Add 10 bit rotate methods.

This initial implementation is based on current unoptimized code in webrtc using just plain for loops.

Bug: libyuv:949
Change-Id: Ic87ee49c3a0b62edbaaa4255c263c1f7be4ea02b
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4110782
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>

10 files changed, 911 insertions, 13 deletions

diff --git a/include/libyuv/rotate.h b/include/libyuv/rotate.h
index 684ed5e6..37460c4a 100644
--- a/include/libyuv/rotate.h
+++ b/include/libyuv/rotate.h
@@ -85,6 +85,60 @@ int I444Rotate(const uint8_t* src_y,
                int height,
                enum RotationMode mode);
 
+// Rotate I010 frame.
+LIBYUV_API
+int I010Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I210 frame.
+LIBYUV_API
+int I210Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
+// Rotate I410 frame.
+LIBYUV_API
+int I410Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode);
+
 // Rotate NV12 input and store in I420.
 LIBYUV_API
 int NV12ToI420Rotate(const uint8_t* src_y,
@@ -156,6 +210,16 @@ void RotatePlane270(const uint8_t* src,
                     int width,
                     int height);
 
+// Rotate a plane by 0, 90, 180, or 270.
+LIBYUV_API
+int RotatePlane_16(const uint16_t* src,
+                   int src_stride,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height,
+                   enum RotationMode mode);
+
 // Rotations for when U and V are interleaved.
 // These functions take one UV input pointer and
 // split the data into two buffers while
diff --git a/include/libyuv/rotate_row.h b/include/libyuv/rotate_row.h
index aa8528a9..b773f886 100644
--- a/include/libyuv/rotate_row.h
+++ b/include/libyuv/rotate_row.h
@@ -215,7 +215,23 @@ void TransposeUVWx16_Any_LSX(const uint8_t* src,
                              uint8_t* dst_b,
                              int dst_stride_b,
                              int width);
+void TransposeWxH_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height);
 
+void TransposeWx8_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width);
+void TransposeWx1_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width);
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/include/libyuv/row.h b/include/libyuv/row.h
index 45b172b2..ff93406b 100644
--- a/include/libyuv/row.h
+++ b/include/libyuv/row.h
@@ -837,6 +837,14 @@ struct YuvConstants {
   free(var##_mem);                  \
   var = 0
 
+#define align_buffer_64_16(var, size)                                          \
+  uint8_t* var##_mem = (uint8_t*)(malloc((size * 2) + 63));       /* NOLINT */ \
+  uint16_t* var = (uint16_t*)(((intptr_t)(var##_mem) + 63) & ~63) /* NOLINT */
+
+#define free_aligned_buffer_64_16(var) \
+  free(var##_mem);                     \
+  var = 0
+
 #if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
 #define OMITFP
 #else
@@ -1934,6 +1942,8 @@ void MirrorSplitUVRow_C(const uint8_t* src_uv,
                         uint8_t* dst_v,
                         int width);
 
+void MirrorRow_16_C(const uint16_t* src, uint16_t* dst, int width);
+
 void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width);
 void ARGBMirrorRow_NEON(const uint8_t* src_argb, uint8_t* dst_argb, int width);
diff --git a/include/libyuv/version.h b/include/libyuv/version.h
index efad56b3..dfcef8d8 100644
--- a/include/libyuv/version.h
+++ b/include/libyuv/version.h
@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1855
+#define LIBYUV_VERSION 1856
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_
diff --git a/source/rotate.cc b/source/rotate.cc
index f1e83cbd..82621cd8 100644
--- a/source/rotate.cc
+++ b/source/rotate.cc
@@ -477,6 +477,120 @@ int RotatePlane(const uint8_t* src,
 }
 
 LIBYUV_API
+void TransposePlane_16(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height) {
+  int i = height;
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8_16_C(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;  // Go down 8 rows.
+    dst += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeWxH_16_C(src, src_stride, dst, dst_stride, width, i);
+  }
+}
+
+static void RotatePlane90_16(const uint16_t* src,
+                             int src_stride,
+                             uint16_t* dst,
+                             int dst_stride,
+                             int width,
+                             int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane_16(src, src_stride, dst, dst_stride, width, height);
+}
+
+static void RotatePlane270_16(const uint16_t* src,
+                              int src_stride,
+                              uint16_t* dst,
+                              int dst_stride,
+                              int width,
+                              int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane_16(src, src_stride, dst, dst_stride, width, height);
+}
+
+static void RotatePlane180_16(const uint16_t* src,
+                              int src_stride,
+                              uint16_t* dst,
+                              int dst_stride,
+                              int width,
+                              int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64_16(row, width);
+  const uint16_t* src_bot = src + src_stride * (height - 1);
+  uint16_t* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    CopyRow_16_C(src, row, width);        // Copy first row into buffer
+    MirrorRow_16_C(src_bot, dst, width);  // Mirror last row into first row
+    MirrorRow_16_C(row, dst_bot, width);  // Mirror buffer into last row
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64_16(row);
+}
+
+LIBYUV_API
+int RotatePlane_16(const uint16_t* src,
+                   int src_stride,
+                   uint16_t* dst,
+                   int dst_stride,
+                   int width,
+                   int height,
+                   enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane_16(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90_16(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src, src_stride, dst, dst_stride, width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
 int I420Rotate(const uint8_t* src_y,
                int src_stride_y,
                const uint8_t* src_u,
@@ -544,6 +658,8 @@ int I420Rotate(const uint8_t* src_y,
   return -1;
 }
 
+// I422 has half width x full height UV planes, so rotate by 90 and 270
+// require scaling to maintain 422 subsampling.
 LIBYUV_API
 int I422Rotate(const uint8_t* src_y,
                int src_stride_y,
@@ -579,31 +695,42 @@ int I422Rotate(const uint8_t* src_y,
 
   switch (mode) {
     case kRotate0:
-      // copy frame
+      // Copy frame.
       CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
       CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
       return 0;
+
+      // Note on temporary Y plane for UV.
+      // Rotation of UV first fits within the Y destination plane rows.
+      // Y plane is width x height
+      // Y plane rotated is height x width
+      // UV plane is (width / 2) x height
+      // UV plane rotated is height x (width / 2)
+      // UV plane rotated+scaled is (height / 2) x width.
+      // UV plane rotated is a temporary that fits within the Y plane rotated.
+
     case kRotate90:
-      // We need to rotate and rescale, we use plane Y as temporal storage.
-      RotatePlane90(src_u, src_stride_u, dst_y, height, halfwidth, height);
-      ScalePlane(dst_y, height, height, halfwidth, dst_u, halfheight,
+      RotatePlane90(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                    height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
                  halfheight, width, kFilterBilinear);
-      RotatePlane90(src_v, src_stride_v, dst_y, height, halfwidth, height);
-      ScalePlane(dst_y, height, height, halfwidth, dst_v, halfheight,
+      RotatePlane90(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                    height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
                  halfheight, width, kFilterLinear);
       RotatePlane90(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
       return 0;
     case kRotate270:
-      // We need to rotate and rescale, we use plane Y as temporal storage.
-      RotatePlane270(src_u, src_stride_u, dst_y, height, halfwidth, height);
-      ScalePlane(dst_y, height, height, halfwidth, dst_u, halfheight,
+      RotatePlane270(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                     height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
                  halfheight, width, kFilterBilinear);
-      RotatePlane270(src_v, src_stride_v, dst_y, height, halfwidth, height);
-      ScalePlane(dst_y, height, height, halfwidth, dst_v, halfheight,
+      RotatePlane270(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                     height);
+      ScalePlane(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
                  halfheight, width, kFilterLinear);
       RotatePlane270(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
-
       return 0;
     case kRotate180:
       RotatePlane180(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
@@ -828,6 +955,228 @@ int Android420ToI420Rotate(const uint8_t* src_y,
   return -1;
 }
 
+LIBYUV_API
+int I010Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v || dst_stride_y < 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I010Copy(src_y, src_stride_y, src_u, src_stride_u, src_v,
+                      src_stride_v, dst_y, dst_stride_y, dst_u, dst_stride_u,
+                      dst_v, dst_stride_v, width, height);
+    case kRotate90:
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                       halfheight);
+      RotatePlane90_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                       halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane270_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        halfheight);
+      RotatePlane270_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        halfheight);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+// I210 has half width x full height UV planes, so rotate by 90 and 270
+// require scaling to maintain 422 subsampling.
+LIBYUV_API
+int I210Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v || dst_stride_y < 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // Copy frame.
+      CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+      CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+      return 0;
+
+      // Note on temporary Y plane for UV.
+      // Rotation of UV first fits within the Y destination plane rows.
+      // Y plane is width x height
+      // Y plane rotated is height x width
+      // UV plane is (width / 2) x height
+      // UV plane rotated is height x (width / 2)
+      // UV plane rotated+scaled is (height / 2) x width.
+      // UV plane rotated is a temporary that fits within the Y plane rotated.
+
+    case kRotate90:
+      RotatePlane90_16(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                       height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                    halfheight, width, kFilterBilinear);
+      RotatePlane90_16(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                       height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                    halfheight, width, kFilterLinear);
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src_u, src_stride_u, dst_y, dst_stride_y, halfwidth,
+                        height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_u, dst_stride_u,
+                    halfheight, width, kFilterBilinear);
+      RotatePlane270_16(src_v, src_stride_v, dst_y, dst_stride_y, halfwidth,
+                        height);
+      ScalePlane_16(dst_y, dst_stride_y, height, halfwidth, dst_v, dst_stride_v,
+                    halfheight, width, kFilterLinear);
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth,
+                        height);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth,
+                        height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I410Rotate(const uint16_t* src_y,
+               int src_stride_y,
+               const uint16_t* src_u,
+               int src_stride_u,
+               const uint16_t* src_v,
+               int src_stride_v,
+               uint16_t* dst_y,
+               int dst_stride_y,
+               uint16_t* dst_u,
+               int dst_stride_u,
+               uint16_t* dst_v,
+               int dst_stride_v,
+               int width,
+               int height,
+               enum RotationMode mode) {
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 || !dst_y ||
+      !dst_u || !dst_v || dst_stride_y < 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      CopyPlane_16(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      CopyPlane_16(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90_16(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+      RotatePlane90_16(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+      RotatePlane90_16(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane270_16(src_u, src_stride_u, dst_u, dst_stride_u, width,
+                        height);
+      RotatePlane270_16(src_v, src_stride_v, dst_v, dst_stride_v, width,
+                        height);
+      return 0;
+    case kRotate180:
+      RotatePlane180_16(src_y, src_stride_y, dst_y, dst_stride_y, width,
+                        height);
+      RotatePlane180_16(src_u, src_stride_u, dst_u, dst_stride_u, width,
+                        height);
+      RotatePlane180_16(src_v, src_stride_v, dst_v, dst_stride_v, width,
+                        height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/source/rotate_common.cc b/source/rotate_common.cc
index ff212ade..f24accfb 100644
--- a/source/rotate_common.cc
+++ b/source/rotate_common.cc
@@ -100,6 +100,72 @@ void TransposeUVWxH_C(const uint8_t* src,
   }
 }
 
+void TransposeWx8_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+void TransposeUVWx8_16_C(const uint16_t* src,
+                         int src_stride,
+                         uint16_t* dst_a,
+                         int dst_stride_a,
+                         uint16_t* dst_b,
+                         int dst_stride_b,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+void TransposeWxH_16_C(const uint16_t* src,
+                       int src_stride,
+                       uint16_t* dst,
+                       int dst_stride,
+                       int width,
+                       int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv
diff --git a/source/row_common.cc b/source/row_common.cc
index 3d1e705e..84afd35b 100644
--- a/source/row_common.cc
+++ b/source/row_common.cc
@@ -2688,6 +2688,19 @@ void MirrorRow_C(const uint8_t* src, uint8_t* dst, int width) {
   }
 }
 
+void MirrorRow_16_C(const uint16_t* src, uint16_t* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
 void MirrorUVRow_C(const uint8_t* src_uv, uint8_t* dst_uv, int width) {
   int x;
   src_uv += (width - 1) << 1;
diff --git a/unit_test/rotate_argb_test.cc b/unit_test/rotate_argb_test.cc
index 01ed69ca..74952c4e 100644
--- a/unit_test/rotate_argb_test.cc
+++ b/unit_test/rotate_argb_test.cc
@@ -225,4 +225,110 @@ TEST_F(LibYUVRotateTest, RotatePlane90_TestStride) {
   free_aligned_buffer_page_end(src_argb);
 }
 
+static void TestRotatePlane_16(int src_width,
+                               int src_height,
+                               int dst_width,
+                               int dst_height,
+                               libyuv::RotationMode mode,
+                               int benchmark_iterations,
+                               int disable_cpu_flags,
+                               int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height < 1) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_stride = src_width;
+  int src_plane_size = src_stride * abs(src_height);
+  align_buffer_page_end_16(src, src_plane_size);
+  for (int i = 0; i < src_plane_size; ++i) {
+    src[i] = fastrand() & 0xff;
+  }
+
+  int dst_stride = dst_width;
+  int dst_plane_size = dst_stride * dst_height;
+  align_buffer_page_end_16(dst_c, dst_plane_size);
+  align_buffer_page_end_16(dst_opt, dst_plane_size);
+  memset(dst_c, 2, dst_plane_size);
+  memset(dst_opt, 3, dst_plane_size);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  RotatePlane_16(src, src_stride, dst_c, dst_stride, src_width, src_height,
+                 mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    RotatePlane_16(src, src_stride, dst_opt, dst_stride, src_width, src_height,
+                   mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_plane_size; ++i) {
+    EXPECT_EQ(dst_c[i], dst_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_c);
+  free_aligned_buffer_page_end_16(dst_opt);
+  free_aligned_buffer_page_end_16(src);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane0_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_width_,
+                     benchmark_height_, kRotate0, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane90_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_height_,
+                     benchmark_width_, kRotate90, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane180_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_width_,
+                     benchmark_height_, kRotate180, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane270_16_Opt) {
+  TestRotatePlane_16(benchmark_width_, benchmark_height_, benchmark_height_,
+                     benchmark_width_, kRotate270, benchmark_iterations_,
+                     disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane0_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_width_ + 1, benchmark_height_ + 1, kRotate0,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane90_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_height_ + 1, benchmark_width_ + 1, kRotate90,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane180_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_width_ + 1, benchmark_height_ + 1, kRotate180,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, RotatePlane270_16_Odd) {
+  TestRotatePlane_16(benchmark_width_ + 1, benchmark_height_ + 1,
+                     benchmark_height_ + 1, benchmark_width_ + 1, kRotate270,
+                     benchmark_iterations_, disable_cpu_flags_,
+                     benchmark_cpu_info_);
+}
+
 }  // namespace libyuv
diff --git a/unit_test/rotate_test.cc b/unit_test/rotate_test.cc
index d3887414..9dec7811 100644
--- a/unit_test/rotate_test.cc
+++ b/unit_test/rotate_test.cc
@@ -596,4 +596,266 @@ TESTAPLANARTOP(Android420, NV21, 2, 1, 0, 2, 2, I420, 2, 2)
 #undef TESTAPLANARTOP
 #undef TESTAPLANARTOPI
 
+static void I010TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i010_y_size = src_width * Abs(src_height);
+  int src_i010_uv_size = ((src_width + 1) / 2) * ((Abs(src_height) + 1) / 2);
+  int src_i010_size = src_i010_y_size + src_i010_uv_size * 2;
+  align_buffer_page_end_16(src_i010, src_i010_size);
+  for (int i = 0; i < src_i010_size; ++i) {
+    src_i010[i] = fastrand() & 0x3ff;
+  }
+
+  int dst_i010_y_size = dst_width * dst_height;
+  int dst_i010_uv_size = ((dst_width + 1) / 2) * ((dst_height + 1) / 2);
+  int dst_i010_size = dst_i010_y_size + dst_i010_uv_size * 2;
+  align_buffer_page_end_16(dst_i010_c, dst_i010_size);
+  align_buffer_page_end_16(dst_i010_opt, dst_i010_size);
+  memset(dst_i010_c, 2, dst_i010_size * 2);
+  memset(dst_i010_opt, 3, dst_i010_size * 2);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I010Rotate(src_i010, src_width, src_i010 + src_i010_y_size,
+             (src_width + 1) / 2, src_i010 + src_i010_y_size + src_i010_uv_size,
+             (src_width + 1) / 2, dst_i010_c, dst_width,
+             dst_i010_c + dst_i010_y_size, (dst_width + 1) / 2,
+             dst_i010_c + dst_i010_y_size + dst_i010_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I010Rotate(
+        src_i010, src_width, src_i010 + src_i010_y_size, (src_width + 1) / 2,
+        src_i010 + src_i010_y_size + src_i010_uv_size, (src_width + 1) / 2,
+        dst_i010_opt, dst_width, dst_i010_opt + dst_i010_y_size,
+        (dst_width + 1) / 2, dst_i010_opt + dst_i010_y_size + dst_i010_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i010_size; ++i) {
+    EXPECT_EQ(dst_i010_c[i], dst_i010_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i010_c);
+  free_aligned_buffer_page_end_16(dst_i010_opt);
+  free_aligned_buffer_page_end_16(src_i010);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate0_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate90_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate180_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I010Rotate270_Opt) {
+  I010TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+static void I210TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i210_y_size = src_width * Abs(src_height);
+  int src_i210_uv_size = ((src_width + 1) / 2) * Abs(src_height);
+  int src_i210_size = src_i210_y_size + src_i210_uv_size * 2;
+  align_buffer_page_end_16(src_i210, src_i210_size);
+  for (int i = 0; i < src_i210_size; ++i) {
+    src_i210[i] = fastrand() & 0x3ff;
+  }
+
+  int dst_i210_y_size = dst_width * dst_height;
+  int dst_i210_uv_size = ((dst_width + 1) / 2) * dst_height;
+  int dst_i210_size = dst_i210_y_size + dst_i210_uv_size * 2;
+  align_buffer_page_end_16(dst_i210_c, dst_i210_size);
+  align_buffer_page_end_16(dst_i210_opt, dst_i210_size);
+  memset(dst_i210_c, 2, dst_i210_size * 2);
+  memset(dst_i210_opt, 3, dst_i210_size * 2);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I210Rotate(src_i210, src_width, src_i210 + src_i210_y_size,
+             (src_width + 1) / 2, src_i210 + src_i210_y_size + src_i210_uv_size,
+             (src_width + 1) / 2, dst_i210_c, dst_width,
+             dst_i210_c + dst_i210_y_size, (dst_width + 1) / 2,
+             dst_i210_c + dst_i210_y_size + dst_i210_uv_size,
+             (dst_width + 1) / 2, src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I210Rotate(
+        src_i210, src_width, src_i210 + src_i210_y_size, (src_width + 1) / 2,
+        src_i210 + src_i210_y_size + src_i210_uv_size, (src_width + 1) / 2,
+        dst_i210_opt, dst_width, dst_i210_opt + dst_i210_y_size,
+        (dst_width + 1) / 2, dst_i210_opt + dst_i210_y_size + dst_i210_uv_size,
+        (dst_width + 1) / 2, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i210_size; ++i) {
+    EXPECT_EQ(dst_i210_c[i], dst_i210_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i210_c);
+  free_aligned_buffer_page_end_16(dst_i210_opt);
+  free_aligned_buffer_page_end_16(src_i210);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate0_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate90_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate180_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I210Rotate270_Opt) {
+  I210TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+static void I410TestRotate(int src_width,
+                           int src_height,
+                           int dst_width,
+                           int dst_height,
+                           libyuv::RotationMode mode,
+                           int benchmark_iterations,
+                           int disable_cpu_flags,
+                           int benchmark_cpu_info) {
+  if (src_width < 1) {
+    src_width = 1;
+  }
+  if (src_height == 0) {
+    src_height = 1;
+  }
+  if (dst_width < 1) {
+    dst_width = 1;
+  }
+  if (dst_height < 1) {
+    dst_height = 1;
+  }
+  int src_i410_y_size = src_width * Abs(src_height);
+  int src_i410_uv_size = src_width * Abs(src_height);
+  int src_i410_size = src_i410_y_size + src_i410_uv_size * 2;
+  align_buffer_page_end_16(src_i410, src_i410_size);
+  for (int i = 0; i < src_i410_size; ++i) {
+    src_i410[i] = fastrand() & 0x3ff;
+  }
+
+  int dst_i410_y_size = dst_width * dst_height;
+  int dst_i410_uv_size = dst_width * dst_height;
+  int dst_i410_size = dst_i410_y_size + dst_i410_uv_size * 2;
+  align_buffer_page_end_16(dst_i410_c, dst_i410_size);
+  align_buffer_page_end_16(dst_i410_opt, dst_i410_size);
+  memset(dst_i410_c, 2, dst_i410_size * 2);
+  memset(dst_i410_opt, 3, dst_i410_size * 2);
+
+  MaskCpuFlags(disable_cpu_flags);  // Disable all CPU optimization.
+  I410Rotate(src_i410, src_width, src_i410 + src_i410_y_size, src_width,
+             src_i410 + src_i410_y_size + src_i410_uv_size, src_width,
+             dst_i410_c, dst_width, dst_i410_c + dst_i410_y_size, dst_width,
+             dst_i410_c + dst_i410_y_size + dst_i410_uv_size, dst_width,
+             src_width, src_height, mode);
+
+  MaskCpuFlags(benchmark_cpu_info);  // Enable all CPU optimization.
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    I410Rotate(src_i410, src_width, src_i410 + src_i410_y_size, src_width,
+               src_i410 + src_i410_y_size + src_i410_uv_size, src_width,
+               dst_i410_opt, dst_width, dst_i410_opt + dst_i410_y_size,
+               dst_width, dst_i410_opt + dst_i410_y_size + dst_i410_uv_size,
+               dst_width, src_width, src_height, mode);
+  }
+
+  // Rotation should be exact.
+  for (int i = 0; i < dst_i410_size; ++i) {
+    EXPECT_EQ(dst_i410_c[i], dst_i410_opt[i]);
+  }
+
+  free_aligned_buffer_page_end_16(dst_i410_c);
+  free_aligned_buffer_page_end_16(dst_i410_opt);
+  free_aligned_buffer_page_end_16(src_i410);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate0_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate0, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate90_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate90, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate180_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_width_,
+                 benchmark_height_, kRotate180, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
+TEST_F(LibYUVRotateTest, I410Rotate270_Opt) {
+  I410TestRotate(benchmark_width_, benchmark_height_, benchmark_height_,
+                 benchmark_width_, kRotate270, benchmark_iterations_,
+                 disable_cpu_flags_, benchmark_cpu_info_);
+}
+
 }  // namespace libyuv
diff --git a/unit_test/unit_test.h b/unit_test/unit_test.h
index 6cc99a65..42e40d12 100644
--- a/unit_test/unit_test.h
+++ b/unit_test/unit_test.h
@@ -78,6 +78,18 @@ static inline bool SizeValid(int src_width,
   free(var##_mem);                        \
   var = 0
 
+#define align_buffer_page_end_16(var, size)                                 \
+  uint8_t* var##_mem =                                                      \
+      reinterpret_cast<uint8_t*>(malloc(((size * 2) + 4095 + 63) & ~4095)); \
+  uint16_t* var = reinterpret_cast<uint16_t*>(                              \
+      (intptr_t)(var##_mem + (((size * 2) + 4095 + 63) & ~4095) -           \
+                 (size * 2)) &                                              \
+      ~63)
+
+#define free_aligned_buffer_page_end_16(var) \
+  free(var##_mem);                           \
+  var = 0
+
 #ifdef WIN32
 static inline double get_time() {
   LARGE_INTEGER t, f;