipRepeatedPixels(argb, ix, xsize)) {
             continue;
           }
-          predict = TransformColor(&tx, argb[ix], 0);
-          ++histo[predict & 0xff];  // blue.
+          ++histo[TransformColorBlue(green_to_blue, red_to_blue, argb[ix])];
         }
       }
       cur_diff =
-        PredictionCostCrossColor(&accumulated_blue_histo[0], &histo[0]);
-      if (tx.green_to_blue_ == prevX.green_to_blue_) {
+          PredictionCostCrossColor(&accumulated_blue_histo[0], &histo[0]);
+      if ((uint8_t)green_to_blue == prevX.green_to_blue_) {
         cur_diff -= 3;  // favor keeping the areas locally similar
       }
-      if (tx.green_to_blue_ == prevY.green_to_blue_) {
+      if ((uint8_t)green_to_blue == prevY.green_to_blue_) {
         cur_diff -= 3;  // favor keeping the areas locally similar
       }
-      if (tx.red_to_blue_ == prevX.red_to_blue_) {
+      if ((uint8_t)red_to_blue == prevX.red_to_blue_) {
         cur_diff -= 3;  // favor keeping the areas locally similar
       }
-      if (tx.red_to_blue_ == prevY.red_to_blue_) {
+      if ((uint8_t)red_to_blue == prevY.red_to_blue_) {
         cur_diff -= 3;  // favor keeping the areas locally similar
       }
-      if (tx.green_to_blue_ == 0) {
+      if (green_to_blue == 0) {
         cur_diff -= 3;
       }
-      if (tx.red_to_blue_ == 0) {
+      if (red_to_blue == 0) {
         cur_diff -= 3;
       }
       if (cur_diff < best_diff) {
         best_diff = cur_diff;
-        best_tx = tx;
+        best_tx.green_to_blue_ = green_to_blue;
+        best_tx.red_to_blue_ = red_to_blue;
       }
     }
   }
@@ -1036,12 +1207,14 @@ static void ConvertBGRAToRGBA4444(const uint32_t* src,
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
-#ifdef ANDROID_WEBP_RGB
-    *dst++ = ((argb >>  0) & 0xf0) | ((argb >> 28) & 0xf);
-    *dst++ = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
+    const uint8_t rg = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
+    const uint8_t ba = ((argb >>  0) & 0xf0) | ((argb >> 28) & 0xf);
+#ifdef WEBP_SWAP_16BIT_CSP
+    *dst++ = ba;
+    *dst++ = rg;
 #else
-    *dst++ = ((argb >> 16) & 0xf0) | ((argb >> 12) & 0xf);
-    *dst++ = ((argb >>  0) & 0xf0) | ((argb >> 28) & 0xf);
+    *dst++ = rg;
+    *dst++ = ba;
 #endif
   }
 }
@@ -1051,12 +1224,14 @@ static void ConvertBGRAToRGB565(const uint32_t* src,
   const uint32_t* const src_end = src + num_pixels;
   while (src < src_end) {
     const uint32_t argb = *src++;
-#ifdef ANDROID_WEBP_RGB
-    *dst++ = ((argb >>  5) & 0xe0) | ((argb >>  3) & 0x1f);
-    *dst++ = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
+    const uint8_t rg = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
+    const uint8_t gb = ((argb >>  5) & 0xe0) | ((argb >>  3) & 0x1f);
+#ifdef WEBP_SWAP_16BIT_CSP
+    *dst++ = gb;
+    *dst++ = rg;
 #else
-    *dst++ = ((argb >> 16) & 0xf8) | ((argb >> 13) & 0x7);
-    *dst++ = ((argb >>  5) & 0xe0) | ((argb >>  3) & 0x1f);
+    *dst++ = rg;
+    *dst++ = gb;
 #endif
   }
 }
@@ -1078,20 +1253,27 @@ static void CopyOrSwap(const uint32_t* src, int num_pixels, uint8_t* dst,
     const uint32_t* const src_end = src + num_pixels;
     while (src < src_end) {
       uint32_t argb = *src++;
+
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
 #if !defined(__BIG_ENDIAN__) && (defined(__i386__) || defined(__x86_64__))
       __asm__ volatile("bswap %0" : "=r"(argb) : "0"(argb));
       *(uint32_t*)dst = argb;
-      dst += sizeof(argb);
 #elif !defined(__BIG_ENDIAN__) && defined(_MSC_VER)
       argb = _byteswap_ulong(argb);
       *(uint32_t*)dst = argb;
-      dst += sizeof(argb);
 #else
-      *dst++ = (argb >> 24) & 0xff;
-      *dst++ = (argb >> 16) & 0xff;
-      *dst++ = (argb >>  8) & 0xff;
-      *dst++ = (argb >>  0) & 0xff;
+      dst[0] = (argb >> 24) & 0xff;
+      dst[1] = (argb >> 16) & 0xff;
+      dst[2] = (argb >>  8) & 0xff;
+      dst[3] = (argb >>  0) & 0xff;
+#endif
+#else   // WEBP_REFERENCE_IMPLEMENTATION
+      dst[0] = (argb >> 24) & 0xff;
+      dst[1] = (argb >> 16) & 0xff;
+      dst[2] = (argb >>  8) & 0xff;
+      dst[3] = (argb >>  0) & 0xff;
 #endif
+      dst += sizeof(argb);
     }
   } else {
     memcpy(dst, src, num_pixels * sizeof(*src));
diff --git a/src/dsp/lossless.h b/src/dsp/lossless.h
index 22a91cbf..f14e2ea7 100644
--- a/src/dsp/lossless.h
+++ b/src/dsp/lossless.h
@@ -59,10 +59,20 @@ static WEBP_INLINE uint32_t VP8LSubSampleSize(uint32_t size,
   return (size + (1 << sampling_bits) - 1) >> sampling_bits;
 }
 
-// Faster logarithm for integers, with the property of log2(0) == 0.
-float VP8LFastLog2(int v);
+// Faster logarithm for integers. Small values use a look-up table.
+#define LOG_LOOKUP_IDX_MAX 256
+extern const float kLog2Table[LOG_LOOKUP_IDX_MAX];
+extern const float kSLog2Table[LOG_LOOKUP_IDX_MAX];
+extern float VP8LFastLog2Slow(int v);
+extern float VP8LFastSLog2Slow(int v);
+static WEBP_INLINE float VP8LFastLog2(int v) {
+  return (v < LOG_LOOKUP_IDX_MAX) ? kLog2Table[v] : VP8LFastLog2Slow(v);
+}
 // Fast calculation of v * log2(v) for integer input.
-static WEBP_INLINE float VP8LFastSLog2(int v) { return VP8LFastLog2(v) * v; }
+static WEBP_INLINE float VP8LFastSLog2(int v) {
+  return (v < LOG_LOOKUP_IDX_MAX) ? kSLog2Table[v] : VP8LFastSLog2Slow(v);
+}
+
 
 // In-place difference of each component with mod 256.
 static WEBP_INLINE uint32_t VP8LSubPixels(uint32_t a, uint32_t b) {
diff --git a/src/dsp/upsampling.c b/src/dsp/upsampling.c
index 4855eb14..91d939cd 100644
--- a/src/dsp/upsampling.c
+++ b/src/dsp/upsampling.c
@@ -328,6 +328,11 @@ void WebPInitUpsamplers(void) {
       WebPInitUpsamplersSSE2();
     }
 #endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      WebPInitUpsamplersNEON();
+    }
+#endif
   }
 #endif  // FANCY_UPSAMPLING
 }
@@ -348,6 +353,11 @@ void WebPInitPremultiply(void) {
       WebPInitPremultiplySSE2();
     }
 #endif
+#if defined(WEBP_USE_NEON)
+    if (VP8GetCPUInfo(kNEON)) {
+      WebPInitPremultiplyNEON();
+    }
+#endif
   }
 #endif  // FANCY_UPSAMPLING
 }
diff --git a/src/dsp/upsampling_neon.c b/src/dsp/upsampling_neon.c
new file mode 100644
index 00000000..00e2f892
--- /dev/null
+++ b/src/dsp/upsampling_neon.c
@@ -0,0 +1,292 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// NEON version of YUV to RGB upsampling functions.
+//
+// Author: mans@mansr.com (Mans Rullgard)
+// Based on SSE code by: somnath@google.com (Somnath Banerjee)
+
+#include "./dsp.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+#if defined(WEBP_USE_NEON)
+
+#include <assert.h>
+#include <arm_neon.h>
+#include <string.h>
+#include "./yuv.h"
+
+#ifdef FANCY_UPSAMPLING
+
+// Loads 9 pixels each from rows r1 and r2 and generates 16 pixels.
+#define UPSAMPLE_16PIXELS(r1, r2, out) {                                \
+  uint8x8_t a = vld1_u8(r1);                                            \
+  uint8x8_t b = vld1_u8(r1 + 1);                                        \
+  uint8x8_t c = vld1_u8(r2);                                            \
+  uint8x8_t d = vld1_u8(r2 + 1);                                        \
+                                                                        \
+  uint16x8_t al = vshll_n_u8(a, 1);                                     \
+  uint16x8_t bl = vshll_n_u8(b, 1);                                     \
+  uint16x8_t cl = vshll_n_u8(c, 1);                                     \
+  uint16x8_t dl = vshll_n_u8(d, 1);                                     \
+                                                                        \
+  uint8x8_t diag1, diag2;                                               \
+  uint16x8_t sl;                                                        \
+                                                                        \
+  /* a + b + c + d */                                                   \
+  sl = vaddl_u8(a,  b);                                                 \
+  sl = vaddw_u8(sl, c);                                                 \
+  sl = vaddw_u8(sl, d);                                                 \
+                                                                        \
+  al = vaddq_u16(sl, al); /* 3a +  b +  c +  d */                       \
+  bl = vaddq_u16(sl, bl); /*  a + 3b +  c +  d */                       \
+                                                                        \
+  al = vaddq_u16(al, dl); /* 3a +  b +  c + 3d */                       \
+  bl = vaddq_u16(bl, cl); /*  a + 3b + 3c +  d */                       \
+                                                                        \
+  diag2 = vshrn_n_u16(al, 3);                                           \
+  diag1 = vshrn_n_u16(bl, 3);                                           \
+                                                                        \
+  a = vrhadd_u8(a, diag1);                                              \
+  b = vrhadd_u8(b, diag2);                                              \
+  c = vrhadd_u8(c, diag2);                                              \
+  d = vrhadd_u8(d, diag1);                                              \
+                                                                        \
+  {                                                                     \
+    const uint8x8x2_t a_b = {{ a, b }};                                 \
+    const uint8x8x2_t c_d = {{ c, d }};                                 \
+    vst2_u8(out,      a_b);                                             \
+    vst2_u8(out + 32, c_d);                                             \
+  }                                                                     \
+}
+
+// Turn the macro into a function for reducing code-size when non-critical
+static void Upsample16Pixels(const uint8_t *r1, const uint8_t *r2,
+                             uint8_t *out) {
+  UPSAMPLE_16PIXELS(r1, r2, out);
+}
+
+#define UPSAMPLE_LAST_BLOCK(tb, bb, num_pixels, out) {                  \
+  uint8_t r1[9], r2[9];                                                 \
+  memcpy(r1, (tb), (num_pixels));                                       \
+  memcpy(r2, (bb), (num_pixels));                                       \
+  /* replicate last byte */                                             \
+  memset(r1 + (num_pixels), r1[(num_pixels) - 1], 9 - (num_pixels));    \
+  memset(r2 + (num_pixels), r2[(num_pixels) - 1], 9 - (num_pixels));    \
+  Upsample16Pixels(r1, r2, out);                                        \
+}
+
+#define CY  76283
+#define CVR 89858
+#define CUG 22014
+#define CVG 45773
+#define CUB 113618
+
+static const int16_t coef[4] = { CVR / 4, CUG, CVG / 2, CUB / 4 };
+
+#define CONVERT8(FMT, XSTEP, N, src_y, src_uv, out, cur_x) {            \
+  int i;                                                                \
+  for (i = 0; i < N; i += 8) {                                          \
+    int off = ((cur_x) + i) * XSTEP;                                    \
+    uint8x8_t y  = vld1_u8(src_y + (cur_x)  + i);                       \
+    uint8x8_t u  = vld1_u8((src_uv) + i);                               \
+    uint8x8_t v  = vld1_u8((src_uv) + i + 16);                          \
+    int16x8_t yy = vreinterpretq_s16_u16(vsubl_u8(y, u16));             \
+    int16x8_t uu = vreinterpretq_s16_u16(vsubl_u8(u, u128));            \
+    int16x8_t vv = vreinterpretq_s16_u16(vsubl_u8(v, u128));            \
+                                                                        \
+    int16x8_t ud = vshlq_n_s16(uu, 1);                                  \
+    int16x8_t vd = vshlq_n_s16(vv, 1);                                  \
+                                                                        \
+    int32x4_t vrl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(vv), 1),  \
+                                     vget_low_s16(vd),  cf16, 0);       \
+    int32x4_t vrh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(vv), 1), \
+                                     vget_high_s16(vd), cf16, 0);       \
+    int16x8_t vr = vcombine_s16(vrshrn_n_s32(vrl, 16),                  \
+                                vrshrn_n_s32(vrh, 16));                 \
+                                                                        \
+    int32x4_t vl = vmovl_s16(vget_low_s16(vv));                         \
+    int32x4_t vh = vmovl_s16(vget_high_s16(vv));                        \
+    int32x4_t ugl = vmlal_lane_s16(vl, vget_low_s16(uu),  cf16, 1);     \
+    int32x4_t ugh = vmlal_lane_s16(vh, vget_high_s16(uu), cf16, 1);     \
+    int32x4_t gcl = vqdmlal_lane_s16(ugl, vget_low_s16(vv),  cf16, 2);  \
+    int32x4_t gch = vqdmlal_lane_s16(ugh, vget_high_s16(vv), cf16, 2);  \
+    int16x8_t gc = vcombine_s16(vrshrn_n_s32(gcl, 16),                  \
+                                vrshrn_n_s32(gch, 16));                 \
+                                                                        \
+    int32x4_t ubl = vqdmlal_lane_s16(vshll_n_s16(vget_low_s16(uu), 1),  \
+                                     vget_low_s16(ud),  cf16, 3);       \
+    int32x4_t ubh = vqdmlal_lane_s16(vshll_n_s16(vget_high_s16(uu), 1), \
+                                     vget_high_s16(ud), cf16, 3);       \
+    int16x8_t ub = vcombine_s16(vrshrn_n_s32(ubl, 16),                  \
+                                vrshrn_n_s32(ubh, 16));                 \
+                                                                        \
+    int32x4_t rl = vaddl_s16(vget_low_s16(yy),  vget_low_s16(vr));      \
+    int32x4_t rh = vaddl_s16(vget_high_s16(yy), vget_high_s16(vr));     \
+    int32x4_t gl = vsubl_s16(vget_low_s16(yy),  vget_low_s16(gc));      \
+    int32x4_t gh = vsubl_s16(vget_high_s16(yy), vget_high_s16(gc));     \
+    int32x4_t bl = vaddl_s16(vget_low_s16(yy),  vget_low_s16(ub));      \
+    int32x4_t bh = vaddl_s16(vget_high_s16(yy), vget_high_s16(ub));     \
+                                                                        \
+    rl = vmulq_lane_s32(rl, cf32, 0);                                   \
+    rh = vmulq_lane_s32(rh, cf32, 0);                                   \
+    gl = vmulq_lane_s32(gl, cf32, 0);                                   \
+    gh = vmulq_lane_s32(gh, cf32, 0);                                   \
+    bl = vmulq_lane_s32(bl, cf32, 0);                                   \
+    bh = vmulq_lane_s32(bh, cf32, 0);                                   \
+                                                                        \
+    y = vqmovun_s16(vcombine_s16(vrshrn_n_s32(rl, 16),                  \
+                                 vrshrn_n_s32(rh, 16)));                \
+    u = vqmovun_s16(vcombine_s16(vrshrn_n_s32(gl, 16),                  \
+                                 vrshrn_n_s32(gh, 16)));                \
+    v = vqmovun_s16(vcombine_s16(vrshrn_n_s32(bl, 16),                  \
+                                 vrshrn_n_s32(bh, 16)));                \
+    STR_ ## FMT(out + off, y, u, v);                                    \
+  }                                                                     \
+}
+
+#define v255 vmov_n_u8(255)
+
+#define STR_Rgb(out, r, g, b) do {                                      \
+  const uint8x8x3_t r_g_b = {{ r, g, b }};                              \
+  vst3_u8(out, r_g_b);                                                  \
+} while (0)
+
+#define STR_Bgr(out, r, g, b) do {                                      \
+  const uint8x8x3_t b_g_r = {{ b, g, r }};                              \
+  vst3_u8(out, b_g_r);                                                  \
+} while (0)
+
+#define STR_Rgba(out, r, g, b) do {                                     \
+  const uint8x8x4_t r_g_b_v255 = {{ r, g, b, v255 }};                   \
+  vst4_u8(out, r_g_b_v255);                                             \
+} while (0)
+
+#define STR_Bgra(out, r, g, b) do {                                     \
+  const uint8x8x4_t b_g_r_v255 = {{ b, g, r, v255 }};                   \
+  vst4_u8(out, b_g_r_v255);                                             \
+} while (0)
+
+#define CONVERT1(FMT, XSTEP, N, src_y, src_uv, rgb, cur_x) {            \
+  int i;                                                                \
+  for (i = 0; i < N; i++) {                                             \
+    int off = ((cur_x) + i) * XSTEP;                                    \
+    int y = src_y[(cur_x) + i];                                         \
+    int u = (src_uv)[i];                                                \
+    int v = (src_uv)[i + 16];                                           \
+    VP8YuvTo ## FMT(y, u, v, rgb + off);                                \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, uv,                  \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  if (top_y) {                                                          \
+    CONVERT8(FMT, XSTEP, len, top_y, uv, top_dst, cur_x)                \
+  }                                                                     \
+  if (bottom_y) {                                                       \
+    CONVERT8(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x)   \
+  }                                                                     \
+}
+
+#define CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, uv,                  \
+                      top_dst, bottom_dst, cur_x, len) {                \
+  if (top_y) {                                                          \
+    CONVERT1(FMT, XSTEP, len, top_y, uv, top_dst, cur_x);               \
+  }                                                                     \
+  if (bottom_y) {                                                       \
+    CONVERT1(FMT, XSTEP, len, bottom_y, (uv) + 32, bottom_dst, cur_x);  \
+  }                                                                     \
+}
+
+#define NEON_UPSAMPLE_FUNC(FUNC_NAME, FMT, XSTEP)                       \
+static void FUNC_NAME(const uint8_t *top_y, const uint8_t *bottom_y,    \
+                      const uint8_t *top_u, const uint8_t *top_v,       \
+                      const uint8_t *cur_u, const uint8_t *cur_v,       \
+                      uint8_t *top_dst, uint8_t *bottom_dst, int len) { \
+  int block;                                                            \
+  /* 16 byte aligned array to cache reconstructed u and v */            \
+  uint8_t uv_buf[2 * 32 + 15];                                          \
+  uint8_t *const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);     \
+  const int uv_len = (len + 1) >> 1;                                    \
+  /* 9 pixels must be read-able for each block */                       \
+  const int num_blocks = (uv_len - 1) >> 3;                             \
+  const int leftover = uv_len - num_blocks * 8;                         \
+  const int last_pos = 1 + 16 * num_blocks;                             \
+                                                                        \
+  const int u_diag = ((top_u[0] + cur_u[0]) >> 1) + 1;                  \
+  const int v_diag = ((top_v[0] + cur_v[0]) >> 1) + 1;                  \
+                                                                        \
+  const int16x4_t cf16 = vld1_s16(coef);                                \
+  const int32x2_t cf32 = vmov_n_s32(CY);                                \
+  const uint8x8_t u16  = vmov_n_u8(16);                                 \
+  const uint8x8_t u128 = vmov_n_u8(128);                                \
+                                                                        \
+  /* Treat the first pixel in regular way */                            \
+  if (top_y) {                                                          \
+    const int u0 = (top_u[0] + u_diag) >> 1;                            \
+    const int v0 = (top_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(top_y[0], u0, v0, top_dst);                         \
+  }                                                                     \
+  if (bottom_y) {                                                       \
+    const int u0 = (cur_u[0] + u_diag) >> 1;                            \
+    const int v0 = (cur_v[0] + v_diag) >> 1;                            \
+    VP8YuvTo ## FMT(bottom_y[0], u0, v0, bottom_dst);                   \
+  }                                                                     \
+                                                                        \
+  for (block = 0; block < num_blocks; ++block) {                        \
+    UPSAMPLE_16PIXELS(top_u, cur_u, r_uv);                              \
+    UPSAMPLE_16PIXELS(top_v, cur_v, r_uv + 16);                         \
+    CONVERT2RGB_8(FMT, XSTEP, top_y, bottom_y, r_uv,                    \
+                  top_dst, bottom_dst, 16 * block + 1, 16);             \
+    top_u += 8;                                                         \
+    cur_u += 8;                                                         \
+    top_v += 8;                                                         \
+    cur_v += 8;                                                         \
+  }                                                                     \
+                                                                        \
+  UPSAMPLE_LAST_BLOCK(top_u, cur_u, leftover, r_uv);                    \
+  UPSAMPLE_LAST_BLOCK(top_v, cur_v, leftover, r_uv + 16);               \
+  CONVERT2RGB_1(FMT, XSTEP, top_y, bottom_y, r_uv,                      \
+                top_dst, bottom_dst, last_pos, len - last_pos);         \
+}
+
+// NEON variants of the fancy upsampler.
+NEON_UPSAMPLE_FUNC(UpsampleRgbLinePairNEON,  Rgb,  3)
+NEON_UPSAMPLE_FUNC(UpsampleBgrLinePairNEON,  Bgr,  3)
+NEON_UPSAMPLE_FUNC(UpsampleRgbaLinePairNEON, Rgba, 4)
+NEON_UPSAMPLE_FUNC(UpsampleBgraLinePairNEON, Bgra, 4)
+
+#endif  // FANCY_UPSAMPLING
+
+#endif   // WEBP_USE_NEON
+
+//------------------------------------------------------------------------------
+
+extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
+
+void WebPInitUpsamplersNEON(void) {
+#if defined(WEBP_USE_NEON)
+  WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairNEON;
+  WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairNEON;
+  WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairNEON;
+  WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairNEON;
+#endif   // WEBP_USE_NEON
+}
+
+void WebPInitPremultiplyNEON(void) {
+#if defined(WEBP_USE_NEON)
+  WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairNEON;
+  WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairNEON;
+#endif   // WEBP_USE_NEON
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
diff --git a/src/dsp/upsampling_sse2.c b/src/dsp/upsampling_sse2.c
index 8cb275a0..ba075d11 100644
--- a/src/dsp/upsampling_sse2.c
+++ b/src/dsp/upsampling_sse2.c
@@ -11,6 +11,10 @@
 
 #include "./dsp.h"
 
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
 #if defined(WEBP_USE_SSE2)
 
 #include <assert.h>
@@ -18,10 +22,6 @@
 #include <string.h>
 #include "./yuv.h"
 
-#if defined(__cplusplus) || defined(c_plusplus)
-extern "C" {
-#endif
-
 #ifdef FANCY_UPSAMPLING
 
 // We compute (9*a + 3*b + 3*c + d + 8) / 16 as follows
@@ -51,12 +51,12 @@ extern "C" {
 
 // pack and store two alterning pixel rows
 #define PACK_AND_STORE(a, b, da, db, out) do {                                 \
-  const __m128i ta = _mm_avg_epu8(a, da);  /* (9a + 3b + 3c +  d + 8) / 16 */  \
-  const __m128i tb = _mm_avg_epu8(b, db);  /* (3a + 9b +  c + 3d + 8) / 16 */  \
-  const __m128i t1 = _mm_unpacklo_epi8(ta, tb);                                \
-  const __m128i t2 = _mm_unpackhi_epi8(ta, tb);                                \
-  _mm_store_si128(((__m128i*)(out)) + 0, t1);                                  \
-  _mm_store_si128(((__m128i*)(out)) + 1, t2);                                  \
+  const __m128i t_a = _mm_avg_epu8(a, da);  /* (9a + 3b + 3c +  d + 8) / 16 */ \
+  const __m128i t_b = _mm_avg_epu8(b, db);  /* (3a + 9b +  c + 3d + 8) / 16 */ \
+  const __m128i t_1 = _mm_unpacklo_epi8(t_a, t_b);                             \
+  const __m128i t_2 = _mm_unpackhi_epi8(t_a, t_b);                             \
+  _mm_store_si128(((__m128i*)(out)) + 0, t_1);                                 \
+  _mm_store_si128(((__m128i*)(out)) + 1, t_2);                                 \
 } while (0)
 
 // Loads 17 pixels each from rows r1 and r2 and generates 32 pixels.
@@ -128,7 +128,7 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
                       const uint8_t* top_u, const uint8_t* top_v,              \
                       const uint8_t* cur_u, const uint8_t* cur_v,              \
                       uint8_t* top_dst, uint8_t* bottom_dst, int len) {        \
-  int b;                                                                       \
+  int block;                                                                   \
   /* 16 byte aligned array to cache reconstructed u and v */                   \
   uint8_t uv_buf[4 * 32 + 15];                                                 \
   uint8_t* const r_uv = (uint8_t*)((uintptr_t)(uv_buf + 15) & ~15);            \
@@ -154,11 +154,11 @@ static void FUNC_NAME(const uint8_t* top_y, const uint8_t* bottom_y,           \
     FUNC(bottom_y[0], u0, v0, bottom_dst);                                     \
   }                                                                            \
                                                                                \
-  for (b = 0; b < num_blocks; ++b) {                                           \
+  for (block = 0; block < num_blocks; ++block) {                               \
     UPSAMPLE_32PIXELS(top_u, cur_u, r_uv + 0 * 32);                            \
     UPSAMPLE_32PIXELS(top_v, cur_v, r_uv + 1 * 32);                            \
     CONVERT2RGB(FUNC, XSTEP, top_y, bottom_y, r_uv, top_dst, bottom_dst,       \
-                32 * b + 1, 32)                                                \
+                32 * block + 1, 32)                                            \
     top_u += 16;                                                               \
     cur_u += 16;                                                               \
     top_v += 16;                                                               \
@@ -184,26 +184,32 @@ SSE2_UPSAMPLE_FUNC(UpsampleBgraLinePairSSE2, VP8YuvToBgra, 4)
 #undef CONVERT2RGB
 #undef SSE2_UPSAMPLE_FUNC
 
+#endif  // FANCY_UPSAMPLING
+
+#endif   // WEBP_USE_SSE2
+
 //------------------------------------------------------------------------------
 
 extern WebPUpsampleLinePairFunc WebPUpsamplers[/* MODE_LAST */];
 
 void WebPInitUpsamplersSSE2(void) {
+#if defined(WEBP_USE_SSE2)
   WebPUpsamplers[MODE_RGB]  = UpsampleRgbLinePairSSE2;
   WebPUpsamplers[MODE_RGBA] = UpsampleRgbaLinePairSSE2;
   WebPUpsamplers[MODE_BGR]  = UpsampleBgrLinePairSSE2;
   WebPUpsamplers[MODE_BGRA] = UpsampleBgraLinePairSSE2;
+#endif   // WEBP_USE_SSE2
 }
 
 void WebPInitPremultiplySSE2(void) {
+#if defined(WEBP_USE_SSE2)
   WebPUpsamplers[MODE_rgbA] = UpsampleRgbaLinePairSSE2;
   WebPUpsamplers[MODE_bgrA] = UpsampleBgraLinePairSSE2;
+#endif   // WEBP_USE_SSE2
 }
 
-#endif  // FANCY_UPSAMPLING
-
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
 
-#endif   // WEBP_USE_SSE2
+
diff --git a/src/dsp/yuv.c b/src/dsp/yuv.c
index 7f05f9a3..38895281 100644
--- a/src/dsp/yuv.c
+++ b/src/dsp/yuv.c
@@ -33,6 +33,7 @@ void VP8YUVInit(void) {
   if (done) {
     return;
   }
+#ifndef USE_YUVj
   for (i = 0; i < 256; ++i) {
     VP8kVToR[i] = (89858 * (i - 128) + YUV_HALF) >> YUV_FIX;
     VP8kUToG[i] = -22014 * (i - 128) + YUV_HALF;
@@ -44,6 +45,20 @@ void VP8YUVInit(void) {
     VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255);
     VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15);
   }
+#else
+  for (i = 0; i < 256; ++i) {
+    VP8kVToR[i] = (91881 * (i - 128) + YUV_HALF) >> YUV_FIX;
+    VP8kUToG[i] = -22554 * (i - 128) + YUV_HALF;
+    VP8kVToG[i] = -46802 * (i - 128);
+    VP8kUToB[i] = (116130 * (i - 128) + YUV_HALF) >> YUV_FIX;
+  }
+  for (i = YUV_RANGE_MIN; i < YUV_RANGE_MAX; ++i) {
+    const int k = i;
+    VP8kClip[i - YUV_RANGE_MIN] = clip(k, 255);
+    VP8kClip4Bits[i - YUV_RANGE_MIN] = clip((k + 8) >> 4, 15);
+  }
+#endif
+
   done = 1;
 }
 
diff --git a/src/dsp/yuv.h b/src/dsp/yuv.h
index ee3587e3..add167ea 100644
--- a/src/dsp/yuv.h
+++ b/src/dsp/yuv.h
@@ -7,6 +7,25 @@
 //
 // inline YUV<->RGB conversion function
 //
+// The exact naming is Y'CbCr, following the ITU-R BT.601 standard.
+// More information at: http://en.wikipedia.org/wiki/YCbCr
+// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16
+// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128
+// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128
+// We use 16bit fixed point operations for RGB->YUV conversion.
+//
+// For the Y'CbCr to RGB conversion, the BT.601 specification reads:
+//   R = 1.164 * (Y-16) + 1.596 * (V-128)
+//   G = 1.164 * (Y-16) - 0.813 * (V-128) - 0.391 * (U-128)
+//   B = 1.164 * (Y-16)                   + 2.018 * (U-128)
+// where Y is in the [16,235] range, and U/V in the [16,240] range.
+// But the common term 1.164 * (Y-16) can be handled as an offset in the
+// VP8kClip[] table. So the formulae should be read as:
+//   R = 1.164 * [Y + 1.371 * (V-128)                  ] - 18.624
+//   G = 1.164 * [Y - 0.698 * (V-128) - 0.336 * (U-128)] - 18.624
+//   B = 1.164 * [Y                   + 1.733 * (U-128)] - 18.624
+// once factorized. Here too, 16bit fixed precision is used.
+//
 // Author: Skal (pascal.massimino@gmail.com)
 
 #ifndef WEBP_DSP_YUV_H_
@@ -14,13 +33,15 @@
 
 #include "../dec/decode_vp8.h"
 
-/*
- * Define ANDROID_WEBP_RGB to enable specific optimizations for Android
- * RGBA_4444 & RGB_565 color support.
- *
- */
-
-#define ANDROID_WEBP_RGB
+#if defined(WEBP_EXPERIMENTAL_FEATURES)
+// Do NOT activate this feature for real compression. This is only experimental!
+// This flag is for comparison purpose against JPEG's "YUVj" natural colorspace.
+// This colorspace is close to Rec.601's Y'CbCr model with the notable
+// difference of allowing larger range for luma/chroma.
+// See http://en.wikipedia.org/wiki/YCbCr#JPEG_conversion paragraph, and its
+// difference with http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
+// #define USE_YUVj
+#endif
 
 //------------------------------------------------------------------------------
 // YUV -> RGB conversion
@@ -53,16 +74,16 @@ static WEBP_INLINE void VP8YuvToRgb565(uint8_t y, uint8_t u, uint8_t v,
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
   const int b_off = VP8kUToB[u];
-#ifdef ANDROID_WEBP_RGB
-  rgb[1] = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
-            (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
-  rgb[0] = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
-            (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
+  const uint8_t rg = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
+                      (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
+  const uint8_t gb = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
+                      (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
+#ifdef WEBP_SWAP_16BIT_CSP
+  rgb[0] = gb;
+  rgb[1] = rg;
 #else
-  rgb[0] = ((VP8kClip[y + r_off - YUV_RANGE_MIN] & 0xf8) |
-            (VP8kClip[y + g_off - YUV_RANGE_MIN] >> 5));
-  rgb[1] = (((VP8kClip[y + g_off - YUV_RANGE_MIN] << 3) & 0xe0) |
-            (VP8kClip[y + b_off - YUV_RANGE_MIN] >> 3));
+  rgb[0] = rg;
+  rgb[1] = gb;
 #endif
 }
 
@@ -77,14 +98,15 @@ static WEBP_INLINE void VP8YuvToRgba4444(uint8_t y, uint8_t u, uint8_t v,
   const int r_off = VP8kVToR[v];
   const int g_off = (VP8kVToG[v] + VP8kUToG[u]) >> YUV_FIX;
   const int b_off = VP8kUToB[u];
-#ifdef ANDROID_WEBP_RGB
-  argb[1] = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
-             VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
-  argb[0] = 0x0f | (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4);
+  const uint8_t rg = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
+                      VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
+  const uint8_t ba = (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4) | 0x0f;
+#ifdef WEBP_SWAP_16BIT_CSP
+  argb[0] = ba;
+  argb[1] = rg;
 #else
-  argb[0] = ((VP8kClip4Bits[y + r_off - YUV_RANGE_MIN] << 4) |
-             VP8kClip4Bits[y + g_off - YUV_RANGE_MIN]);
-  argb[1] = 0x0f | (VP8kClip4Bits[y + b_off - YUV_RANGE_MIN] << 4);
+  argb[0] = rg;
+  argb[1] = ba;
 #endif
 }
 
@@ -115,18 +137,14 @@ void VP8YUVInit(void);
 
 //------------------------------------------------------------------------------
 // RGB -> YUV conversion
-// The exact naming is Y'CbCr, following the ITU-R BT.601 standard.
-// More information at: http://en.wikipedia.org/wiki/YCbCr
-// Y = 0.2569 * R + 0.5044 * G + 0.0979 * B + 16
-// U = -0.1483 * R - 0.2911 * G + 0.4394 * B + 128
-// V = 0.4394 * R - 0.3679 * G - 0.0715 * B + 128
-// We use 16bit fixed point operations.
 
 static WEBP_INLINE int VP8ClipUV(int v) {
-   v = (v + (257 << (YUV_FIX + 2 - 1))) >> (YUV_FIX + 2);
-   return ((v & ~0xff) == 0) ? v : (v < 0) ? 0 : 255;
+  v = (v + (257 << (YUV_FIX + 2 - 1))) >> (YUV_FIX + 2);
+  return ((v & ~0xff) == 0) ? v : (v < 0) ? 0 : 255;
 }
 
+#ifndef USE_YUVj
+
 static WEBP_INLINE int VP8RGBToY(int r, int g, int b) {
   const int kRound = (1 << (YUV_FIX - 1)) + (16 << YUV_FIX);
   const int luma = 16839 * r + 33059 * g + 6420 * b;
@@ -134,13 +152,38 @@ static WEBP_INLINE int VP8RGBToY(int r, int g, int b) {
 }
 
 static WEBP_INLINE int VP8RGBToU(int r, int g, int b) {
-  return VP8ClipUV(-9719 * r - 19081 * g + 28800 * b);
+  const int u = -9719 * r - 19081 * g + 28800 * b;
+  return VP8ClipUV(u);
+}
+
+static WEBP_INLINE int VP8RGBToV(int r, int g, int b) {
+  const int v = +28800 * r - 24116 * g - 4684 * b;
+  return VP8ClipUV(v);
+}
+
+#else
+
+// This JPEG-YUV colorspace, only for comparison!
+// These are also 16-bit precision coefficients from Rec.601, but with full
+// [0..255] output range.
+static WEBP_INLINE int VP8RGBToY(int r, int g, int b) {
+  const int kRound = (1 << (YUV_FIX - 1));
+  const int luma = 19595 * r + 38470 * g + 7471 * b;
+  return (luma + kRound) >> YUV_FIX;  // no need to clip
+}
+
+static WEBP_INLINE int VP8RGBToU(int r, int g, int b) {
+  const int u = -11058 * r - 21710 * g + 32768 * b;
+  return VP8ClipUV(u);
 }
 
 static WEBP_INLINE int VP8RGBToV(int r, int g, int b) {
-  return VP8ClipUV(+28800 * r - 24116 * g - 4684 * b);
+  const int v = 32768 * r - 27439 * g - 5329 * b;
+  return VP8ClipUV(v);
 }
 
+#endif    // USE_YUVj
+
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/enc/Android.mk b/src/enc/Android.mk
index 7f38d40d..615a0013 100644
--- a/src/enc/Android.mk
+++ b/src/enc/Android.mk
@@ -30,19 +30,16 @@ LOCAL_SRC_FILES := \
         quant.c \
         syntax.c \
         tree.c \
+        token.c \
         vp8l.c \
         webpenc.c \
         ../dsp/cpu.c \
-        ../dsp/dec.c \
-        ../dsp/dec_neon.c \
-        ../dsp/dec_sse2.c \
+        ../dsp/cpu-features.c \
         ../dsp/enc.c \
+        ../dsp/enc_neon.c \
         ../dsp/enc_sse2.c \
         ../dsp/lossless.c \
-        ../dsp/upsampling.c \
-        ../dsp/upsampling_sse2.c \
         ../dsp/yuv.c \
-        ../utils/bit_reader.c \
         ../utils/bit_writer.c \
         ../utils/color_cache.c \
         ../utils/filters.c \
@@ -53,7 +50,7 @@ LOCAL_SRC_FILES := \
         ../utils/thread.c \
         ../utils/utils.c
 
-LOCAL_CFLAGS := -DANDROID
+LOCAL_CFLAGS := -DANDROID -DWEBP_SWAP_16BIT_CSP
 
 LOCAL_C_INCLUDES += \
         $(LOCAL_PATH) \
diff --git a/src/enc/alpha.c b/src/enc/alpha.c
index 376f786c..f902a032 100644
--- a/src/enc/alpha.c
+++ b/src/enc/alpha.c
@@ -79,18 +79,17 @@ static int EncodeLossless(const uint8_t* const data, int width, int height,
   WebPConfigInit(&config);
   config.lossless = 1;
   config.method = effort_level;  // impact is very small
-  // Set moderate default quality setting for alpha. Higher qualities (80 and
-  // above) could be very slow.
-  config.quality = 10.f + 15.f * effort_level;
-  if (config.quality > 100.f) config.quality = 100.f;
+  // Set a moderate default quality setting for alpha.
+  config.quality = 5.f * effort_level;
+  assert(config.quality >= 0 && config.quality <= 100.f);
 
   ok = VP8LBitWriterInit(&tmp_bw, (width * height) >> 3);
   ok = ok && (VP8LEncodeStream(&config, &picture, &tmp_bw) == VP8_ENC_OK);
   WebPPictureFree(&picture);
   if (ok) {
-    const uint8_t* const data = VP8LBitWriterFinish(&tmp_bw);
-    const size_t data_size = VP8LBitWriterNumBytes(&tmp_bw);
-    VP8BitWriterAppend(bw, data, data_size);
+    const uint8_t* const buffer = VP8LBitWriterFinish(&tmp_bw);
+    const size_t buffer_size = VP8LBitWriterNumBytes(&tmp_bw);
+    VP8BitWriterAppend(bw, buffer, buffer_size);
   }
   VP8LBitWriterDestroy(&tmp_bw);
   return ok && !bw->error_;
@@ -287,42 +286,80 @@ static int EncodeAlpha(VP8Encoder* const enc,
 //------------------------------------------------------------------------------
 // Main calls
 
+static int CompressAlphaJob(VP8Encoder* const enc, void* dummy) {
+  const WebPConfig* config = enc->config_;
+  uint8_t* alpha_data = NULL;
+  size_t alpha_size = 0;
+  const int effort_level = config->method;  // maps to [0..6]
+  const WEBP_FILTER_TYPE filter =
+      (config->alpha_filtering == 0) ? WEBP_FILTER_NONE :
+      (config->alpha_filtering == 1) ? WEBP_FILTER_FAST :
+                                       WEBP_FILTER_BEST;
+  if (!EncodeAlpha(enc, config->alpha_quality, config->alpha_compression,
+                   filter, effort_level, &alpha_data, &alpha_size)) {
+    return 0;
+  }
+  if (alpha_size != (uint32_t)alpha_size) {  // Sanity check.
+    free(alpha_data);
+    return 0;
+  }
+  enc->alpha_data_size_ = (uint32_t)alpha_size;
+  enc->alpha_data_ = alpha_data;
+  (void)dummy;
+  return 1;
+}
+
 void VP8EncInitAlpha(VP8Encoder* const enc) {
   enc->has_alpha_ = WebPPictureHasTransparency(enc->pic_);
   enc->alpha_data_ = NULL;
   enc->alpha_data_size_ = 0;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    WebPWorkerInit(worker);
+    worker->data1 = enc;
+    worker->data2 = NULL;
+    worker->hook = (WebPWorkerHook)CompressAlphaJob;
+  }
 }
 
-int VP8EncFinishAlpha(VP8Encoder* const enc) {
+int VP8EncStartAlpha(VP8Encoder* const enc) {
   if (enc->has_alpha_) {
-    const WebPConfig* config = enc->config_;
-    uint8_t* tmp_data = NULL;
-    size_t tmp_size = 0;
-    const int effort_level = config->method;  // maps to [0..6]
-    const WEBP_FILTER_TYPE filter =
-        (config->alpha_filtering == 0) ? WEBP_FILTER_NONE :
-        (config->alpha_filtering == 1) ? WEBP_FILTER_FAST :
-                                         WEBP_FILTER_BEST;
-
-    if (!EncodeAlpha(enc, config->alpha_quality, config->alpha_compression,
-                     filter, effort_level, &tmp_data, &tmp_size)) {
-      return 0;
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPWorkerReset(worker)) {    // Makes sure worker is good to go.
+        return 0;
+      }
+      WebPWorkerLaunch(worker);
+      return 1;
+    } else {
+      return CompressAlphaJob(enc, NULL);   // just do the job right away
     }
-    if (tmp_size != (uint32_t)tmp_size) {  // Sanity check.
-      free(tmp_data);
-      return 0;
+  }
+  return 1;
+}
+
+int VP8EncFinishAlpha(VP8Encoder* const enc) {
+  if (enc->has_alpha_) {
+    if (enc->thread_level_ > 0) {
+      WebPWorker* const worker = &enc->alpha_worker_;
+      if (!WebPWorkerSync(worker)) return 0;  // error
     }
-    enc->alpha_data_size_ = (uint32_t)tmp_size;
-    enc->alpha_data_ = tmp_data;
   }
   return WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
 }
 
-void VP8EncDeleteAlpha(VP8Encoder* const enc) {
+int VP8EncDeleteAlpha(VP8Encoder* const enc) {
+  int ok = 1;
+  if (enc->thread_level_ > 0) {
+    WebPWorker* const worker = &enc->alpha_worker_;
+    ok = WebPWorkerSync(worker);  // finish anything left in flight
+    WebPWorkerEnd(worker);  // still need to end the worker, even if !ok
+  }
   free(enc->alpha_data_);
   enc->alpha_data_ = NULL;
   enc->alpha_data_size_ = 0;
   enc->has_alpha_ = 0;
+  return ok;
 }
 
 #if defined(__cplusplus) || defined(c_plusplus)
diff --git a/src/enc/analysis.c b/src/enc/analysis.c
index 22cfb492..221e9d06 100644
--- a/src/enc/analysis.c
+++ b/src/enc/analysis.c
@@ -23,10 +23,6 @@ extern "C" {
 
 #define MAX_ITERS_K_MEANS  6
 
-static int ClipAlpha(int alpha) {
-  return alpha < 0 ? 0 : alpha > 255 ? 255 : alpha;
-}
-
 //------------------------------------------------------------------------------
 // Smooth the segment map by replacing isolated block by the majority of its
 // neighbours.
@@ -72,50 +68,10 @@ static void SmoothSegmentMap(VP8Encoder* const enc) {
 }
 
 //------------------------------------------------------------------------------
-// Finalize Segment probability based on the coding tree
-
-static int GetProba(int a, int b) {
-  int proba;
-  const int total = a + b;
-  if (total == 0) return 255;  // that's the default probability.
-  proba = (255 * a + total / 2) / total;
-  return proba;
-}
-
-static void SetSegmentProbas(VP8Encoder* const enc) {
-  int p[NUM_MB_SEGMENTS] = { 0 };
-  int n;
-
-  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
-    const VP8MBInfo* const mb = &enc->mb_info_[n];
-    p[mb->segment_]++;
-  }
-  if (enc->pic_->stats) {
-    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
-      enc->pic_->stats->segment_size[n] = p[n];
-    }
-  }
-  if (enc->segment_hdr_.num_segments_ > 1) {
-    uint8_t* const probas = enc->proba_.segments_;
-    probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
-    probas[1] = GetProba(p[0], p[1]);
-    probas[2] = GetProba(p[2], p[3]);
-
-    enc->segment_hdr_.update_map_ =
-        (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
-    enc->segment_hdr_.size_ =
-      p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
-      p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
-      p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
-      p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
-  } else {
-    enc->segment_hdr_.update_map_ = 0;
-    enc->segment_hdr_.size_ = 0;
-  }
-}
+// set segment susceptibility alpha_ / beta_
 
 static WEBP_INLINE int clip(int v, int m, int M) {
-  return v < m ? m : v > M ? M : v;
+  return (v < m) ? m : (v > M) ? M : v;
 }
 
 static void SetSegmentAlphas(VP8Encoder* const enc,
@@ -142,22 +98,63 @@ static void SetSegmentAlphas(VP8Encoder* const enc,
 }
 
 //------------------------------------------------------------------------------
+// Compute susceptibility based on DCT-coeff histograms:
+// the higher, the "easier" the macroblock is to compress.
+
+#define MAX_ALPHA 255                // 8b of precision for susceptibilities.
+#define ALPHA_SCALE (2 * MAX_ALPHA)  // scaling factor for alpha.
+#define DEFAULT_ALPHA (-1)
+#define IS_BETTER_ALPHA(alpha, best_alpha) ((alpha) > (best_alpha))
+
+static int FinalAlphaValue(int alpha) {
+  alpha = MAX_ALPHA - alpha;
+  return clip(alpha, 0, MAX_ALPHA);
+}
+
+static int GetAlpha(const VP8Histogram* const histo) {
+  int max_value = 0, last_non_zero = 1;
+  int k;
+  int alpha;
+  for (k = 0; k <= MAX_COEFF_THRESH; ++k) {
+    const int value = histo->distribution[k];
+    if (value > 0) {
+      if (value > max_value) max_value = value;
+      last_non_zero = k;
+    }
+  }
+  // 'alpha' will later be clipped to [0..MAX_ALPHA] range, clamping outer
+  // values which happen to be mostly noise. This leaves the maximum precision
+  // for handling the useful small values which contribute most.
+  alpha = (max_value > 1) ? ALPHA_SCALE * last_non_zero / max_value : 0;
+  return alpha;
+}
+
+static void MergeHistograms(const VP8Histogram* const in,
+                            VP8Histogram* const out) {
+  int i;
+  for (i = 0; i <= MAX_COEFF_THRESH; ++i) {
+    out->distribution[i] += in->distribution[i];
+  }
+}
+
+//------------------------------------------------------------------------------
 // Simplified k-Means, to assign Nb segments based on alpha-histogram
 
-static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
+static void AssignSegments(VP8Encoder* const enc,
+                           const int alphas[MAX_ALPHA + 1]) {
   const int nb = enc->segment_hdr_.num_segments_;
   int centers[NUM_MB_SEGMENTS];
   int weighted_average = 0;
-  int map[256];
+  int map[MAX_ALPHA + 1];
   int a, n, k;
-  int min_a = 0, max_a = 255, range_a;
+  int min_a = 0, max_a = MAX_ALPHA, range_a;
   // 'int' type is ok for histo, and won't overflow
   int accum[NUM_MB_SEGMENTS], dist_accum[NUM_MB_SEGMENTS];
 
   // bracket the input
-  for (n = 0; n < 256 && alphas[n] == 0; ++n) {}
+  for (n = 0; n <= MAX_ALPHA && alphas[n] == 0; ++n) {}
   min_a = n;
-  for (n = 255; n > min_a && alphas[n] == 0; --n) {}
+  for (n = MAX_ALPHA; n > min_a && alphas[n] == 0; --n) {}
   max_a = n;
   range_a = max_a - min_a;
 
@@ -210,7 +207,7 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
     VP8MBInfo* const mb = &enc->mb_info_[n];
     const int alpha = mb->alpha_;
     mb->segment_ = map[alpha];
-    mb->alpha_ = centers[map[alpha]];     // just for the record.
+    mb->alpha_ = centers[map[alpha]];  // for the record.
   }
 
   if (nb > 1) {
@@ -218,7 +215,6 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
     if (smooth) SmoothSegmentMap(enc);
   }
 
-  SetSegmentProbas(enc);                             // Assign final proba
   SetSegmentAlphas(enc, centers, weighted_average);  // pick some alphas.
 }
 
@@ -227,24 +223,32 @@ static void AssignSegments(VP8Encoder* const enc, const int alphas[256]) {
 // susceptibility and set best modes for this macroblock.
 // Segment assignment is done later.
 
-// Number of modes to inspect for alpha_ evaluation. For high-quality settings,
-// we don't need to test all the possible modes during the analysis phase.
+// Number of modes to inspect for alpha_ evaluation. For high-quality settings
+// (method >= FAST_ANALYSIS_METHOD) we don't need to test all the possible modes
+// during the analysis phase.
+#define FAST_ANALYSIS_METHOD 4  // method above which we do partial analysis
 #define MAX_INTRA16_MODE 2
 #define MAX_INTRA4_MODE  2
 #define MAX_UV_MODE      2
 
 static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_INTRA16_MODE : 4;
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA16_MODE
+                                                  : NUM_PRED_MODES;
   int mode;
-  int best_alpha = -1;
+  int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
 
   VP8MakeLuma16Preds(it);
   for (mode = 0; mode < max_mode; ++mode) {
-    const int alpha = VP8CollectHistogram(it->yuv_in_ + Y_OFF,
-                                          it->yuv_p_ + VP8I16ModeOffsets[mode],
-                                          0, 16);
-    if (alpha > best_alpha) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+
+    VP8CollectHistogram(it->yuv_in_ + Y_OFF,
+                        it->yuv_p_ + VP8I16ModeOffsets[mode],
+                        0, 16, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
       best_alpha = alpha;
       best_mode = mode;
     }
@@ -256,46 +260,63 @@ static int MBAnalyzeBestIntra16Mode(VP8EncIterator* const it) {
 static int MBAnalyzeBestIntra4Mode(VP8EncIterator* const it,
                                    int best_alpha) {
   uint8_t modes[16];
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_INTRA4_MODE : NUM_BMODES;
-  int i4_alpha = 0;
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_INTRA4_MODE
+                                                  : NUM_BMODES;
+  int i4_alpha;
+  VP8Histogram total_histo = { { 0 } };
+  int cur_histo = 0;
+
   VP8IteratorStartI4(it);
   do {
     int mode;
-    int best_mode_alpha = -1;
+    int best_mode_alpha = DEFAULT_ALPHA;
+    VP8Histogram histos[2];
     const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
 
     VP8MakeIntra4Preds(it);
     for (mode = 0; mode < max_mode; ++mode) {
-      const int alpha = VP8CollectHistogram(src,
-                                            it->yuv_p_ + VP8I4ModeOffsets[mode],
-                                            0, 1);
-      if (alpha > best_mode_alpha) {
+      int alpha;
+
+      memset(&histos[cur_histo], 0, sizeof(histos[cur_histo]));
+      VP8CollectHistogram(src, it->yuv_p_ + VP8I4ModeOffsets[mode],
+                          0, 1, &histos[cur_histo]);
+      alpha = GetAlpha(&histos[cur_histo]);
+      if (IS_BETTER_ALPHA(alpha, best_mode_alpha)) {
         best_mode_alpha = alpha;
         modes[it->i4_] = mode;
+        cur_histo ^= 1;   // keep track of best histo so far.
       }
     }
-    i4_alpha += best_mode_alpha;
+    // accumulate best histogram
+    MergeHistograms(&histos[cur_histo ^ 1], &total_histo);
     // Note: we reuse the original samples for predictors
   } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
 
-  if (i4_alpha > best_alpha) {
+  i4_alpha = GetAlpha(&total_histo);
+  if (IS_BETTER_ALPHA(i4_alpha, best_alpha)) {
     VP8SetIntra4Mode(it, modes);
-    best_alpha = ClipAlpha(i4_alpha);
+    best_alpha = i4_alpha;
   }
   return best_alpha;
 }
 
 static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
-  int best_alpha = -1;
+  int best_alpha = DEFAULT_ALPHA;
   int best_mode = 0;
-  const int max_mode = (it->enc_->method_ >= 3) ? MAX_UV_MODE : 4;
+  const int max_mode =
+      (it->enc_->method_ >= FAST_ANALYSIS_METHOD) ? MAX_UV_MODE
+                                                  : NUM_PRED_MODES;
   int mode;
   VP8MakeChroma8Preds(it);
   for (mode = 0; mode < max_mode; ++mode) {
-    const int alpha = VP8CollectHistogram(it->yuv_in_ + U_OFF,
-                                          it->yuv_p_ + VP8UVModeOffsets[mode],
-                                          16, 16 + 4 + 4);
-    if (alpha > best_alpha) {
+    VP8Histogram histo = { { 0 } };
+    int alpha;
+    VP8CollectHistogram(it->yuv_in_ + U_OFF,
+                        it->yuv_p_ + VP8UVModeOffsets[mode],
+                        16, 16 + 4 + 4, &histo);
+    alpha = GetAlpha(&histo);
+    if (IS_BETTER_ALPHA(alpha, best_alpha)) {
       best_alpha = alpha;
       best_mode = mode;
     }
@@ -305,7 +326,8 @@ static int MBAnalyzeBestUVMode(VP8EncIterator* const it) {
 }
 
 static void MBAnalyze(VP8EncIterator* const it,
-                      int alphas[256], int* const uv_alpha) {
+                      int alphas[MAX_ALPHA + 1],
+                      int* const alpha, int* const uv_alpha) {
   const VP8Encoder* const enc = it->enc_;
   int best_alpha, best_uv_alpha;
 
@@ -314,7 +336,7 @@ static void MBAnalyze(VP8EncIterator* const it,
   VP8SetSegment(it, 0);      // default segment, spec-wise.
 
   best_alpha = MBAnalyzeBestIntra16Mode(it);
-  if (enc->method_ != 3) {
+  if (enc->method_ >= 5) {
     // We go and make a fast decision for intra4/intra16.
     // It's usually not a good and definitive pick, but helps seeding the stats
     // about level bit-cost.
@@ -324,10 +346,22 @@ static void MBAnalyze(VP8EncIterator* const it,
   best_uv_alpha = MBAnalyzeBestUVMode(it);
 
   // Final susceptibility mix
-  best_alpha = (best_alpha + best_uv_alpha + 1) / 2;
+  best_alpha = (3 * best_alpha + best_uv_alpha + 2) >> 2;
+  best_alpha = FinalAlphaValue(best_alpha);
   alphas[best_alpha]++;
+  it->mb_->alpha_ = best_alpha;   // for later remapping.
+
+  // Accumulate for later complexity analysis.
+  *alpha += best_alpha;   // mixed susceptibility (not just luma)
   *uv_alpha += best_uv_alpha;
-  it->mb_->alpha_ = best_alpha;   // Informative only.
+}
+
+static void DefaultMBInfo(VP8MBInfo* const mb) {
+  mb->type_ = 1;     // I16x16
+  mb->uv_mode_ = 0;
+  mb->skip_ = 0;     // not skipped
+  mb->segment_ = 0;  // default segment
+  mb->alpha_ = 0;
 }
 
 //------------------------------------------------------------------------------
@@ -340,22 +374,43 @@ static void MBAnalyze(VP8EncIterator* const it,
 // and decide intra4/intra16, but that's usually almost always a bad choice at
 // this stage.
 
+static void ResetAllMBInfo(VP8Encoder* const enc) {
+  int n;
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    DefaultMBInfo(&enc->mb_info_[n]);
+  }
+  // Default susceptibilities.
+  enc->dqm_[0].alpha_ = 0;
+  enc->dqm_[0].beta_ = 0;
+  // Note: we can't compute this alpha_ / uv_alpha_.
+  WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+}
+
 int VP8EncAnalyze(VP8Encoder* const enc) {
   int ok = 1;
-  int alphas[256] = { 0 };
-  VP8EncIterator it;
-
-  VP8IteratorInit(enc, &it);
+  const int do_segments =
+      enc->config_->emulate_jpeg_size ||   // We need the complexity evaluation.
+      (enc->segment_hdr_.num_segments_ > 1) ||
+      (enc->method_ == 0);  // for method 0, we need preds_[] to be filled.
+  enc->alpha_ = 0;
   enc->uv_alpha_ = 0;
-  do {
-    VP8IteratorImport(&it);
-    MBAnalyze(&it, alphas, &enc->uv_alpha_);
-    ok = VP8IteratorProgress(&it, 20);
-    // Let's pretend we have perfect lossless reconstruction.
-  } while (ok && VP8IteratorNext(&it, it.yuv_in_));
-  enc->uv_alpha_ /= enc->mb_w_ * enc->mb_h_;
-  if (ok) AssignSegments(enc, alphas);
-
+  if (do_segments) {
+    int alphas[MAX_ALPHA + 1] = { 0 };
+    VP8EncIterator it;
+
+    VP8IteratorInit(enc, &it);
+    do {
+      VP8IteratorImport(&it);
+      MBAnalyze(&it, alphas, &enc->alpha_, &enc->uv_alpha_);
+      ok = VP8IteratorProgress(&it, 20);
+      // Let's pretend we have perfect lossless reconstruction.
+    } while (ok && VP8IteratorNext(&it, it.yuv_in_));
+    enc->alpha_ /= enc->mb_w_ * enc->mb_h_;
+    enc->uv_alpha_ /= enc->mb_w_ * enc->mb_h_;
+    if (ok) AssignSegments(enc, alphas);
+  } else {   // Use only one default segment.
+    ResetAllMBInfo(enc);
+  }
   return ok;
 }
 
diff --git a/src/enc/backward_references.c b/src/enc/backward_references.c
index b8c8ece8..cf027875 100644
--- a/src/enc/backward_references.c
+++ b/src/enc/backward_references.c
@@ -141,21 +141,35 @@ static void HashChainInsert(HashChain* const p,
   p->hash_to_first_index_[hash_code] = pos;
 }
 
+static void GetParamsForHashChainFindCopy(int quality, int xsize,
+                                          int* window_size, int* iter_pos,
+                                          int* iter_limit) {
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  // Limit the backward-ref window size for lower qualities.
+  const int max_window_size = (quality > 50) ? WINDOW_SIZE
+                            : (quality > 25) ? (xsize << 8)
+                            : (xsize << 4);
+  assert(xsize > 0);
+  *window_size = (max_window_size > WINDOW_SIZE) ? WINDOW_SIZE
+               : max_window_size;
+  *iter_pos = 5 + (quality >> 3);
+  *iter_limit = -quality * iter_mult;
+}
+
 static int HashChainFindCopy(const HashChain* const p,
-                             int quality, int index, int xsize,
+                             int base_position, int xsize,
                              const uint32_t* const argb, int maxlen,
+                             int window_size, int iter_pos, int iter_limit,
                              int* const distance_ptr,
                              int* const length_ptr) {
-  const uint64_t hash_code = GetPixPairHash64(&argb[index]);
+  const uint64_t hash_code = GetPixPairHash64(&argb[base_position]);
   int prev_length = 0;
   int64_t best_val = 0;
   int best_length = 0;
   int best_distance = 0;
-  const uint32_t* const argb_start = argb + index;
-  const int iter_min_mult = (quality < 50) ? 2 : (quality < 75) ? 4 : 8;
-  const int iter_min = -quality * iter_min_mult;
-  int iter_cnt = 10 + (quality >> 1);
-  const int min_pos = (index > WINDOW_SIZE) ? index - WINDOW_SIZE : 0;
+  const uint32_t* const argb_start = argb + base_position;
+  const int min_pos =
+      (base_position > window_size) ? base_position - window_size : 0;
   int pos;
 
   assert(xsize > 0);
@@ -164,12 +178,12 @@ static int HashChainFindCopy(const HashChain* const p,
        pos = p->chain_[pos]) {
     int64_t val;
     int curr_length;
-    if (iter_cnt < 0) {
-      if (iter_cnt < iter_min || best_val >= 0xff0000) {
+    if (iter_pos < 0) {
+      if (iter_pos < iter_limit || best_val >= 0xff0000) {
         break;
       }
     }
-    --iter_cnt;
+    --iter_pos;
     if (best_length != 0 &&
         argb[pos + best_length - 1] != argb_start[best_length - 1]) {
       continue;
@@ -180,9 +194,9 @@ static int HashChainFindCopy(const HashChain* const p,
     }
     val = 65536 * curr_length;
     // Favoring 2d locality here gives savings for certain images.
-    if (index - pos < 9 * xsize) {
-      const int y = (index - pos) / xsize;
-      int x = (index - pos) % xsize;
+    if (base_position - pos < 9 * xsize) {
+      const int y = (base_position - pos) / xsize;
+      int x = (base_position - pos) % xsize;
       if (x > xsize / 2) {
         x = xsize - x;
       }
@@ -198,7 +212,7 @@ static int HashChainFindCopy(const HashChain* const p,
       prev_length = curr_length;
       best_val = val;
       best_length = curr_length;
-      best_distance = index - pos;
+      best_distance = base_position - pos;
       if (curr_length >= MAX_LENGTH) {
         break;
       }
@@ -257,6 +271,9 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
   const int pix_count = xsize * ysize;
   HashChain* const hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
 
   if (hash_chain == NULL) return 0;
   if (use_color_cache) {
@@ -267,6 +284,8 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
   if (!HashChainInit(hash_chain, pix_count)) goto Error;
 
   refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, &window_size, &iter_pos,
+                                &iter_limit);
   for (i = 0; i < pix_count; ) {
     // Alternative#1: Code the pixels starting at 'i' using backward reference.
     int offset = 0;
@@ -276,7 +295,8 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
       if (maxlen > MAX_LENGTH) {
         maxlen = MAX_LENGTH;
       }
-      HashChainFindCopy(hash_chain, quality, i, xsize, argb, maxlen,
+      HashChainFindCopy(hash_chain, i, xsize, argb, maxlen,
+                        window_size, iter_pos, iter_limit,
                         &offset, &len);
     }
     if (len >= MIN_LENGTH) {
@@ -291,8 +311,9 @@ static int BackwardReferencesHashChain(int xsize, int ysize,
         if (maxlen > MAX_LENGTH) {
           maxlen = MAX_LENGTH;
         }
-        HashChainFindCopy(hash_chain, quality,
-                          i + 1, xsize, argb, maxlen, &offset2, &len2);
+        HashChainFindCopy(hash_chain, i + 1, xsize, argb, maxlen,
+                          window_size, iter_pos, iter_limit,
+                          &offset2, &len2);
         if (len2 > len + 1) {
           const uint32_t pixel = argb[i];
           // Alternative#2 is a better match. So push pixel at 'i' as literal.
@@ -362,7 +383,8 @@ typedef struct {
 
 static int BackwardReferencesTraceBackwards(
     int xsize, int ysize, int recursive_cost_model,
-    const uint32_t* const argb, int cache_bits, VP8LBackwardRefs* const refs);
+    const uint32_t* const argb, int quality, int cache_bits,
+    VP8LBackwardRefs* const refs);
 
 static void ConvertPopulationCountTableToBitEstimates(
     int num_symbols, const int population_counts[], double output[]) {
@@ -387,17 +409,16 @@ static void ConvertPopulationCountTableToBitEstimates(
 
 static int CostModelBuild(CostModel* const m, int xsize, int ysize,
                           int recursion_level, const uint32_t* const argb,
-                          int cache_bits) {
+                          int quality, int cache_bits) {
   int ok = 0;
   VP8LHistogram histo;
   VP8LBackwardRefs refs;
-  const int quality = 100;
 
   if (!VP8LBackwardRefsAlloc(&refs, xsize * ysize)) goto Error;
 
   if (recursion_level > 0) {
     if (!BackwardReferencesTraceBackwards(xsize, ysize, recursion_level - 1,
-                                          argb, cache_bits, &refs)) {
+                                          argb, quality, cache_bits, &refs)) {
       goto Error;
     }
   } else {
@@ -452,20 +473,23 @@ static WEBP_INLINE double GetDistanceCost(const CostModel* const m,
 
 static int BackwardReferencesHashChainDistanceOnly(
     int xsize, int ysize, int recursive_cost_model, const uint32_t* const argb,
-    int cache_bits, uint32_t* const dist_array) {
+    int quality, int cache_bits, uint32_t* const dist_array) {
   int i;
   int ok = 0;
   int cc_init = 0;
-  const int quality = 100;
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
-  double* const cost =
-      (double*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
+  float* const cost =
+      (float*)WebPSafeMalloc((uint64_t)pix_count, sizeof(*cost));
   CostModel* cost_model = (CostModel*)malloc(sizeof(*cost_model));
   HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
   const double mul0 = (recursive_cost_model != 0) ? 1.0 : 0.68;
   const double mul1 = (recursive_cost_model != 0) ? 1.0 : 0.82;
+  const int min_distance_code = 2;  // TODO(vikasa): tune as function of quality
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
 
   if (cost == NULL || cost_model == NULL || hash_chain == NULL) goto Error;
 
@@ -477,15 +501,17 @@ static int BackwardReferencesHashChainDistanceOnly(
   }
 
   if (!CostModelBuild(cost_model, xsize, ysize, recursive_cost_model, argb,
-                      cache_bits)) {
+                      quality, cache_bits)) {
     goto Error;
   }
 
-  for (i = 0; i < pix_count; ++i) cost[i] = 1e100;
+  for (i = 0; i < pix_count; ++i) cost[i] = 1e38f;
 
   // We loop one pixel at a time, but store all currently best points to
   // non-processed locations from this point.
   dist_array[0] = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, &window_size, &iter_pos,
+                                &iter_limit);
   for (i = 0; i < pix_count; ++i) {
     double prev_cost = 0.0;
     int shortmax;
@@ -500,7 +526,8 @@ static int BackwardReferencesHashChainDistanceOnly(
         if (maxlen > pix_count - i) {
           maxlen = pix_count - i;
         }
-        HashChainFindCopy(hash_chain, quality, i, xsize, argb, maxlen,
+        HashChainFindCopy(hash_chain, i, xsize, argb, maxlen,
+                          window_size, iter_pos, iter_limit,
                           &offset, &len);
       }
       if (len >= MIN_LENGTH) {
@@ -509,16 +536,15 @@ static int BackwardReferencesHashChainDistanceOnly(
             prev_cost + GetDistanceCost(cost_model, code);
         int k;
         for (k = 1; k < len; ++k) {
-          const double cost_val =
-              distance_cost + GetLengthCost(cost_model, k);
+          const double cost_val = distance_cost + GetLengthCost(cost_model, k);
           if (cost[i + k] > cost_val) {
-            cost[i + k] = cost_val;
+            cost[i + k] = (float)cost_val;
             dist_array[i + k] = k + 1;
           }
         }
         // This if is for speedup only. It roughly doubles the speed, and
         // makes compression worse by .1 %.
-        if (len >= 128 && code < 2) {
+        if (len >= 128 && code <= min_distance_code) {
           // Long copy for short distances, let's skip the middle
           // lookups for better copies.
           // 1) insert the hashes.
@@ -529,10 +555,10 @@ static int BackwardReferencesHashChainDistanceOnly(
           }
           // 2) Add to the hash_chain (but cannot add the last pixel)
           {
-            const int last = (len < pix_count - 1 - i) ? len
-                                                       : pix_count - 1 - i;
-            for (k = 0; k < last; ++k) {
-              HashChainInsert(hash_chain, &argb[i + k], i + k);
+            const int last = (len + i < pix_count - 1) ? len + i
+                                                       : pix_count - 1;
+            for (k = i; k < last; ++k) {
+              HashChainInsert(hash_chain, &argb[k], k);
             }
           }
           // 3) jump.
@@ -554,7 +580,7 @@ static int BackwardReferencesHashChainDistanceOnly(
         cost_val += GetLiteralCost(cost_model, argb[i]) * mul1;
       }
       if (cost[i] > cost_val) {
-        cost[i] = cost_val;
+        cost[i] = (float)cost_val;
         dist_array[i] = 1;  // only one is inserted.
       }
       if (use_color_cache) VP8LColorCacheInsert(&hashers, argb[i]);
@@ -572,40 +598,30 @@ Error:
   return ok;
 }
 
-static int TraceBackwards(const uint32_t* const dist_array,
-                          int dist_array_size,
-                          uint32_t** const chosen_path,
-                          int* const chosen_path_size) {
-  int i;
-  // Count how many.
-  int count = 0;
-  for (i = dist_array_size - 1; i >= 0; ) {
-    int k = dist_array[i];
-    assert(k >= 1);
-    ++count;
-    i -= k;
-  }
-  // Allocate.
-  *chosen_path_size = count;
-  *chosen_path =
-      (uint32_t*)WebPSafeMalloc((uint64_t)count, sizeof(**chosen_path));
-  if (*chosen_path == NULL) return 0;
-
-  // Write in reverse order.
-  for (i = dist_array_size - 1; i >= 0; ) {
-    int k = dist_array[i];
-    assert(k >= 1);
-    (*chosen_path)[--count] = k;
-    i -= k;
-  }
-  return 1;
+// We pack the path at the end of *dist_array and return
+// a pointer to this part of the array. Example:
+// dist_array = [1x2xx3x2] => packed [1x2x1232], chosen_path = [1232]
+static void TraceBackwards(uint32_t* const dist_array,
+                           int dist_array_size,
+                           uint32_t** const chosen_path,
+                           int* const chosen_path_size) {
+  uint32_t* path = dist_array + dist_array_size;
+  uint32_t* cur = dist_array + dist_array_size - 1;
+  while (cur >= dist_array) {
+    const int k = *cur;
+    --path;
+    *path = k;
+    cur -= k;
+  }
+  *chosen_path = path;
+  *chosen_path_size = (int)(dist_array + dist_array_size - path);
 }
 
 static int BackwardReferencesHashChainFollowChosenPath(
-    int xsize, int ysize, const uint32_t* const argb, int cache_bits,
+    int xsize, int ysize, const uint32_t* const argb,
+    int quality, int cache_bits,
     const uint32_t* const chosen_path, int chosen_path_size,
     VP8LBackwardRefs* const refs) {
-  const int quality = 100;
   const int pix_count = xsize * ysize;
   const int use_color_cache = (cache_bits > 0);
   int size = 0;
@@ -614,6 +630,9 @@ static int BackwardReferencesHashChainFollowChosenPath(
   int ix;
   int ok = 0;
   int cc_init = 0;
+  int window_size = WINDOW_SIZE;
+  int iter_pos = 1;
+  int iter_limit = -1;
   HashChain* hash_chain = (HashChain*)malloc(sizeof(*hash_chain));
   VP8LColorCache hashers;
 
@@ -626,13 +645,16 @@ static int BackwardReferencesHashChainFollowChosenPath(
   }
 
   refs->size = 0;
+  GetParamsForHashChainFindCopy(quality, xsize, &window_size, &iter_pos,
+                                &iter_limit);
   for (ix = 0; ix < chosen_path_size; ++ix, ++size) {
     int offset = 0;
     int len = 0;
     int maxlen = chosen_path[ix];
     if (maxlen != 1) {
-      HashChainFindCopy(hash_chain, quality,
-                        i, xsize, argb, maxlen, &offset, &len);
+      HashChainFindCopy(hash_chain, i, xsize, argb, maxlen,
+                        window_size, iter_pos, iter_limit,
+                        &offset, &len);
       assert(len == maxlen);
       refs->refs[size] = PixOrCopyCreateCopy(offset, len);
       if (use_color_cache) {
@@ -675,7 +697,7 @@ Error:
 static int BackwardReferencesTraceBackwards(int xsize, int ysize,
                                             int recursive_cost_model,
                                             const uint32_t* const argb,
-                                            int cache_bits,
+                                            int quality, int cache_bits,
                                             VP8LBackwardRefs* const refs) {
   int ok = 0;
   const int dist_array_size = xsize * ysize;
@@ -687,22 +709,18 @@ static int BackwardReferencesTraceBackwards(int xsize, int ysize,
   if (dist_array == NULL) goto Error;
 
   if (!BackwardReferencesHashChainDistanceOnly(
-      xsize, ysize, recursive_cost_model, argb, cache_bits, dist_array)) {
-    goto Error;
-  }
-  if (!TraceBackwards(dist_array, dist_array_size,
-                      &chosen_path, &chosen_path_size)) {
+      xsize, ysize, recursive_cost_model, argb, quality, cache_bits,
+      dist_array)) {
     goto Error;
   }
-  free(dist_array);   // no need to retain this memory any longer
-  dist_array = NULL;
+  TraceBackwards(dist_array, dist_array_size, &chosen_path, &chosen_path_size);
   if (!BackwardReferencesHashChainFollowChosenPath(
-      xsize, ysize, argb, cache_bits, chosen_path, chosen_path_size, refs)) {
+      xsize, ysize, argb, quality, cache_bits, chosen_path, chosen_path_size,
+      refs)) {
     goto Error;
   }
   ok = 1;
  Error:
-  free(chosen_path);
   free(dist_array);
   return ok;
 }
@@ -762,8 +780,8 @@ int VP8LGetBackwardReferences(int width, int height,
 
   // Choose appropriate backward reference.
   if (lz77_is_useful) {
-    // TraceBackwards is costly. Run it for higher qualities.
-    const int try_lz77_trace_backwards = (quality >= 75);
+    // TraceBackwards is costly. Don't execute it at lower quality (q <= 10).
+    const int try_lz77_trace_backwards = (quality > 10);
     *best = refs_lz77;   // default guess: lz77 is better
     VP8LClearBackwardRefs(&refs_rle);
     if (try_lz77_trace_backwards) {
@@ -772,8 +790,8 @@ int VP8LGetBackwardReferences(int width, int height,
       if (!VP8LBackwardRefsAlloc(&refs_trace, num_pix)) {
         goto End;
       }
-      if (BackwardReferencesTraceBackwards(
-          width, height, recursion_level, argb, cache_bits, &refs_trace)) {
+      if (BackwardReferencesTraceBackwards(width, height, recursion_level, argb,
+                                           quality, cache_bits, &refs_trace)) {
         VP8LClearBackwardRefs(&refs_lz77);
         *best = refs_trace;
       }
diff --git a/src/enc/backward_references.h b/src/enc/backward_references.h
index cda7c2b1..d1f6e7d6 100644
--- a/src/enc/backward_references.h
+++ b/src/enc/backward_references.h
@@ -35,7 +35,8 @@ extern "C" {
 #if defined(__GNUC__) && \
     ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
 static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
-  return n == 0 ? -1 : 31 ^ __builtin_clz(n);
+  assert(n != 0);
+  return 31 ^ __builtin_clz(n);
 }
 #elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
 #include <intrin.h>
@@ -43,15 +44,18 @@ static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
 
 static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
   unsigned long first_set_bit;
-  return _BitScanReverse(&first_set_bit, n) ? first_set_bit : -1;
+  assert(n != 0);
+  _BitScanReverse(&first_set_bit, n);
+  return first_set_bit;
 }
 #else
+// Returns (int)floor(log2(n)). n must be > 0.
 static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
   int log = 0;
   uint32_t value = n;
   int i;
 
-  if (value == 0) return -1;
+  assert(n != 0);
   for (i = 4; i >= 0; --i) {
     const int shift = (1 << i);
     const uint32_t x = value >> shift;
@@ -65,11 +69,11 @@ static WEBP_INLINE int BitsLog2Floor(uint32_t n) {
 #endif
 
 static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
-  const int floor = BitsLog2Floor(n);
+  const int log_floor = BitsLog2Floor(n);
   if (n == (n & ~(n - 1)))  // zero or a power of two.
-    return floor;
+    return log_floor;
   else
-    return floor + 1;
+    return log_floor + 1;
 }
 
 // Splitting of distance and length codes into prefixes and
@@ -78,16 +82,17 @@ static WEBP_INLINE int VP8LBitsLog2Ceiling(uint32_t n) {
 static WEBP_INLINE void PrefixEncode(int distance, int* const code,
                                      int* const extra_bits_count,
                                      int* const extra_bits_value) {
-  // Collect the two most significant bits where the highest bit is 1.
-  const int highest_bit = BitsLog2Floor(--distance);
-  // & 0x3f is to make behavior well defined when highest_bit
-  // does not exist or is the least significant bit.
-  const int second_highest_bit =
-      (distance >> ((highest_bit - 1) & 0x3f)) & 1;
-  *extra_bits_count = (highest_bit > 0) ? (highest_bit - 1) : 0;
-  *extra_bits_value = distance & ((1 << *extra_bits_count) - 1);
-  *code = (highest_bit > 0) ? (2 * highest_bit + second_highest_bit)
-                            : (highest_bit == 0) ? 1 : 0;
+  if (distance > 2) {  // Collect the two most significant bits.
+    const int highest_bit = BitsLog2Floor(--distance);
+    const int second_highest_bit = (distance >> (highest_bit - 1)) & 1;
+    *extra_bits_count = highest_bit - 1;
+    *extra_bits_value = distance & ((1 << *extra_bits_count) - 1);
+    *code = 2 * highest_bit + second_highest_bit;
+  } else {
+    *extra_bits_count = 0;
+    *extra_bits_value = 0;
+    *code = (distance == 2) ? 1 : 0;
+  }
 }
 
 // -----------------------------------------------------------------------------
diff --git a/src/enc/config.c b/src/enc/config.c
index b05328d1..263db324 100644
--- a/src/enc/config.c
+++ b/src/enc/config.c
@@ -31,9 +31,9 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->target_PSNR = 0.;
   config->method = 4;
   config->sns_strength = 50;
-  config->filter_strength = 20;   // default: light filtering
+  config->filter_strength = 60;   // rather high filtering, helps w/ gradients.
   config->filter_sharpness = 0;
-  config->filter_type = 0;        // default: simple
+  config->filter_type = 1;        // default: strong (so U/V is filtered too)
   config->partitions = 0;
   config->segments = 4;
   config->pass = 1;
@@ -46,6 +46,9 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->alpha_quality = 100;
   config->lossless = 0;
   config->image_hint = WEBP_HINT_DEFAULT;
+  config->emulate_jpeg_size = 0;
+  config->thread_level = 0;
+  config->low_memory = 0;
 
   // TODO(skal): tune.
   switch (preset) {
@@ -122,6 +125,12 @@ int WebPValidateConfig(const WebPConfig* config) {
     return 0;
   if (config->image_hint >= WEBP_HINT_LAST)
     return 0;
+  if (config->emulate_jpeg_size < 0 || config->emulate_jpeg_size > 1)
+    return 0;
+  if (config->thread_level < 0 || config->thread_level > 1)
+    return 0;
+  if (config->low_memory < 0 || config->low_memory > 1)
+    return 0;
   return 1;
 }
 
diff --git a/src/enc/frame.c b/src/enc/frame.c
index bdd36006..445d8570 100644
--- a/src/enc/frame.c
+++ b/src/enc/frame.c
@@ -45,10 +45,10 @@ const uint8_t VP8EncBands[16 + 1] = {
   0  // sentinel
 };
 
-static const uint8_t kCat3[] = { 173, 148, 140 };
-static const uint8_t kCat4[] = { 176, 155, 140, 135 };
-static const uint8_t kCat5[] = { 180, 157, 141, 134, 130 };
-static const uint8_t kCat6[] =
+const uint8_t VP8Cat3[] = { 173, 148, 140 };
+const uint8_t VP8Cat4[] = { 176, 155, 140, 135 };
+const uint8_t VP8Cat5[] = { 180, 157, 141, 134, 130 };
+const uint8_t VP8Cat6[] =
     { 254, 254, 243, 230, 196, 177, 153, 140, 133, 130, 129 };
 
 //------------------------------------------------------------------------------
@@ -113,14 +113,15 @@ static int Record(int bit, proba_t* const stats) {
 // Note: no need to record the fixed probas.
 static int RecordCoeffs(int ctx, const VP8Residual* const res) {
   int n = res->first;
-  proba_t* s = res->stats[VP8EncBands[n]][ctx];
+  // should be stats[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  proba_t* s = res->stats[n][ctx];
   if (res->last  < 0) {
     Record(0, s + 0);
     return 0;
   }
   while (n <= res->last) {
     int v;
-    Record(1, s + 0);
+    Record(1, s + 0);  // order of record doesn't matter
     while ((v = res->coeffs[n++]) == 0) {
       Record(0, s + 1);
       s = res->stats[VP8EncBands[n]][0];
@@ -174,8 +175,7 @@ static int BranchCost(int nb, int total, int proba) {
   return nb * VP8BitCost(1, proba) + (total - nb) * VP8BitCost(0, proba);
 }
 
-static int FinalizeTokenProbas(VP8Encoder* const enc) {
-  VP8Proba* const proba = &enc->proba_;
+static int FinalizeTokenProbas(VP8Proba* const proba) {
   int has_changed = 0;
   int size = 0;
   int t, b, c, p;
@@ -212,6 +212,47 @@ static int FinalizeTokenProbas(VP8Encoder* const enc) {
 }
 
 //------------------------------------------------------------------------------
+// Finalize Segment probability based on the coding tree
+
+static int GetProba(int a, int b) {
+  const int total = a + b;
+  return (total == 0) ? 255     // that's the default probability.
+                      : (255 * a + total / 2) / total;  // rounded proba
+}
+
+static void SetSegmentProbas(VP8Encoder* const enc) {
+  int p[NUM_MB_SEGMENTS] = { 0 };
+  int n;
+
+  for (n = 0; n < enc->mb_w_ * enc->mb_h_; ++n) {
+    const VP8MBInfo* const mb = &enc->mb_info_[n];
+    p[mb->segment_]++;
+  }
+  if (enc->pic_->stats != NULL) {
+    for (n = 0; n < NUM_MB_SEGMENTS; ++n) {
+      enc->pic_->stats->segment_size[n] = p[n];
+    }
+  }
+  if (enc->segment_hdr_.num_segments_ > 1) {
+    uint8_t* const probas = enc->proba_.segments_;
+    probas[0] = GetProba(p[0] + p[1], p[2] + p[3]);
+    probas[1] = GetProba(p[0], p[1]);
+    probas[2] = GetProba(p[2], p[3]);
+
+    enc->segment_hdr_.update_map_ =
+        (probas[0] != 255) || (probas[1] != 255) || (probas[2] != 255);
+    enc->segment_hdr_.size_ =
+        p[0] * (VP8BitCost(0, probas[0]) + VP8BitCost(0, probas[1])) +
+        p[1] * (VP8BitCost(0, probas[0]) + VP8BitCost(1, probas[1])) +
+        p[2] * (VP8BitCost(1, probas[0]) + VP8BitCost(0, probas[2])) +
+        p[3] * (VP8BitCost(1, probas[0]) + VP8BitCost(1, probas[2]));
+  } else {
+    enc->segment_hdr_.update_map_ = 0;
+    enc->segment_hdr_.size_ = 0;
+  }
+}
+
+//------------------------------------------------------------------------------
 // helper functions for residuals struct VP8Residual.
 
 static void InitResidual(int first, int coeff_type,
@@ -239,18 +280,19 @@ static void SetResidualCoeffs(const int16_t* const coeffs,
 //------------------------------------------------------------------------------
 // Mode costs
 
-static int GetResidualCost(int ctx, const VP8Residual* const res) {
+static int GetResidualCost(int ctx0, const VP8Residual* const res) {
   int n = res->first;
-  int p0 = res->prob[VP8EncBands[n]][ctx][0];
-  const uint16_t* t = res->cost[VP8EncBands[n]][ctx];
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  int p0 = res->prob[n][ctx0][0];
+  const uint16_t* t = res->cost[n][ctx0];
   int cost;
 
   if (res->last < 0) {
     return VP8BitCost(0, p0);
   }
   cost = 0;
-  while (n <= res->last) {
-    const int v = res->coeffs[n];
+  while (n < res->last) {
+    int v = res->coeffs[n];
     const int b = VP8EncBands[n + 1];
     ++n;
     if (v == 0) {
@@ -259,19 +301,28 @@ static int GetResidualCost(int ctx, const VP8Residual* const res) {
       t = res->cost[b][0];
       continue;
     }
+    v = abs(v);
     cost += VP8BitCost(1, p0);
-    if (2u >= (unsigned int)(v + 1)) {   // v = -1 or 1
-      // short-case for "VP8LevelCost(t, 1)" (256 is VP8LevelFixedCosts[1]):
-      cost += 256 + t[1];
-      p0 = res->prob[b][1][0];
-      t = res->cost[b][1];
-    } else {
-      cost += VP8LevelCost(t, abs(v));
-      p0 = res->prob[b][2][0];
-      t = res->cost[b][2];
+    cost += VP8LevelCost(t, v);
+    {
+      const int ctx = (v == 1) ? 1 : 2;
+      p0 = res->prob[b][ctx][0];
+      t = res->cost[b][ctx];
+    }
+  }
+  // Last coefficient is always non-zero
+  {
+    const int v = abs(res->coeffs[n]);
+    assert(v != 0);
+    cost += VP8BitCost(1, p0);
+    cost += VP8LevelCost(t, v);
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = (v == 1) ? 1 : 2;
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
     }
   }
-  if (n < 16) cost += VP8BitCost(0, p0);
   return cost;
 }
 
@@ -342,7 +393,8 @@ int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd) {
 
 static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
   int n = res->first;
-  const uint8_t* p = res->prob[VP8EncBands[n]][ctx];
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  const uint8_t* p = res->prob[n][ctx];
   if (!VP8PutBit(bw, res->last >= 0, p[0])) {
     return 0;
   }
@@ -371,30 +423,30 @@ static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
       } else {
         int mask;
         const uint8_t* tab;
-        if (v < 3 + (8 << 1)) {          // kCat3  (3b)
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
           VP8PutBit(bw, 0, p[8]);
           VP8PutBit(bw, 0, p[9]);
           v -= 3 + (8 << 0);
           mask = 1 << 2;
-          tab = kCat3;
-        } else if (v < 3 + (8 << 2)) {   // kCat4  (4b)
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
           VP8PutBit(bw, 0, p[8]);
           VP8PutBit(bw, 1, p[9]);
           v -= 3 + (8 << 1);
           mask = 1 << 3;
-          tab = kCat4;
-        } else if (v < 3 + (8 << 3)) {   // kCat5  (5b)
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
           VP8PutBit(bw, 1, p[8]);
           VP8PutBit(bw, 0, p[10]);
           v -= 3 + (8 << 2);
           mask = 1 << 4;
-          tab = kCat5;
-        } else {                         // kCat6 (11b)
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
           VP8PutBit(bw, 1, p[8]);
           VP8PutBit(bw, 1, p[10]);
           v -= 3 + (8 << 3);
           mask = 1 << 10;
-          tab = kCat6;
+          tab = VP8Cat6;
         }
         while (mask) {
           VP8PutBit(bw, !!(v & mask), *tab++);
@@ -411,8 +463,7 @@ static int PutCoeffs(VP8BitWriter* const bw, int ctx, const VP8Residual* res) {
   return 1;
 }
 
-static void CodeResiduals(VP8BitWriter* const bw,
-                          VP8EncIterator* const it,
+static void CodeResiduals(VP8BitWriter* const bw, VP8EncIterator* const it,
                           const VP8ModeScore* const rd) {
   int x, y, ch;
   VP8Residual res;
@@ -512,146 +563,23 @@ static void RecordResiduals(VP8EncIterator* const it,
 //------------------------------------------------------------------------------
 // Token buffer
 
-#ifdef USE_TOKEN_BUFFER
-
-void VP8TBufferInit(VP8TBuffer* const b) {
-  b->rows_ = NULL;
-  b->tokens_ = NULL;
-  b->last_ = &b->rows_;
-  b->left_ = 0;
-  b->error_ = 0;
-}
-
-int VP8TBufferNewPage(VP8TBuffer* const b) {
-  VP8Tokens* const page = b->error_ ? NULL : (VP8Tokens*)malloc(sizeof(*page));
-  if (page == NULL) {
-    b->error_ = 1;
-    return 0;
-  }
-  *b->last_ = page;
-  b->last_ = &page->next_;
-  b->left_ = MAX_NUM_TOKEN;
-  b->tokens_ = page->tokens_;
-  return 1;
-}
-
-void VP8TBufferClear(VP8TBuffer* const b) {
-  if (b != NULL) {
-    const VP8Tokens* p = b->rows_;
-    while (p != NULL) {
-      const VP8Tokens* const next = p->next_;
-      free((void*)p);
-      p = next;
-    }
-    VP8TBufferInit(b);
-  }
-}
-
-int VP8EmitTokens(const VP8TBuffer* const b, VP8BitWriter* const bw,
-                  const uint8_t* const probas) {
-  VP8Tokens* p = b->rows_;
-  if (b->error_) return 0;
-  while (p != NULL) {
-    const int N = (p->next_ == NULL) ? b->left_ : 0;
-    int n = MAX_NUM_TOKEN;
-    while (n-- > N) {
-      VP8PutBit(bw, (p->tokens_[n] >> 15) & 1, probas[p->tokens_[n] & 0x7fff]);
-    }
-    p = p->next_;
-  }
-  return 1;
-}
-
-#define TOKEN_ID(b, ctx, p) ((p) + NUM_PROBAS * ((ctx) + (b) * NUM_CTX))
+#if !defined(DISABLE_TOKEN_BUFFER)
 
-static int RecordCoeffTokens(int ctx, const VP8Residual* const res,
-                             VP8TBuffer* tokens) {
-  int n = res->first;
-  int b = VP8EncBands[n];
-  if (!VP8AddToken(tokens, res->last >= 0, TOKEN_ID(b, ctx, 0))) {
-    return 0;
-  }
-
-  while (n < 16) {
-    const int c = res->coeffs[n++];
-    const int sign = c < 0;
-    int v = sign ? -c : c;
-    const int base_id = TOKEN_ID(b, ctx, 0);
-    if (!VP8AddToken(tokens, v != 0, base_id + 1)) {
-      b = VP8EncBands[n];
-      ctx = 0;
-      continue;
-    }
-    if (!VP8AddToken(tokens, v > 1, base_id + 2)) {
-      b = VP8EncBands[n];
-      ctx = 1;
-    } else {
-      if (!VP8AddToken(tokens, v > 4, base_id + 3)) {
-        if (VP8AddToken(tokens, v != 2, base_id + 4))
-          VP8AddToken(tokens, v == 4, base_id + 5);
-      } else if (!VP8AddToken(tokens, v > 10, base_id + 6)) {
-        if (!VP8AddToken(tokens, v > 6, base_id + 7)) {
-//          VP8AddToken(tokens, v == 6, 159);
-        } else {
-//          VP8AddToken(tokens, v >= 9, 165);
-//          VP8AddToken(tokens, !(v & 1), 145);
-        }
-      } else {
-        int mask;
-        const uint8_t* tab;
-        if (v < 3 + (8 << 1)) {          // kCat3  (3b)
-          VP8AddToken(tokens, 0, base_id + 8);
-          VP8AddToken(tokens, 0, base_id + 9);
-          v -= 3 + (8 << 0);
-          mask = 1 << 2;
-          tab = kCat3;
-        } else if (v < 3 + (8 << 2)) {   // kCat4  (4b)
-          VP8AddToken(tokens, 0, base_id + 8);
-          VP8AddToken(tokens, 1, base_id + 9);
-          v -= 3 + (8 << 1);
-          mask = 1 << 3;
-          tab = kCat4;
-        } else if (v < 3 + (8 << 3)) {   // kCat5  (5b)
-          VP8AddToken(tokens, 1, base_id + 8);
-          VP8AddToken(tokens, 0, base_id + 10);
-          v -= 3 + (8 << 2);
-          mask = 1 << 4;
-          tab = kCat5;
-        } else {                         // kCat6 (11b)
-          VP8AddToken(tokens, 1, base_id + 8);
-          VP8AddToken(tokens, 1, base_id + 10);
-          v -= 3 + (8 << 3);
-          mask = 1 << 10;
-          tab = kCat6;
-        }
-        while (mask) {
-          // VP8AddToken(tokens, !!(v & mask), *tab++);
-          mask >>= 1;
-        }
-      }
-      ctx = 2;
-    }
-    b = VP8EncBands[n];
-    // VP8PutBitUniform(bw, sign);
-    if (n == 16 || !VP8AddToken(tokens, n <= res->last, TOKEN_ID(b, ctx, 0))) {
-      return 1;   // EOB
-    }
-  }
-  return 1;
-}
-
-static void RecordTokens(VP8EncIterator* const it,
-                         const VP8ModeScore* const rd, VP8TBuffer tokens[2]) {
+static void RecordTokens(VP8EncIterator* const it, const VP8ModeScore* const rd,
+                         VP8TBuffer* const tokens) {
   int x, y, ch;
   VP8Residual res;
   VP8Encoder* const enc = it->enc_;
 
   VP8IteratorNzToBytes(it);
   if (it->mb_->type_ == 1) {   // i16x16
+    const int ctx = it->top_nz_[8] + it->left_nz_[8];
     InitResidual(0, 1, enc, &res);
     SetResidualCoeffs(rd->y_dc_levels, &res);
-// TODO(skal): FIX ->    it->top_nz_[8] = it->left_nz_[8] =
-      RecordCoeffTokens(it->top_nz_[8] + it->left_nz_[8], &res, &tokens[0]);
+    it->top_nz_[8] = it->left_nz_[8] =
+        VP8RecordCoeffTokens(ctx, 1,
+                             res.first, res.last, res.coeffs, tokens);
+    RecordCoeffs(ctx, &res);
     InitResidual(1, 0, enc, &res);
   } else {
     InitResidual(0, 3, enc, &res);
@@ -663,7 +591,9 @@ static void RecordTokens(VP8EncIterator* const it,
       const int ctx = it->top_nz_[x] + it->left_nz_[y];
       SetResidualCoeffs(rd->y_ac_levels[x + y * 4], &res);
       it->top_nz_[x] = it->left_nz_[y] =
-          RecordCoeffTokens(ctx, &res, &tokens[0]);
+          VP8RecordCoeffTokens(ctx, res.coeff_type,
+                               res.first, res.last, res.coeffs, tokens);
+      RecordCoeffs(ctx, &res);
     }
   }
 
@@ -675,13 +605,16 @@ static void RecordTokens(VP8EncIterator* const it,
         const int ctx = it->top_nz_[4 + ch + x] + it->left_nz_[4 + ch + y];
         SetResidualCoeffs(rd->uv_levels[ch * 2 + x + y * 2], &res);
         it->top_nz_[4 + ch + x] = it->left_nz_[4 + ch + y] =
-            RecordCoeffTokens(ctx, &res, &tokens[1]);
+            VP8RecordCoeffTokens(ctx, 2,
+                                 res.first, res.last, res.coeffs, tokens);
+        RecordCoeffs(ctx, &res);
       }
     }
   }
+  VP8IteratorBytesToNz(it);
 }
 
-#endif    // USE_TOKEN_BUFFER
+#endif    // !DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // ExtraInfo map / Debug function
@@ -697,7 +630,10 @@ static void SetBlock(uint8_t* p, int value, int size) {
 #endif
 
 static void ResetSSE(VP8Encoder* const enc) {
-  memset(enc->sse_, 0, sizeof(enc->sse_));
+  enc->sse_[0] = 0;
+  enc->sse_[1] = 0;
+  enc->sse_[2] = 0;
+  // Note: enc->sse_[3] is managed by alpha.c
   enc->sse_count_ = 0;
 }
 
@@ -736,6 +672,7 @@ static void StoreSideInfo(const VP8EncIterator* const it) {
         const int b = (int)((it->luma_bits_ + it->uv_bits_ + 7) >> 3);
         *info = (b > 255) ? 255 : b; break;
       }
+      case 7: *info = mb->alpha_; break;
       default: *info = 0; break;
     };
   }
@@ -747,99 +684,13 @@ static void StoreSideInfo(const VP8EncIterator* const it) {
 }
 
 //------------------------------------------------------------------------------
-// Main loops
-//
-//  VP8EncLoop(): does the final bitstream coding.
-
-static void ResetAfterSkip(VP8EncIterator* const it) {
-  if (it->mb_->type_ == 1) {
-    *it->nz_ = 0;  // reset all predictors
-    it->left_nz_[8] = 0;
-  } else {
-    *it->nz_ &= (1 << 24);  // preserve the dc_nz bit
-  }
-}
-
-int VP8EncLoop(VP8Encoder* const enc) {
-  int i, s, p;
-  int ok = 1;
-  VP8EncIterator it;
-  VP8ModeScore info;
-  const int dont_use_skip = !enc->proba_.use_skip_proba_;
-  const int rd_opt = enc->rd_opt_level_;
-  const int kAverageBytesPerMB = 5;     // TODO: have a kTable[quality/10]
-  const int bytes_per_parts =
-    enc->mb_w_ * enc->mb_h_ * kAverageBytesPerMB / enc->num_parts_;
-
-  // Initialize the bit-writers
-  for (p = 0; p < enc->num_parts_; ++p) {
-    VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
-  }
-
-  ResetStats(enc);
-  ResetSSE(enc);
-
-  VP8IteratorInit(enc, &it);
-  VP8InitFilter(&it);
-  do {
-    VP8IteratorImport(&it);
-    // Warning! order is important: first call VP8Decimate() and
-    // *then* decide how to code the skip decision if there's one.
-    if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
-      CodeResiduals(it.bw_, &it, &info);
-    } else {   // reset predictors after a skip
-      ResetAfterSkip(&it);
-    }
-#ifdef WEBP_EXPERIMENTAL_FEATURES
-    if (enc->use_layer_) {
-      VP8EncCodeLayerBlock(&it);
-    }
-#endif
-    StoreSideInfo(&it);
-    VP8StoreFilterStats(&it);
-    VP8IteratorExport(&it);
-    ok = VP8IteratorProgress(&it, 20);
-  } while (ok && VP8IteratorNext(&it, it.yuv_out_));
-
-  if (ok) {      // Finalize the partitions, check for extra errors.
-    for (p = 0; p < enc->num_parts_; ++p) {
-      VP8BitWriterFinish(enc->parts_ + p);
-      ok &= !enc->parts_[p].error_;
-    }
-  }
-
-  if (ok) {      // All good. Finish up.
-    if (enc->pic_->stats) {           // finalize byte counters...
-      for (i = 0; i <= 2; ++i) {
-        for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
-          enc->residual_bytes_[i][s] = (int)((it.bit_count_[s][i] + 7) >> 3);
-        }
-      }
-    }
-    VP8AdjustFilterStrength(&it);     // ...and store filter stats.
-  } else {
-    // Something bad happened -> need to do some memory cleanup.
-    VP8EncFreeBitWriters(enc);
-  }
-
-  return ok;
-}
-
-//------------------------------------------------------------------------------
-//  VP8StatLoop(): only collect statistics (number of skips, token usage, ...)
-//                 This is used for deciding optimal probabilities. It also
-//                 modifies the quantizer value if some target (size, PNSR)
-//                 was specified.
+//  StatLoop(): only collect statistics (number of skips, token usage, ...).
+//  This is used for deciding optimal probabilities. It also modifies the
+//  quantizer value if some target (size, PNSR) was specified.
 
 #define kHeaderSizeEstimate (15 + 20 + 10)      // TODO: fix better
 
-static int OneStatPass(VP8Encoder* const enc, float q, int rd_opt, int nb_mbs,
-                       float* const PSNR, int percent_delta) {
-  VP8EncIterator it;
-  uint64_t size = 0;
-  uint64_t distortion = 0;
-  const uint64_t pixel_count = nb_mbs * 384;
-
+static void SetLoopParams(VP8Encoder* const enc, float q) {
   // Make sure the quality parameter is inside valid bounds
   if (q < 0.) {
     q = 0;
@@ -848,10 +699,23 @@ static int OneStatPass(VP8Encoder* const enc, float q, int rd_opt, int nb_mbs,
   }
 
   VP8SetSegmentParams(enc, q);      // setup segment quantizations and filters
+  SetSegmentProbas(enc);            // compute segment probabilities
 
   ResetStats(enc);
   ResetTokenStats(enc);
 
+  ResetSSE(enc);
+}
+
+static int OneStatPass(VP8Encoder* const enc, float q, VP8RDLevel rd_opt,
+                       int nb_mbs, float* const PSNR, int percent_delta) {
+  VP8EncIterator it;
+  uint64_t size = 0;
+  uint64_t distortion = 0;
+  const uint64_t pixel_count = nb_mbs * 384;
+
+  SetLoopParams(enc, q);
+
   VP8IteratorInit(enc, &it);
   do {
     VP8ModeScore info;
@@ -867,7 +731,7 @@ static int OneStatPass(VP8Encoder* const enc, float q, int rd_opt, int nb_mbs,
       return 0;
   } while (VP8IteratorNext(&it, it.yuv_out_) && --nb_mbs > 0);
   size += FinalizeSkipProba(enc);
-  size += FinalizeTokenProbas(enc);
+  size += FinalizeTokenProbas(&enc->proba_);
   size += enc->segment_hdr_.size_;
   size = ((size + 1024) >> 11) + kHeaderSizeEstimate;
 
@@ -880,10 +744,10 @@ static int OneStatPass(VP8Encoder* const enc, float q, int rd_opt, int nb_mbs,
 // successive refinement increments.
 static const int dqs[] = { 20, 15, 10, 8, 6, 4, 2, 1, 0 };
 
-int VP8StatLoop(VP8Encoder* const enc) {
-  const int do_search =
-    (enc->config_->target_size > 0 || enc->config_->target_PSNR > 0);
-  const int fast_probe = (enc->method_ < 2 && !do_search);
+static int StatLoop(VP8Encoder* const enc) {
+  const int method = enc->method_;
+  const int do_search = enc->do_search_;
+  const int fast_probe = ((method == 0 || method == 3) && !do_search);
   float q = enc->config_->quality;
   const int max_passes = enc->config_->pass;
   const int task_percent = 20;
@@ -894,12 +758,18 @@ int VP8StatLoop(VP8Encoder* const enc) {
 
   // Fast mode: quick analysis pass over few mbs. Better than nothing.
   nb_mbs = enc->mb_w_ * enc->mb_h_;
-  if (fast_probe && nb_mbs > 100) nb_mbs = 100;
+  if (fast_probe) {
+    if (method == 3) {  // we need more stats for method 3 to be reliable.
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 1 : 100;
+    } else {
+      nb_mbs = (nb_mbs > 200) ? nb_mbs >> 2 : 50;
+    }
+  }
 
   // No target size: just do several pass without changing 'q'
   if (!do_search) {
     for (pass = 0; pass < max_passes; ++pass) {
-      const int rd_opt = (enc->method_ > 2);
+      const VP8RDLevel rd_opt = (method >= 3) ? RD_OPT_BASIC : RD_OPT_NONE;
       if (!OneStatPass(enc, q, rd_opt, nb_mbs, NULL, percent_per_pass)) {
         return 0;
       }
@@ -907,15 +777,14 @@ int VP8StatLoop(VP8Encoder* const enc) {
   } else {
     // binary search for a size close to target
     for (pass = 0; pass < max_passes && (dqs[pass] > 0); ++pass) {
-      const int rd_opt = 1;
       float PSNR;
       int criterion;
-      const int size = OneStatPass(enc, q, rd_opt, nb_mbs, &PSNR,
+      const int size = OneStatPass(enc, q, RD_OPT_BASIC, nb_mbs, &PSNR,
                                    percent_per_pass);
 #if DEBUG_SEARCH
       printf("#%d size=%d PSNR=%.2f q=%.2f\n", pass, size, PSNR, q);
 #endif
-      if (!size) return 0;
+      if (size == 0) return 0;
       if (enc->config_->target_PSNR > 0) {
         criterion = (PSNR < enc->config_->target_PSNR);
       } else {
@@ -929,10 +798,173 @@ int VP8StatLoop(VP8Encoder* const enc) {
       }
     }
   }
+  VP8CalculateLevelCosts(&enc->proba_);  // finalize costs
   return WebPReportProgress(enc->pic_, final_percent, &enc->percent_);
 }
 
 //------------------------------------------------------------------------------
+// Main loops
+//
+
+static const int kAverageBytesPerMB[8] = { 50, 24, 16, 9, 7, 5, 3, 2 };
+
+static int PreLoopInitialize(VP8Encoder* const enc) {
+  int p;
+  int ok = 1;
+  const int average_bytes_per_MB = kAverageBytesPerMB[enc->base_quant_ >> 4];
+  const int bytes_per_parts =
+      enc->mb_w_ * enc->mb_h_ * average_bytes_per_MB / enc->num_parts_;
+  // Initialize the bit-writers
+  for (p = 0; ok && p < enc->num_parts_; ++p) {
+    ok = VP8BitWriterInit(enc->parts_ + p, bytes_per_parts);
+  }
+  if (!ok) VP8EncFreeBitWriters(enc);  // malloc error occurred
+  return ok;
+}
+
+static int PostLoopFinalize(VP8EncIterator* const it, int ok) {
+  VP8Encoder* const enc = it->enc_;
+  if (ok) {      // Finalize the partitions, check for extra errors.
+    int p;
+    for (p = 0; p < enc->num_parts_; ++p) {
+      VP8BitWriterFinish(enc->parts_ + p);
+      ok &= !enc->parts_[p].error_;
+    }
+  }
+
+  if (ok) {      // All good. Finish up.
+    if (enc->pic_->stats) {           // finalize byte counters...
+      int i, s;
+      for (i = 0; i <= 2; ++i) {
+        for (s = 0; s < NUM_MB_SEGMENTS; ++s) {
+          enc->residual_bytes_[i][s] = (int)((it->bit_count_[s][i] + 7) >> 3);
+        }
+      }
+    }
+    VP8AdjustFilterStrength(it);     // ...and store filter stats.
+  } else {
+    // Something bad happened -> need to do some memory cleanup.
+    VP8EncFreeBitWriters(enc);
+  }
+  return ok;
+}
+
+//------------------------------------------------------------------------------
+//  VP8EncLoop(): does the final bitstream coding.
+
+static void ResetAfterSkip(VP8EncIterator* const it) {
+  if (it->mb_->type_ == 1) {
+    *it->nz_ = 0;  // reset all predictors
+    it->left_nz_[8] = 0;
+  } else {
+    *it->nz_ &= (1 << 24);  // preserve the dc_nz bit
+  }
+}
+
+int VP8EncLoop(VP8Encoder* const enc) {
+  VP8EncIterator it;
+  int ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  StatLoop(enc);  // stats-collection loop
+
+  VP8IteratorInit(enc, &it);
+  VP8InitFilter(&it);
+  do {
+    VP8ModeScore info;
+    const int dont_use_skip = !enc->proba_.use_skip_proba_;
+    const VP8RDLevel rd_opt = enc->rd_opt_level_;
+
+    VP8IteratorImport(&it);
+    // Warning! order is important: first call VP8Decimate() and
+    // *then* decide how to code the skip decision if there's one.
+    if (!VP8Decimate(&it, &info, rd_opt) || dont_use_skip) {
+      CodeResiduals(it.bw_, &it, &info);
+    } else {   // reset predictors after a skip
+      ResetAfterSkip(&it);
+    }
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (enc->use_layer_) {
+      VP8EncCodeLayerBlock(&it);
+    }
+#endif
+    StoreSideInfo(&it);
+    VP8StoreFilterStats(&it);
+    VP8IteratorExport(&it);
+    ok = VP8IteratorProgress(&it, 20);
+  } while (ok && VP8IteratorNext(&it, it.yuv_out_));
+
+  return PostLoopFinalize(&it, ok);
+}
+
+//------------------------------------------------------------------------------
+// Single pass using Token Buffer.
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  int ok;
+  // refresh the proba 8 times per pass
+  const int max_count = (enc->mb_w_ * enc->mb_h_) >> 3;
+  int cnt = max_count;
+  VP8EncIterator it;
+  VP8Proba* const proba = &enc->proba_;
+  const VP8RDLevel rd_opt = enc->rd_opt_level_;
+
+  assert(enc->num_parts_ == 1);
+  assert(enc->use_tokens_);
+  assert(proba->use_skip_proba_ == 0);
+  assert(rd_opt >= RD_OPT_BASIC);   // otherwise, token-buffer won't be useful
+  assert(!enc->do_search_);         // TODO(skal): handle pass and dichotomy
+
+  SetLoopParams(enc, enc->config_->quality);
+
+  ok = PreLoopInitialize(enc);
+  if (!ok) return 0;
+
+  VP8IteratorInit(enc, &it);
+  VP8InitFilter(&it);
+  do {
+    VP8ModeScore info;
+    VP8IteratorImport(&it);
+    if (--cnt < 0) {
+      FinalizeTokenProbas(proba);
+      VP8CalculateLevelCosts(proba);  // refresh cost tables for rd-opt
+      cnt = max_count;
+    }
+    VP8Decimate(&it, &info, rd_opt);
+    RecordTokens(&it, &info, &enc->tokens_);
+#ifdef WEBP_EXPERIMENTAL_FEATURES
+    if (enc->use_layer_) {
+      VP8EncCodeLayerBlock(&it);
+    }
+#endif
+    StoreSideInfo(&it);
+    VP8StoreFilterStats(&it);
+    VP8IteratorExport(&it);
+    ok = VP8IteratorProgress(&it, 20);
+  } while (ok && VP8IteratorNext(&it, it.yuv_out_));
+
+  ok = ok && WebPReportProgress(enc->pic_, enc->percent_ + 20, &enc->percent_);
+
+  if (ok) {
+    FinalizeTokenProbas(proba);
+    ok = VP8EmitTokens(&enc->tokens_, enc->parts_ + 0,
+                       (const uint8_t*)proba->coeffs_, 1);
+  }
+
+  return PostLoopFinalize(&it, ok);
+}
+
+#else
+
+int VP8EncTokenLoop(VP8Encoder* const enc) {
+  (void)enc;
+  return 0;   // we shouldn't be here.
+}
+
+#endif    // DISABLE_TOKEN_BUFFER
+
+//------------------------------------------------------------------------------
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
diff --git a/src/enc/histogram.c b/src/enc/histogram.c
index ca838e06..bc3d3e51 100644
--- a/src/enc/histogram.c
+++ b/src/enc/histogram.c
@@ -55,9 +55,9 @@ VP8LHistogramSet* VP8LAllocateHistogramSet(int size, int cache_bits) {
   int i;
   VP8LHistogramSet* set;
   VP8LHistogram* bulk;
-  const uint64_t total_size = (uint64_t)sizeof(*set)
-                            + size * sizeof(*set->histograms)
-                            + size * sizeof(**set->histograms);
+  const uint64_t total_size = sizeof(*set)
+                            + (uint64_t)size * sizeof(*set->histograms)
+                            + (uint64_t)size * sizeof(**set->histograms);
   uint8_t* memory = (uint8_t*)WebPSafeMalloc(total_size, sizeof(*memory));
   if (memory == NULL) return NULL;
 
@@ -249,14 +249,15 @@ static uint32_t MyRand(uint32_t *seed) {
 }
 
 static int HistogramCombine(const VP8LHistogramSet* const in,
-                            VP8LHistogramSet* const out, int num_pairs) {
+                            VP8LHistogramSet* const out, int iter_mult,
+                            int num_pairs, int num_tries_no_success) {
   int ok = 0;
   int i, iter;
   uint32_t seed = 0;
   int tries_with_no_success = 0;
-  const int min_cluster_size = 2;
   int out_size = in->size;
-  const int outer_iters = in->size * 3;
+  const int outer_iters = in->size * iter_mult;
+  const int min_cluster_size = 2;
   VP8LHistogram* const histos = (VP8LHistogram*)malloc(2 * sizeof(*histos));
   VP8LHistogram* cur_combo = histos + 0;    // trial merged histogram
   VP8LHistogram* best_combo = histos + 1;   // best merged histogram so far
@@ -271,12 +272,12 @@ static int HistogramCombine(const VP8LHistogramSet* const in,
 
   // Collapse similar histograms in 'out'.
   for (iter = 0; iter < outer_iters && out_size >= min_cluster_size; ++iter) {
-    // We pick the best pair to be combined out of 'inner_iters' pairs.
     double best_cost_diff = 0.;
     int best_idx1 = 0, best_idx2 = 1;
     int j;
+    const int num_tries = (num_pairs < out_size) ? num_pairs : out_size;
     seed += iter;
-    for (j = 0; j < num_pairs; ++j) {
+    for (j = 0; j < num_tries; ++j) {
       double curr_cost_diff;
       // Choose two histograms at random and try to combine them.
       const uint32_t idx1 = MyRand(&seed) % out_size;
@@ -315,7 +316,7 @@ static int HistogramCombine(const VP8LHistogramSet* const in,
       }
       tries_with_no_success = 0;
     }
-    if (++tries_with_no_success >= 50) {
+    if (++tries_with_no_success >= num_tries_no_success) {
       break;
     }
   }
@@ -384,8 +385,13 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
   int ok = 0;
   const int histo_xsize = histo_bits ? VP8LSubSampleSize(xsize, histo_bits) : 1;
   const int histo_ysize = histo_bits ? VP8LSubSampleSize(ysize, histo_bits) : 1;
-  const int num_histo_pairs = 10 + quality / 2;  // For HistogramCombine().
   const int histo_image_raw_size = histo_xsize * histo_ysize;
+
+  // Heuristic params for HistogramCombine().
+  const int num_tries_no_success = 8 + (quality >> 1);
+  const int iter_mult = (quality < 27) ? 1 : 1 + ((quality - 27) >> 4);
+  const int num_pairs = (quality < 25) ? 10 : (5 * quality) >> 3;
+
   VP8LHistogramSet* const image_out =
       VP8LAllocateHistogramSet(histo_image_raw_size, cache_bits);
   if (image_out == NULL) return 0;
@@ -393,7 +399,8 @@ int VP8LGetHistoImageSymbols(int xsize, int ysize,
   // Build histogram image.
   HistogramBuildImage(xsize, histo_bits, refs, image_out);
   // Collapse similar histograms.
-  if (!HistogramCombine(image_out, image_in, num_histo_pairs)) {
+  if (!HistogramCombine(image_out, image_in, iter_mult, num_pairs,
+                        num_tries_no_success)) {
     goto Error;
   }
   // Find the optimal map from original histograms to the final ones.
diff --git a/src/enc/picture.c b/src/enc/picture.c
index 44eed060..739a7aa2 100644
--- a/src/enc/picture.c
+++ b/src/enc/picture.c
@@ -290,8 +290,11 @@ int WebPPictureView(const WebPPicture* src,
     dst->y = src->y + top * src->y_stride + left;
     dst->u = src->u + (top >> 1) * src->uv_stride + (left >> 1);
     dst->v = src->v + (top >> 1) * src->uv_stride + (left >> 1);
+    dst->y_stride = src->y_stride;
+    dst->uv_stride = src->uv_stride;
     if (src->a != NULL) {
       dst->a = src->a + top * src->a_stride + left;
+      dst->a_stride = src->a_stride;
     }
 #ifdef WEBP_EXPERIMENTAL_FEATURES
     if (src->u0 != NULL) {
@@ -299,10 +302,12 @@ int WebPPictureView(const WebPPicture* src,
           IS_YUV_CSP(dst->colorspace, WEBP_YUV422) ? (left >> 1) : left;
       dst->u0 = src->u0 + top * src->uv0_stride + left_pos;
       dst->v0 = src->v0 + top * src->uv0_stride + left_pos;
+      dst->uv0_stride = src->uv0_stride;
     }
 #endif
   } else {
     dst->argb = src->argb + top * src->argb_stride + left;
+    dst->argb_stride = src->argb_stride;
   }
   return 1;
 }
@@ -801,11 +806,11 @@ int WebPPictureYUVAToARGB(WebPPicture* picture) {
     // Insert alpha values if needed, in replacement for the default 0xff ones.
     if (picture->colorspace & WEBP_CSP_ALPHA_BIT) {
       for (y = 0; y < height; ++y) {
-        uint32_t* const dst = picture->argb + y * picture->argb_stride;
+        uint32_t* const argb_dst = picture->argb + y * picture->argb_stride;
         const uint8_t* const src = picture->a + y * picture->a_stride;
         int x;
         for (x = 0; x < width; ++x) {
-          dst[x] = (dst[x] & 0x00ffffffu) | (src[x] << 24);
+          argb_dst[x] = (argb_dst[x] & 0x00ffffffu) | (src[x] << 24);
         }
       }
     }
@@ -906,67 +911,135 @@ void WebPCleanupTransparentArea(WebPPicture* pic) {
 #undef SIZE
 #undef SIZE2
 
+//------------------------------------------------------------------------------
+// local-min distortion
+//
+// For every pixel in the *reference* picture, we search for the local best
+// match in the compressed image. This is not a symmetrical measure.
+
+// search radius. Shouldn't be too large.
+#define RADIUS 2
+
+static float AccumulateLSIM(const uint8_t* src, int src_stride,
+                            const uint8_t* ref, int ref_stride,
+                            int w, int h) {
+  int x, y;
+  double total_sse = 0.;
+  for (y = 0; y < h; ++y) {
+    const int y_0 = (y - RADIUS < 0) ? 0 : y - RADIUS;
+    const int y_1 = (y + RADIUS + 1 >= h) ? h : y + RADIUS + 1;
+    for (x = 0; x < w; ++x) {
+      const int x_0 = (x - RADIUS < 0) ? 0 : x - RADIUS;
+      const int x_1 = (x + RADIUS + 1 >= w) ? w : x + RADIUS + 1;
+      double best_sse = 255. * 255.;
+      const double value = (double)ref[y * ref_stride + x];
+      int i, j;
+      for (j = y_0; j < y_1; ++j) {
+        const uint8_t* s = src + j * src_stride;
+        for (i = x_0; i < x_1; ++i) {
+          const double sse = (double)(s[i] - value) * (s[i] - value);
+          if (sse < best_sse) best_sse = sse;
+        }
+      }
+      total_sse += best_sse;
+    }
+  }
+  return (float)total_sse;
+}
+#undef RADIUS
 
 //------------------------------------------------------------------------------
 // Distortion
 
 // Max value returned in case of exact similarity.
 static const double kMinDistortion_dB = 99.;
+static float GetPSNR(const double v) {
+  return (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
+                          : kMinDistortion_dB);
+}
 
-int WebPPictureDistortion(const WebPPicture* pic1, const WebPPicture* pic2,
+int WebPPictureDistortion(const WebPPicture* src, const WebPPicture* ref,
                           int type, float result[5]) {
-  int c;
   DistoStats stats[5];
   int has_alpha;
+  int uv_w, uv_h;
 
-  if (pic1 == NULL || pic2 == NULL ||
-      pic1->width != pic2->width || pic1->height != pic2->height ||
-      pic1->y == NULL || pic2->y == NULL ||
-      pic1->u == NULL || pic2->u == NULL ||
-      pic1->v == NULL || pic2->v == NULL ||
+  if (src == NULL || ref == NULL ||
+      src->width != ref->width || src->height != ref->height ||
+      src->y == NULL || ref->y == NULL ||
+      src->u == NULL || ref->u == NULL ||
+      src->v == NULL || ref->v == NULL ||
       result == NULL) {
     return 0;
   }
   // TODO(skal): provide distortion for ARGB too.
-  if (pic1->use_argb == 1 || pic1->use_argb != pic2->use_argb) {
+  if (src->use_argb == 1 || src->use_argb != ref->use_argb) {
     return 0;
   }
 
-  has_alpha = !!(pic1->colorspace & WEBP_CSP_ALPHA_BIT);
-  if (has_alpha != !!(pic2->colorspace & WEBP_CSP_ALPHA_BIT) ||
-      (has_alpha && (pic1->a == NULL || pic2->a == NULL))) {
+  has_alpha = !!(src->colorspace & WEBP_CSP_ALPHA_BIT);
+  if (has_alpha != !!(ref->colorspace & WEBP_CSP_ALPHA_BIT) ||
+      (has_alpha && (src->a == NULL || ref->a == NULL))) {
     return 0;
   }
 
   memset(stats, 0, sizeof(stats));
-  VP8SSIMAccumulatePlane(pic1->y, pic1->y_stride,
-                         pic2->y, pic2->y_stride,
-                         pic1->width, pic1->height, &stats[0]);
-  VP8SSIMAccumulatePlane(pic1->u, pic1->uv_stride,
-                         pic2->u, pic2->uv_stride,
-                         (pic1->width + 1) >> 1, (pic1->height + 1) >> 1,
-                         &stats[1]);
-  VP8SSIMAccumulatePlane(pic1->v, pic1->uv_stride,
-                         pic2->v, pic2->uv_stride,
-                         (pic1->width + 1) >> 1, (pic1->height + 1) >> 1,
-                         &stats[2]);
-  if (has_alpha) {
-    VP8SSIMAccumulatePlane(pic1->a, pic1->a_stride,
-                           pic2->a, pic2->a_stride,
-                           pic1->width, pic1->height, &stats[3]);
-  }
-  for (c = 0; c <= 4; ++c) {
-    if (type == 1) {
-      const double v = VP8SSIMGet(&stats[c]);
-      result[c] = (float)((v < 1.) ? -10.0 * log10(1. - v)
-                                   : kMinDistortion_dB);
-    } else {
-      const double v = VP8SSIMGetSquaredError(&stats[c]);
-      result[c] = (float)((v > 0.) ? -4.3429448 * log(v / (255 * 255.))
-                                   : kMinDistortion_dB);
+
+  uv_w = HALVE(src->width);
+  uv_h = HALVE(src->height);
+  if (type >= 2) {
+    float sse[4];
+    sse[0] = AccumulateLSIM(src->y, src->y_stride,
+                            ref->y, ref->y_stride, src->width, src->height);
+    sse[1] = AccumulateLSIM(src->u, src->uv_stride,
+                            ref->u, ref->uv_stride, uv_w, uv_h);
+    sse[2] = AccumulateLSIM(src->v, src->uv_stride,
+                            ref->v, ref->uv_stride, uv_w, uv_h);
+    sse[3] = has_alpha ? AccumulateLSIM(src->a, src->a_stride,
+                                        ref->a, ref->a_stride,
+                                        src->width, src->height)
+                       : 0.f;
+    result[0] = GetPSNR(sse[0] / (src->width * src->height));
+    result[1] = GetPSNR(sse[1] / (uv_w * uv_h));
+    result[2] = GetPSNR(sse[2] / (uv_w * uv_h));
+    result[3] = GetPSNR(sse[3] / (src->width * src->height));
+    {
+      double total_sse = sse[0] + sse[1] + sse[2];
+      int total_pixels = src->width * src->height + 2 * uv_w * uv_h;
+      if (has_alpha) {
+        total_pixels += src->width * src->height;
+        total_sse += sse[3];
+      }
+      result[4] = GetPSNR(total_sse / total_pixels);
+    }
+  } else {
+    int c;
+    VP8SSIMAccumulatePlane(src->y, src->y_stride,
+                           ref->y, ref->y_stride,
+                           src->width, src->height, &stats[0]);
+    VP8SSIMAccumulatePlane(src->u, src->uv_stride,
+                           ref->u, ref->uv_stride,
+                           uv_w, uv_h, &stats[1]);
+    VP8SSIMAccumulatePlane(src->v, src->uv_stride,
+                           ref->v, ref->uv_stride,
+                           uv_w, uv_h, &stats[2]);
+    if (has_alpha) {
+      VP8SSIMAccumulatePlane(src->a, src->a_stride,
+                             ref->a, ref->a_stride,
+                             src->width, src->height, &stats[3]);
+    }
+    for (c = 0; c <= 4; ++c) {
+      if (type == 1) {
+        const double v = VP8SSIMGet(&stats[c]);
+        result[c] = (float)((v < 1.) ? -10.0 * log10(1. - v)
+                                     : kMinDistortion_dB);
+      } else {
+        const double v = VP8SSIMGetSquaredError(&stats[c]);
+        result[c] = GetPSNR(v);
+      }
+      // Accumulate forward
+      if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
     }
-    // Accumulate forward
-    if (c < 4) VP8SSIMAddStats(&stats[c], &stats[4]);
   }
   return 1;
 }
diff --git a/src/enc/quant.c b/src/enc/quant.c
index ea153849..dcfd4d16 100644
--- a/src/enc/quant.c
+++ b/src/enc/quant.c
@@ -27,6 +27,8 @@
 #define SNS_TO_DQ 0.9     // Scaling constant between the sns value and the QP
                           // power-law modulation. Must be strictly less than 1.
 
+#define I4_PENALTY 4000   // Rate-penalty for quick i4/i16 decision
+
 #define MULT_8B(a, b) (((a) * (b) + 128) >> 8)
 
 #if defined(__cplusplus) || defined(c_plusplus)
@@ -224,28 +226,90 @@ static void SetupFilterStrength(VP8Encoder* const enc) {
 // We want to emulate jpeg-like behaviour where the expected "good" quality
 // is around q=75. Internally, our "good" middle is around c=50. So we
 // map accordingly using linear piece-wise function
-static double QualityToCompression(double q) {
-  const double c = q / 100.;
-  return (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.;
+static double QualityToCompression(double c) {
+  const double linear_c = (c < 0.75) ? c * (2. / 3.) : 2. * c - 1.;
+  // The file size roughly scales as pow(quantizer, 3.). Actually, the
+  // exponent is somewhere between 2.8 and 3.2, but we're mostly interested
+  // in the mid-quant range. So we scale the compressibility inversely to
+  // this power-law: quant ~= compression ^ 1/3. This law holds well for
+  // low quant. Finer modelling for high-quant would make use of kAcTable[]
+  // more explicitly.
+  const double v = pow(linear_c, 1 / 3.);
+  return v;
+}
+
+static double QualityToJPEGCompression(double c, double alpha) {
+  // We map the complexity 'alpha' and quality setting 'c' to a compression
+  // exponent empirically matched to the compression curve of libjpeg6b.
+  // On average, the WebP output size will be roughly similar to that of a
+  // JPEG file compressed with same quality factor.
+  const double amin = 0.30;
+  const double amax = 0.85;
+  const double exp_min = 0.4;
+  const double exp_max = 0.9;
+  const double slope = (exp_min - exp_max) / (amax - amin);
+  // Linearly interpolate 'expn' from exp_min to exp_max
+  // in the [amin, amax] range.
+  const double expn = (alpha > amax) ? exp_min
+                    : (alpha < amin) ? exp_max
+                    : exp_max + slope * (alpha - amin);
+  const double v = pow(c, expn);
+  return v;
+}
+
+static int SegmentsAreEquivalent(const VP8SegmentInfo* const S1,
+                                 const VP8SegmentInfo* const S2) {
+  return (S1->quant_ == S2->quant_) && (S1->fstrength_ == S2->fstrength_);
+}
+
+static void SimplifySegments(VP8Encoder* const enc) {
+  int map[NUM_MB_SEGMENTS] = { 0, 1, 2, 3 };
+  const int num_segments = enc->segment_hdr_.num_segments_;
+  int num_final_segments = 1;
+  int s1, s2;
+  for (s1 = 1; s1 < num_segments; ++s1) {    // find similar segments
+    const VP8SegmentInfo* const S1 = &enc->dqm_[s1];
+    int found = 0;
+    // check if we already have similar segment
+    for (s2 = 0; s2 < num_final_segments; ++s2) {
+      const VP8SegmentInfo* const S2 = &enc->dqm_[s2];
+      if (SegmentsAreEquivalent(S1, S2)) {
+        found = 1;
+        break;
+      }
+    }
+    map[s1] = s2;
+    if (!found) {
+      if (num_final_segments != s1) {
+        enc->dqm_[num_final_segments] = enc->dqm_[s1];
+      }
+      ++num_final_segments;
+    }
+  }
+  if (num_final_segments < num_segments) {  // Remap
+    int i = enc->mb_w_ * enc->mb_h_;
+    while (i-- > 0) enc->mb_info_[i].segment_ = map[enc->mb_info_[i].segment_];
+    enc->segment_hdr_.num_segments_ = num_final_segments;
+    // Replicate the trailing segment infos (it's mostly cosmetics)
+    for (i = num_final_segments; i < num_segments; ++i) {
+      enc->dqm_[i] = enc->dqm_[num_final_segments - 1];
+    }
+  }
 }
 
 void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
   int i;
   int dq_uv_ac, dq_uv_dc;
-  const int num_segments = enc->config_->segments;
+  const int num_segments = enc->segment_hdr_.num_segments_;
   const double amp = SNS_TO_DQ * enc->config_->sns_strength / 100. / 128.;
-  const double c_base = QualityToCompression(quality);
+  const double Q = quality / 100.;
+  const double c_base = enc->config_->emulate_jpeg_size ?
+      QualityToJPEGCompression(Q, enc->alpha_ / 255.) :
+      QualityToCompression(Q);
   for (i = 0; i < num_segments; ++i) {
-    // The file size roughly scales as pow(quantizer, 3.). Actually, the
-    // exponent is somewhere between 2.8 and 3.2, but we're mostly interested
-    // in the mid-quant range. So we scale the compressibility inversely to
-    // this power-law: quant ~= compression ^ 1/3. This law holds well for
-    // low quant. Finer modelling for high-quant would make use of kAcTable[]
-    // more explicitely.
-    // Additionally, we modulate the base exponent 1/3 to accommodate for the
-    // quantization susceptibility and allow denser segments to be quantized
-    // more.
-    const double expn = (1. - amp * enc->dqm_[i].alpha_) / 3.;
+    // We modulate the base coefficient to accommodate for the quantization
+    // susceptibility and allow denser segments to be quantized more.
+    const double expn = 1. - amp * enc->dqm_[i].alpha_;
     const double c = pow(c_base, expn);
     const int q = (int)(127. * (1. - c));
     assert(expn > 0.);
@@ -281,9 +345,11 @@ void VP8SetSegmentParams(VP8Encoder* const enc, float quality) {
   enc->dq_uv_dc_ = dq_uv_dc;
   enc->dq_uv_ac_ = dq_uv_ac;
 
-  SetupMatrices(enc);
-
   SetupFilterStrength(enc);   // initialize segments' filtering, eventually
+
+  if (num_segments > 1) SimplifySegments(enc);
+
+  SetupMatrices(enc);         // finalize quantization matrices
 }
 
 //------------------------------------------------------------------------------
@@ -709,7 +775,7 @@ static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
   int mode;
 
   rd->mode_i16 = -1;
-  for (mode = 0; mode < 4; ++mode) {
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
     uint8_t* const tmp_dst = it->yuv_out2_ + Y_OFF;  // scratch buffer
     int nz;
 
@@ -838,7 +904,7 @@ static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
 
   rd->mode_uv = -1;
   InitScore(&rd_best);
-  for (mode = 0; mode < 4; ++mode) {
+  for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
     VP8ModeScore rd_uv;
 
     // Reconstruct
@@ -867,10 +933,10 @@ static void PickBestUV(VP8EncIterator* const it, VP8ModeScore* const rd) {
 
 static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
   const VP8Encoder* const enc = it->enc_;
-  const int i16 = (it->mb_->type_ == 1);
+  const int is_i16 = (it->mb_->type_ == 1);
   int nz = 0;
 
-  if (i16) {
+  if (is_i16) {
     nz = ReconstructIntra16(it, rd, it->yuv_out_ + Y_OFF, it->preds_[0]);
   } else {
     VP8IteratorStartI4(it);
@@ -889,11 +955,66 @@ static void SimpleQuantize(VP8EncIterator* const it, VP8ModeScore* const rd) {
   rd->nz = nz;
 }
 
+// Refine intra16/intra4 sub-modes based on distortion only (not rate).
+static void DistoRefine(VP8EncIterator* const it, int try_both_i4_i16) {
+  const int is_i16 = (it->mb_->type_ == 1);
+  score_t best_score = MAX_COST;
+
+  if (try_both_i4_i16 || is_i16) {
+    int mode;
+    int best_mode = -1;
+    for (mode = 0; mode < NUM_PRED_MODES; ++mode) {
+      const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
+      const uint8_t* const src = it->yuv_in_ + Y_OFF;
+      const score_t score = VP8SSE16x16(src, ref);
+      if (score < best_score) {
+        best_mode = mode;
+        best_score = score;
+      }
+    }
+    VP8SetIntra16Mode(it, best_mode);
+  }
+  if (try_both_i4_i16 || !is_i16) {
+    uint8_t modes_i4[16];
+    // We don't evaluate the rate here, but just account for it through a
+    // constant penalty (i4 mode usually needs more bits compared to i16).
+    score_t score_i4 = (score_t)I4_PENALTY;
+
+    VP8IteratorStartI4(it);
+    do {
+      int mode;
+      int best_sub_mode = -1;
+      score_t best_sub_score = MAX_COST;
+      const uint8_t* const src = it->yuv_in_ + Y_OFF + VP8Scan[it->i4_];
+
+      // TODO(skal): we don't really need the prediction pixels here,
+      // but just the distortion against 'src'.
+      VP8MakeIntra4Preds(it);
+      for (mode = 0; mode < NUM_BMODES; ++mode) {
+        const uint8_t* const ref = it->yuv_p_ + VP8I4ModeOffsets[mode];
+        const score_t score = VP8SSE4x4(src, ref);
+        if (score < best_sub_score) {
+          best_sub_mode = mode;
+          best_sub_score = score;
+        }
+      }
+      modes_i4[it->i4_] = best_sub_mode;
+      score_i4 += best_sub_score;
+      if (score_i4 >= best_score) break;
+    } while (VP8IteratorRotateI4(it, it->yuv_in_ + Y_OFF));
+    if (score_i4 < best_score) {
+      VP8SetIntra4Mode(it, modes_i4);
+    }
+  }
+}
+
 //------------------------------------------------------------------------------
 // Entry point
 
-int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt) {
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt) {
   int is_skipped;
+  const int method = it->enc_->method_;
 
   InitScore(rd);
 
@@ -902,22 +1023,21 @@ int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt) {
   VP8MakeLuma16Preds(it);
   VP8MakeChroma8Preds(it);
 
-  // for rd_opt = 2, we perform trellis-quant on the final decision only.
-  // for rd_opt > 2, we use it for every scoring (=much slower).
-  if (rd_opt > 0) {
-    it->do_trellis_ = (rd_opt > 2);
+  if (rd_opt > RD_OPT_NONE) {
+    it->do_trellis_ = (rd_opt >= RD_OPT_TRELLIS_ALL);
     PickBestIntra16(it, rd);
-    if (it->enc_->method_ >= 2) {
+    if (method >= 2) {
       PickBestIntra4(it, rd);
     }
     PickBestUV(it, rd);
-    if (rd_opt == 2) {
+    if (rd_opt == RD_OPT_TRELLIS) {   // finish off with trellis-optim now
       it->do_trellis_ = 1;
       SimpleQuantize(it, rd);
     }
   } else {
-    // TODO: for method_ == 2, pick the best intra4/intra16 based on SSE
-    it->do_trellis_ = (it->enc_->method_ == 2);
+    // For method == 2, pick the best intra4/intra16 based on SSE (~tad slower).
+    // For method <= 1, we refine intra4 or intra16 (but don't re-examine mode).
+    DistoRefine(it, (method >= 2));
     SimpleQuantize(it, rd);
   }
   is_skipped = (rd->nz == 0);
diff --git a/src/enc/syntax.c b/src/enc/syntax.c
index 99c21fec..7639d2fe 100644
--- a/src/enc/syntax.c
+++ b/src/enc/syntax.c
@@ -11,8 +11,9 @@
 
 #include <assert.h>
 
-#include "./vp8enci.h"
+#include "../utils/utils.h"
 #include "webp/format_constants.h"
+#include "./vp8enci.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
@@ -21,18 +22,6 @@ extern "C" {
 //------------------------------------------------------------------------------
 // Helper functions
 
-// TODO(later): Move to webp/format_constants.h?
-static void PutLE24(uint8_t* const data, uint32_t val) {
-  data[0] = (val >>  0) & 0xff;
-  data[1] = (val >>  8) & 0xff;
-  data[2] = (val >> 16) & 0xff;
-}
-
-static void PutLE32(uint8_t* const data, uint32_t val) {
-  PutLE24(data, val);
-  data[3] = (val >> 24) & 0xff;
-}
-
 static int IsVP8XNeeded(const VP8Encoder* const enc) {
   return !!enc->has_alpha_;  // Currently the only case when VP8X is needed.
                              // This could change in the future.
@@ -73,7 +62,7 @@ static WebPEncodingError PutVP8XHeader(const VP8Encoder* const enc) {
   assert(pic->width <= MAX_CANVAS_SIZE && pic->height <= MAX_CANVAS_SIZE);
 
   if (enc->has_alpha_) {
-    flags |= ALPHA_FLAG_BIT;
+    flags |= ALPHA_FLAG;
   }
 
   PutLE32(vp8x + TAG_SIZE,              VP8X_CHUNK_SIZE);
@@ -327,7 +316,9 @@ static size_t GeneratePartition0(VP8Encoder* const enc) {
 
   PutSegmentHeader(bw, enc);
   PutFilterHeader(bw, &enc->filter_hdr_);
-  VP8PutValue(bw, enc->config_->partitions, 2);
+  VP8PutValue(bw, enc->num_parts_ == 8 ? 3 :
+                  enc->num_parts_ == 4 ? 2 :
+                  enc->num_parts_ == 2 ? 1 : 0, 2);
   PutQuant(bw, enc);
   VP8PutBitUniform(bw, 0);   // no proba update
   VP8WriteProbas(bw, &enc->proba_);
diff --git a/src/enc/token.c b/src/enc/token.c
new file mode 100644
index 00000000..4e2f6c00
--- /dev/null
+++ b/src/enc/token.c
@@ -0,0 +1,254 @@
+// Copyright 2011 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// Paginated token buffer
+//
+//  A 'token' is a bit value associated with a probability, either fixed
+// or a later-to-be-determined after statistics have been collected.
+// For dynamic probability, we just record the slot id (idx) for the probability
+// value in the final probability array (uint8_t* probas in VP8EmitTokens).
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include <assert.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./vp8enci.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// we use pages to reduce the number of memcpy()
+#define MAX_NUM_TOKEN 8192          // max number of token per page
+#define FIXED_PROBA_BIT (1u << 14)
+
+struct VP8Tokens {
+  uint16_t tokens_[MAX_NUM_TOKEN];  // bit#15: bit
+                                    // bit #14: constant proba or idx
+                                    // bits 0..13: slot or constant proba
+  VP8Tokens* next_;
+};
+
+//------------------------------------------------------------------------------
+
+void VP8TBufferInit(VP8TBuffer* const b) {
+  b->tokens_ = NULL;
+  b->pages_ = NULL;
+  b->last_page_ = &b->pages_;
+  b->left_ = 0;
+  b->error_ = 0;
+}
+
+void VP8TBufferClear(VP8TBuffer* const b) {
+  if (b != NULL) {
+    const VP8Tokens* p = b->pages_;
+    while (p != NULL) {
+      const VP8Tokens* const next = p->next_;
+      free((void*)p);
+      p = next;
+    }
+    VP8TBufferInit(b);
+  }
+}
+
+static int TBufferNewPage(VP8TBuffer* const b) {
+  VP8Tokens* const page = b->error_ ? NULL : (VP8Tokens*)malloc(sizeof(*page));
+  if (page == NULL) {
+    b->error_ = 1;
+    return 0;
+  }
+  *b->last_page_ = page;
+  b->last_page_ = &page->next_;
+  b->left_ = MAX_NUM_TOKEN;
+  b->tokens_ = page->tokens_;
+  page->next_ = NULL;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#define TOKEN_ID(t, b, ctx, p) \
+    ((p) + NUM_PROBAS * ((ctx) + NUM_CTX * ((b) + NUM_BANDS * (t))))
+
+static WEBP_INLINE int AddToken(VP8TBuffer* const b,
+                                int bit, uint32_t proba_idx) {
+  assert(proba_idx < FIXED_PROBA_BIT);
+  assert(bit == 0 || bit == 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | proba_idx;
+  }
+  return bit;
+}
+
+static WEBP_INLINE void AddConstantToken(VP8TBuffer* const b,
+                                         int bit, int proba) {
+  assert(proba < 256);
+  assert(bit == 0 || bit == 1);
+  if (b->left_ > 0 || TBufferNewPage(b)) {
+    const int slot = --b->left_;
+    b->tokens_[slot] = (bit << 15) | FIXED_PROBA_BIT | proba;
+  }
+}
+
+int VP8RecordCoeffTokens(int ctx, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens) {
+  int n = first;
+  uint32_t base_id = TOKEN_ID(coeff_type, n, ctx, 0);
+  if (!AddToken(tokens, last >= 0, base_id + 0)) {
+    return 0;
+  }
+
+  while (n < 16) {
+    const int c = coeffs[n++];
+    const int sign = c < 0;
+    int v = sign ? -c : c;
+    if (!AddToken(tokens, v != 0, base_id + 1)) {
+      ctx = 0;
+      base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+      continue;
+    }
+    if (!AddToken(tokens, v > 1, base_id + 2)) {
+      ctx = 1;
+    } else {
+      if (!AddToken(tokens, v > 4, base_id + 3)) {
+        if (AddToken(tokens, v != 2, base_id + 4))
+          AddToken(tokens, v == 4, base_id + 5);
+      } else if (!AddToken(tokens, v > 10, base_id + 6)) {
+        if (!AddToken(tokens, v > 6, base_id + 7)) {
+          AddConstantToken(tokens, v == 6, 159);
+        } else {
+          AddConstantToken(tokens, v >= 9, 165);
+          AddConstantToken(tokens, !(v & 1), 145);
+        }
+      } else {
+        int mask;
+        const uint8_t* tab;
+        if (v < 3 + (8 << 1)) {          // VP8Cat3  (3b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 0, base_id + 9);
+          v -= 3 + (8 << 0);
+          mask = 1 << 2;
+          tab = VP8Cat3;
+        } else if (v < 3 + (8 << 2)) {   // VP8Cat4  (4b)
+          AddToken(tokens, 0, base_id + 8);
+          AddToken(tokens, 1, base_id + 9);
+          v -= 3 + (8 << 1);
+          mask = 1 << 3;
+          tab = VP8Cat4;
+        } else if (v < 3 + (8 << 3)) {   // VP8Cat5  (5b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 0, base_id + 10);
+          v -= 3 + (8 << 2);
+          mask = 1 << 4;
+          tab = VP8Cat5;
+        } else {                         // VP8Cat6 (11b)
+          AddToken(tokens, 1, base_id + 8);
+          AddToken(tokens, 1, base_id + 10);
+          v -= 3 + (8 << 3);
+          mask = 1 << 10;
+          tab = VP8Cat6;
+        }
+        while (mask) {
+          AddConstantToken(tokens, !!(v & mask), *tab++);
+          mask >>= 1;
+        }
+      }
+      ctx = 2;
+    }
+    AddConstantToken(tokens, sign, 128);
+    base_id = TOKEN_ID(coeff_type, VP8EncBands[n], ctx, 0);
+    if (n == 16 || !AddToken(tokens, n <= last, base_id + 0)) {
+      return 1;   // EOB
+    }
+  }
+  return 1;
+}
+
+#undef TOKEN_ID
+
+//------------------------------------------------------------------------------
+// This function works, but isn't currently used. Saved for later.
+
+#if 0
+
+static void Record(int bit, proba_t* const stats) {
+  proba_t p = *stats;
+  if (p >= 0xffff0000u) {               // an overflow is inbound.
+    p = ((p + 1u) >> 1) & 0x7fff7fffu;  // -> divide the stats by 2.
+  }
+  // record bit count (lower 16 bits) and increment total count (upper 16 bits).
+  p += 0x00010000u + bit;
+  *stats = p;
+}
+
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats) {
+  const VP8Tokens* p = b->pages_;
+  while (p != NULL) {
+    const int N = (p->next_ == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      if (!(token & FIXED_PROBA_BIT)) {
+        Record((token >> 15) & 1, stats + (token & 0x3fffu));
+      }
+    }
+    p = p->next_;
+  }
+}
+
+#endif   // 0
+
+//------------------------------------------------------------------------------
+// Final coding pass, with known probabilities
+
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass) {
+  const VP8Tokens* p = b->pages_;
+  (void)final_pass;
+  if (b->error_) return 0;
+  while (p != NULL) {
+    const VP8Tokens* const next = p->next_;
+    const int N = (next == NULL) ? b->left_ : 0;
+    int n = MAX_NUM_TOKEN;
+    while (n-- > N) {
+      const uint16_t token = p->tokens_[n];
+      const int bit = (token >> 15) & 1;
+      if (token & FIXED_PROBA_BIT) {
+        VP8PutBit(bw, bit, token & 0xffu);  // constant proba
+      } else {
+        VP8PutBit(bw, bit, probas[token & 0x3fffu]);
+      }
+    }
+    if (final_pass) free((void*)p);
+    p = next;
+  }
+  if (final_pass) b->pages_ = NULL;
+  return 1;
+}
+
+//------------------------------------------------------------------------------
+
+#else     // DISABLE_TOKEN_BUFFER
+
+void VP8TBufferInit(VP8TBuffer* const b) {
+  (void)b;
+}
+void VP8TBufferClear(VP8TBuffer* const b) {
+  (void)b;
+}
+
+#endif    // !DISABLE_TOKEN_BUFFER
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
diff --git a/src/enc/vp8enci.h b/src/enc/vp8enci.h
index a0d9001f..bf902e8f 100644
--- a/src/enc/vp8enci.h
+++ b/src/enc/vp8enci.h
@@ -13,9 +13,10 @@
 #define WEBP_ENC_VP8ENCI_H_
 
 #include <string.h>     // for memcpy()
+#include "webp/encode.h"
 #include "../dsp/dsp.h"
 #include "../utils/bit_writer.h"
-#include "webp/encode.h"
+#include "../utils/thread.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
@@ -27,10 +28,7 @@ extern "C" {
 // version numbers
 #define ENC_MAJ_VERSION 0
 #define ENC_MIN_VERSION 2
-#define ENC_REV_VERSION 0
-
-// size of histogram used by CollectHistogram.
-#define MAX_COEFF_THRESH   64
+#define ENC_REV_VERSION 1
 
 // intra prediction modes
 enum { B_DC_PRED = 0,   // 4x4 modes
@@ -47,7 +45,8 @@ enum { B_DC_PRED = 0,   // 4x4 modes
 
        // Luma16 or UV modes
        DC_PRED = B_DC_PRED, V_PRED = B_VE_PRED,
-       H_PRED = B_HE_PRED, TM_PRED = B_TM_PRED
+       H_PRED = B_HE_PRED, TM_PRED = B_TM_PRED,
+       NUM_PRED_MODES = 4
      };
 
 enum { NUM_MB_SEGMENTS = 4,
@@ -60,6 +59,13 @@ enum { NUM_MB_SEGMENTS = 4,
        MAX_VARIABLE_LEVEL = 67  // last (inclusive) level with variable cost
      };
 
+typedef enum {   // Rate-distortion optimization levels
+  RD_OPT_NONE        = 0,  // no rd-opt
+  RD_OPT_BASIC       = 1,  // basic scoring (no trellis)
+  RD_OPT_TRELLIS     = 2,  // perform trellis-quant on the final decision only
+  RD_OPT_TRELLIS_ALL = 3   // trellis-quant for every scoring (much slower)
+} VP8RDLevel;
+
 // YUV-cache parameters. Cache is 16-pixels wide.
 // The original or reconstructed samples can be accessed using VP8Scan[]
 // The predicted blocks can be accessed using offsets to yuv_p_ and
@@ -160,7 +166,17 @@ typedef int64_t score_t;     // type used for scores, rate, distortion
 static WEBP_INLINE int QUANTDIV(int n, int iQ, int B) {
   return (n * iQ + B) >> QFIX;
 }
-extern const uint8_t VP8Zigzag[16];
+
+// size of histogram used by CollectHistogram.
+#define MAX_COEFF_THRESH   31
+typedef struct VP8Histogram VP8Histogram;
+struct VP8Histogram {
+  // TODO(skal): we only need to store the max_value and last_non_zero actually.
+  int distribution[MAX_COEFF_THRESH + 1];
+};
+
+// Uncomment the following to remove token-buffer code:
+// #define DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // Headers
@@ -314,44 +330,37 @@ void VP8SetSegment(const VP8EncIterator* const it, int segment);
 //------------------------------------------------------------------------------
 // Paginated token buffer
 
-// WIP: #define USE_TOKEN_BUFFER
-
-#ifdef USE_TOKEN_BUFFER
-
-#define MAX_NUM_TOKEN 2048
-
-typedef struct VP8Tokens VP8Tokens;
-struct VP8Tokens {
-  uint16_t tokens_[MAX_NUM_TOKEN];  // bit#15: bit, bits 0..14: slot
-  int left_;
-  VP8Tokens* next_;
-};
+typedef struct VP8Tokens VP8Tokens;  // struct details in token.c
 
 typedef struct {
-  VP8Tokens* rows_;
-  uint16_t* tokens_;    // set to (*last_)->tokens_
-  VP8Tokens** last_;
-  int left_;
-  int error_;  // true in case of malloc error
+#if !defined(DISABLE_TOKEN_BUFFER)
+  VP8Tokens* pages_;        // first page
+  VP8Tokens** last_page_;   // last page
+  uint16_t* tokens_;        // set to (*last_page_)->tokens_
+  int left_;          // how many free tokens left before the page is full.
+#endif
+  int error_;         // true in case of malloc error
 } VP8TBuffer;
 
 void VP8TBufferInit(VP8TBuffer* const b);    // initialize an empty buffer
-int VP8TBufferNewPage(VP8TBuffer* const b);  // allocate a new page
-void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate memory
-
-int VP8EmitTokens(const VP8TBuffer* const b, VP8BitWriter* const bw,
-                  const uint8_t* const probas);
-
-static WEBP_INLINE int VP8AddToken(VP8TBuffer* const b,
-                                   int bit, int proba_idx) {
-  if (b->left_ > 0 || VP8TBufferNewPage(b)) {
-    const int slot = --b->left_;
-    b->tokens_[slot] = (bit << 15) | proba_idx;
-  }
-  return bit;
-}
+void VP8TBufferClear(VP8TBuffer* const b);   // de-allocate pages memory
 
-#endif  // USE_TOKEN_BUFFER
+#if !defined(DISABLE_TOKEN_BUFFER)
+
+// Finalizes bitstream when probabilities are known.
+// Deletes the allocated token memory if final_pass is true.
+int VP8EmitTokens(VP8TBuffer* const b, VP8BitWriter* const bw,
+                  const uint8_t* const probas, int final_pass);
+
+// record the coding of coefficients without knowing the probabilities yet
+int VP8RecordCoeffTokens(int ctx, int coeff_type, int first, int last,
+                         const int16_t* const coeffs,
+                         VP8TBuffer* const tokens);
+
+// unused for now
+void VP8TokenToStats(const VP8TBuffer* const b, proba_t* const stats);
+
+#endif  // !DISABLE_TOKEN_BUFFER
 
 //------------------------------------------------------------------------------
 // VP8Encoder
@@ -376,6 +385,7 @@ struct VP8Encoder {
   // per-partition boolean decoders.
   VP8BitWriter bw_;                         // part0
   VP8BitWriter parts_[MAX_NUM_PARTITIONS];  // token partitions
+  VP8TBuffer tokens_;                       // token buffer
 
   int percent_;                             // for progress
 
@@ -383,6 +393,7 @@ struct VP8Encoder {
   int has_alpha_;
   uint8_t* alpha_data_;       // non-NULL if transparency is present
   uint32_t alpha_data_size_;
+  WebPWorker alpha_worker_;
 
   // enhancement layer
   int use_layer_;
@@ -394,6 +405,7 @@ struct VP8Encoder {
   VP8SegmentInfo dqm_[NUM_MB_SEGMENTS];
   int base_quant_;                 // nominal quantizer value. Only used
                                    // for relative coding of segments' quant.
+  int alpha_;                      // global susceptibility (<=> complexity)
   int uv_alpha_;                   // U/V quantization susceptibility
   // global offset of quantizers, shared by all segments
   int dq_y1_dc_;
@@ -409,9 +421,12 @@ struct VP8Encoder {
   int      block_count_[3];
 
   // quality/speed settings
-  int method_;              // 0=fastest, 6=best/slowest.
-  int rd_opt_level_;        // Deduced from method_.
-  int max_i4_header_bits_;  // partition #0 safeness factor
+  int method_;               // 0=fastest, 6=best/slowest.
+  VP8RDLevel rd_opt_level_;  // Deduced from method_.
+  int max_i4_header_bits_;   // partition #0 safeness factor
+  int thread_level_;         // derived from config->thread_level
+  int do_search_;            // derived from config->target_XXX
+  int use_tokens_;           // if true, use token buffer
 
   // Memory
   VP8MBInfo* mb_info_;   // contextual macroblock infos (mb_w_ + 1)
@@ -455,6 +470,11 @@ void VP8EncFreeBitWriters(VP8Encoder* const enc);
 
   // in frame.c
 extern const uint8_t VP8EncBands[16 + 1];
+extern const uint8_t VP8Cat3[];
+extern const uint8_t VP8Cat4[];
+extern const uint8_t VP8Cat5[];
+extern const uint8_t VP8Cat6[];
+
 // Form all the four Intra16x16 predictions in the yuv_p_ cache
 void VP8MakeLuma16Preds(const VP8EncIterator* const it);
 // Form all the four Chroma8x8 predictions in the yuv_p_ cache
@@ -466,9 +486,9 @@ void VP8MakeIntra4Preds(const VP8EncIterator* const it);
 int VP8GetCostLuma16(VP8EncIterator* const it, const VP8ModeScore* const rd);
 int VP8GetCostLuma4(VP8EncIterator* const it, const int16_t levels[16]);
 int VP8GetCostUV(VP8EncIterator* const it, const VP8ModeScore* const rd);
-// Main stat / coding passes
+// Main coding calls
 int VP8EncLoop(VP8Encoder* const enc);
-int VP8StatLoop(VP8Encoder* const enc);
+int VP8EncTokenLoop(VP8Encoder* const enc);
 
   // in webpenc.c
 // Assign an error code to a picture. Return false for convenience.
@@ -485,12 +505,14 @@ int VP8EncAnalyze(VP8Encoder* const enc);
 // Sets up segment's quantization values, base_quant_ and filter strengths.
 void VP8SetSegmentParams(VP8Encoder* const enc, float quality);
 // Pick best modes and fills the levels. Returns true if skipped.
-int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd, int rd_opt);
+int VP8Decimate(VP8EncIterator* const it, VP8ModeScore* const rd,
+                VP8RDLevel rd_opt);
 
   // in alpha.c
 void VP8EncInitAlpha(VP8Encoder* const enc);    // initialize alpha compression
+int VP8EncStartAlpha(VP8Encoder* const enc);    // start alpha coding process
 int VP8EncFinishAlpha(VP8Encoder* const enc);   // finalize compressed data
-void VP8EncDeleteAlpha(VP8Encoder* const enc);  // delete compressed data
+int VP8EncDeleteAlpha(VP8Encoder* const enc);   // delete compressed data
 
   // in layer.c
 void VP8EncInitLayer(VP8Encoder* const enc);     // init everything
diff --git a/src/enc/vp8l.c b/src/enc/vp8l.c
index 41aa62b7..2798d670 100644
--- a/src/enc/vp8l.c
+++ b/src/enc/vp8l.c
@@ -37,7 +37,8 @@ extern "C" {
 static int CompareColors(const void* p1, const void* p2) {
   const uint32_t a = *(const uint32_t*)p1;
   const uint32_t b = *(const uint32_t*)p2;
-  return (a < b) ? -1 : (a > b) ? 1 : 0;
+  assert(a != b);
+  return (a < b) ? -1 : 1;
 }
 
 // If number of colors in the image is less than or equal to MAX_PALETTE_SIZE,
@@ -220,7 +221,7 @@ static int GetHuffBitLengthsAndCodes(
   }
 
   // Create Huffman trees.
-  for (i = 0; i < histogram_image_size; ++i) {
+  for (i = 0; ok && (i < histogram_image_size); ++i) {
     HuffmanTreeCode* const codes = &huffman_codes[5 * i];
     VP8LHistogram* const histo = histogram_image->histograms[i];
     ok = ok && VP8LCreateHuffmanTree(histo->literal_, 15, codes + 0);
@@ -231,7 +232,11 @@ static int GetHuffBitLengthsAndCodes(
   }
 
  End:
-  if (!ok) free(mem_buf);
+  if (!ok) {
+    free(mem_buf);
+    // If one VP8LCreateHuffmanTree() above fails, we need to clean up behind.
+    memset(huffman_codes, 0, 5 * histogram_image_size * sizeof(*huffman_codes));
+  }
   return ok;
 }
 
@@ -406,9 +411,10 @@ static int StoreHuffmanCode(VP8LBitWriter* const bw,
 }
 
 static void WriteHuffmanCode(VP8LBitWriter* const bw,
-                             const HuffmanTreeCode* const code, int index) {
-  const int depth = code->code_lengths[index];
-  const int symbol = code->codes[index];
+                             const HuffmanTreeCode* const code,
+                             int code_index) {
+  const int depth = code->code_lengths[code_index];
+  const int symbol = code->codes[code_index];
   VP8LWriteBits(bw, depth, symbol);
 }
 
@@ -529,7 +535,12 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
                                 sizeof(*histogram_symbols));
   assert(histogram_bits >= MIN_HUFFMAN_BITS);
   assert(histogram_bits <= MAX_HUFFMAN_BITS);
-  if (histogram_image == NULL || histogram_symbols == NULL) goto Error;
+
+  if (histogram_image == NULL || histogram_symbols == NULL) {
+    free(histogram_image);
+    free(histogram_symbols);
+    return 0;
+  }
 
   // Calculate backward references from ARGB image.
   if (!VP8LGetBackwardReferences(width, height, argb, quality, cache_bits,
@@ -571,10 +582,10 @@ static int EncodeImageInternal(VP8LBitWriter* const bw,
       uint32_t i;
       if (histogram_argb == NULL) goto Error;
       for (i = 0; i < histogram_image_xysize; ++i) {
-        const int index = histogram_symbols[i] & 0xffff;
-        histogram_argb[i] = 0xff000000 | (index << 8);
-        if (index >= max_index) {
-          max_index = index + 1;
+        const int symbol_index = histogram_symbols[i] & 0xffff;
+        histogram_argb[i] = 0xff000000 | (symbol_index << 8);
+        if (symbol_index >= max_index) {
+          max_index = symbol_index + 1;
         }
       }
       histogram_image_size = max_index;
@@ -706,13 +717,6 @@ static int ApplyCrossColorFilter(const VP8LEncoder* const enc,
 
 // -----------------------------------------------------------------------------
 
-static void PutLE32(uint8_t* const data, uint32_t val) {
-  data[0] = (val >>  0) & 0xff;
-  data[1] = (val >>  8) & 0xff;
-  data[2] = (val >> 16) & 0xff;
-  data[3] = (val >> 24) & 0xff;
-}
-
 static WebPEncodingError WriteRiffHeader(const WebPPicture* const pic,
                                          size_t riff_size, size_t vp8l_size) {
   uint8_t riff[RIFF_HEADER_SIZE + CHUNK_HEADER_SIZE + VP8L_SIGNATURE_SIZE] = {
@@ -807,30 +811,24 @@ static WebPEncodingError AllocateTransformBuffer(VP8LEncoder* const enc,
   return err;
 }
 
-// Bundles multiple (2, 4 or 8) pixels into a single pixel.
-// Returns the new xsize.
-static void BundleColorMap(const WebPPicture* const pic,
-                           int xbits, uint32_t* bundled_argb, int xs) {
-  int y;
-  const int bit_depth = 1 << (3 - xbits);
-  uint32_t code = 0;
-  const uint32_t* argb = pic->argb;
-  const int width = pic->width;
-  const int height = pic->height;
-
-  for (y = 0; y < height; ++y) {
-    int x;
+// Bundles multiple (1, 2, 4 or 8) pixels into a single pixel.
+static void BundleColorMap(const uint8_t* const row, int width,
+                           int xbits, uint32_t* const dst) {
+  int x;
+  if (xbits > 0) {
+    const int bit_depth = 1 << (3 - xbits);
+    const int mask = (1 << xbits) - 1;
+    uint32_t code = 0xff000000;
     for (x = 0; x < width; ++x) {
-      const int mask = (1 << xbits) - 1;
       const int xsub = x & mask;
       if (xsub == 0) {
-        code = 0;
+        code = 0xff000000;
       }
-      // TODO(vikasa): simplify the bundling logic.
-      code |= (argb[x] & 0xff00) << (bit_depth * xsub);
-      bundled_argb[y * xs + (x >> xbits)] = 0xff000000 | code;
+      code |= row[x] << (8 + bit_depth * xsub);
+      dst[x >> xbits] = code;
     }
-    argb += pic->argb_stride;
+  } else {
+    for (x = 0; x < width; ++x) dst[x] = 0xff000000 | (row[x] << 8);
   }
 }
 
@@ -842,24 +840,43 @@ static WebPEncodingError ApplyPalette(VP8LBitWriter* const bw,
   WebPEncodingError err = VP8_ENC_OK;
   int i, x, y;
   const WebPPicture* const pic = enc->pic_;
-  uint32_t* argb = pic->argb;
+  uint32_t* src = pic->argb;
+  uint32_t* dst;
   const int width = pic->width;
   const int height = pic->height;
   uint32_t* const palette = enc->palette_;
   const int palette_size = enc->palette_size_;
+  uint8_t* row = NULL;
+  int xbits;
 
   // Replace each input pixel by corresponding palette index.
+  // This is done line by line.
+  if (palette_size <= 4) {
+    xbits = (palette_size <= 2) ? 3 : 2;
+  } else {
+    xbits = (palette_size <= 16) ? 1 : 0;
+  }
+
+  err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
+  if (err != VP8_ENC_OK) goto Error;
+  dst = enc->argb_;
+
+  row = WebPSafeMalloc((uint64_t)width, sizeof(*row));
+  if (row == NULL) return VP8_ENC_ERROR_OUT_OF_MEMORY;
+
   for (y = 0; y < height; ++y) {
     for (x = 0; x < width; ++x) {
-      const uint32_t pix = argb[x];
+      const uint32_t pix = src[x];
       for (i = 0; i < palette_size; ++i) {
         if (pix == palette[i]) {
-          argb[x] = 0xff000000u | (i << 8);
+          row[x] = i;
           break;
         }
       }
     }
-    argb += pic->argb_stride;
+    BundleColorMap(row, width, xbits, dst);
+    src += pic->argb_stride;
+    dst += enc->current_width_;
   }
 
   // Save palette to bitstream.
@@ -875,20 +892,8 @@ static WebPEncodingError ApplyPalette(VP8LBitWriter* const bw,
     goto Error;
   }
 
-  if (palette_size <= 16) {
-    // Image can be packed (multiple pixels per uint32_t).
-    int xbits = 1;
-    if (palette_size <= 2) {
-      xbits = 3;
-    } else if (palette_size <= 4) {
-      xbits = 2;
-    }
-    err = AllocateTransformBuffer(enc, VP8LSubSampleSize(width, xbits), height);
-    if (err != VP8_ENC_OK) goto Error;
-    BundleColorMap(pic, xbits, enc->argb_, enc->current_width_);
-  }
-
  Error:
+  free(row);
   return err;
 }
 
@@ -898,13 +903,13 @@ static int GetHistoBits(const WebPConfig* const config,
                         const WebPPicture* const pic) {
   const int width = pic->width;
   const int height = pic->height;
-  const size_t hist_size = sizeof(VP8LHistogram);
+  const uint64_t hist_size = sizeof(VP8LHistogram);
   // Make tile size a function of encoding method (Range: 0 to 6).
   int histo_bits = 7 - config->method;
   while (1) {
-    const size_t huff_image_size = VP8LSubSampleSize(width, histo_bits) *
-                                   VP8LSubSampleSize(height, histo_bits) *
-                                   hist_size;
+    const uint64_t huff_image_size = VP8LSubSampleSize(width, histo_bits) *
+                                     VP8LSubSampleSize(height, histo_bits) *
+                                     hist_size;
     if (huff_image_size <= MAX_HUFF_IMAGE_SIZE) break;
     ++histo_bits;
   }
diff --git a/src/enc/webpenc.c b/src/enc/webpenc.c
index 3c275589..5e13ddc1 100644
--- a/src/enc/webpenc.c
+++ b/src/enc/webpenc.c
@@ -93,34 +93,53 @@ static void ResetBoundaryPredictions(VP8Encoder* const enc) {
   enc->nz_[-1] = 0;   // constant
 }
 
-// Map configured quality level to coding tools used.
-//-------------+---+---+---+---+---+---+
-//   Quality   | 0 | 1 | 2 | 3 | 4 | 5 +
-//-------------+---+---+---+---+---+---+
-// dynamic prob| ~ | x | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// rd-opt modes|   |   | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// fast i4/i16 | x | x |   |   |   |   |
-//-------------+---+---+---+---+---+---+
-// rd-opt i4/16|   |   | x | x | x | x |
-//-------------+---+---+---+---+---+---+
-// Trellis     |   | x |   |   | x | x |
-//-------------+---+---+---+---+---+---+
-// full-SNS    |   |   |   |   |   | x |
-//-------------+---+---+---+---+---+---+
+// Mapping from config->method_ to coding tools used.
+//-------------------+---+---+---+---+---+---+---+
+//   Method          | 0 | 1 | 2 | 3 |(4)| 5 | 6 |
+//-------------------+---+---+---+---+---+---+---+
+// fast probe        | x |   |   | x |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// dynamic proba     | ~ | x | x | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// fast mode analysis|   |   |   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// basic rd-opt      |   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// disto-score i4/16 |   |   | x |   |   |   |   |
+//-------------------+---+---+---+---+---+---+---+
+// rd-opt i4/16      |   |   | ~ | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// token buffer (opt)|   |   |   | x | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
+// Trellis           |   |   |   |   |   | x |Ful|
+//-------------------+---+---+---+---+---+---+---+
+// full-SNS          |   |   |   |   | x | x | x |
+//-------------------+---+---+---+---+---+---+---+
 
 static void MapConfigToTools(VP8Encoder* const enc) {
-  const int method = enc->config_->method;
-  const int limit = 100 - enc->config_->partition_limit;
+  const WebPConfig* const config = enc->config_;
+  const int method = config->method;
+  const int limit = 100 - config->partition_limit;
   enc->method_ = method;
-  enc->rd_opt_level_ = (method >= 6) ? 3
-                     : (method >= 5) ? 2
-                     : (method >= 3) ? 1
-                     : 0;
+  enc->rd_opt_level_ = (method >= 6) ? RD_OPT_TRELLIS_ALL
+                     : (method >= 5) ? RD_OPT_TRELLIS
+                     : (method >= 3) ? RD_OPT_BASIC
+                     : RD_OPT_NONE;
   enc->max_i4_header_bits_ =
       256 * 16 * 16 *                 // upper bound: up to 16bit per 4x4 block
       (limit * limit) / (100 * 100);  // ... modulated with a quadratic curve.
+
+  enc->thread_level_ = config->thread_level;
+
+  enc->do_search_ = (config->target_size > 0 || config->target_PSNR > 0);
+  if (!config->low_memory) {
+#if !defined(DISABLE_TOKEN_BUFFER)
+    enc->use_tokens_ = (method >= 3) && !enc->do_search_;
+#endif
+    if (enc->use_tokens_) {
+      enc->num_parts_ = 1;   // doesn't work with multi-partition
+    }
+  }
 }
 
 // Memory scaling with dimensions:
@@ -259,17 +278,21 @@ static VP8Encoder* InitVP8Encoder(const WebPConfig* const config,
   VP8EncInitLayer(enc);
 #endif
 
+  VP8TBufferInit(&enc->tokens_);
   return enc;
 }
 
-static void DeleteVP8Encoder(VP8Encoder* enc) {
+static int DeleteVP8Encoder(VP8Encoder* enc) {
+  int ok = 1;
   if (enc != NULL) {
-    VP8EncDeleteAlpha(enc);
+    ok = VP8EncDeleteAlpha(enc);
 #ifdef WEBP_EXPERIMENTAL_FEATURES
     VP8EncDeleteLayer(enc);
 #endif
+    VP8TBufferClear(&enc->tokens_);
     free(enc);
   }
+  return ok;
 }
 
 //------------------------------------------------------------------------------
@@ -332,7 +355,7 @@ int WebPReportProgress(const WebPPicture* const pic,
 //------------------------------------------------------------------------------
 
 int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
-  int ok;
+  int ok = 0;
 
   if (pic == NULL)
     return 0;
@@ -361,19 +384,26 @@ int WebPEncode(const WebPConfig* config, WebPPicture* pic) {
     enc = InitVP8Encoder(config, pic);
     if (enc == NULL) return 0;  // pic->error is already set.
     // Note: each of the tasks below account for 20% in the progress report.
-    ok = VP8EncAnalyze(enc)
-      && VP8StatLoop(enc)
-      && VP8EncLoop(enc)
-      && VP8EncFinishAlpha(enc)
+    ok = VP8EncAnalyze(enc);
+
+    // Analysis is done, proceed to actual coding.
+    ok = ok && VP8EncStartAlpha(enc);   // possibly done in parallel
+    if (!enc->use_tokens_) {
+      ok = VP8EncLoop(enc);
+    } else {
+      ok = VP8EncTokenLoop(enc);
+    }
+    ok = ok && VP8EncFinishAlpha(enc);
 #ifdef WEBP_EXPERIMENTAL_FEATURES
-      && VP8EncFinishLayer(enc)
+    ok = ok && VP8EncFinishLayer(enc);
 #endif
-      && VP8EncWrite(enc);
+
+    ok = ok && VP8EncWrite(enc);
     StoreStats(enc);
     if (!ok) {
       VP8EncFreeBitWriters(enc);
     }
-    DeleteVP8Encoder(enc);
+    ok &= DeleteVP8Encoder(enc);  // must always be called, even if !ok
   } else {
     if (pic->argb == NULL)
       return WebPEncodingSetError(pic, VP8_ENC_ERROR_NULL_PARAMETER);
diff --git a/src/utils/bit_reader.c b/src/utils/bit_reader.c
index 1afb1db8..d6cfd864 100644
--- a/src/utils/bit_reader.c
+++ b/src/utils/bit_reader.c
@@ -15,7 +15,11 @@
 extern "C" {
 #endif
 
-#define MK(X) (((bit_t)(X) << (BITS)) | (MASK))
+#ifndef USE_RIGHT_JUSTIFY
+#define MK(X) (((range_t)(X) << (BITS)) | (MASK))
+#else
+#define MK(X) ((range_t)(X))
+#endif
 
 //------------------------------------------------------------------------------
 // VP8BitReader
@@ -29,7 +33,7 @@ void VP8InitBitReader(VP8BitReader* const br,
   br->buf_     = start;
   br->buf_end_ = end;
   br->value_   = 0;
-  br->missing_ = 8;   // to load the very first 8bits
+  br->bits_    = -8;   // to load the very first 8bits
   br->eof_     = 0;
 }
 
@@ -46,7 +50,7 @@ const uint8_t kVP8Log2Range[128] = {
 };
 
 // range = (range << kVP8Log2Range[range]) + trailing 1's
-const bit_t kVP8NewRange[128] = {
+const range_t kVP8NewRange[128] = {
   MK(127), MK(127), MK(191), MK(127), MK(159), MK(191), MK(223), MK(127),
   MK(143), MK(159), MK(175), MK(191), MK(207), MK(223), MK(239), MK(127),
   MK(135), MK(143), MK(151), MK(159), MK(167), MK(175), MK(183), MK(191),
@@ -71,9 +75,19 @@ void VP8LoadFinalBytes(VP8BitReader* const br) {
   assert(br != NULL && br->buf_ != NULL);
   // Only read 8bits at a time
   if (br->buf_ < br->buf_end_) {
-    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 + br->missing_);
-    br->missing_ -= 8;
-  } else {
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= (bit_t)(*br->buf_++) << ((BITS) - 8 - br->bits_);
+#else
+    br->value_ = (bit_t)(*br->buf_++) | (br->value_ << 8);
+#endif
+    br->bits_ += 8;
+  } else if (!br->eof_) {
+#ifdef USE_RIGHT_JUSTIFY
+    // These are not strictly needed, but it makes the behaviour
+    // consistent for both USE_RIGHT_JUSTIFY and !USE_RIGHT_JUSTIFY.
+    br->value_ <<= 8;
+    br->bits_ += 8;
+#endif
     br->eof_ = 1;
   }
 }
@@ -99,6 +113,10 @@ int32_t VP8GetSignedValue(VP8BitReader* const br, int bits) {
 
 #define MAX_NUM_BIT_READ 25
 
+#define LBITS 64      // Number of bits prefetched.
+#define WBITS 32      // Minimum number of bytes needed after VP8LFillBitWindow.
+#define LOG8_WBITS 4  // Number of bytes needed to store WBITS bits.
+
 static const uint32_t kBitMask[MAX_NUM_BIT_READ] = {
   0, 1, 3, 7, 15, 31, 63, 127, 255, 511, 1023, 2047, 4095, 8191, 16383, 32767,
   65535, 131071, 262143, 524287, 1048575, 2097151, 4194303, 8388607, 16777215
@@ -120,7 +138,7 @@ void VP8LInitBitReader(VP8LBitReader* const br,
   br->eos_ = 0;
   br->error_ = 0;
   for (i = 0; i < sizeof(br->val_) && i < br->len_; ++i) {
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << (8 * i);
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (8 * i);
     ++br->pos_;
   }
 }
@@ -135,91 +153,56 @@ void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
   br->len_ = len;
 }
 
+// If not at EOS, reload up to LBITS byte-by-byte
 static void ShiftBytes(VP8LBitReader* const br) {
   while (br->bit_pos_ >= 8 && br->pos_ < br->len_) {
     br->val_ >>= 8;
-    br->val_ |= ((uint64_t)br->buf_[br->pos_]) << 56;
+    br->val_ |= ((vp8l_val_t)br->buf_[br->pos_]) << (LBITS - 8);
     ++br->pos_;
     br->bit_pos_ -= 8;
   }
 }
 
 void VP8LFillBitWindow(VP8LBitReader* const br) {
-  if (br->bit_pos_ >= 32) {
-#if defined(__x86_64__) || defined(_M_X64)
-    if (br->pos_ + 8 < br->len_) {
-      br->val_ >>= 32;
+  if (br->bit_pos_ >= WBITS) {
+#if (defined(__x86_64__) || defined(_M_X64))
+    if (br->pos_ + sizeof(br->val_) < br->len_) {
+      br->val_ >>= WBITS;
+      br->bit_pos_ -= WBITS;
       // The expression below needs a little-endian arch to work correctly.
       // This gives a large speedup for decoding speed.
-      br->val_ |= *(const uint64_t *)(br->buf_ + br->pos_) << 32;
-      br->pos_ += 4;
-      br->bit_pos_ -= 32;
-    } else {
-      // Slow path.
-      ShiftBytes(br);
+      br->val_ |= *(const vp8l_val_t*)(br->buf_ + br->pos_) << (LBITS - WBITS);
+      br->pos_ += LOG8_WBITS;
+      return;
     }
-#else
-    // Always the slow path.
-    ShiftBytes(br);
 #endif
-  }
-  if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
-    br->eos_ = 1;
-  }
-}
-
-uint32_t VP8LReadOneBit(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  // Flag an error at end_of_stream.
-  if (!br->eos_) {
-    ++br->bit_pos_;
-    if (br->bit_pos_ >= 32) {
-      ShiftBytes(br);
-    }
-    // After this last bit is read, check if eos needs to be flagged.
-    if (br->pos_ == br->len_ && br->bit_pos_ == 64) {
+    ShiftBytes(br);       // Slow path.
+    if (br->pos_ == br->len_ && br->bit_pos_ == LBITS) {
       br->eos_ = 1;
     }
-  } else {
-    br->error_ = 1;
   }
-  return val;
 }
 
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits) {
-  uint32_t val = 0;
   assert(n_bits >= 0);
   // Flag an error if end_of_stream or n_bits is more than allowed limit.
   if (!br->eos_ && n_bits < MAX_NUM_BIT_READ) {
+    const uint32_t val =
+        (uint32_t)(br->val_ >> br->bit_pos_) & kBitMask[n_bits];
+    const int new_bits = br->bit_pos_ + n_bits;
+    br->bit_pos_ = new_bits;
     // If this read is going to cross the read buffer, set the eos flag.
     if (br->pos_ == br->len_) {
-      if ((br->bit_pos_ + n_bits) >= 64) {
+      if (new_bits >= LBITS) {
         br->eos_ = 1;
-        if ((br->bit_pos_ + n_bits) > 64) return val;
-      }
-    }
-    val = (br->val_ >> br->bit_pos_) & kBitMask[n_bits];
-    br->bit_pos_ += n_bits;
-    if (br->bit_pos_ >= 40) {
-      if (br->pos_ + 5 < br->len_) {
-        br->val_ >>= 40;
-        br->val_ |=
-            (((uint64_t)br->buf_[br->pos_ + 0]) << 24) |
-            (((uint64_t)br->buf_[br->pos_ + 1]) << 32) |
-            (((uint64_t)br->buf_[br->pos_ + 2]) << 40) |
-            (((uint64_t)br->buf_[br->pos_ + 3]) << 48) |
-            (((uint64_t)br->buf_[br->pos_ + 4]) << 56);
-        br->pos_ += 5;
-        br->bit_pos_ -= 40;
-      }
-      if (br->bit_pos_ >= 8) {
-        ShiftBytes(br);
       }
     }
+    ShiftBytes(br);
+    return val;
   } else {
     br->error_ = 1;
+    return 0;
   }
-  return val;
 }
 
 //------------------------------------------------------------------------------
diff --git a/src/utils/bit_reader.h b/src/utils/bit_reader.h
index 11a40a55..d98332e1 100644
--- a/src/utils/bit_reader.h
+++ b/src/utils/bit_reader.h
@@ -24,11 +24,72 @@
 extern "C" {
 #endif
 
-#define BITS 32     // can be 32, 16 or 8
-#define MASK ((((bit_t)1) << (BITS)) - 1)
+// The Boolean decoder needs to maintain infinite precision on the value_ field.
+// However, since range_ is only 8bit, we only need an active window of 8 bits
+// for value_. Left bits (MSB) gets zeroed and shifted away when value_ falls
+// below 128, range_ is updated, and fresh bits read from the bitstream are
+// brought in as LSB.
+// To avoid reading the fresh bits one by one (slow), we cache a few of them
+// ahead (actually, we cache BITS of them ahead. See below). There's two
+// strategies regarding how to shift these looked-ahead fresh bits into the
+// 8bit window of value_: either we shift them in, while keeping the position of
+// the window fixed. Or we slide the window to the right while keeping the cache
+// bits at a fixed, right-justified, position.
+//
+//  Example, for BITS=16: here is the content of value_ for both strategies:
+//
+//          !USE_RIGHT_JUSTIFY            ||        USE_RIGHT_JUSTIFY
+//                                        ||
+//   <- 8b -><- 8b -><- BITS bits  ->     ||  <- 8b+3b -><- 8b -><- 13 bits ->
+//   [unused][value_][cached bits][0]     ||  [unused...][value_][cached bits]
+//  [........00vvvvvvBBBBBBBBBBBBB000]LSB || [...........00vvvvvvBBBBBBBBBBBBB]
+//                                        ||
+// After calling VP8Shift(), where we need to shift away two zeros:
+//  [........vvvvvvvvBBBBBBBBBBB00000]LSB || [.............vvvvvvvvBBBBBBBBBBB]
+//                                        ||
+// Just before we need to call VP8LoadNewBytes(), the situation is:
+//  [........vvvvvv000000000000000000]LSB || [..........................vvvvvv]
+//                                        ||
+// And just after calling VP8LoadNewBytes():
+//  [........vvvvvvvvBBBBBBBBBBBBBBBB]LSB || [........vvvvvvvvBBBBBBBBBBBBBBBB]
+//
+// -> we're back to height active 'value_' bits (marked 'v') and BITS cached
+// bits (marked 'B')
+//
+// The right-justify strategy tends to use less shifts and is often faster.
+
+//------------------------------------------------------------------------------
+// BITS can be either 32, 24, 16 or 8.
+// Pick values that fit natural register size.
+
+#if !defined(WEBP_REFERENCE_IMPLEMENTATION)
+
+#define USE_RIGHT_JUSTIFY
+
+#if defined(__i386__) || defined(_M_IX86)      // x86 32bit
+#define BITS 16
+#elif defined(__arm__) || defined(_M_ARM)     // ARM
+#define BITS 24
+#else                      // reasonable default
+#define BITS 24
+#endif
+
+#else     // reference choices
+
+#define USE_RIGHT_JUSTIFY
+#define BITS 8
+
+#endif
+
+//------------------------------------------------------------------------------
+// Derived types and constants
+
 #if (BITS == 32)
 typedef uint64_t bit_t;   // natural register type
 typedef uint32_t lbit_t;  // natural type for memory I/O
+#elif (BITS == 24)
+typedef uint32_t bit_t;
+typedef uint32_t lbit_t;
 #elif (BITS == 16)
 typedef uint32_t bit_t;
 typedef uint16_t lbit_t;
@@ -37,8 +98,15 @@ typedef uint32_t bit_t;
 typedef uint8_t lbit_t;
 #endif
 
+#ifndef USE_RIGHT_JUSTIFY
+typedef bit_t range_t;     // type for storing range_
+#define MASK ((((bit_t)1) << (BITS)) - 1)
+#else
+typedef uint32_t range_t;  // range_ only uses 8bits here. No need for bit_t.
+#endif
+
 //------------------------------------------------------------------------------
-// Bitreader and code-tree reader
+// Bitreader
 
 typedef struct VP8BitReader VP8BitReader;
 struct VP8BitReader {
@@ -47,9 +115,9 @@ struct VP8BitReader {
   int eof_;                   // true if input is exhausted
 
   // boolean decoder
-  bit_t range_;            // current range minus 1. In [127, 254] interval.
-  bit_t value_;            // current value
-  int missing_;            // number of missing bits in value_ (8bit)
+  range_t range_;            // current range minus 1. In [127, 254] interval.
+  bit_t value_;              // current value
+  int bits_;                 // number of valid bits left
 };
 
 // Initialize the bit reader and the boolean decoder.
@@ -67,12 +135,12 @@ int32_t VP8GetSignedValue(VP8BitReader* const br, int num_bits);
 
 // Read a bit with proba 'prob'. Speed-critical function!
 extern const uint8_t kVP8Log2Range[128];
-extern const bit_t kVP8NewRange[128];
+extern const range_t kVP8NewRange[128];
 
 void VP8LoadFinalBytes(VP8BitReader* const br);    // special case for the tail
 
 static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
-  assert(br && br->buf_);
+  assert(br != NULL && br->buf_ != NULL);
   // Read 'BITS' bits at a time if possible.
   if (br->buf_ + sizeof(lbit_t) <= br->buf_end_) {
     // convert memory type to register type (with some zero'ing!)
@@ -80,64 +148,103 @@ static WEBP_INLINE void VP8LoadNewBytes(VP8BitReader* const br) {
     lbit_t in_bits = *(lbit_t*)br->buf_;
     br->buf_ += (BITS) >> 3;
 #if !defined(__BIG_ENDIAN__)
-#if (BITS == 32)
+#if (BITS == 32) || (BITS == 24)
 #if defined(__i386__) || defined(__x86_64__)
     __asm__ volatile("bswap %k0" : "=r"(in_bits) : "0"(in_bits));
-    bits = (bit_t)in_bits;   // 32b -> 64b zero-extension
+    bits = (bit_t)in_bits;   // 24b/32b -> 32b/64b zero-extension
 #elif defined(_MSC_VER)
     bits = _byteswap_ulong(in_bits);
 #else
     bits = (bit_t)(in_bits >> 24) | ((in_bits >> 8) & 0xff00)
          | ((in_bits << 8) & 0xff0000)  | (in_bits << 24);
 #endif  // x86
+#if (BITS == 24)
+    bits >>= 8;
+#endif
 #elif (BITS == 16)
     // gcc will recognize a 'rorw $8, ...' here:
     bits = (bit_t)(in_bits >> 8) | ((in_bits & 0xff) << 8);
+#else   // BITS == 8
+    bits = (bit_t)in_bits;
 #endif
-#else    // LITTLE_ENDIAN
+#else    // BIG_ENDIAN
     bits = (bit_t)in_bits;
 #endif
-    br->value_ |= bits << br->missing_;
-    br->missing_ -= (BITS);
+#ifndef USE_RIGHT_JUSTIFY
+    br->value_ |= bits << (-br->bits_);
+#else
+    br->value_ = bits | (br->value_ << (BITS));
+#endif
+    br->bits_ += (BITS);
   } else {
     VP8LoadFinalBytes(br);    // no need to be inlined
   }
 }
 
-static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, bit_t split) {
-  const bit_t value_split = split | (MASK);
-  if (br->missing_ > 0) {  // Make sure we have a least BITS bits in 'value_'
+static WEBP_INLINE int VP8BitUpdate(VP8BitReader* const br, range_t split) {
+  if (br->bits_ < 0) {  // Make sure we have a least BITS bits in 'value_'
     VP8LoadNewBytes(br);
   }
-  if (br->value_ > value_split) {
-    br->range_ -= value_split + 1;
-    br->value_ -= value_split + 1;
+#ifndef USE_RIGHT_JUSTIFY
+  split |= (MASK);
+  if (br->value_ > split) {
+    br->range_ -= split + 1;
+    br->value_ -= split + 1;
     return 1;
   } else {
-    br->range_ = value_split;
+    br->range_ = split;
     return 0;
   }
+#else
+  {
+    const int pos = br->bits_;
+    const range_t value = (range_t)(br->value_ >> pos);
+    if (value > split) {
+      br->range_ -= split + 1;
+      br->value_ -= (bit_t)(split + 1) << pos;
+      return 1;
+    } else {
+      br->range_ = split;
+      return 0;
+    }
+  }
+#endif
 }
 
 static WEBP_INLINE void VP8Shift(VP8BitReader* const br) {
+#ifndef USE_RIGHT_JUSTIFY
   // range_ is in [0..127] interval here.
-  const int idx = br->range_ >> (BITS);
+  const bit_t idx = br->range_ >> (BITS);
   const int shift = kVP8Log2Range[idx];
   br->range_ = kVP8NewRange[idx];
   br->value_ <<= shift;
-  br->missing_ += shift;
+  br->bits_ -= shift;
+#else
+  const int shift = kVP8Log2Range[br->range_];
+  assert(br->range_ < (range_t)128);
+  br->range_ = kVP8NewRange[br->range_];
+  br->bits_ -= shift;
+#endif
 }
-
 static WEBP_INLINE int VP8GetBit(VP8BitReader* const br, int prob) {
+#ifndef USE_RIGHT_JUSTIFY
   // It's important to avoid generating a 64bit x 64bit multiply here.
   // We just need an 8b x 8b after all.
-  const bit_t split =
-      (bit_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
+  const range_t split =
+      (range_t)((uint32_t)(br->range_ >> (BITS)) * prob) << ((BITS) - 8);
+  const int bit = VP8BitUpdate(br, split);
+  if (br->range_ <= (((range_t)0x7e << (BITS)) | (MASK))) {
+    VP8Shift(br);
+  }
+  return bit;
+#else
+  const range_t split = (br->range_ * prob) >> 8;
   const int bit = VP8BitUpdate(br, split);
-  if (br->range_ <= (((bit_t)0x7e << (BITS)) | (MASK))) {
+  if (br->range_ <= (range_t)0x7e) {
     VP8Shift(br);
   }
   return bit;
+#endif
 }
 
 static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
@@ -149,16 +256,18 @@ static WEBP_INLINE int VP8GetSigned(VP8BitReader* const br, int v) {
 
 
 // -----------------------------------------------------------------------------
-// Bitreader
+// Bitreader for lossless format
+
+typedef uint64_t vp8l_val_t;  // right now, this bit-reader can only use 64bit.
 
 typedef struct {
-  uint64_t       val_;
-  const uint8_t* buf_;
-  size_t         len_;
-  size_t         pos_;
-  int            bit_pos_;
-  int            eos_;
-  int            error_;
+  vp8l_val_t     val_;        // pre-fetched bits
+  const uint8_t* buf_;        // input byte buffer
+  size_t         len_;        // buffer length
+  size_t         pos_;        // byte position in buf_
+  int            bit_pos_;    // current bit-reading position in val_
+  int            eos_;        // bitstream is finished
+  int            error_;      // an error occurred (buffer overflow attempt...)
 } VP8LBitReader;
 
 void VP8LInitBitReader(VP8LBitReader* const br,
@@ -174,17 +283,14 @@ void VP8LBitReaderSetBuffer(VP8LBitReader* const br,
 // Flags eos if this read attempt is going to cross the read buffer.
 uint32_t VP8LReadBits(VP8LBitReader* const br, int n_bits);
 
-// Reads one bit from Read Buffer. Flags an error in case end_of_stream.
-// Flags eos after reading last bit from the buffer.
-uint32_t VP8LReadOneBit(VP8LBitReader* const br);
-
-// VP8LReadOneBitUnsafe is faster than VP8LReadOneBit, but it can be called only
-// 32 times after the last VP8LFillBitWindow. Any subsequent calls
-// (without VP8LFillBitWindow) will return invalid data.
-static WEBP_INLINE uint32_t VP8LReadOneBitUnsafe(VP8LBitReader* const br) {
-  const uint32_t val = (br->val_ >> br->bit_pos_) & 1;
-  ++br->bit_pos_;
-  return val;
+// Return the prefetched bits, so they can be looked up.
+static WEBP_INLINE uint32_t VP8LPrefetchBits(VP8LBitReader* const br) {
+  return (uint32_t)(br->val_ >> br->bit_pos_);
+}
+
+// Discard 'num_bits' bits from the cache.
+static WEBP_INLINE void VP8LDiscardBits(VP8LBitReader* const br, int num_bits) {
+  br->bit_pos_ += num_bits;
 }
 
 // Advances the Read buffer by 4 bytes to make room for reading next 32 bits.
diff --git a/src/utils/huffman_encode.c b/src/utils/huffman_encode.c
index 2686c665..49187592 100644
--- a/src/utils/huffman_encode.c
+++ b/src/utils/huffman_encode.c
@@ -138,13 +138,8 @@ static int CompareHuffmanTrees(const void* ptr1, const void* ptr2) {
   } else if (t1->total_count_ < t2->total_count_) {
     return 1;
   } else {
-    if (t1->value_ < t2->value_) {
-      return -1;
-    }
-    if (t1->value_ > t2->value_) {
-      return 1;
-    }
-    return 0;
+    assert(t1->value_ != t2->value_);
+    return (t1->value_ < t2->value_) ? -1 : 1;
   }
 }
 
@@ -193,6 +188,10 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
     }
   }
 
+  if (tree_size_orig == 0) {   // pretty optimal already!
+    return 1;
+  }
+
   // 3 * tree_size is enough to cover all the nodes representing a
   // population and all the inserted nodes combining two existing nodes.
   // The tree pool needs 2 * (tree_size_orig - 1) entities, and the
@@ -234,7 +233,7 @@ static int GenerateOptimalTree(const int* const histogram, int histogram_size,
         tree_pool[tree_pool_size++] = tree[tree_size - 1];
         tree_pool[tree_pool_size++] = tree[tree_size - 2];
         count = tree_pool[tree_pool_size - 1].total_count_ +
-            tree_pool[tree_pool_size - 2].total_count_;
+                tree_pool[tree_pool_size - 2].total_count_;
         tree_size -= 2;
         {
           // Search for the insertion point.
diff --git a/src/utils/quant_levels.c b/src/utils/quant_levels.c
index f6884392..649aae65 100644
--- a/src/utils/quant_levels.c
+++ b/src/utils/quant_levels.c
@@ -140,15 +140,6 @@ int QuantizeLevels(uint8_t* const data, int width, int height,
   return 1;
 }
 
-int DequantizeLevels(uint8_t* const data, int width, int height) {
-  if (data == NULL || width <= 0 || height <= 0) return 0;
-  // TODO(skal): implement gradient smoothing.
-  (void)data;
-  (void)width;
-  (void)height;
-  return 1;
-}
-
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/utils/quant_levels.h b/src/utils/quant_levels.h
index 8dd3afee..f86cbe59 100644
--- a/src/utils/quant_levels.h
+++ b/src/utils/quant_levels.h
@@ -27,11 +27,6 @@ extern "C" {
 int QuantizeLevels(uint8_t* const data, int width, int height, int num_levels,
                    uint64_t* const sse);
 
-// Apply post-processing to input 'data' of size 'width'x'height' assuming
-// that the source was quantized to a reduced number of levels.
-// Returns false in case of error (data is NULL, invalid parameters, ...).
-int DequantizeLevels(uint8_t* const data, int width, int height);
-
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif
diff --git a/src/utils/quant_levels_dec.c b/src/utils/quant_levels_dec.c
new file mode 100644
index 00000000..95142b1b
--- /dev/null
+++ b/src/utils/quant_levels_dec.c
@@ -0,0 +1,28 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// TODO(skal): implement gradient smoothing.
+//
+// Author: Skal (pascal.massimino@gmail.com)
+
+#include "./quant_levels_dec.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+int DequantizeLevels(uint8_t* const data, int width, int height) {
+  if (data == NULL || width <= 0 || height <= 0) return 0;
+  (void)data;
+  (void)width;
+  (void)height;
+  return 1;
+}
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
diff --git a/src/utils/quant_levels_dec.h b/src/utils/quant_levels_dec.h
new file mode 100644
index 00000000..de60ef95
--- /dev/null
+++ b/src/utils/quant_levels_dec.h
@@ -0,0 +1,30 @@
+// Copyright 2013 Google Inc. All Rights Reserved.
+//
+// This code is licensed under the same terms as WebM:
+//  Software License Agreement:  http://www.webmproject.org/license/software/
+//  Additional IP Rights Grant:  http://www.webmproject.org/license/additional/
+// -----------------------------------------------------------------------------
+//
+// Alpha plane de-quantization utility
+//
+// Author:  Vikas Arora (vikasa@google.com)
+
+#ifndef WEBP_UTILS_QUANT_LEVELS_DEC_H_
+#define WEBP_UTILS_QUANT_LEVELS_DEC_H_
+
+#include "webp/types.h"
+
+#if defined(__cplusplus) || defined(c_plusplus)
+extern "C" {
+#endif
+
+// Apply post-processing to input 'data' of size 'width'x'height' assuming
+// that the source was quantized to a reduced number of levels.
+// Returns false in case of error (data is NULL, invalid parameters, ...).
+int DequantizeLevels(uint8_t* const data, int width, int height);
+
+#if defined(__cplusplus) || defined(c_plusplus)
+}    // extern "C"
+#endif
+
+#endif  /* WEBP_UTILS_QUANT_LEVELS_DEC_H_ */
diff --git a/src/utils/thread.c b/src/utils/thread.c
index ce89cf9d..a14af559 100644
--- a/src/utils/thread.c
+++ b/src/utils/thread.c
@@ -9,10 +9,6 @@
 //
 // Author: Skal (pascal.massimino@gmail.com)
 
-#ifdef HAVE_CONFIG_H
-#include "config.h"
-#endif
-
 #include <assert.h>
 #include <string.h>   // for memset()
 #include "./thread.h"
diff --git a/src/utils/thread.h b/src/utils/thread.h
index 3191890b..9afe0967 100644
--- a/src/utils/thread.h
+++ b/src/utils/thread.h
@@ -12,6 +12,10 @@
 #ifndef WEBP_UTILS_THREAD_H_
 #define WEBP_UTILS_THREAD_H_
 
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
 #endif
@@ -63,13 +67,13 @@ typedef struct {
 
 // Must be called first, before any other method.
 void WebPWorkerInit(WebPWorker* const worker);
-// Must be called initialize the object and spawn the thread. Re-entrant.
+// Must be called to initialize the object and spawn the thread. Re-entrant.
 // Will potentially launch the thread. Returns false in case of error.
 int WebPWorkerReset(WebPWorker* const worker);
-// Make sure the previous work is finished. Returns true if worker->had_error
-// was not set and not error condition was triggered by the working thread.
+// Makes sure the previous work is finished. Returns true if worker->had_error
+// was not set and no error condition was triggered by the working thread.
 int WebPWorkerSync(WebPWorker* const worker);
-// Trigger the thread to call hook() with data1 and data2 argument. These
+// Triggers the thread to call hook() with data1 and data2 argument. These
 // hook/data1/data2 can be changed at any time before calling this function,
 // but not be changed afterward until the next call to WebPWorkerSync().
 void WebPWorkerLaunch(WebPWorker* const worker);
diff --git a/src/utils/utils.c b/src/utils/utils.c
index 673b7e28..b1db2f9d 100644
--- a/src/utils/utils.c
+++ b/src/utils/utils.c
@@ -19,7 +19,8 @@ extern "C" {
 //------------------------------------------------------------------------------
 // Checked memory allocation
 
-static int CheckSizeArguments(uint64_t nmemb, size_t size) {
+// Returns 0 in case of overflow of nmemb * size.
+static int CheckSizeArgumentsOverflow(uint64_t nmemb, size_t size) {
   const uint64_t total_size = nmemb * size;
   if (nmemb == 0) return 1;
   if ((uint64_t)size > WEBP_MAX_ALLOCABLE_MEMORY / nmemb) return 0;
@@ -28,12 +29,14 @@ static int CheckSizeArguments(uint64_t nmemb, size_t size) {
 }
 
 void* WebPSafeMalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
   return malloc((size_t)(nmemb * size));
 }
 
 void* WebPSafeCalloc(uint64_t nmemb, size_t size) {
-  if (!CheckSizeArguments(nmemb, size)) return NULL;
+  if (!CheckSizeArgumentsOverflow(nmemb, size)) return NULL;
+  assert(nmemb * size > 0);
   return calloc((size_t)nmemb, size);
 }
 
diff --git a/src/utils/utils.h b/src/utils/utils.h
index aa445695..32dfb8a9 100644
--- a/src/utils/utils.h
+++ b/src/utils/utils.h
@@ -7,11 +7,14 @@
 //
 // Misc. common utility functions
 //
-// Author: Skal (pascal.massimino@gmail.com)
+// Authors: Skal (pascal.massimino@gmail.com)
+//          Urvang (urvang@google.com)
 
 #ifndef WEBP_UTILS_UTILS_H_
 #define WEBP_UTILS_UTILS_H_
 
+#include <assert.h>
+
 #include "webp/types.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
@@ -36,6 +39,40 @@ void* WebPSafeMalloc(uint64_t nmemb, size_t size);
 void* WebPSafeCalloc(uint64_t nmemb, size_t size);
 
 //------------------------------------------------------------------------------
+// Reading/writing data.
+
+// Read 16, 24 or 32 bits stored in little-endian order.
+static WEBP_INLINE int GetLE16(const uint8_t* const data) {
+  return (int)(data[0] << 0) | (data[1] << 8);
+}
+
+static WEBP_INLINE int GetLE24(const uint8_t* const data) {
+  return GetLE16(data) | (data[2] << 16);
+}
+
+static WEBP_INLINE uint32_t GetLE32(const uint8_t* const data) {
+  return (uint32_t)GetLE16(data) | (GetLE16(data + 2) << 16);
+}
+
+// Store 16, 24 or 32 bits in little-endian order.
+static WEBP_INLINE void PutLE16(uint8_t* const data, int val) {
+  assert(val < (1 << 16));
+  data[0] = (val >> 0);
+  data[1] = (val >> 8);
+}
+
+static WEBP_INLINE void PutLE24(uint8_t* const data, int val) {
+  assert(val < (1 << 24));
+  PutLE16(data, val & 0xffff);
+  data[2] = (val >> 16);
+}
+
+static WEBP_INLINE void PutLE32(uint8_t* const data, uint32_t val) {
+  PutLE16(data, (int)(val & 0xffff));
+  PutLE16(data + 2, (int)(val >> 16));
+}
+
+//------------------------------------------------------------------------------
 
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"