Implements RGB to YV12 conversions in YASM.

This change implements two conversions (RGB32-to-YV12 and RGB24-to-YV12) in YASM. This change uses 8-bit fixed-point calculation and converts up to four pixels at once.

BUG=none
TEST=YUVConvertTest.SideBySideRGB
Review URL: http://codereview.chromium.org/7003082

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@100352 0039d316-1c4b-4281-b951-d872f2087c98

12 files changed, 1942 insertions, 9 deletions

diff --git a/media/base/cpu_features.h b/media/base/cpu_features.h
index c5bdcd7..c2762d8 100644
--- a/media/base/cpu_features.h
+++ b/media/base/cpu_features.h
@@ -13,6 +13,9 @@ namespace media {
 // Returns true if CPU has SSE2 support.
 bool hasSSE2();
 
+// Returns true if CPU supports SSE2, SSE3, and SSSE3.
+bool hasSSSE3();
+
 }  // namespace media
 
 #endif  // MEDIA_BASE_CPU_FEATURES_H_
diff --git a/media/base/cpu_features_arm.cc b/media/base/cpu_features_arm.cc
index e86fc97..a0d5c68 100644
--- a/media/base/cpu_features_arm.cc
+++ b/media/base/cpu_features_arm.cc
@@ -10,4 +10,8 @@ bool hasSSE2() {
   return false;
 }
 
+bool hasSSSE3() {
+  return false;
+}
+
 }  // namespace media
diff --git a/media/base/cpu_features_x86.cc b/media/base/cpu_features_x86.cc
index 98dfba1..bf7d05d 100644
--- a/media/base/cpu_features_x86.cc
+++ b/media/base/cpu_features_x86.cc
@@ -12,12 +12,6 @@
 
 namespace media {
 
-#if defined(ARCH_CPU_X86_64)
-/* All x86_64 machines have SSE2, so don't even bother checking. */
-bool hasSSE2() {
-  return true;
-}
-#else
 #ifdef _MSC_VER
 static inline void getcpuid(int info_type, int info[4]) {
   __asm {
@@ -33,6 +27,7 @@ static inline void getcpuid(int info_type, int info[4]) {
 #else
 static inline void getcpuid(int info_type, int info[4]) {
   // We save and restore ebx, so this code can be compatible with -fPIC
+#if defined(__i386__)
   asm volatile (
         "pushl %%ebx      \n\t"
         "cpuid            \n\t"
@@ -41,14 +36,34 @@ static inline void getcpuid(int info_type, int info[4]) {
         : "=a"(info[0]), "=r"(info[1]), "=c"(info[2]), "=d"(info[3])
         : "a"(info_type)
                 );
+#else
+  // We can use cpuid instruction without saving ebx on gcc x86-64 because it
+  // does not use ebx (or rbx) as a GOT register.
+  asm volatile (
+      "cpuid            \n\t"
+      : "=a"(info[0]), "=r"(info[1]), "=c"(info[2]), "=d"(info[3])
+      : "a"(info_type)
+  );
+#endif
 }
 #endif
 
 bool hasSSE2() {
+#if defined(ARCH_CPU_X86_64)
+  /* All x86_64 machines have SSE2, so don't even bother checking. */
+  return true;
+#else
   int cpu_info[4] = { 0 };
   getcpuid(1, cpu_info);
   return (cpu_info[3] & (1<<26)) != 0;
-}
 #endif
+}
+
+bool hasSSSE3() {
+  int cpu_info[4] = { 0 };
+  getcpuid(1, cpu_info);
+  return (cpu_info[3] & 0x04000000) != 0 && (cpu_info[2] & 0x00000001) != 0 &&
+      (cpu_info[2] & 0x00000200) != 0;
+}
 
 }  // namespace media
diff --git a/media/base/simd/convert_rgb_to_yuv.cc b/media/base/simd/convert_rgb_to_yuv.cc
new file mode 100644
index 0000000..2bd6930
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv.cc
@@ -0,0 +1,101 @@
+// Copyright (c) 2011 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "media/base/simd/convert_rgb_to_yuv.h"
+
+#include "build/build_config.h"
+#include "media/base/cpu_features.h"
+#include "media/base/simd/convert_rgb_to_yuv_ssse3.h"
+
+namespace media {
+
+void ConvertRGB32ToYUV_SSSE3(const uint8* rgbframe,
+                             uint8* yplane,
+                             uint8* uplane,
+                             uint8* vplane,
+                             int width,
+                             int height,
+                             int rgbstride,
+                             int ystride,
+                             int uvstride) {
+#ifdef ENABLE_SUBSAMPLING
+  for (; height >= 2; height -= 2) {
+    ConvertARGBToYUVEven_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    ConvertARGBToYUVOdd_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    uplane += uvstride;
+    vplane += uvstride;
+  }
+
+  if (height)
+    ConvertARGBToYUVEven_SSSE3(rgbframe, yplane, uplane, vplane, width);
+#else
+  for (; height >= 2; height -= 2) {
+    ConvertARGBToYUVRow_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    ConvertARGBToYUVRow_SSSE3(rgbframe, yplane, NULL, NULL, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    uplane += uvstride;
+    vplane += uvstride;
+  }
+
+  if (height)
+    ConvertARGBToYUVRow_SSSE3(rgbframe, yplane, uplane, vplane, width);
+#endif
+}
+
+void ConvertRGB24ToYUV_SSSE3(const uint8* rgbframe,
+                             uint8* yplane,
+                             uint8* uplane,
+                             uint8* vplane,
+                             int width,
+                             int height,
+                             int rgbstride,
+                             int ystride,
+                             int uvstride) {
+#ifdef ENABLE_SUBSAMPLING
+  for (; height >= 2; height -= 2) {
+    ConvertRGBToYUVEven_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    ConvertRGBToYUVOdd_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    uplane += uvstride;
+    vplane += uvstride;
+  }
+
+  if (height)
+    ConvertRGBToYUVEven_SSSE3(rgbframe, yplane, uplane, vplane, width);
+#else
+  for (; height >= 2; height -= 2) {
+    ConvertRGBToYUVRow_SSSE3(rgbframe, yplane, uplane, vplane, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    ConvertRGBToYUVRow_SSSE3(rgbframe, yplane, NULL, NULL, width);
+    rgbframe += rgbstride;
+    yplane += ystride;
+
+    uplane += uvstride;
+    vplane += uvstride;
+  }
+
+  if (height)
+    ConvertRGBToYUVRow_SSSE3(rgbframe, yplane, uplane, vplane, width);
+#endif
+}
+
+}  // namespace media
diff --git a/media/base/simd/convert_rgb_to_yuv.h b/media/base/simd/convert_rgb_to_yuv.h
new file mode 100644
index 0000000..e16fa51
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv.h
@@ -0,0 +1,40 @@
+// Copyright (c) 2011 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_H_
+#define MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_H_
+
+#include "base/basictypes.h"
+#include "media/base/yuv_convert.h"
+
+namespace media {
+
+// Converts an ARGB image to a YV12 image. This function calls ASM functions
+// implemented in "convert_rgb_to_yuv_ssse3.asm" to convert the specified ARGB
+// image to a YV12 image.
+void ConvertRGB32ToYUV_SSSE3(const uint8* rgbframe,
+                             uint8* yplane,
+                             uint8* uplane,
+                             uint8* vplane,
+                             int width,
+                             int height,
+                             int rgbstride,
+                             int ystride,
+                             int uvstride);
+
+// Converts an RGB image to a YV12 image. This function is almost same as
+// ConvertRGB32ToYUV_SSSE3 except its first argument is a pointer to RGB pixels.
+void ConvertRGB24ToYUV_SSSE3(const uint8* rgbframe,
+                             uint8* yplane,
+                             uint8* uplane,
+                             uint8* vplane,
+                             int width,
+                             int height,
+                             int rgbstride,
+                             int ystride,
+                             int uvstride);
+
+}  // namespace media
+
+#endif  // MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_H_
diff --git a/media/base/simd/convert_rgb_to_yuv_ssse3.asm b/media/base/simd/convert_rgb_to_yuv_ssse3.asm
new file mode 100644
index 0000000..f445e98
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv_ssse3.asm
@@ -0,0 +1,317 @@
+; Copyright (c) 2011 The Chromium Authors. All rights reserved.
+; Use of this source code is governed by a BSD-style license that can be
+; found in the LICENSE file.
+
+%include "x86inc.asm"
+
+;
+; This file uses SSE, SSE2, SSE3, and SSSE3, which are supported by all ATOM
+; processors.
+;
+  SECTION_TEXT
+  CPU       SSE, SSE3, SSE3, SSSE3
+
+;
+; XMM registers representing constants. We must not use these registers as
+; destination operands.
+; for (int i = 0; i < 16; i += 4) {
+;   xmm7.b[i] = 25;  xmm7.b[i+1] = 2;   xmm7.b[i+2] = 66;  xmm7.b[i+3] = 0;
+;   xmm6.b[i] = 0;   xmm6.b[i+1] = 127; xmm6.b[i+2] = 0;   xmm6.b[i+3] = 0;
+;   xmm5.b[i] = 112; xmm5.b[i+1] = -74; xmm5.b[i+2] = -38; xmm5.b[i+3] = 0;
+;   xmm4.b[i] = -18; xmm4.b[i+1] = -94; xmm4.b[i+2] = 112; xmm4.b[i+3] = 0;
+; }
+;
+%define XMM_CONST_Y0    xmm7
+%define XMM_CONST_Y1    xmm6
+%define XMM_CONST_U     xmm5
+%define XMM_CONST_V     xmm4
+%define XMM_CONST_128   xmm3
+
+;
+; LOAD_XMM %1 (xmm), %2 (imm32)
+; Loads an immediate value to an XMM register.
+;   %1.d[0] = %1.d[1] =  %1.d[2] =  %1.d[3] = %2;
+;
+%macro LOAD_XMM 2
+  mov       TEMPd, %2
+  movd      %1, TEMPd
+  pshufd    %1, %1, 00000000B
+%endmacro
+
+;
+; UNPACKRGB %1 (xmm), %2 (imm8)
+; Unpacks one RGB pixel in the specified XMM register.
+;   for (int i = 15; i > %2; --i) %1.b[i] = %1.b[i - 1];
+;   %1.b[%2] = 0;
+;   for (int i = %2 - 1; i >= 0; --i) %1.b[i] = %1.b[i];
+;
+%macro UNPACKRGB 2
+  movdqa    xmm1, %1
+  psrldq    xmm1, %2
+  pslldq    xmm1, %2
+  pxor      %1, xmm1
+  pslldq    xmm1, 1
+  por       %1, xmm1
+%endmacro
+
+;
+; READ_ARGB %1 (xmm), %2 (imm)
+; Read the specified number of ARGB (or RGB) pixels from the source and store
+; them to the destination xmm register. If the input format is RGB, we read RGB
+; pixels and convert them to ARGB pixels. (For this case, the alpha values of
+; the output pixels become 0.)
+;
+%macro READ_ARGB 2
+
+%if PIXELSIZE == 4
+
+  ; Read ARGB pixels from the source. (This macro assumes the input buffer may
+  ; not be aligned to a 16-byte boundary.)
+%if %2 == 1
+  movd      %1, DWORD [ARGBq + WIDTHq * 4 * 2]
+%elif %2 == 2
+  movq      %1, QWORD [ARGBq + WIDTHq * 4 * 2]
+%elif %2 == 4
+  movdqu    %1, DQWORD [ARGBq + WIDTHq * 4 * 2]
+%else
+%error unsupported number of pixels.
+%endif
+
+%elif PIXELSIZE == 3
+
+  ; Read RGB pixels from the source and convert them to ARGB pixels.
+%if %2 == 1
+  ; Read one RGB pixel and convert it to one ARGB pixel.
+  ; Save the WIDTH register to xmm1. (This macro needs to break it.)
+  MOVq      xmm1, WIDTHq
+
+  ; Once read three bytes from the source to TEMPd, and copy it to the
+  ; destination xmm register.
+  lea       WIDTHq, [WIDTHq + WIDTHq * 2]
+  movzx     TEMPd, BYTE [ARGBq + WIDTHq * 2 + 2]
+  shl       TEMPd, 16
+  mov       TEMPw, WORD [ARGBq + WIDTHq * 2]
+  movd      %1, TEMPd
+
+  ; Restore the WIDTH register.
+  MOVq      WIDTHq, xmm1
+%elif %2 == 2
+  ; Read two RGB pixels and convert them to two ARGB pixels.
+  ; Read six bytes from the source to the destination xmm register.
+  mov       TEMPq, WIDTHq
+  lea       TEMPq, [TEMPq + TEMPq * 2]
+  movd      %1, DWORD [ARGBq + TEMPq * 2]
+  pinsrw    %1, WORD [ARGBq + TEMPq * 2 + 4], 3
+
+  ; Fill the alpha values of these RGB pixels with 0 and convert them to two
+  ; ARGB pixels.
+  UNPACKRGB %1, 3
+%elif %2 == 4
+  ; Read four RGB pixels and convert them to four ARGB pixels.
+  ; Read twelve bytes from the source to the destination xmm register.
+  mov       TEMPq, WIDTHq
+  lea       TEMPq, [TEMPq + TEMPq * 2]
+  movq      %1, QWORD [ARGBq + TEMPq * 2]
+  movd      xmm1, DWORD [ARGBq + TEMPq * 2 + 8]
+  shufps    %1, xmm1, 01000100B
+
+  ; Fill the alpha values of these RGB pixels with 0 and convert them to four
+  ; ARGB pixels.
+  UNPACKRGB %1, 3
+  UNPACKRGB %1, 4 + 3
+  UNPACKRGB %1, 4 + 4 + 3
+%else
+%error unsupported number of pixels.
+%endif
+
+%else
+%error unsupported PIXELSIZE value.
+%endif
+
+%endmacro
+
+;
+; CALC_Y %1 (xmm), %2 (xmm)
+; Calculates four Y values from four ARGB pixels stored in %2.
+;   %1.b[0] = ToByte((25 * B(0) + 129 * G(0) + 66 * R(0) + 128) / 256 + 16);
+;   %1.b[1] = ToByte((25 * B(1) + 129 * G(1) + 66 * R(1) + 128) / 256 + 16);
+;   %1.b[2] = ToByte((25 * B(2) + 129 * G(2) + 66 * R(2) + 128) / 256 + 16);
+;   %1.b[3] = ToByte((25 * B(3) + 129 * G(3) + 66 * R(3) + 128) / 256 + 16);
+;
+%macro CALC_Y 2
+  ; To avoid signed saturation, we divide this conversion formula into two
+  ; formulae and store their results into two XMM registers %1 and xmm2.
+  ; %1.w[0]   = 25  * %2.b[0]  + 2   * %2.b[1]  + 66  * %2.b[2]  + 0 * %2.b[3];
+  ; %1.w[1]   = 25  * %2.b[4]  + 2   * %2.b[5]  + 66  * %2.b[6]  + 0 * %2.b[7];
+  ; %1.w[2]   = 25  * %2.b[8]  + 2   * %2.b[9]  + 66  * %2.b[10] + 0 * %2.b[11];
+  ; %1.w[3]   = 25  * %2.b[12] + 2   * %2.b[13] + 66  * %2.b[14] + 0 * %2.b[15];
+  ; xmm2.w[0] = 0   * %2.b[0]  + 127 * %2.b[1]  + 0   * %2.b[2]  + 0 * %2.b[3];
+  ; xmm2.w[1] = 0   * %2.b[4]  + 127 * %2.b[5]  + 0   * %2.b[6]  + 0 * %2.b[7];
+  ; xmm2.w[2] = 0   * %2.b[8]  + 127 * %2.b[9]  + 0   * %2.b[10] + 0 * %2.b[11];
+  ; xmm2.w[3] = 0   * %2.b[12] + 127 * %2.b[13] + 0   * %2.b[14] + 0 * %2.b[15];
+  movdqa    %1, %2
+  pmaddubsw %1, XMM_CONST_Y0
+  phaddsw   %1, %1
+  movdqa    xmm2, %2
+  pmaddubsw xmm2, XMM_CONST_Y1
+  phaddsw   xmm2, xmm2
+
+  ; %1.b[0] = ToByte((%1.w[0] + xmm2.w[0] + 128) / 256 + 16);
+  ; %1.b[1] = ToByte((%1.w[1] + xmm2.w[1] + 128) / 256 + 16);
+  ; %1.b[2] = ToByte((%1.w[2] + xmm2.w[2] + 128) / 256 + 16);
+  ; %1.b[3] = ToByte((%1.w[3] + xmm2.w[3] + 128) / 256 + 16);
+  paddw     %1, xmm2
+  movdqa    xmm2, XMM_CONST_128
+  paddw     %1, xmm2
+  psrlw     %1, 8
+  psrlw     xmm2, 3
+  paddw     %1, xmm2
+  packuswb  %1, %1
+%endmacro
+
+;
+; INIT_UV %1 (r32), %2 (reg) %3 (imm)
+;
+%macro INIT_UV 3
+
+%if SUBSAMPLING == 1 && LINE == 1
+%if %3 == 1 || %3 == 2
+  movzx     %1, BYTE [%2 + WIDTHq]
+%elif %3 == 4
+  movzx     %1, WORD [%2 + WIDTHq]
+%else
+%error unsupported number of pixels.
+%endif
+%endif
+
+%endmacro
+
+;
+; CALC_UV %1 (xmm), %2 (xmm), %3 (xmm), %4 (r32)
+; Calculates two U (or V) values from four ARGB pixels stored in %2.
+; if %3 == XMM_CONST_U
+; if (SUBSAMPLING) {
+;   %1.b[0] = ToByte((112 * B(0) - 74 * G(0) - 38 * R(0) + 128) / 256 + 128);
+;   %1.b[0] = ToByte((112 * B(0) - 74 * G(0) - 38 * R(0) + 128) / 256 + 128);
+;   %1.b[1] = ToByte((112 * B(2) - 74 * G(2) - 38 * R(2) + 128) / 256 + 128);
+;   %1.b[1] = ToByte((112 * B(2) - 74 * G(2) - 38 * R(2) + 128) / 256 + 128);
+; } else {
+;   %1.b[0] = ToByte((112 * B(0) - 74 * G(0) - 38 * R(0) + 128) / 256 + 128);
+;   %1.b[1] = ToByte((112 * B(2) - 74 * G(2) - 38 * R(2) + 128) / 256 + 128);
+; }
+; if %3 == XMM_CONST_V
+;   %1.b[0] = ToByte((-18 * B(0) - 94 * G(0) + 112 * R(0) + 128) / 256 + 128);
+;   %1.b[1] = ToByte((-18 * B(2) - 94 * G(2) + 112 * R(2) + 128) / 256 + 128);
+;
+%macro CALC_UV 4
+  ; for (int i = 0; i < 4; ++i) {
+  ;   %1.w[i] = 0;
+  ;   for (int j = 0; j < 4; ++j)
+  ;     %1.w[i] += %3.b[i * 4 + j] + %2.b[i * 4 + j];
+  ; }
+  movdqa    %1, %2
+  pmaddubsw %1, %3
+  phaddsw   %1, %1
+
+%if SUBSAMPLING == 1
+  ; %1.w[0] = (%1.w[0] + %1.w[1] + 1) / 2;
+  ; %1.w[1] = (%1.w[1] + %1.w[0] + 1) / 2;
+  ; %1.w[2] = (%1.w[2] + %1.w[3] + 1) / 2;
+  ; %1.w[3] = (%1.w[3] + %1.w[2] + 1) / 2;
+  pshuflw   xmm2, %1, 10110001B
+  pavgw     %1, xmm2
+%endif
+
+  ; %1.b[0] = ToByte((%1.w[0] + 128) / 256 + 128);
+  ; %1.b[1] = ToByte((%1.w[2] + 128) / 256 + 128);
+  pshuflw   %1, %1, 10001000B
+  paddw     %1, XMM_CONST_128
+  psraw     %1, 8
+  paddw     %1, XMM_CONST_128
+  packuswb  %1, %1
+
+%if SUBSAMPLING == 1 && LINE == 1
+  ; %1.b[0] = (%1.b[0] + %3.b[0] + 1) / 2;
+  ; %1.b[1] = (%1.b[1] + %3.b[1] + 1) / 2;
+  movd      xmm2, %4
+  pavgb     %1, xmm2
+%endif
+%endmacro
+
+;
+; extern "C" void ConvertARGBToYUVRow_SSSE3(const uint8* argb,
+;                                           uint8* y,
+;                                           uint8* u,
+;                                           uint8* v,
+;                                           int width);
+;
+%define SYMBOL          ConvertARGBToYUVRow_SSSE3
+%define PIXELSIZE       4
+%define SUBSAMPLING     0
+%define LINE            0
+%include "convert_rgb_to_yuv_ssse3.inc"
+
+;
+; extern "C" void ConvertRGBToYUVRow_SSSE3(const uint8* rgb,
+;                                          uint8* y,
+;                                          uint8* u,
+;                                          uint8* v,
+;                                          int width);
+;
+%define SYMBOL          ConvertRGBToYUVRow_SSSE3
+%define PIXELSIZE       3
+%define SUBSAMPLING     0
+%define LINE            0
+%include "convert_rgb_to_yuv_ssse3.inc"
+
+;
+; extern "C" void ConvertARGBToYUVEven_SSSE3(const uint8* argb,
+;                                            uint8* y,
+;                                            uint8* u,
+;                                            uint8* v,
+;                                            int width);
+;
+%define SYMBOL          ConvertARGBToYUVEven_SSSE3
+%define PIXELSIZE       4
+%define SUBSAMPLING     1
+%define LINE            0
+%include "convert_rgb_to_yuv_ssse3.inc"
+
+;
+; extern "C" void ConvertARGBToYUVOdd_SSSE3(const uint8* argb,
+;                                           uint8* y,
+;                                           uint8* u,
+;                                           uint8* v,
+;                                           int width);
+;
+%define SYMBOL          ConvertARGBToYUVOdd_SSSE3
+%define PIXELSIZE       4
+%define SUBSAMPLING     1
+%define LINE            1
+%include "convert_rgb_to_yuv_ssse3.inc"
+
+;
+; extern "C" void ConvertRGBToYUVEven_SSSE3(const uint8* rgb,
+;                                           uint8* y,
+;                                           uint8* u,
+;                                           uint8* v,
+;                                           int width);
+;
+%define SYMBOL          ConvertRGBToYUVEven_SSSE3
+%define PIXELSIZE       3
+%define SUBSAMPLING     1
+%define LINE            0
+%include "convert_rgb_to_yuv_ssse3.inc"
+
+;
+; extern "C" void ConvertRGBToYUVOdd_SSSE3(const uint8* rgb,
+;                                          uint8* y,
+;                                          uint8* u,
+;                                          uint8* v,
+;                                          int width);
+;
+%define SYMBOL          ConvertRGBToYUVOdd_SSSE3
+%define PIXELSIZE       3
+%define SUBSAMPLING     1
+%define LINE            1
+%include "convert_rgb_to_yuv_ssse3.inc"
diff --git a/media/base/simd/convert_rgb_to_yuv_ssse3.h b/media/base/simd/convert_rgb_to_yuv_ssse3.h
new file mode 100644
index 0000000..9f7a372
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv_ssse3.h
@@ -0,0 +1,57 @@
+// Copyright (c) 2011 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_SSSE3_H_
+#define MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_SSSE3_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The header file for ASM functions that convert a row of RGB pixels with SSSE3
+// instructions so we can call them from C++ code. These functions are
+// implemented in "convert_rgb_to_yuv_ssse3.asm".
+
+// Convert a row of 24-bit RGB pixels to YV12 pixels.
+void ConvertRGBToYUVRow_SSSE3(const uint8* rgb,
+                              uint8* y,
+                              uint8* u,
+                              uint8* v,
+                              int width);
+
+// Convert a row of 32-bit RGB pixels to YV12 pixels.
+void ConvertARGBToYUVRow_SSSE3(const uint8* argb,
+                               uint8* y,
+                               uint8* u,
+                               uint8* v,
+                               int width);
+
+// Convert a row of 24-bit RGB pixels to YV12 pixels.
+void ConvertRGBToYUVEven_SSSE3(const uint8* rgb,
+                               uint8* y,
+                               uint8* u,
+                               uint8* v,
+                               int width);
+void ConvertRGBToYUVOdd_SSSE3(const uint8* rgb,
+                              uint8* y,
+                              uint8* u,
+                              uint8* v,
+                              int width);
+
+// Convert a row of 32-bit RGB pixels to YV12 pixels.
+void ConvertARGBToYUVEven_SSSE3(const uint8* argb,
+                                uint8* y,
+                                uint8* u,
+                                uint8* v,
+                                int width);
+void ConvertARGBToYUVOdd_SSSE3(const uint8* argb,
+                               uint8* y,
+                               uint8* u,
+                               uint8* v,
+                               int width);
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // MEDIA_BASE_SIMD_CONVERT_RGB_TO_YUV_SSSE3_H_
diff --git a/media/base/simd/convert_rgb_to_yuv_ssse3.inc b/media/base/simd/convert_rgb_to_yuv_ssse3.inc
new file mode 100644
index 0000000..35c0db9
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv_ssse3.inc
@@ -0,0 +1,200 @@
+; Copyright (c) 2011 The Chromium Authors. All rights reserved.
+; Use of this source code is governed by a BSD-style license that can be
+; found in the LICENSE file.
+
+;
+; void SYMBOL(const uint8* argb, uint8* y, uint8* u, uint8* v, int width);
+;
+; The main code that converts RGB pixels to YUV pixels. This function roughly
+; consists of three parts: converting one ARGB pixel to YUV pixels, converting
+; two ARGB pixels to YUV pixels, and converting four ARGB pixels to YUV pixels.
+; To write the structure of this function in C, it becomes the snippet listed
+; below.
+;
+;   if (width & 1) {
+;     --width;
+;     // Convert one ARGB pixel to one Y pixel, one U pixel, and one V pixel.
+;   }
+;
+;   if (width & 2) {
+;     width -= 2;
+;     // Convert two ARGB pixels to two Y pixels, one U pixel, and one V pixel.
+;   }
+;
+;   while (width) {
+;     width -= 4;
+;     // Convert four ARGB pixels to four Y pixels, two U pixels, and two V
+;     // pixels.
+;   }
+;
+  global    mangle(SYMBOL) PRIVATE
+  align     function_align
+
+mangle(SYMBOL):
+  %assign stack_offset 0
+  PROLOGUE 5, 6, 8, ARGB, Y, U, V, WIDTH, TEMP
+
+  ; Initialize constants used in this function. (We use immediates to avoid
+  ; dependency onto GOT.)
+  LOAD_XMM  XMM_CONST_Y0, 0x00420219
+  LOAD_XMM  XMM_CONST_Y1, 0x00007F00
+  LOAD_XMM  XMM_CONST_U, 0x00DAB670
+  LOAD_XMM  XMM_CONST_V, 0x0070A2EE
+  LOAD_XMM  XMM_CONST_128, 0x00800080
+
+.convert_one_pixel:
+  ; Divide the input width by two so it represents the offsets for u[] and v[].
+  ; When the width is odd, We read the rightmost ARGB pixel and convert its
+  ; colorspace to YUV. This code stores one Y pixel, one U pixel, and one V
+  ; pixel.
+  sar       WIDTHq, 1
+  jnc       .convert_two_pixels
+
+  ; Read one ARGB (or RGB) pixel.
+  READ_ARGB xmm0, 1
+
+  ; Calculate y[0] from one RGB pixel read above.
+  CALC_Y    xmm1, xmm0
+  movd      TEMPd, xmm1
+  mov       BYTE [Yq + WIDTHq * 2], TEMPb
+
+  ; Calculate u[0] from one RGB pixel read above. If this is an odd line, the
+  ; output pixel contains the U value calculated in the previous call. We also
+  ; read this pixel and calculate their average.
+  INIT_UV   TEMPd, Uq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_U, TEMPd
+  movd      TEMPd, xmm1
+  mov       BYTE [Uq + WIDTHq], TEMPb
+
+  ; Calculate v[0] from one RGB pixel. Same as u[0], we read the result of the
+  ; previous call and get their average.
+  INIT_UV   TEMPd, Uq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_V, TEMPd
+  movd      TEMPd, xmm1
+  mov       BYTE [Vq + WIDTHq], TEMPb
+
+.convert_two_pixels:
+  ; If the input width is not a multiple of four, read the rightmost two ARGB
+  ; pixels and convert their colorspace to YUV. This code stores two Y pixels,
+  ; one U pixel, and one V pixel.
+  test      WIDTHb, 2 / 2
+  jz        .convert_four_pixels
+  sub       WIDTHb, 2 / 2
+
+  ; Read two ARGB (or RGB) pixels.
+  READ_ARGB xmm0, 2
+
+  ; Calculate r[0] and r[1] from two RGB pixels read above.
+  CALC_Y    xmm1, xmm0
+  movd      TEMPd, xmm1
+  mov       WORD [Yq + WIDTHq * 2], TEMPw
+
+  ; Skip calculating u and v if the output buffer is NULL.
+  test      Uq, Uq
+  jz        .convert_four_pixels
+
+  ; Calculate u[0] from two RGB pixels read above. (For details, read the above
+  ; comment in .convert_one_pixel).
+  INIT_UV   TEMPd, Uq, 2
+  CALC_UV   xmm1, xmm0, XMM_CONST_U, TEMPd
+  movd      TEMPd, xmm1
+  mov       BYTE [Uq + WIDTHq], TEMPb
+
+  ; Calculate v[0] from two RGB pixels read above.
+  INIT_UV   TEMPd, Vq, 2
+  CALC_UV   xmm1, xmm0, XMM_CONST_V, TEMPd
+  movd      TEMPd, xmm1
+  mov       BYTE [Vq + WIDTHq], TEMPb
+
+.convert_four_pixels:
+  ; Read four ARGB pixels and convert their colorspace to YUV. This code stores
+  ; four Y pixels, two U pixels, and two V pixels.
+  test      WIDTHq, WIDTHq
+  jz        .convert_finish
+
+%if PIXELSIZE == 4
+  ; Check if the input buffer is aligned to a 16-byte boundary and use movdqa
+  ; for reading the ARGB pixels.
+  test      ARGBw, 15
+  jnz       .convert_four_pixels_unaligned
+
+.convert_four_pixels_aligned:
+  sub       WIDTHq, 4 / 2
+
+  ; Read four ARGB pixels. (We can use movdqa here since we have checked if the
+  ; source address is aligned.)
+  movdqa    xmm0, DQWORD [ARGBq + WIDTHq * 4 * 2]
+
+  ; Calculate y[0], y[1], y[2],and, y[3] from the input ARGB pixels.
+  CALC_Y    xmm1, xmm0
+  movd      DWORD [Yq + WIDTHq * 2], xmm1
+
+%if SUBSAMPLING == 0
+  ; Skip calculating u and v if the output buffer is NULL, which means we are
+  ; converting an odd line. (When we enable subsampling, these buffers must
+  ; contain the u and v values for the previous call, i.e. these variables must
+  ; not be NULL.)
+  test      Uq, Uq
+  jz        .convert_four_pixels_aligned_next
+%endif
+
+  ; Calculate u[0] and u[1] from four ARGB pixels read above.
+  INIT_UV   TEMPd, Uq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_U, TEMPd
+  movd      TEMPd, xmm1
+  mov       WORD [Uq + WIDTHq], TEMPw
+
+  ; Calculate v[0] and v[1] from four ARGB pixels read above.
+  INIT_UV   TEMPd, Vq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_V, TEMPd
+  movd      TEMPd, xmm1
+  mov       WORD [Vq + WIDTHq], TEMPw
+
+%if SUBSAMPLING == 0
+.convert_four_pixels_aligned_next:
+%endif
+
+  test      WIDTHq, WIDTHq
+  jnz       .convert_four_pixels_aligned
+
+  jmp       .convert_finish
+%endif
+
+.convert_four_pixels_unaligned:
+  sub       WIDTHq, 4 / 2
+
+  ; Read four ARGB (or RGB) pixels.
+  READ_ARGB xmm0, 4
+
+  ; Calculate y[0], y[1], y[2],and, y[3] from the input ARGB pixels.
+  CALC_Y    xmm1, xmm0
+  movd      DWORD [Yq + WIDTHq * 2], xmm1
+
+%if SUBSAMPLING == 0
+  ; Skip calculating u and v if the output buffer is NULL.
+  test      Uq, Uq
+  jz        .convert_four_pixels_unaligned_next
+%endif
+
+  ; Calculate u[0] and u[1] from the input ARGB pixels.
+  INIT_UV   TEMPd, Uq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_U, TEMPd
+  movd      TEMPd, xmm1
+  mov       WORD [Uq + WIDTHq], TEMPw
+
+  ; Calculate v[0] and v[1] from the input ARGB pixels.
+  INIT_UV   TEMPd, Vq, 4
+  CALC_UV   xmm1, xmm0, XMM_CONST_V, TEMPd
+  movd      TEMPd, xmm1
+  mov       WORD [Vq + WIDTHq], TEMPw
+
+%if SUBSAMPLING == 0
+.convert_four_pixels_unaligned_next:
+%endif
+
+  test      WIDTHq, WIDTHq
+  jnz       .convert_four_pixels_unaligned
+
+.convert_finish:
+  ; Just exit this function since this is a void function.
+  RET
diff --git a/media/base/simd/convert_rgb_to_yuv_unittest.cc b/media/base/simd/convert_rgb_to_yuv_unittest.cc
new file mode 100644
index 0000000..5dcd9e8
--- /dev/null
+++ b/media/base/simd/convert_rgb_to_yuv_unittest.cc
@@ -0,0 +1,124 @@
+// Copyright (c) 2011 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "base/scoped_ptr.h"
+#include "base/stringprintf.h"
+#include "media/base/cpu_features.h"
+#include "media/base/simd/convert_rgb_to_yuv.h"
+#include "testing/gtest/include/gtest/gtest.h"
+
+namespace {
+
+// Reference code that converts RGB pixels to YUV pixels.
+int ConvertRGBToY(const uint8* rgb) {
+  int y = 25 * rgb[0] + 129 * rgb[1] + 66 * rgb[2];
+  y = ((y + 128) >> 8) + 16;
+  return std::max(0, std::min(255, y));
+}
+
+int ConvertRGBToU(const uint8* rgb, int size, bool subsampling) {
+  int u = 0;
+  if (!subsampling) {
+    u = 112 * rgb[0] - 74 * rgb[1] - 38 * rgb[2];
+  } else {
+    int u0 = 112 * rgb[0] - 74 * rgb[1] - 38 * rgb[2];
+    int u1 = 112 * rgb[size] - 74 * rgb[size + 1] - 38 * rgb[size + 2];
+    u = (u0 + u1 + 1) / 2;
+  }
+  u = ((u + 128) >> 8) + 128;
+  return std::max(0, std::min(255, u));
+}
+
+int ConvertRGBToV(const uint8* rgb, int size, bool subsampling) {
+  int v = 0;
+  if (!subsampling) {
+    v = -18 * rgb[0] - 94 * rgb[1] + 112 * rgb[2];
+  } else {
+    int v0 = -18 * rgb[0] - 94 * rgb[1] + 112 * rgb[2];
+    int v1 = -18 * rgb[size] - 94 * rgb[size + 1] + 112 * rgb[size + 2];
+    v = (v0 + v1 + 1) / 2;
+  }
+  v = ((v + 128) >> 8) + 128;
+  return std::max(0, std::min(255, v));
+}
+
+}  // namespace
+
+// A side-by-side test that verifies our ASM functions that convert RGB pixels
+// to YUV pixels can output the expected results. This test converts RGB pixels
+// to YUV pixels with our ASM functions (which use SSE, SSE2, SSE3, and SSSE3)
+// and compare the output YUV pixels with the ones calculated with out reference
+// functions implemented in C++.
+TEST(YUVConvertTest, SideBySideRGB) {
+  // We skip this test on PCs which does not support SSE3 because this test
+  // needs it.
+  if (!media::hasSSSE3())
+    return;
+
+  // This test checks a subset of all RGB values so this test does not take so
+  // long time.
+  const int kStep = 8;
+  const int kWidth = 256 / kStep;
+
+#ifdef ENABLE_SUBSAMPLING
+  const bool kSubsampling = true;
+#else
+  const bool kSubsampling = false;
+#endif
+
+  for (int size = 3; size <= 4; ++size) {
+    // Create the output buffers.
+    scoped_array<uint8> rgb(new uint8[kWidth * size]);
+    scoped_array<uint8> y(new uint8[kWidth]);
+    scoped_array<uint8> u(new uint8[kWidth / 2]);
+    scoped_array<uint8> v(new uint8[kWidth / 2]);
+
+    // Choose the function that converts from RGB pixels to YUV ones.
+    void (*convert)(const uint8*, uint8*, uint8*, uint8*,
+                    int, int, int, int, int) = NULL;
+    if (size == 3)
+      convert = media::ConvertRGB24ToYUV_SSSE3;
+    else
+      convert = media::ConvertRGB32ToYUV_SSSE3;
+
+    for (int r = 0; r < kWidth; ++r) {
+      for (int g = 0; g < kWidth; ++g) {
+
+        // Fill the input pixels.
+        for (int b = 0; b < kWidth; ++b) {
+          rgb[b * size + 0] = b * kStep;
+          rgb[b * size + 1] = g * kStep;
+          rgb[b * size + 2] = r * kStep;
+          if (size == 4)
+            rgb[b * size + 3] = 255;
+        }
+
+        // Convert the input RGB pixels to YUV ones.
+        convert(rgb.get(), y.get(), u.get(), v.get(), kWidth, 1, kWidth * size,
+                kWidth, kWidth / 2);
+
+        // Check the output Y pixels.
+        for (int i = 0; i < kWidth; ++i) {
+          const uint8* p = &rgb[i * size];
+          SCOPED_TRACE(base::StringPrintf("r=%d,g=%d,b=%d", p[2], p[1], p[0]));
+          EXPECT_EQ(ConvertRGBToY(p), y[i]);
+        }
+
+        // Check the output U pixels.
+        for (int i = 0; i < kWidth / 2; ++i) {
+          const uint8* p = &rgb[i * 2 * size];
+          SCOPED_TRACE(base::StringPrintf("r=%d,g=%d,b=%d", p[2], p[1], p[0]));
+          EXPECT_EQ(ConvertRGBToU(p, size, kSubsampling), u[i]);
+        }
+
+        // Check the output V pixels.
+        for (int i = 0; i < kWidth / 2; ++i) {
+          const uint8* p = &rgb[i * 2 * size];
+          SCOPED_TRACE(base::StringPrintf("r=%d,g=%d,b=%d", p[2], p[1], p[0]));
+          EXPECT_EQ(ConvertRGBToV(p, size, kSubsampling), v[i]);
+        }
+      }
+    }
+  }
+}
diff --git a/media/base/simd/x86inc.asm b/media/base/simd/x86inc.asm
new file mode 100644
index 0000000..956b999
--- /dev/null
+++ b/media/base/simd/x86inc.asm
@@ -0,0 +1,1001 @@
+;*****************************************************************************
+;* x86inc.asm
+;*****************************************************************************
+;* Copyright (C) 2005-2011 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Jason Garrett-Glaser <darkshikari@gmail.com>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+%ifndef MEDIA_BASE_SIMD_X86INC_ASM_
+%define MEDIA_BASE_SIMD_X86INC_ASM_
+
+%define program_name ff
+
+%ifdef ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64
+    %else
+        %define UNIX64
+    %endif
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; FIXME: All of the 64bit asm functions that take a stride as an argument
+; via register, assume that the high dword of that register is filled with 0.
+; This is true in practice (since we never do any 64bit arithmetic on strides,
+; and x264's strides are all positive), but is not guaranteed by the ABI.
+
+; Name of the .rodata section.
+; Kludge: Something on OS X fails to align .rodata even given an align attribute,
+; so use a different read-only section.
+%ifdef CHROMIUM
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho64
+        SECTION .text align=%1
+    %elifidn __OUTPUT_FORMAT__,macho
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+%else
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+%endif
+
+; aout does not support align=
+%macro SECTION_TEXT 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .text align=%1
+    %endif
+%endmacro
+
+%ifdef WIN64
+    %define PIC
+%elifndef ARCH_X86_64
+; x86_32 doesn't require PIC.
+; Some distros prefer shared objects to be PIC, but nothing breaks if
+; the code contains a few textrels, so we'll skip that complexity.
+    %undef PIC
+%endif
+%ifdef PIC
+    default rel
+%endif
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,0, dst, src, tmp
+; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE
+
+; REP_RET:
+; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
+; which are slow when a normal ret follows a branch.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 6
+    %define r%1q %2
+    %define r%1d %3
+    %define r%1w %4
+    %define r%1b %5
+    %define r%1m %6
+    %ifid %6 ; i.e. it's a register
+        %define r%1mp %2
+    %elifdef ARCH_X86_64 ; memory
+        %define r%1mp qword %6
+    %else
+        %define r%1mp dword %6
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 2
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1b %2
+    %define e%1b %2
+%ifndef ARCH_X86_64
+    %define r%1  e%1
+%endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al
+DECLARE_REG_SIZE bx, bl
+DECLARE_REG_SIZE cx, cl
+DECLARE_REG_SIZE dx, dl
+DECLARE_REG_SIZE si, sil
+DECLARE_REG_SIZE di, dil
+DECLARE_REG_SIZE bp, bpl
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9
+
+%ifdef ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %assign stack_offset stack_offset+gprsize
+%endmacro
+
+%macro POP 1
+    pop %1
+    %assign stack_offset stack_offset-gprsize
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assert failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %assign n_arg_names %%i
+%endmacro
+
+%ifdef WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0, rcx, ecx, cx,  cl,  ecx
+DECLARE_REG 1, rdx, edx, dx,  dl,  edx
+DECLARE_REG 2, r8,  r8d, r8w, r8b, r8d
+DECLARE_REG 3, r9,  r9d, r9w, r9b, r9d
+DECLARE_REG 4, rdi, edi, di,  dil, [rsp + stack_offset + 40]
+DECLARE_REG 5, rsi, esi, si,  sil, [rsp + stack_offset + 48]
+DECLARE_REG 6, rax, eax, ax,  al,  [rsp + stack_offset + 56]
+%define r7m [rsp + stack_offset + 64]
+%define r8m [rsp + stack_offset + 72]
+
+%macro LOAD_IF_USED 2 ; reg_id, number_of_args
+    %if %1 < %2
+        mov r%1, [rsp + stack_offset + 8 + %1*8]
+    %endif
+%endmacro
+
+%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+    ASSERT %2 >= %1
+    %assign regs_used %2
+    ASSERT regs_used <= 7
+    %if regs_used > 4
+        push r4
+        push r5
+        %assign stack_offset stack_offset+16
+    %endif
+    WIN64_SPILL_XMM %3
+    LOAD_IF_USED 4, %1
+    LOAD_IF_USED 5, %1
+    LOAD_IF_USED 6, %1
+    DEFINE_ARGS %4
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 6
+        sub rsp, (xmm_regs_used-6)*16+16
+        %assign stack_offset stack_offset+(xmm_regs_used-6)*16+16
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa [rsp + (%%i-6)*16+8], xmm %+ %%i
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %if xmm_regs_used > 6
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+8]
+        %endrep
+        add %1, (xmm_regs_used-6)*16+16
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign xmm_regs_used 0
+%endmacro
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    %if regs_used > 4
+        pop r5
+        pop r4
+    %endif
+    ret
+%endmacro
+
+%macro REP_RET 0
+    %if regs_used > 4 || xmm_regs_used > 6
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%elifdef ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0, rdi, edi, di,  dil, edi
+DECLARE_REG 1, rsi, esi, si,  sil, esi
+DECLARE_REG 2, rdx, edx, dx,  dl,  edx
+DECLARE_REG 3, rcx, ecx, cx,  cl,  ecx
+DECLARE_REG 4, r8,  r8d, r8w, r8b, r8d
+DECLARE_REG 5, r9,  r9d, r9w, r9b, r9d
+DECLARE_REG 6, rax, eax, ax,  al,  [rsp + stack_offset + 8]
+%define r7m [rsp + stack_offset + 16]
+%define r8m [rsp + stack_offset + 24]
+
+%macro LOAD_IF_USED 2 ; reg_id, number_of_args
+    %if %1 < %2
+        mov r%1, [rsp - 40 + %1*8]
+    %endif
+%endmacro
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    ASSERT %2 >= %1
+    ASSERT %2 <= 7
+    LOAD_IF_USED 6, %1
+    DEFINE_ARGS %4
+%endmacro
+
+%macro RET 0
+    ret
+%endmacro
+
+%macro REP_RET 0
+    rep ret
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+; Begin chromium edits
+%ifdef CHROMIUM
+; Change the order of registers so we can get the lower 8-bit or the 5th and 6th
+; arguments.
+DECLARE_REG 0, esi, esi, si, null, [esp + stack_offset + 4]
+DECLARE_REG 1, edi, edi, di, null, [esp + stack_offset + 8]
+DECLARE_REG 2, ecx, ecx, cx, cl,   [esp + stack_offset + 12]
+DECLARE_REG 3, edx, edx, dx, dl,   [esp + stack_offset + 16]
+DECLARE_REG 4, eax, eax, ax, al,   [esp + stack_offset + 20]
+DECLARE_REG 5, ebx, ebx, bx, bl,   [esp + stack_offset + 24]
+%else
+DECLARE_REG 0, eax, eax, ax, al,   [esp + stack_offset + 4]
+DECLARE_REG 1, ecx, ecx, cx, cl,   [esp + stack_offset + 8]
+DECLARE_REG 2, edx, edx, dx, dl,   [esp + stack_offset + 12]
+DECLARE_REG 3, ebx, ebx, bx, bl,   [esp + stack_offset + 16]
+DECLARE_REG 4, esi, esi, si, null, [esp + stack_offset + 20]
+DECLARE_REG 5, edi, edi, di, null, [esp + stack_offset + 24]
+%endif
+; End chromium edits
+DECLARE_REG 6, ebp, ebp, bp, null, [esp + stack_offset + 28]
+%define r7m [esp + stack_offset + 32]
+%define r8m [esp + stack_offset + 36]
+%define rsp esp
+
+%macro PUSH_IF_USED 1 ; reg_id
+    %if %1 < regs_used
+        push r%1
+        %assign stack_offset stack_offset+4
+    %endif
+%endmacro
+
+%macro POP_IF_USED 1 ; reg_id
+    %if %1 < regs_used
+        pop r%1
+    %endif
+%endmacro
+
+%macro LOAD_IF_USED 2 ; reg_id, number_of_args
+    %if %1 < %2
+        mov r%1, [esp + stack_offset + 4 + %1*4]
+    %endif
+%endmacro
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    ASSERT %2 >= %1
+    %assign regs_used %2
+    ASSERT regs_used <= 7
+%ifdef CHROMIUM
+    PUSH_IF_USED 0
+    PUSH_IF_USED 1
+    PUSH_IF_USED 5
+%else
+    PUSH_IF_USED 3
+    PUSH_IF_USED 4
+    PUSH_IF_USED 5
+%endif
+    PUSH_IF_USED 6
+    LOAD_IF_USED 0, %1
+    LOAD_IF_USED 1, %1
+    LOAD_IF_USED 2, %1
+    LOAD_IF_USED 3, %1
+    LOAD_IF_USED 4, %1
+    LOAD_IF_USED 5, %1
+    LOAD_IF_USED 6, %1
+    DEFINE_ARGS %4
+%endmacro
+
+%macro RET 0
+    POP_IF_USED 6
+%ifdef CHROMIUM
+    POP_IF_USED 5
+    POP_IF_USED 1
+    POP_IF_USED 0
+%else
+    POP_IF_USED 5
+    POP_IF_USED 4
+    POP_IF_USED 3
+%endif
+    ret
+%endmacro
+
+%macro REP_RET 0
+    %if regs_used > 3
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%endif ;======================================================================
+
+%ifndef WIN64
+%macro WIN64_SPILL_XMM 1
+%endmacro
+%macro WIN64_RESTORE_XMM 1
+%endmacro
+%endif
+
+
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Symbol prefix for C linkage
+%macro cglobal 1-2+
+    %xdefine %1 mangle(program_name %+ _ %+ %1)
+    %xdefine %1.skip_prologue %1 %+ .skip_prologue
+    %ifidn __OUTPUT_FORMAT__,elf
+        global %1:function hidden
+    %else
+        global %1
+    %endif
+    align function_align
+    %1:
+    RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %assign stack_offset 0
+    %if %0 > 1
+        PROLOGUE %2
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(program_name %+ _ %+ %1)
+    extern %1
+%endmacro
+
+;like cextern, but without the prefix
+%macro cextern_naked 1
+    %xdefine %1 mangle(%1)
+    extern %1
+%endmacro
+
+%macro const 2+
+    %xdefine %1 mangle(program_name %+ _ %+ %1)
+    global %1
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is
+; executable by default.
+%ifidn __OUTPUT_FORMAT__,elf
+SECTION .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+
+; merge mmx and sse*
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+    CAT_XDEFINE m, %%i, mm %+ %%i
+    CAT_XDEFINE nmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    %rep 8
+    CAT_UNDEF m, %%i
+    CAT_UNDEF nmm, %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro INIT_XMM 0
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM
+    %define mmsize 16
+    %define num_mmregs 8
+    %ifdef ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, xmm %+ %%i
+    CAT_XDEFINE nxmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro INIT_AVX 0
+    INIT_XMM
+    %assign avx_enabled 1
+    %define PALIGNR PALIGNR_SSSE3
+    %define RESET_MM_PERMUTATION INIT_AVX
+%endmacro
+
+%macro INIT_YMM 0
+    %assign avx_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM
+    %define mmsize 32
+    %define num_mmregs 8
+    %ifdef ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova vmovaps
+    %define movu vmovups
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, ymm %+ %%i
+    CAT_XDEFINE nymm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+INIT_MMX
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+%rep %0/2
+    %xdefine tmp%2 m%2
+    %xdefine ntmp%2 nm%2
+    %rotate 2
+%endrep
+%rep %0/2
+    %xdefine m%1 tmp%2
+    %xdefine nm%1 ntmp%2
+    %undef tmp%2
+    %undef ntmp%2
+    %rotate 2
+%endrep
+%endmacro
+
+%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
+%rep %0-1
+%ifdef m%1
+    %xdefine tmp m%1
+    %xdefine m%1 m%2
+    %xdefine m%2 tmp
+    CAT_XDEFINE n, m%1, %1
+    CAT_XDEFINE n, m%2, %2
+%else
+    ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
+    ; Be careful using this mode in nested macros though, as in some cases there may be
+    ; other copies of m# that have already been dereferenced and don't get updated correctly.
+    %xdefine %%n1 n %+ %1
+    %xdefine %%n2 n %+ %2
+    %xdefine tmp m %+ %%n1
+    CAT_XDEFINE m, %%n1, m %+ %%n2
+    CAT_XDEFINE m, %%n2, tmp
+    CAT_XDEFINE n, m %+ %%n1, %%n1
+    CAT_XDEFINE n, m %+ %%n2, %%n2
+%endif
+    %undef tmp
+    %rotate 1
+%endrep
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function and given the
+; function name, then any later calls to that function will automatically
+; load the permutation, so values can be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 1 ; name to save as
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE %1_m, %%i, m %+ %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, %1_m %+ %%i
+    CAT_XDEFINE n, m %+ %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro call 1
+    call %1
+    %ifdef %1_m0
+        LOAD_MM_PERMUTATION %1
+    %endif
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+%assign i i+1
+%endrep
+%undef i
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 0 if 3-operand (xmm, xmm, xmm), 1 if 4-operand (xmm, xmm, xmm, imm)
+;%4 == number of operands given
+;%5+: operands
+%macro RUN_AVX_INSTR 6-7+
+    %if sizeof%5==32
+        v%1 %5, %6, %7
+    %else
+        %if sizeof%5==8
+            %define %%regmov movq
+        %elif %2
+            %define %%regmov movaps
+        %else
+            %define %%regmov movdqa
+        %endif
+
+        %if %4>=3+%3
+            %ifnidn %5, %6
+                %if avx_enabled && sizeof%5==16
+                    v%1 %5, %6, %7
+                %else
+                    %%regmov %5, %6
+                    %1 %5, %7
+                %endif
+            %else
+                %1 %5, %7
+            %endif
+        %elif %3
+            %1 %5, %6, %7
+        %else
+            %1 %5, %6
+        %endif
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 0 if 3-operand (xmm, xmm, xmm), 1 if 4-operand (xmm, xmm, xmm, imm)
+%macro AVX_INSTR 3
+    %macro %1 2-8 fnord, fnord, fnord, %1, %2, %3
+        %ifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR %6, %7, %8, 3, %1, %2, %3
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+AVX_INSTR addpd, 1, 0
+AVX_INSTR addps, 1, 0
+AVX_INSTR addsd, 1, 0
+AVX_INSTR addss, 1, 0
+AVX_INSTR addsubpd, 1, 0
+AVX_INSTR addsubps, 1, 0
+AVX_INSTR andpd, 1, 0
+AVX_INSTR andps, 1, 0
+AVX_INSTR andnpd, 1, 0
+AVX_INSTR andnps, 1, 0
+AVX_INSTR blendpd, 1, 0
+AVX_INSTR blendps, 1, 0
+AVX_INSTR blendvpd, 1, 0
+AVX_INSTR blendvps, 1, 0
+AVX_INSTR cmppd, 1, 0
+AVX_INSTR cmpps, 1, 0
+AVX_INSTR cmpsd, 1, 0
+AVX_INSTR cmpss, 1, 0
+AVX_INSTR divpd, 1, 0
+AVX_INSTR divps, 1, 0
+AVX_INSTR divsd, 1, 0
+AVX_INSTR divss, 1, 0
+AVX_INSTR dppd, 1, 0
+AVX_INSTR dpps, 1, 0
+AVX_INSTR haddpd, 1, 0
+AVX_INSTR haddps, 1, 0
+AVX_INSTR hsubpd, 1, 0
+AVX_INSTR hsubps, 1, 0
+AVX_INSTR maxpd, 1, 0
+AVX_INSTR maxps, 1, 0
+AVX_INSTR maxsd, 1, 0
+AVX_INSTR maxss, 1, 0
+AVX_INSTR minpd, 1, 0
+AVX_INSTR minps, 1, 0
+AVX_INSTR minsd, 1, 0
+AVX_INSTR minss, 1, 0
+AVX_INSTR mpsadbw, 0, 1
+AVX_INSTR mulpd, 1, 0
+AVX_INSTR mulps, 1, 0
+AVX_INSTR mulsd, 1, 0
+AVX_INSTR mulss, 1, 0
+AVX_INSTR orpd, 1, 0
+AVX_INSTR orps, 1, 0
+AVX_INSTR packsswb, 0, 0
+AVX_INSTR packssdw, 0, 0
+AVX_INSTR packuswb, 0, 0
+AVX_INSTR packusdw, 0, 0
+AVX_INSTR paddb, 0, 0
+AVX_INSTR paddw, 0, 0
+AVX_INSTR paddd, 0, 0
+AVX_INSTR paddq, 0, 0
+AVX_INSTR paddsb, 0, 0
+AVX_INSTR paddsw, 0, 0
+AVX_INSTR paddusb, 0, 0
+AVX_INSTR paddusw, 0, 0
+AVX_INSTR palignr, 0, 1
+AVX_INSTR pand, 0, 0
+AVX_INSTR pandn, 0, 0
+AVX_INSTR pavgb, 0, 0
+AVX_INSTR pavgw, 0, 0
+AVX_INSTR pblendvb, 0, 0
+AVX_INSTR pblendw, 0, 1
+AVX_INSTR pcmpestri, 0, 0
+AVX_INSTR pcmpestrm, 0, 0
+AVX_INSTR pcmpistri, 0, 0
+AVX_INSTR pcmpistrm, 0, 0
+AVX_INSTR pcmpeqb, 0, 0
+AVX_INSTR pcmpeqw, 0, 0
+AVX_INSTR pcmpeqd, 0, 0
+AVX_INSTR pcmpeqq, 0, 0
+AVX_INSTR pcmpgtb, 0, 0
+AVX_INSTR pcmpgtw, 0, 0
+AVX_INSTR pcmpgtd, 0, 0
+AVX_INSTR pcmpgtq, 0, 0
+AVX_INSTR phaddw, 0, 0
+AVX_INSTR phaddd, 0, 0
+AVX_INSTR phaddsw, 0, 0
+AVX_INSTR phsubw, 0, 0
+AVX_INSTR phsubd, 0, 0
+AVX_INSTR phsubsw, 0, 0
+AVX_INSTR pmaddwd, 0, 0
+AVX_INSTR pmaddubsw, 0, 0
+AVX_INSTR pmaxsb, 0, 0
+AVX_INSTR pmaxsw, 0, 0
+AVX_INSTR pmaxsd, 0, 0
+AVX_INSTR pmaxub, 0, 0
+AVX_INSTR pmaxuw, 0, 0
+AVX_INSTR pmaxud, 0, 0
+AVX_INSTR pminsb, 0, 0
+AVX_INSTR pminsw, 0, 0
+AVX_INSTR pminsd, 0, 0
+AVX_INSTR pminub, 0, 0
+AVX_INSTR pminuw, 0, 0
+AVX_INSTR pminud, 0, 0
+AVX_INSTR pmulhuw, 0, 0
+AVX_INSTR pmulhrsw, 0, 0
+AVX_INSTR pmulhw, 0, 0
+AVX_INSTR pmullw, 0, 0
+AVX_INSTR pmulld, 0, 0
+AVX_INSTR pmuludq, 0, 0
+AVX_INSTR pmuldq, 0, 0
+AVX_INSTR por, 0, 0
+AVX_INSTR psadbw, 0, 0
+AVX_INSTR pshufb, 0, 0
+AVX_INSTR psignb, 0, 0
+AVX_INSTR psignw, 0, 0
+AVX_INSTR psignd, 0, 0
+AVX_INSTR psllw, 0, 0
+AVX_INSTR pslld, 0, 0
+AVX_INSTR psllq, 0, 0
+AVX_INSTR pslldq, 0, 0
+AVX_INSTR psraw, 0, 0
+AVX_INSTR psrad, 0, 0
+AVX_INSTR psrlw, 0, 0
+AVX_INSTR psrld, 0, 0
+AVX_INSTR psrlq, 0, 0
+AVX_INSTR psrldq, 0, 0
+AVX_INSTR psubb, 0, 0
+AVX_INSTR psubw, 0, 0
+AVX_INSTR psubd, 0, 0
+AVX_INSTR psubq, 0, 0
+AVX_INSTR psubsb, 0, 0
+AVX_INSTR psubsw, 0, 0
+AVX_INSTR psubusb, 0, 0
+AVX_INSTR psubusw, 0, 0
+AVX_INSTR punpckhbw, 0, 0
+AVX_INSTR punpckhwd, 0, 0
+AVX_INSTR punpckhdq, 0, 0
+AVX_INSTR punpckhqdq, 0, 0
+AVX_INSTR punpcklbw, 0, 0
+AVX_INSTR punpcklwd, 0, 0
+AVX_INSTR punpckldq, 0, 0
+AVX_INSTR punpcklqdq, 0, 0
+AVX_INSTR pxor, 0, 0
+AVX_INSTR shufps, 0, 1
+AVX_INSTR subpd, 1, 0
+AVX_INSTR subps, 1, 0
+AVX_INSTR subsd, 1, 0
+AVX_INSTR subss, 1, 0
+AVX_INSTR unpckhpd, 1, 0
+AVX_INSTR unpckhps, 1, 0
+AVX_INSTR unpcklpd, 1, 0
+AVX_INSTR unpcklps, 1, 0
+AVX_INSTR xorpd, 1, 0
+AVX_INSTR xorps, 1, 0
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 1, 0
+AVX_INSTR pfsub, 1, 0
+AVX_INSTR pfmul, 1, 0
+
+;=============================================================================
+; Chromium extensions
+;=============================================================================
+
+%ifdef CHROMIUM
+;
+; LOAD_SYM %1 (reg), %2 (sym)
+; Copies the address to a local symbol to the specified register.
+;
+
+%macro LOAD_SYM 2
+
+%ifdef MACHO
+  call      %%geteip
+  add       %1, %2 - $
+  jmp       %%end
+%%geteip:
+  mov       %1, [rsp]
+  ret
+%%end:
+
+%else
+  lea       %1, [%2]
+%endif
+
+%endmacro
+
+;
+; MOVq %1 (xmm), %2 (reg)
+; MOVq %1 (reg), %2 (xmm)
+; Copies a general-purpose register to an XMM register, and vice versa.
+;
+%macro MOVq 2
+%if gprsize == 8
+  movq      %1, %2
+%else
+  movd      %1, %2
+%endif
+%endmacro
+
+;
+; PRIVATE
+; A flag representing the specified symbol is a private symbol. This define adds
+; a hidden flag on Linux and a private_extern flag on Mac. (We can use this
+; private_extern flag only on the latest yasm.)
+;
+%ifdef MACHO
+%define PRIVATE :private_extern
+%elifdef ELF
+%define PRIVATE :hidden
+%else
+%define PRIVATE
+%endif
+
+%endif  ; CHROMIUM
+
+%endif  ; MEDIA_BASE_SIMD_X86INC_ASM_
diff --git a/media/base/yuv_convert.cc b/media/base/yuv_convert.cc
index d332065..3a27ca4 100644
--- a/media/base/yuv_convert.cc
+++ b/media/base/yuv_convert.cc
@@ -19,6 +19,7 @@
 
 #include "build/build_config.h"
 #include "media/base/cpu_features.h"
+#include "media/base/simd/convert_rgb_to_yuv.h"
 #include "media/base/yuv_convert_internal.h"
 #include "media/base/yuv_row.h"
 
@@ -369,8 +370,10 @@ void ConvertRGB32ToYUV(const uint8* rgbframe,
     // TODO(hclam): Implement a NEON version.
     convert_proc = &ConvertRGB32ToYUV_C;
 #else
-    // For x86 processors, check if SSE2 is supported.
-    if (hasSSE2())
+    // For x86 processors, check if SSSE3 (or SSE2) is supported.
+    if (hasSSSE3())
+      convert_proc = &ConvertRGB32ToYUV_SSSE3;
+    else if (hasSSE2())
       convert_proc = &ConvertRGB32ToYUV_SSE2;
     else
       convert_proc = &ConvertRGB32ToYUV_C;
@@ -390,8 +393,21 @@ void ConvertRGB24ToYUV(const uint8* rgbframe,
                        int rgbstride,
                        int ystride,
                        int uvstride) {
+#if defined(ARCH_CPU_ARM_FAMILY)
   ConvertRGB24ToYUV_C(rgbframe, yplane, uplane, vplane, width, height,
                       rgbstride, ystride, uvstride);
+#else
+  static void (*convert_proc)(const uint8*, uint8*, uint8*, uint8*,
+                              int, int, int, int, int) = NULL;
+  if (!convert_proc) {
+    if (hasSSSE3())
+      convert_proc = &ConvertRGB24ToYUV_SSSE3;
+    else
+      convert_proc = &ConvertRGB24ToYUV_C;
+  }
+  convert_proc(rgbframe, yplane, uplane, vplane, width, height,
+               rgbstride, ystride, uvstride);
+#endif
 }
 
 void ConvertYUY2ToYUV(const uint8* src,
diff --git a/media/media.gyp b/media/media.gyp
index fcc9b53..dfe4bbf 100644
--- a/media/media.gyp
+++ b/media/media.gyp
@@ -380,9 +380,63 @@
              },
           },
         }],
+        [ 'target_arch=="ia32" or target_arch=="x64"', {
+          'sources': [
+            'base/simd/convert_rgb_to_yuv_ssse3.asm',
+          ],
+        }],
+        [ 'OS=="win"', {
+          'variables': {
+            'yasm_flags': [
+              '-DWIN32',
+              '-DMSVC',
+              '-DCHROMIUM',
+              '-Isimd',
+            ],
+          },
+        }],
+        [ 'OS=="mac"', {
+          'variables': {
+            'yasm_flags': [
+              '-DPREFIX',
+              '-DMACHO',
+              '-DCHROMIUM',
+              '-Isimd',
+            ],
+          },
+        }],
+        [ 'OS=="linux"', {
+          'variables': {
+            'conditions': [
+              [ 'target_arch=="ia32"', {
+                'yasm_flags': [
+                  '-DX86_32',
+                  '-DELF',
+                  '-DCHROMIUM',
+                  '-Isimd',
+                ],
+              }, {
+                'yasm_flags': [
+                  '-DARCH_X86_64',
+                  '-DELF',
+                  '-DPIC',
+                  '-DCHROMIUM',
+                  '-Isimd',
+                ],
+              }],
+            ],
+          },
+        }],
       ],
       'sources': [
         'base/yuv_convert_sse2.cc',
+        'base/simd/convert_rgb_to_yuv.cc',
+      ],
+      'variables': {
+        'yasm_output_path': '<(SHARED_INTERMEDIATE_DIR)/media',
+      },
+      'includes': [
+        '../third_party/yasm/yasm_compile.gypi',
       ],
     },
     {
@@ -460,6 +514,7 @@
         'base/pts_stream_unittest.cc',
         'base/run_all_unittests.cc',
         'base/seekable_buffer_unittest.cc',
+        'base/simd/convert_rgb_to_yuv_unittest.cc',
         'base/state_matrix_unittest.cc',
         'base/test_data_util.cc',
         'base/test_data_util.h',