MMX optimized YUV conversion in GCC and Visual C.

Unittest updated with 640x360 image converted offline with same code.

Review URL: http://codereview.chromium.org/43003

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@12292 0039d316-1c4b-4281-b951-d872f2087c98

2 files changed, 528 insertions, 64 deletions

diff --git a/media/base/yuv_convert.cc b/media/base/yuv_convert.cc
index 5b61bf0..4b044d4 100644
--- a/media/base/yuv_convert.cc
+++ b/media/base/yuv_convert.cc
@@ -27,28 +27,504 @@
 // An article on optimizing YUV conversion using tables instead of multiplies
 //   http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf
 //
-// Implimentation note
-//   To improve performance, the 298 x C subexpression is 'hoisted'.
-//   The shift by 8 is done on the subexpression, saving 2 shifts per pixel,
-//   but introducing a half value error. (0.4% error)
+// Implimentation notes
+//   This version uses MMX for Visual C and GCC, which should cover all
+//   current platforms.  C++ is included for reference and future platforms.
+//
 //   ARGB pixel format is assumed, which on little endian is stored as BGRA.
+//   The alpha is filled in, allowing the application to use RGBA or RGB32.
+//   The row based conversion allows for a future YV16 version, and simplifies
+//   the platform specific portion of the code.
+//
+//  The Visual C assembler is considered the source.
+//  The GCC asm was created by compiling with Visual C and disassembling
+//  with GNU objdump.
+//    cl /c /Ox yuv_convert.cc
+//    objdump -d yuv_convert.o
+//  The code almost copy/pasted in, except the table lookups, which produced
+//    movq   0x800(,%eax,8),%mm0
+//  and needed to be changed to cdecl style table names
+//   "movq   _coefficients_RGB_U(,%eax,8),%mm0\n"
+//  extern "C" was used to avoid name mangling.
+//
+//  Once compiled with both MinGW GCC and Visual C on PC, performance should
+//    be identical.  A small difference will occur in the C++ calling code,
+//    depending on the frame size.
+//  To confirm the same code is being generated
+//    g++ -O3 -c yuv_convert.cc
+//    dumpbin -disasm yuv_convert.o >gcc.txt
+//    cl /Ox /c yuv_convert.cc
+//    dumpbin -disasm yuv_convert.obj >vc.txt
+//    and compare the files.
+//
+//  The GCC function label is inside the assembler to avoid a stack frame
+//    push ebp, that may vary depending on compile options.
+
+#include "media/base/yuv_convert.h"
 
 #ifdef _OPENMP
 #include <omp.h>
 #endif
 
-// DCHECK affects performance in release build.
-// OFFICIAL_BUILD is okay, but this ifdef makes DCHECK only work in debug.
-#ifndef NDEBUG
+#ifdef _DEBUG
 #include "base/logging.h"
 #else
 #define DCHECK(a)
 #endif
 
-#include "media/base/yuv_convert.h"
+// MMX for Windows, Linux and Mac.
+// C++ code provided as a fall back.
+// Compile with /DUSE_MMX=0
+#ifndef USE_MMX
+#define USE_MMX 1
+#endif
 
 namespace media {
 
+#if USE_MMX
+
+#define RGBY(i) { \
+  static_cast<int16>(1.164 * 64 * (i - 16) + 0.5), \
+  static_cast<int16>(1.164 * 64 * (i - 16) + 0.5), \
+  static_cast<int16>(1.164 * 64 * (i - 16) + 0.5), \
+  0 \
+}
+
+#define RGBU(i) { \
+  static_cast<int16>(2.018 * 64 * (i - 128) + 0.5), \
+  static_cast<int16>(-0.391 * 64 * (i - 128) + 0.5), \
+  0, \
+  static_cast<int16>(256 * 64 - 1) \
+}
+
+#define RGBV(i) { \
+  0, \
+  static_cast<int16>(-0.813 * 64 * (i - 128) + 0.5), \
+  static_cast<int16>(1.596 * 64 * (i - 128) + 0.5), \
+  0 \
+}
+
+#if defined(_MSC_VER)
+#define MMX_ALIGNED(var) __declspec(align(16)) var
+#else
+#define MMX_ALIGNED(var) var __attribute__((aligned(16)))
+#endif
+
+extern "C" {
+MMX_ALIGNED(int16 coefficients_RGB_Y[256][4]) = {
+  RGBY(0x00), RGBY(0x01), RGBY(0x02), RGBY(0x03),
+  RGBY(0x04), RGBY(0x05), RGBY(0x06), RGBY(0x07),
+  RGBY(0x08), RGBY(0x09), RGBY(0x0A), RGBY(0x0B),
+  RGBY(0x0C), RGBY(0x0D), RGBY(0x0E), RGBY(0x0F),
+  RGBY(0x10), RGBY(0x11), RGBY(0x12), RGBY(0x13),
+  RGBY(0x14), RGBY(0x15), RGBY(0x16), RGBY(0x17),
+  RGBY(0x18), RGBY(0x19), RGBY(0x1A), RGBY(0x1B),
+  RGBY(0x1C), RGBY(0x1D), RGBY(0x1E), RGBY(0x1F),
+  RGBY(0x20), RGBY(0x21), RGBY(0x22), RGBY(0x23),
+  RGBY(0x24), RGBY(0x25), RGBY(0x26), RGBY(0x27),
+  RGBY(0x28), RGBY(0x29), RGBY(0x2A), RGBY(0x2B),
+  RGBY(0x2C), RGBY(0x2D), RGBY(0x2E), RGBY(0x2F),
+  RGBY(0x30), RGBY(0x31), RGBY(0x32), RGBY(0x33),
+  RGBY(0x34), RGBY(0x35), RGBY(0x36), RGBY(0x37),
+  RGBY(0x38), RGBY(0x39), RGBY(0x3A), RGBY(0x3B),
+  RGBY(0x3C), RGBY(0x3D), RGBY(0x3E), RGBY(0x3F),
+  RGBY(0x40), RGBY(0x41), RGBY(0x42), RGBY(0x43),
+  RGBY(0x44), RGBY(0x45), RGBY(0x46), RGBY(0x47),
+  RGBY(0x48), RGBY(0x49), RGBY(0x4A), RGBY(0x4B),
+  RGBY(0x4C), RGBY(0x4D), RGBY(0x4E), RGBY(0x4F),
+  RGBY(0x50), RGBY(0x51), RGBY(0x52), RGBY(0x53),
+  RGBY(0x54), RGBY(0x55), RGBY(0x56), RGBY(0x57),
+  RGBY(0x58), RGBY(0x59), RGBY(0x5A), RGBY(0x5B),
+  RGBY(0x5C), RGBY(0x5D), RGBY(0x5E), RGBY(0x5F),
+  RGBY(0x60), RGBY(0x61), RGBY(0x62), RGBY(0x63),
+  RGBY(0x64), RGBY(0x65), RGBY(0x66), RGBY(0x67),
+  RGBY(0x68), RGBY(0x69), RGBY(0x6A), RGBY(0x6B),
+  RGBY(0x6C), RGBY(0x6D), RGBY(0x6E), RGBY(0x6F),
+  RGBY(0x70), RGBY(0x71), RGBY(0x72), RGBY(0x73),
+  RGBY(0x74), RGBY(0x75), RGBY(0x76), RGBY(0x77),
+  RGBY(0x78), RGBY(0x79), RGBY(0x7A), RGBY(0x7B),
+  RGBY(0x7C), RGBY(0x7D), RGBY(0x7E), RGBY(0x7F),
+  RGBY(0x80), RGBY(0x81), RGBY(0x82), RGBY(0x83),
+  RGBY(0x84), RGBY(0x85), RGBY(0x86), RGBY(0x87),
+  RGBY(0x88), RGBY(0x89), RGBY(0x8A), RGBY(0x8B),
+  RGBY(0x8C), RGBY(0x8D), RGBY(0x8E), RGBY(0x8F),
+  RGBY(0x90), RGBY(0x91), RGBY(0x92), RGBY(0x93),
+  RGBY(0x94), RGBY(0x95), RGBY(0x96), RGBY(0x97),
+  RGBY(0x98), RGBY(0x99), RGBY(0x9A), RGBY(0x9B),
+  RGBY(0x9C), RGBY(0x9D), RGBY(0x9E), RGBY(0x9F),
+  RGBY(0xA0), RGBY(0xA1), RGBY(0xA2), RGBY(0xA3),
+  RGBY(0xA4), RGBY(0xA5), RGBY(0xA6), RGBY(0xA7),
+  RGBY(0xA8), RGBY(0xA9), RGBY(0xAA), RGBY(0xAB),
+  RGBY(0xAC), RGBY(0xAD), RGBY(0xAE), RGBY(0xAF),
+  RGBY(0xB0), RGBY(0xB1), RGBY(0xB2), RGBY(0xB3),
+  RGBY(0xB4), RGBY(0xB5), RGBY(0xB6), RGBY(0xB7),
+  RGBY(0xB8), RGBY(0xB9), RGBY(0xBA), RGBY(0xBB),
+  RGBY(0xBC), RGBY(0xBD), RGBY(0xBE), RGBY(0xBF),
+  RGBY(0xC0), RGBY(0xC1), RGBY(0xC2), RGBY(0xC3),
+  RGBY(0xC4), RGBY(0xC5), RGBY(0xC6), RGBY(0xC7),
+  RGBY(0xC8), RGBY(0xC9), RGBY(0xCA), RGBY(0xCB),
+  RGBY(0xCC), RGBY(0xCD), RGBY(0xCE), RGBY(0xCF),
+  RGBY(0xD0), RGBY(0xD1), RGBY(0xD2), RGBY(0xD3),
+  RGBY(0xD4), RGBY(0xD5), RGBY(0xD6), RGBY(0xD7),
+  RGBY(0xD8), RGBY(0xD9), RGBY(0xDA), RGBY(0xDB),
+  RGBY(0xDC), RGBY(0xDD), RGBY(0xDE), RGBY(0xDF),
+  RGBY(0xE0), RGBY(0xE1), RGBY(0xE2), RGBY(0xE3),
+  RGBY(0xE4), RGBY(0xE5), RGBY(0xE6), RGBY(0xE7),
+  RGBY(0xE8), RGBY(0xE9), RGBY(0xEA), RGBY(0xEB),
+  RGBY(0xEC), RGBY(0xED), RGBY(0xEE), RGBY(0xEF),
+  RGBY(0xF0), RGBY(0xF1), RGBY(0xF2), RGBY(0xF3),
+  RGBY(0xF4), RGBY(0xF5), RGBY(0xF6), RGBY(0xF7),
+  RGBY(0xF8), RGBY(0xF9), RGBY(0xFA), RGBY(0xFB),
+  RGBY(0xFC), RGBY(0xFD), RGBY(0xFE), RGBY(0xFF),
+};
+
+MMX_ALIGNED(int16 coefficients_RGB_U[256][4]) = {
+  RGBU(0x00), RGBU(0x01), RGBU(0x02), RGBU(0x03),
+  RGBU(0x04), RGBU(0x05), RGBU(0x06), RGBU(0x07),
+  RGBU(0x08), RGBU(0x09), RGBU(0x0A), RGBU(0x0B),
+  RGBU(0x0C), RGBU(0x0D), RGBU(0x0E), RGBU(0x0F),
+  RGBU(0x10), RGBU(0x11), RGBU(0x12), RGBU(0x13),
+  RGBU(0x14), RGBU(0x15), RGBU(0x16), RGBU(0x17),
+  RGBU(0x18), RGBU(0x19), RGBU(0x1A), RGBU(0x1B),
+  RGBU(0x1C), RGBU(0x1D), RGBU(0x1E), RGBU(0x1F),
+  RGBU(0x20), RGBU(0x21), RGBU(0x22), RGBU(0x23),
+  RGBU(0x24), RGBU(0x25), RGBU(0x26), RGBU(0x27),
+  RGBU(0x28), RGBU(0x29), RGBU(0x2A), RGBU(0x2B),
+  RGBU(0x2C), RGBU(0x2D), RGBU(0x2E), RGBU(0x2F),
+  RGBU(0x30), RGBU(0x31), RGBU(0x32), RGBU(0x33),
+  RGBU(0x34), RGBU(0x35), RGBU(0x36), RGBU(0x37),
+  RGBU(0x38), RGBU(0x39), RGBU(0x3A), RGBU(0x3B),
+  RGBU(0x3C), RGBU(0x3D), RGBU(0x3E), RGBU(0x3F),
+  RGBU(0x40), RGBU(0x41), RGBU(0x42), RGBU(0x43),
+  RGBU(0x44), RGBU(0x45), RGBU(0x46), RGBU(0x47),
+  RGBU(0x48), RGBU(0x49), RGBU(0x4A), RGBU(0x4B),
+  RGBU(0x4C), RGBU(0x4D), RGBU(0x4E), RGBU(0x4F),
+  RGBU(0x50), RGBU(0x51), RGBU(0x52), RGBU(0x53),
+  RGBU(0x54), RGBU(0x55), RGBU(0x56), RGBU(0x57),
+  RGBU(0x58), RGBU(0x59), RGBU(0x5A), RGBU(0x5B),
+  RGBU(0x5C), RGBU(0x5D), RGBU(0x5E), RGBU(0x5F),
+  RGBU(0x60), RGBU(0x61), RGBU(0x62), RGBU(0x63),
+  RGBU(0x64), RGBU(0x65), RGBU(0x66), RGBU(0x67),
+  RGBU(0x68), RGBU(0x69), RGBU(0x6A), RGBU(0x6B),
+  RGBU(0x6C), RGBU(0x6D), RGBU(0x6E), RGBU(0x6F),
+  RGBU(0x70), RGBU(0x71), RGBU(0x72), RGBU(0x73),
+  RGBU(0x74), RGBU(0x75), RGBU(0x76), RGBU(0x77),
+  RGBU(0x78), RGBU(0x79), RGBU(0x7A), RGBU(0x7B),
+  RGBU(0x7C), RGBU(0x7D), RGBU(0x7E), RGBU(0x7F),
+  RGBU(0x80), RGBU(0x81), RGBU(0x82), RGBU(0x83),
+  RGBU(0x84), RGBU(0x85), RGBU(0x86), RGBU(0x87),
+  RGBU(0x88), RGBU(0x89), RGBU(0x8A), RGBU(0x8B),
+  RGBU(0x8C), RGBU(0x8D), RGBU(0x8E), RGBU(0x8F),
+  RGBU(0x90), RGBU(0x91), RGBU(0x92), RGBU(0x93),
+  RGBU(0x94), RGBU(0x95), RGBU(0x96), RGBU(0x97),
+  RGBU(0x98), RGBU(0x99), RGBU(0x9A), RGBU(0x9B),
+  RGBU(0x9C), RGBU(0x9D), RGBU(0x9E), RGBU(0x9F),
+  RGBU(0xA0), RGBU(0xA1), RGBU(0xA2), RGBU(0xA3),
+  RGBU(0xA4), RGBU(0xA5), RGBU(0xA6), RGBU(0xA7),
+  RGBU(0xA8), RGBU(0xA9), RGBU(0xAA), RGBU(0xAB),
+  RGBU(0xAC), RGBU(0xAD), RGBU(0xAE), RGBU(0xAF),
+  RGBU(0xB0), RGBU(0xB1), RGBU(0xB2), RGBU(0xB3),
+  RGBU(0xB4), RGBU(0xB5), RGBU(0xB6), RGBU(0xB7),
+  RGBU(0xB8), RGBU(0xB9), RGBU(0xBA), RGBU(0xBB),
+  RGBU(0xBC), RGBU(0xBD), RGBU(0xBE), RGBU(0xBF),
+  RGBU(0xC0), RGBU(0xC1), RGBU(0xC2), RGBU(0xC3),
+  RGBU(0xC4), RGBU(0xC5), RGBU(0xC6), RGBU(0xC7),
+  RGBU(0xC8), RGBU(0xC9), RGBU(0xCA), RGBU(0xCB),
+  RGBU(0xCC), RGBU(0xCD), RGBU(0xCE), RGBU(0xCF),
+  RGBU(0xD0), RGBU(0xD1), RGBU(0xD2), RGBU(0xD3),
+  RGBU(0xD4), RGBU(0xD5), RGBU(0xD6), RGBU(0xD7),
+  RGBU(0xD8), RGBU(0xD9), RGBU(0xDA), RGBU(0xDB),
+  RGBU(0xDC), RGBU(0xDD), RGBU(0xDE), RGBU(0xDF),
+  RGBU(0xE0), RGBU(0xE1), RGBU(0xE2), RGBU(0xE3),
+  RGBU(0xE4), RGBU(0xE5), RGBU(0xE6), RGBU(0xE7),
+  RGBU(0xE8), RGBU(0xE9), RGBU(0xEA), RGBU(0xEB),
+  RGBU(0xEC), RGBU(0xED), RGBU(0xEE), RGBU(0xEF),
+  RGBU(0xF0), RGBU(0xF1), RGBU(0xF2), RGBU(0xF3),
+  RGBU(0xF4), RGBU(0xF5), RGBU(0xF6), RGBU(0xF7),
+  RGBU(0xF8), RGBU(0xF9), RGBU(0xFA), RGBU(0xFB),
+  RGBU(0xFC), RGBU(0xFD), RGBU(0xFE), RGBU(0xFF),
+};
+
+MMX_ALIGNED(int16 coefficients_RGB_V[256][4]) = {
+  RGBV(0x00), RGBV(0x01), RGBV(0x02), RGBV(0x03),
+  RGBV(0x04), RGBV(0x05), RGBV(0x06), RGBV(0x07),
+  RGBV(0x08), RGBV(0x09), RGBV(0x0A), RGBV(0x0B),
+  RGBV(0x0C), RGBV(0x0D), RGBV(0x0E), RGBV(0x0F),
+  RGBV(0x10), RGBV(0x11), RGBV(0x12), RGBV(0x13),
+  RGBV(0x14), RGBV(0x15), RGBV(0x16), RGBV(0x17),
+  RGBV(0x18), RGBV(0x19), RGBV(0x1A), RGBV(0x1B),
+  RGBV(0x1C), RGBV(0x1D), RGBV(0x1E), RGBV(0x1F),
+  RGBV(0x20), RGBV(0x21), RGBV(0x22), RGBV(0x23),
+  RGBV(0x24), RGBV(0x25), RGBV(0x26), RGBV(0x27),
+  RGBV(0x28), RGBV(0x29), RGBV(0x2A), RGBV(0x2B),
+  RGBV(0x2C), RGBV(0x2D), RGBV(0x2E), RGBV(0x2F),
+  RGBV(0x30), RGBV(0x31), RGBV(0x32), RGBV(0x33),
+  RGBV(0x34), RGBV(0x35), RGBV(0x36), RGBV(0x37),
+  RGBV(0x38), RGBV(0x39), RGBV(0x3A), RGBV(0x3B),
+  RGBV(0x3C), RGBV(0x3D), RGBV(0x3E), RGBV(0x3F),
+  RGBV(0x40), RGBV(0x41), RGBV(0x42), RGBV(0x43),
+  RGBV(0x44), RGBV(0x45), RGBV(0x46), RGBV(0x47),
+  RGBV(0x48), RGBV(0x49), RGBV(0x4A), RGBV(0x4B),
+  RGBV(0x4C), RGBV(0x4D), RGBV(0x4E), RGBV(0x4F),
+  RGBV(0x50), RGBV(0x51), RGBV(0x52), RGBV(0x53),
+  RGBV(0x54), RGBV(0x55), RGBV(0x56), RGBV(0x57),
+  RGBV(0x58), RGBV(0x59), RGBV(0x5A), RGBV(0x5B),
+  RGBV(0x5C), RGBV(0x5D), RGBV(0x5E), RGBV(0x5F),
+  RGBV(0x60), RGBV(0x61), RGBV(0x62), RGBV(0x63),
+  RGBV(0x64), RGBV(0x65), RGBV(0x66), RGBV(0x67),
+  RGBV(0x68), RGBV(0x69), RGBV(0x6A), RGBV(0x6B),
+  RGBV(0x6C), RGBV(0x6D), RGBV(0x6E), RGBV(0x6F),
+  RGBV(0x70), RGBV(0x71), RGBV(0x72), RGBV(0x73),
+  RGBV(0x74), RGBV(0x75), RGBV(0x76), RGBV(0x77),
+  RGBV(0x78), RGBV(0x79), RGBV(0x7A), RGBV(0x7B),
+  RGBV(0x7C), RGBV(0x7D), RGBV(0x7E), RGBV(0x7F),
+  RGBV(0x80), RGBV(0x81), RGBV(0x82), RGBV(0x83),
+  RGBV(0x84), RGBV(0x85), RGBV(0x86), RGBV(0x87),
+  RGBV(0x88), RGBV(0x89), RGBV(0x8A), RGBV(0x8B),
+  RGBV(0x8C), RGBV(0x8D), RGBV(0x8E), RGBV(0x8F),
+  RGBV(0x90), RGBV(0x91), RGBV(0x92), RGBV(0x93),
+  RGBV(0x94), RGBV(0x95), RGBV(0x96), RGBV(0x97),
+  RGBV(0x98), RGBV(0x99), RGBV(0x9A), RGBV(0x9B),
+  RGBV(0x9C), RGBV(0x9D), RGBV(0x9E), RGBV(0x9F),
+  RGBV(0xA0), RGBV(0xA1), RGBV(0xA2), RGBV(0xA3),
+  RGBV(0xA4), RGBV(0xA5), RGBV(0xA6), RGBV(0xA7),
+  RGBV(0xA8), RGBV(0xA9), RGBV(0xAA), RGBV(0xAB),
+  RGBV(0xAC), RGBV(0xAD), RGBV(0xAE), RGBV(0xAF),
+  RGBV(0xB0), RGBV(0xB1), RGBV(0xB2), RGBV(0xB3),
+  RGBV(0xB4), RGBV(0xB5), RGBV(0xB6), RGBV(0xB7),
+  RGBV(0xB8), RGBV(0xB9), RGBV(0xBA), RGBV(0xBB),
+  RGBV(0xBC), RGBV(0xBD), RGBV(0xBE), RGBV(0xBF),
+  RGBV(0xC0), RGBV(0xC1), RGBV(0xC2), RGBV(0xC3),
+  RGBV(0xC4), RGBV(0xC5), RGBV(0xC6), RGBV(0xC7),
+  RGBV(0xC8), RGBV(0xC9), RGBV(0xCA), RGBV(0xCB),
+  RGBV(0xCC), RGBV(0xCD), RGBV(0xCE), RGBV(0xCF),
+  RGBV(0xD0), RGBV(0xD1), RGBV(0xD2), RGBV(0xD3),
+  RGBV(0xD4), RGBV(0xD5), RGBV(0xD6), RGBV(0xD7),
+  RGBV(0xD8), RGBV(0xD9), RGBV(0xDA), RGBV(0xDB),
+  RGBV(0xDC), RGBV(0xDD), RGBV(0xDE), RGBV(0xDF),
+  RGBV(0xE0), RGBV(0xE1), RGBV(0xE2), RGBV(0xE3),
+  RGBV(0xE4), RGBV(0xE5), RGBV(0xE6), RGBV(0xE7),
+  RGBV(0xE8), RGBV(0xE9), RGBV(0xEA), RGBV(0xEB),
+  RGBV(0xEC), RGBV(0xED), RGBV(0xEE), RGBV(0xEF),
+  RGBV(0xF0), RGBV(0xF1), RGBV(0xF2), RGBV(0xF3),
+  RGBV(0xF4), RGBV(0xF5), RGBV(0xF6), RGBV(0xF7),
+  RGBV(0xF8), RGBV(0xF9), RGBV(0xFA), RGBV(0xFB),
+  RGBV(0xFC), RGBV(0xFD), RGBV(0xFE), RGBV(0xFF),
+};
+
+#undef RGBY
+#undef RGBU
+#undef RGBV
+#undef MMX_ALIGNED
+
+#if defined(_MSC_VER)
+
+// Warning C4799: function has no EMMS instruction.
+#pragma warning(disable: 4799)
+
+__declspec(naked)
+void ConvertYV12ToRGB32Row(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb_buf,
+                           size_t width) {
+  __asm {
+    pushad
+
+    mov       edx, [esp + 32 + 4]   // Y
+    mov       edi, [esp + 32 + 8]   // U
+    mov       esi, [esp + 32 + 12]  // V
+    mov       ebp, [esp + 32 + 16]  // rgb
+    mov       ecx, [esp + 32 + 20]  // width
+    shr       ecx, 1
+    xor       eax, eax
+    xor       ebx, ebx
+
+ wloop :
+    mov       al, [edi]
+    mov       bl, [esi]
+    movq      mm0, [coefficients_RGB_U + 8 * eax]
+    paddsw    mm0, [coefficients_RGB_V + 8 * ebx]
+
+    mov       al, [edx]
+    mov       bl, [edx + 1]
+    movq      mm1, [coefficients_RGB_Y + 8 * eax]
+    movq      mm2, [coefficients_RGB_Y + 8 * ebx]
+
+    paddsw    mm1, mm0
+    paddsw    mm2, mm0
+    psraw     mm1, 6
+    psraw     mm2, 6
+    packuswb  mm1, mm2
+
+    movntq    [ebp], mm1  // NOLINT
+
+    add       ebp, 8
+    add       edx, 2
+    add       edi, 1
+    add       esi, 1
+
+    sub       ecx, 1
+    jnz       wloop
+
+    popad
+    ret
+  }
+}
+
+
+#elif defined(__linux)
+
+void ConvertYV12ToRGB32Row(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb_buf,
+                           size_t width) __attribute__((noinline));
+
+  asm(
+"ConvertYV12ToRGB32Row:\n"
+  "pusha\n"
+  "mov    0x24(%esp),%edx\n"
+  "mov    0x28(%esp),%edi\n"
+  "mov    0x2c(%esp),%esi\n"
+  "mov    0x30(%esp),%ebp\n"
+  "mov    0x34(%esp),%ecx\n"
+  "shr    %ecx\n"
+  "xor    %eax,%eax\n"
+  "xor    %ebx,%ebx\n"
+"1:\n"
+  "mov    (%edi),%al\n"
+  "mov    (%esi),%bl\n"
+  "movq   coefficients_RGB_U(,%eax,8),%mm0\n"
+  "paddsw coefficients_RGB_V(,%ebx,8),%mm0\n"
+  "mov    (%edx),%al\n"
+  "mov    0x1(%edx),%bl\n"
+  "movq   coefficients_RGB_Y(,%eax,8),%mm1\n"
+  "movq   coefficients_RGB_Y(,%ebx,8),%mm2\n"
+  "paddsw %mm0,%mm1\n"
+  "paddsw %mm0,%mm2\n"
+  "psraw  $0x6,%mm1\n"
+  "psraw  $0x6,%mm2\n"
+  "packuswb %mm2,%mm1\n"
+  "movntq %mm1,0x0(%ebp)\n"
+  "add    $0x8,%ebp\n"
+  "add    $0x2,%edx\n"
+  "add    $0x1,%edi\n"
+  "add    $0x1,%esi\n"
+  "sub    $0x1,%ecx\n"
+  "jne    1b\n"
+  "popa\n"
+  "ret\n"
+  "nop\n"  // pad function to 0x70 bytes
+  "nop\n"
+  "nop\n"
+);
+
+#else
+
+
+void ConvertYV12ToRGB32Row(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb_buf,
+                           size_t width) __attribute__((noinline));
+
+  asm(
+"_ConvertYV12ToRGB32Row:\n"
+  "pusha\n"
+  "mov    0x24(%esp),%edx\n"
+  "mov    0x28(%esp),%edi\n"
+  "mov    0x2c(%esp),%esi\n"
+  "mov    0x30(%esp),%ebp\n"
+  "mov    0x34(%esp),%ecx\n"
+  "shr    %ecx\n"
+  "xor    %eax,%eax\n"
+  "xor    %ebx,%ebx\n"
+"1:\n"
+  "mov    (%edi),%al\n"
+  "mov    (%esi),%bl\n"
+  "movq   _coefficients_RGB_U(,%eax,8),%mm0\n"
+  "paddsw _coefficients_RGB_V(,%ebx,8),%mm0\n"
+  "mov    (%edx),%al\n"
+  "mov    0x1(%edx),%bl\n"
+  "movq   _coefficients_RGB_Y(,%eax,8),%mm1\n"
+  "movq   _coefficients_RGB_Y(,%ebx,8),%mm2\n"
+  "paddsw %mm0,%mm1\n"
+  "paddsw %mm0,%mm2\n"
+  "psraw  $0x6,%mm1\n"
+  "psraw  $0x6,%mm2\n"
+  "packuswb %mm2,%mm1\n"
+  "movntq %mm1,0x0(%ebp)\n"
+  "add    $0x8,%ebp\n"
+  "add    $0x2,%edx\n"
+  "add    $0x1,%edi\n"
+  "add    $0x1,%esi\n"
+  "sub    $0x1,%ecx\n"
+  "jne    1b\n"
+  "popa\n"
+  "ret\n"
+  "nop\n"  // pad function to 0x70 bytes
+  "nop\n"
+  "nop\n"
+);
+
+#endif  // MSC_VER
+}  // extern "C"
+
+#else  // USE_MMX
+
+void ConvertYV12ToRGB32Row(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb_buf,
+                           size_t width);
+#endif
+
+// Convert a frame of YUV to 32 bit ARGB.
+void ConvertYV12ToRGB32(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        size_t width,
+                        size_t height,
+                        int y_pitch,
+                        int uv_pitch,
+                        int rgb_pitch) {
+  // Image must be multiple of 2 in width.
+  DCHECK((width & 1) == 0);
+  // Check alignment. Use memalign to allocate the buffer if you hit this
+  // check:
+  DCHECK((reinterpret_cast<uintptr_t>(rgb_buf) & 7) == 0);
+#ifdef _OPENMP
+#pragma omp parallel for
+#endif
+  for (int y = 0; y < static_cast<int>(height); ++y) {
+    uint8* d1 = rgb_buf + y * rgb_pitch;
+    const uint8* y_ptr = y_buf + y * y_pitch;
+    const uint8* u_ptr = u_buf + y/2 * uv_pitch;
+    const uint8* v_ptr = v_buf + y/2 * uv_pitch;
+
+    ConvertYV12ToRGB32Row(y_ptr,
+                          u_ptr,
+                          v_ptr,
+                          d1,
+                          width);
+  }
+#if USE_MMX
+#if defined(_MSC_VER)
+  __asm emms;
+#else
+  asm("emms");
+#endif
+#endif
+}
+
+//------------------------------------------------------------------------------
+// This is pure C code
+
+#if !USE_MMX
+
+// Reference version of YUV converter.
 static const int kClipTableSize = 256;
 static const int kClipOverflow = 128;
 
@@ -133,56 +609,40 @@ static inline uint32 clip(int32 value) {
   return static_cast<uint32>(g_rgb_clip_table[((value) >> 8) + kClipOverflow]);
 }
 
-// Convert a frame of YUV to 32 bit ARGB.
-void ConvertYV12ToRGB32(const uint8* yplane,
-                        const uint8* uplane,
-                        const uint8* vplane,
-                        uint8* rgbframe,
-                        size_t frame_width,
-                        size_t frame_height,
-                        int ystride,
-                        int uvstride,
-                        int rgbstride) {
-  // Image must be multiple of 2 in both dimensions.
-  DCHECK((frame_width & 1) == 0);
-  DCHECK((frame_height & 1) == 0);
-  // OpenMP note - Index in OpenMP must have signed integral type.
-#ifdef _OPENMP
-#pragma omp parallel for
-#endif
-  for (int32 y = 0; y < static_cast<int32>(frame_height); ++y) {
-    const uint8* yrow = yplane + ystride  * y;
-    const uint8* urow = uplane + uvstride * (y >> 1);
-    const uint8* vrow = vplane + uvstride * (y >> 1);
-    uint8* rgbrow  = rgbframe + rgbstride * y;
-
-    for (int32 x = 0; x < static_cast<int32>(frame_width); x += 2) {
-      uint8 u = urow[x >> 1];
-      uint8 v = vrow[x >> 1];
-      int32 D = static_cast<int32>(u) - 128;
-      int32 E = static_cast<int32>(v) - 128;
-
-      int32 Cb =   (516 * D + 128);
-      int32 Cg = (- 100 * D - 208 * E + 128);
-      int32 Cr =             (409 * E + 128);
-
-      uint8 y0 = yrow[x];
-      int32 C298a = ((static_cast<int32>(y0) - 16) * 298 + 128);
-      *reinterpret_cast<uint32*>(rgbrow) = clip(C298a + Cb)
-                                        | (clip(C298a + Cg) << 8)
-                                        | (clip(C298a + Cr) << 16)
-                                        | 0xff000000;
-
-      uint8 y1 = yrow[x + 1];
-      int32 C298b = ((static_cast<int32>(y1) - 16) * 298 + 128);
-      *reinterpret_cast<uint32*>(rgbrow + 4) = clip(C298b + Cb)
-                                            | (clip(C298b + Cg) << 8)
-                                            | (clip(C298b + Cr) << 16)
-                                            | 0xff000000;
-
-      rgbrow += 8;  // Advance 2 pixels.
-    }
+void ConvertYV12ToRGB32Row(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb_buf,
+                           size_t width) {
+  for (int32 x = 0; x < static_cast<int32>(width); x += 2) {
+    uint8 u = u_buf[x >> 1];
+    uint8 v = v_buf[x >> 1];
+    int32 D = static_cast<int32>(u) - 128;
+    int32 E = static_cast<int32>(v) - 128;
+
+    int32 Cb =   (516 * D + 128);
+    int32 Cg = (- 100 * D - 208 * E + 128);
+    int32 Cr =             (409 * E + 128);
+
+    uint8 y0 = y_buf[x];
+    int32 C298a = ((static_cast<int32>(y0) - 16) * 298 + 128);
+    *reinterpret_cast<uint32*>(rgb_buf) = clip(C298a + Cb)
+                                       | (clip(C298a + Cg) << 8)
+                                       | (clip(C298a + Cr) << 16)
+                                       | 0xff000000;
+
+    uint8 y1 = y_buf[x + 1];
+    int32 C298b = ((static_cast<int32>(y1) - 16) * 298 + 128);
+    *reinterpret_cast<uint32*>(rgb_buf + 4) = clip(C298b + Cb)
+                                           | (clip(C298b + Cg) << 8)
+                                           | (clip(C298b + Cr) << 16)
+                                           | 0xff000000;
+
+    rgb_buf += 8;  // Advance 2 pixels.
   }
 }
+#endif
+
 
 }  // namespace media
+
diff --git a/media/base/yuv_convert_unittest.cc b/media/base/yuv_convert_unittest.cc
index 59ceb14..cc63538 100644
--- a/media/base/yuv_convert_unittest.cc
+++ b/media/base/yuv_convert_unittest.cc
@@ -7,9 +7,15 @@
 #include "media/base/yuv_convert.h"
 #include "testing/gtest/include/gtest/gtest.h"
 
+// Reference images were created with the following steps
+// ffmpeg -vframes 25 -i bali.mov -vcodec rawvideo -pix_fmt yuv420p -an
+//   bali.1280_720.yuv
+// yuvhalf -skip 24 bali.1280_720.yuv bali.640_360.yuv
+// yuvtool bali.640_360.yuv bali.640_360.rgb
+
 // Size of raw image.
-static const int kWidth = 1280;
-static const int kHeight = 720;
+static const int kWidth = 640;
+static const int kHeight = 360;
 static const int kBpp = 4;
 
 TEST(YuvConvertTest, Basic) {
@@ -19,7 +25,7 @@ TEST(YuvConvertTest, Basic) {
   yuv_url = yuv_url.Append(FILE_PATH_LITERAL("media"))
                    .Append(FILE_PATH_LITERAL("test"))
                    .Append(FILE_PATH_LITERAL("data"))
-                   .Append(FILE_PATH_LITERAL("yuv_file"));
+                   .Append(FILE_PATH_LITERAL("bali.640_360.yuv"));
   const size_t size_of_yuv = kWidth * kHeight * 12 / 8;
   uint8* yuv_bytes = new uint8[size_of_yuv];
   EXPECT_EQ(static_cast<int>(size_of_yuv),
@@ -28,15 +34,13 @@ TEST(YuvConvertTest, Basic) {
                                 static_cast<int>(size_of_yuv)));
 
   // Read RGB reference data from file.
-  // To keep the file smaller, only the first quarter of the file
-  // is checked into SVN.
   FilePath rgb_url;
   EXPECT_TRUE(PathService::Get(base::DIR_SOURCE_ROOT, &rgb_url));
   rgb_url = rgb_url.Append(FILE_PATH_LITERAL("media"))
                    .Append(FILE_PATH_LITERAL("test"))
                    .Append(FILE_PATH_LITERAL("data"))
-                   .Append(FILE_PATH_LITERAL("rgb_file"));
-  const size_t size_of_rgb = kWidth * kHeight * kBpp / 4;
+                   .Append(FILE_PATH_LITERAL("bali.640_360.rgb"));
+  const size_t size_of_rgb = kWidth * kHeight * kBpp;
   uint8* rgb_bytes = new uint8[size_of_rgb];
   EXPECT_EQ(static_cast<int>(size_of_rgb),
             file_util::ReadFile(rgb_url,