src/media/base/yuv_convert.cc

   1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 // This webpage shows layout of YV12 and other YUV formats
   6 // http://www.fourcc.org/yuv.php
   7 // The actual conversion is best described here
   8 // http://en.wikipedia.org/wiki/YUV
   9 // An article on optimizing YUV conversion using tables instead of multiplies
  10 // http://lestourtereaux.free.fr/papers/data/yuvrgb.pdf
  11 //
  12 // YV12 is a full plane of Y and a half height, half width chroma planes
  13 // YV16 is a full plane of Y and a full height, half width chroma planes
  14 //
  15 // ARGB pixel format is output, which on little endian is stored as BGRA.
  16 // The alpha is set to 255, allowing the application to use RGBA or RGB32.
  17
  18 #include "media/base/yuv_convert.h"
  19
  20 #include "base/cpu.h"
  21 #include "base/logging.h"
  22 #include "base/memory/scoped_ptr.h"
  23 #include "base/third_party/dynamic_annotations/dynamic_annotations.h"
  24 #include "build/build_config.h"
  25 #include "media/base/simd/convert_rgb_to_yuv.h"
  26 #include "media/base/simd/convert_yuv_to_rgb.h"
  27 #include "media/base/simd/filter_yuv.h"
  28 #include "media/base/simd/yuv_to_rgb_table.h"
  29
  30 #if defined(ARCH_CPU_X86_FAMILY)
  31 #if defined(COMPILER_MSVC)
  32 #include <intrin.h>
  33 #else
  34 #include <mmintrin.h>
  35 #endif
  36 #endif
  37
  38 // Assembly functions are declared without namespace.
  39 extern "C" { void EmptyRegisterState_MMX(); }  // extern "C"
  40
  41 namespace media {
  42
  43 typedef void (*FilterYUVRowsProc)(uint8*, const uint8*, const uint8*, int, int);
  44
  45 typedef void (*ConvertRGBToYUVProc)(const uint8*,
  46                                     uint8*,
  47                                     uint8*,
  48                                     uint8*,
  49                                     int,
  50                                     int,
  51                                     int,
  52                                     int,
  53                                     int);
  54
  55 typedef void (*ConvertYUVToRGB32Proc)(const uint8*,
  56                                       const uint8*,
  57                                       const uint8*,
  58                                       uint8*,
  59                                       int,
  60                                       int,
  61                                       int,
  62                                       int,
  63                                       int,
  64                                       YUVType);
  65
  66 typedef void (*ConvertYUVAToARGBProc)(const uint8*,
  67                                       const uint8*,
  68                                       const uint8*,
  69                                       const uint8*,
  70                                       uint8*,
  71                                       int,
  72                                       int,
  73                                       int,
  74                                       int,
  75                                       int,
  76                                       int,
  77                                       YUVType);
  78
  79 typedef void (*ConvertYUVToRGB32RowProc)(const uint8*,
  80                                          const uint8*,
  81                                          const uint8*,
  82                                          uint8*,
  83                                          ptrdiff_t,
  84                                          const int16[1024][4]);
  85
  86 typedef void (*ConvertYUVAToARGBRowProc)(const uint8*,
  87                                          const uint8*,
  88                                          const uint8*,
  89                                          const uint8*,
  90                                          uint8*,
  91                                          ptrdiff_t,
  92                                          const int16[1024][4]);
  93
  94 typedef void (*ScaleYUVToRGB32RowProc)(const uint8*,
  95                                        const uint8*,
  96                                        const uint8*,
  97                                        uint8*,
  98                                        ptrdiff_t,
  99                                        ptrdiff_t,
 100                                        const int16[1024][4]);
 101
 102 static FilterYUVRowsProc g_filter_yuv_rows_proc_ = NULL;
 103 static ConvertYUVToRGB32RowProc g_convert_yuv_to_rgb32_row_proc_ = NULL;
 104 static ScaleYUVToRGB32RowProc g_scale_yuv_to_rgb32_row_proc_ = NULL;
 105 static ScaleYUVToRGB32RowProc g_linear_scale_yuv_to_rgb32_row_proc_ = NULL;
 106 static ConvertRGBToYUVProc g_convert_rgb32_to_yuv_proc_ = NULL;
 107 static ConvertRGBToYUVProc g_convert_rgb24_to_yuv_proc_ = NULL;
 108 static ConvertYUVToRGB32Proc g_convert_yuv_to_rgb32_proc_ = NULL;
 109 static ConvertYUVAToARGBProc g_convert_yuva_to_argb_proc_ = NULL;
 110
 111 // Empty SIMD registers state after using them.
 112 void EmptyRegisterStateStub() {}
 113 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 114 void EmptyRegisterStateIntrinsic() { _mm_empty(); }
 115 #endif
 116 typedef void (*EmptyRegisterStateProc)();
 117 static EmptyRegisterStateProc g_empty_register_state_proc_ = NULL;
 118
 119 // Get the appropriate value to bitshift by for vertical indices.
 120 int GetVerticalShift(YUVType type) {
 121   switch (type) {
 122     case YV16:
 123       return 0;
 124     case YV12:
 125     case YV12J:
 126       return 1;
 127   }
 128   NOTREACHED();
 129   return 0;
 130 }
 131
 132 const int16 (&GetLookupTable(YUVType type))[1024][4] {
 133   switch (type) {
 134     case YV12:
 135     case YV16:
 136       return kCoefficientsRgbY;
 137     case YV12J:
 138       return kCoefficientsRgbY_JPEG;
 139   }
 140   NOTREACHED();
 141   return kCoefficientsRgbY;
 142 }
 143
 144 void InitializeCPUSpecificYUVConversions() {
 145   CHECK(!g_filter_yuv_rows_proc_);
 146   CHECK(!g_convert_yuv_to_rgb32_row_proc_);
 147   CHECK(!g_scale_yuv_to_rgb32_row_proc_);
 148   CHECK(!g_linear_scale_yuv_to_rgb32_row_proc_);
 149   CHECK(!g_convert_rgb32_to_yuv_proc_);
 150   CHECK(!g_convert_rgb24_to_yuv_proc_);
 151   CHECK(!g_convert_yuv_to_rgb32_proc_);
 152   CHECK(!g_convert_yuva_to_argb_proc_);
 153   CHECK(!g_empty_register_state_proc_);
 154
 155   g_filter_yuv_rows_proc_ = FilterYUVRows_C;
 156   g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_C;
 157   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_C;
 158   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_C;
 159   g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_C;
 160   g_convert_rgb24_to_yuv_proc_ = ConvertRGB24ToYUV_C;
 161   g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_C;
 162   g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_C;
 163   g_empty_register_state_proc_ = EmptyRegisterStateStub;
 164
 165   // Assembly code confuses MemorySanitizer.
 166 #if defined(ARCH_CPU_X86_FAMILY) && !defined(MEMORY_SANITIZER)
 167   g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_MMX;
 168
 169 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 170   g_empty_register_state_proc_ = EmptyRegisterStateIntrinsic;
 171 #else
 172   g_empty_register_state_proc_ = EmptyRegisterState_MMX;
 173 #endif
 174
 175   g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_SSE;
 176   g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_SSE;
 177
 178   g_filter_yuv_rows_proc_ = FilterYUVRows_SSE2;
 179   g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_SSE2;
 180
 181 #if defined(ARCH_CPU_X86_64)
 182   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE2_X64;
 183
 184   // Technically this should be in the MMX section, but MSVC will optimize out
 185   // the export of LinearScaleYUVToRGB32Row_MMX, which is required by the unit
 186   // tests, if that decision can be made at compile time.  Since all X64 CPUs
 187   // have SSE2, we can hack around this by making the selection here.
 188   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_MMX_X64;
 189 #else
 190   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE;
 191   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_SSE;
 192 #endif
 193
 194   base::CPU cpu;
 195   if (cpu.has_ssse3()) {
 196     g_convert_rgb24_to_yuv_proc_ = &ConvertRGB24ToYUV_SSSE3;
 197
 198     // TODO(hclam): Add ConvertRGB32ToYUV_SSSE3 when the cyan problem is solved.
 199     // See: crbug.com/100462
 200   }
 201 #endif
 202 }
 203
 204 // Empty SIMD registers state after using them.
 205 void EmptyRegisterState() { g_empty_register_state_proc_(); }
 206
 207 // 16.16 fixed point arithmetic
 208 const int kFractionBits = 16;
 209 const int kFractionMax = 1 << kFractionBits;
 210 const int kFractionMask = ((1 << kFractionBits) - 1);
 211
 212 // Scale a frame of YUV to 32 bit ARGB.
 213 void ScaleYUVToRGB32(const uint8* y_buf,
 214                      const uint8* u_buf,
 215                      const uint8* v_buf,
 216                      uint8* rgb_buf,
 217                      int source_width,
 218                      int source_height,
 219                      int width,
 220                      int height,
 221                      int y_pitch,
 222                      int uv_pitch,
 223                      int rgb_pitch,
 224                      YUVType yuv_type,
 225                      Rotate view_rotate,
 226                      ScaleFilter filter) {
 227   // Handle zero sized sources and destinations.
 228   if ((yuv_type == YV12 && (source_width < 2 || source_height < 2)) ||
 229       (yuv_type == YV16 && (source_width < 2 || source_height < 1)) ||
 230       width == 0 || height == 0)
 231     return;
 232
 233   // 4096 allows 3 buffers to fit in 12k.
 234   // Helps performance on CPU with 16K L1 cache.
 235   // Large enough for 3830x2160 and 30" displays which are 2560x1600.
 236   const int kFilterBufferSize = 4096;
 237   // Disable filtering if the screen is too big (to avoid buffer overflows).
 238   // This should never happen to regular users: they don't have monitors
 239   // wider than 4096 pixels.
 240   // TODO(fbarchard): Allow rotated videos to filter.
 241   if (source_width > kFilterBufferSize || view_rotate)
 242     filter = FILTER_NONE;
 243
 244   unsigned int y_shift = GetVerticalShift(yuv_type);
 245   // Diagram showing origin and direction of source sampling.
 246   // ->0   4<-
 247   // 7       3
 248   //
 249   // 6       5
 250   // ->1   2<-
 251   // Rotations that start at right side of image.
 252   if ((view_rotate == ROTATE_180) || (view_rotate == ROTATE_270) ||
 253       (view_rotate == MIRROR_ROTATE_0) || (view_rotate == MIRROR_ROTATE_90)) {
 254     y_buf += source_width - 1;
 255     u_buf += source_width / 2 - 1;
 256     v_buf += source_width / 2 - 1;
 257     source_width = -source_width;
 258   }
 259   // Rotations that start at bottom of image.
 260   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_180) ||
 261       (view_rotate == MIRROR_ROTATE_90) || (view_rotate == MIRROR_ROTATE_180)) {
 262     y_buf += (source_height - 1) * y_pitch;
 263     u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 264     v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 265     source_height = -source_height;
 266   }
 267
 268   int source_dx = source_width * kFractionMax / width;
 269
 270   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_270)) {
 271     int tmp = height;
 272     height = width;
 273     width = tmp;
 274     tmp = source_height;
 275     source_height = source_width;
 276     source_width = tmp;
 277     int source_dy = source_height * kFractionMax / height;
 278     source_dx = ((source_dy >> kFractionBits) * y_pitch) << kFractionBits;
 279     if (view_rotate == ROTATE_90) {
 280       y_pitch = -1;
 281       uv_pitch = -1;
 282       source_height = -source_height;
 283     } else {
 284       y_pitch = 1;
 285       uv_pitch = 1;
 286     }
 287   }
 288
 289   // Need padding because FilterRows() will write 1 to 16 extra pixels
 290   // after the end for SSE2 version.
 291   uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
 292   uint8* ybuf =
 293       reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
 294   uint8* ubuf = ybuf + kFilterBufferSize;
 295   uint8* vbuf = ubuf + kFilterBufferSize;
 296
 297   // TODO(fbarchard): Fixed point math is off by 1 on negatives.
 298
 299   // We take a y-coordinate in [0,1] space in the source image space, and
 300   // transform to a y-coordinate in [0,1] space in the destination image space.
 301   // Note that the coordinate endpoints lie on pixel boundaries, not on pixel
 302   // centers: e.g. a two-pixel-high image will have pixel centers at 0.25 and
 303   // 0.75.  The formula is as follows (in fixed-point arithmetic):
 304   //   y_dst = dst_height * ((y_src + 0.5) / src_height)
 305   //   dst_pixel = clamp([0, dst_height - 1], floor(y_dst - 0.5))
 306   // Implement this here as an accumulator + delta, to avoid expensive math
 307   // in the loop.
 308   int source_y_subpixel_accum =
 309       ((kFractionMax / 2) * source_height) / height - (kFractionMax / 2);
 310   int source_y_subpixel_delta = ((1 << kFractionBits) * source_height) / height;
 311
 312   // TODO(fbarchard): Split this into separate function for better efficiency.
 313   for (int y = 0; y < height; ++y) {
 314     uint8* dest_pixel = rgb_buf + y * rgb_pitch;
 315     int source_y_subpixel = source_y_subpixel_accum;
 316     source_y_subpixel_accum += source_y_subpixel_delta;
 317     if (source_y_subpixel < 0)
 318       source_y_subpixel = 0;
 319     else if (source_y_subpixel > ((source_height - 1) << kFractionBits))
 320       source_y_subpixel = (source_height - 1) << kFractionBits;
 321
 322     const uint8* y_ptr = NULL;
 323     const uint8* u_ptr = NULL;
 324     const uint8* v_ptr = NULL;
 325     // Apply vertical filtering if necessary.
 326     // TODO(fbarchard): Remove memcpy when not necessary.
 327     if (filter & media::FILTER_BILINEAR_V) {
 328       int source_y = source_y_subpixel >> kFractionBits;
 329       y_ptr = y_buf + source_y * y_pitch;
 330       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 331       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 332
 333       // Vertical scaler uses 16.8 fixed point.
 334       int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
 335       if (source_y_fraction != 0) {
 336         g_filter_yuv_rows_proc_(
 337             ybuf, y_ptr, y_ptr + y_pitch, source_width, source_y_fraction);
 338       } else {
 339         memcpy(ybuf, y_ptr, source_width);
 340       }
 341       y_ptr = ybuf;
 342       ybuf[source_width] = ybuf[source_width - 1];
 343
 344       int uv_source_width = (source_width + 1) / 2;
 345       int source_uv_fraction;
 346
 347       // For formats with half-height UV planes, each even-numbered pixel row
 348       // should not interpolate, since the next row to interpolate from should
 349       // be a duplicate of the current row.
 350       if (y_shift && (source_y & 0x1) == 0)
 351         source_uv_fraction = 0;
 352       else
 353         source_uv_fraction = source_y_fraction;
 354
 355       if (source_uv_fraction != 0) {
 356         g_filter_yuv_rows_proc_(
 357             ubuf, u_ptr, u_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 358         g_filter_yuv_rows_proc_(
 359             vbuf, v_ptr, v_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 360       } else {
 361         memcpy(ubuf, u_ptr, uv_source_width);
 362         memcpy(vbuf, v_ptr, uv_source_width);
 363       }
 364       u_ptr = ubuf;
 365       v_ptr = vbuf;
 366       ubuf[uv_source_width] = ubuf[uv_source_width - 1];
 367       vbuf[uv_source_width] = vbuf[uv_source_width - 1];
 368     } else {
 369       // Offset by 1/2 pixel for center sampling.
 370       int source_y = (source_y_subpixel + (kFractionMax / 2)) >> kFractionBits;
 371       y_ptr = y_buf + source_y * y_pitch;
 372       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 373       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 374     }
 375     if (source_dx == kFractionMax) {  // Not scaled
 376       g_convert_yuv_to_rgb32_row_proc_(
 377           y_ptr, u_ptr, v_ptr, dest_pixel, width, kCoefficientsRgbY);
 378     } else {
 379       if (filter & FILTER_BILINEAR_H) {
 380         g_linear_scale_yuv_to_rgb32_row_proc_(y_ptr,
 381                                               u_ptr,
 382                                               v_ptr,
 383                                               dest_pixel,
 384                                               width,
 385                                               source_dx,
 386                                               kCoefficientsRgbY);
 387       } else {
 388         g_scale_yuv_to_rgb32_row_proc_(y_ptr,
 389                                        u_ptr,
 390                                        v_ptr,
 391                                        dest_pixel,
 392                                        width,
 393                                        source_dx,
 394                                        kCoefficientsRgbY);
 395       }
 396     }
 397   }
 398
 399   g_empty_register_state_proc_();
 400 }
 401
 402 // Scale a frame of YV12 to 32 bit ARGB for a specific rectangle.
 403 void ScaleYUVToRGB32WithRect(const uint8* y_buf,
 404                              const uint8* u_buf,
 405                              const uint8* v_buf,
 406                              uint8* rgb_buf,
 407                              int source_width,
 408                              int source_height,
 409                              int dest_width,
 410                              int dest_height,
 411                              int dest_rect_left,
 412                              int dest_rect_top,
 413                              int dest_rect_right,
 414                              int dest_rect_bottom,
 415                              int y_pitch,
 416                              int uv_pitch,
 417                              int rgb_pitch) {
 418   // This routine doesn't currently support up-scaling.
 419   CHECK_LE(dest_width, source_width);
 420   CHECK_LE(dest_height, source_height);
 421
 422   // Sanity-check the destination rectangle.
 423   DCHECK(dest_rect_left >= 0 && dest_rect_right <= dest_width);
 424   DCHECK(dest_rect_top >= 0 && dest_rect_bottom <= dest_height);
 425   DCHECK(dest_rect_right > dest_rect_left);
 426   DCHECK(dest_rect_bottom > dest_rect_top);
 427
 428   // Fixed-point value of vertical and horizontal scale down factor.
 429   // Values are in the format 16.16.
 430   int y_step = kFractionMax * source_height / dest_height;
 431   int x_step = kFractionMax * source_width / dest_width;
 432
 433   // Determine the coordinates of the rectangle in 16.16 coords.
 434   // NB: Our origin is the *center* of the top/left pixel, NOT its top/left.
 435   // If we're down-scaling by more than a factor of two, we start with a 50%
 436   // fraction to avoid degenerating to point-sampling - we should really just
 437   // fix the fraction at 50% for all pixels in that case.
 438   int source_left = dest_rect_left * x_step;
 439   int source_right = (dest_rect_right - 1) * x_step;
 440   if (x_step < kFractionMax * 2) {
 441     source_left += ((x_step - kFractionMax) / 2);
 442     source_right += ((x_step - kFractionMax) / 2);
 443   } else {
 444     source_left += kFractionMax / 2;
 445     source_right += kFractionMax / 2;
 446   }
 447   int source_top = dest_rect_top * y_step;
 448   if (y_step < kFractionMax * 2) {
 449     source_top += ((y_step - kFractionMax) / 2);
 450   } else {
 451     source_top += kFractionMax / 2;
 452   }
 453
 454   // Determine the parts of the Y, U and V buffers to interpolate.
 455   int source_y_left = source_left >> kFractionBits;
 456   int source_y_right =
 457       std::min((source_right >> kFractionBits) + 2, source_width + 1);
 458
 459   int source_uv_left = source_y_left / 2;
 460   int source_uv_right = std::min((source_right >> (kFractionBits + 1)) + 2,
 461                                  (source_width + 1) / 2);
 462
 463   int source_y_width = source_y_right - source_y_left;
 464   int source_uv_width = source_uv_right - source_uv_left;
 465
 466   // Determine number of pixels in each output row.
 467   int dest_rect_width = dest_rect_right - dest_rect_left;
 468
 469   // Intermediate buffer for vertical interpolation.
 470   // 4096 bytes allows 3 buffers to fit in 12k, which fits in a 16K L1 cache,
 471   // and is bigger than most users will generally need.
 472   // The buffer is 16-byte aligned and padded with 16 extra bytes; some of the
 473   // FilterYUVRowProcs have alignment requirements, and the SSE version can
 474   // write up to 16 bytes past the end of the buffer.
 475   const int kFilterBufferSize = 4096;
 476   const bool kAvoidUsingOptimizedFilter = source_width > kFilterBufferSize;
 477   uint8 yuv_temp[16 + kFilterBufferSize * 3 + 16];
 478   // memset() yuv_temp to 0 to avoid bogus warnings when running on Valgrind.
 479   if (RunningOnValgrind())
 480     memset(yuv_temp, 0, sizeof(yuv_temp));
 481   uint8* y_temp = reinterpret_cast<uint8*>(
 482       reinterpret_cast<uintptr_t>(yuv_temp + 15) & ~15);
 483   uint8* u_temp = y_temp + kFilterBufferSize;
 484   uint8* v_temp = u_temp + kFilterBufferSize;
 485
 486   // Move to the top-left pixel of output.
 487   rgb_buf += dest_rect_top * rgb_pitch;
 488   rgb_buf += dest_rect_left * 4;
 489
 490   // For each destination row perform interpolation and color space
 491   // conversion to produce the output.
 492   for (int row = dest_rect_top; row < dest_rect_bottom; ++row) {
 493     // Round the fixed-point y position to get the current row.
 494     int source_row = source_top >> kFractionBits;
 495     int source_uv_row = source_row / 2;
 496     DCHECK(source_row < source_height);
 497
 498     // Locate the first row for each plane for interpolation.
 499     const uint8* y0_ptr = y_buf + y_pitch * source_row + source_y_left;
 500     const uint8* u0_ptr = u_buf + uv_pitch * source_uv_row + source_uv_left;
 501     const uint8* v0_ptr = v_buf + uv_pitch * source_uv_row + source_uv_left;
 502     const uint8* y1_ptr = NULL;
 503     const uint8* u1_ptr = NULL;
 504     const uint8* v1_ptr = NULL;
 505
 506     // Locate the second row for interpolation, being careful not to overrun.
 507     if (source_row + 1 >= source_height) {
 508       y1_ptr = y0_ptr;
 509     } else {
 510       y1_ptr = y0_ptr + y_pitch;
 511     }
 512     if (source_uv_row + 1 >= (source_height + 1) / 2) {
 513       u1_ptr = u0_ptr;
 514       v1_ptr = v0_ptr;
 515     } else {
 516       u1_ptr = u0_ptr + uv_pitch;
 517       v1_ptr = v0_ptr + uv_pitch;
 518     }
 519
 520     if (!kAvoidUsingOptimizedFilter) {
 521       // Vertical scaler uses 16.8 fixed point.
 522       int fraction = (source_top & kFractionMask) >> 8;
 523       g_filter_yuv_rows_proc_(
 524           y_temp + source_y_left, y0_ptr, y1_ptr, source_y_width, fraction);
 525       g_filter_yuv_rows_proc_(
 526           u_temp + source_uv_left, u0_ptr, u1_ptr, source_uv_width, fraction);
 527       g_filter_yuv_rows_proc_(
 528           v_temp + source_uv_left, v0_ptr, v1_ptr, source_uv_width, fraction);
 529
 530       // Perform horizontal interpolation and color space conversion.
 531       // TODO(hclam): Use the MMX version after more testing.
 532       LinearScaleYUVToRGB32RowWithRange_C(y_temp,
 533                                           u_temp,
 534                                           v_temp,
 535                                           rgb_buf,
 536                                           dest_rect_width,
 537                                           source_left,
 538                                           x_step,
 539                                           kCoefficientsRgbY);
 540     } else {
 541       // If the frame is too large then we linear scale a single row.
 542       LinearScaleYUVToRGB32RowWithRange_C(y0_ptr,
 543                                           u0_ptr,
 544                                           v0_ptr,
 545                                           rgb_buf,
 546                                           dest_rect_width,
 547                                           source_left,
 548                                           x_step,
 549                                           kCoefficientsRgbY);
 550     }
 551
 552     // Advance vertically in the source and destination image.
 553     source_top += y_step;
 554     rgb_buf += rgb_pitch;
 555   }
 556
 557   g_empty_register_state_proc_();
 558 }
 559
 560 void ConvertRGB32ToYUV(const uint8* rgbframe,
 561                        uint8* yplane,
 562                        uint8* uplane,
 563                        uint8* vplane,
 564                        int width,
 565                        int height,
 566                        int rgbstride,
 567                        int ystride,
 568                        int uvstride) {
 569   g_convert_rgb32_to_yuv_proc_(rgbframe,
 570                                yplane,
 571                                uplane,
 572                                vplane,
 573                                width,
 574                                height,
 575                                rgbstride,
 576                                ystride,
 577                                uvstride);
 578 }
 579
 580 void ConvertRGB24ToYUV(const uint8* rgbframe,
 581                        uint8* yplane,
 582                        uint8* uplane,
 583                        uint8* vplane,
 584                        int width,
 585                        int height,
 586                        int rgbstride,
 587                        int ystride,
 588                        int uvstride) {
 589   g_convert_rgb24_to_yuv_proc_(rgbframe,
 590                                yplane,
 591                                uplane,
 592                                vplane,
 593                                width,
 594                                height,
 595                                rgbstride,
 596                                ystride,
 597                                uvstride);
 598 }
 599
 600 void ConvertYUY2ToYUV(const uint8* src,
 601                       uint8* yplane,
 602                       uint8* uplane,
 603                       uint8* vplane,
 604                       int width,
 605                       int height) {
 606   for (int i = 0; i < height / 2; ++i) {
 607     for (int j = 0; j < (width / 2); ++j) {
 608       yplane[0] = src[0];
 609       *uplane = src[1];
 610       yplane[1] = src[2];
 611       *vplane = src[3];
 612       src += 4;
 613       yplane += 2;
 614       uplane++;
 615       vplane++;
 616     }
 617     for (int j = 0; j < (width / 2); ++j) {
 618       yplane[0] = src[0];
 619       yplane[1] = src[2];
 620       src += 4;
 621       yplane += 2;
 622     }
 623   }
 624 }
 625
 626 void ConvertNV21ToYUV(const uint8* src,
 627                       uint8* yplane,
 628                       uint8* uplane,
 629                       uint8* vplane,
 630                       int width,
 631                       int height) {
 632   int y_plane_size = width * height;
 633   memcpy(yplane, src, y_plane_size);
 634
 635   src += y_plane_size;
 636   int u_plane_size = y_plane_size >> 2;
 637   for (int i = 0; i < u_plane_size; ++i) {
 638     *vplane++ = *src++;
 639     *uplane++ = *src++;
 640   }
 641 }
 642
 643 void ConvertYUVToRGB32(const uint8* yplane,
 644                        const uint8* uplane,
 645                        const uint8* vplane,
 646                        uint8* rgbframe,
 647                        int width,
 648                        int height,
 649                        int ystride,
 650                        int uvstride,
 651                        int rgbstride,
 652                        YUVType yuv_type) {
 653   g_convert_yuv_to_rgb32_proc_(yplane,
 654                                uplane,
 655                                vplane,
 656                                rgbframe,
 657                                width,
 658                                height,
 659                                ystride,
 660                                uvstride,
 661                                rgbstride,
 662                                yuv_type);
 663 }
 664
 665 void ConvertYUVAToARGB(const uint8* yplane,
 666                        const uint8* uplane,
 667                        const uint8* vplane,
 668                        const uint8* aplane,
 669                        uint8* rgbframe,
 670                        int width,
 671                        int height,
 672                        int ystride,
 673                        int uvstride,
 674                        int astride,
 675                        int rgbstride,
 676                        YUVType yuv_type) {
 677   g_convert_yuva_to_argb_proc_(yplane,
 678                                uplane,
 679                                vplane,
 680                                aplane,
 681                                rgbframe,
 682                                width,
 683                                height,
 684                                ystride,
 685                                uvstride,
 686                                astride,
 687                                rgbstride,
 688                                yuv_type);
 689 }
 690
 691 }  // namespace media