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*,
  44                                   const uint8*,
  45                                   const uint8*,
  46                                   int,
  47                                   uint8);
  48
  49 typedef void (*ConvertRGBToYUVProc)(const uint8*,
  50                                     uint8*,
  51                                     uint8*,
  52                                     uint8*,
  53                                     int,
  54                                     int,
  55                                     int,
  56                                     int,
  57                                     int);
  58
  59 typedef void (*ConvertYUVToRGB32Proc)(const uint8*,
  60                                       const uint8*,
  61                                       const uint8*,
  62                                       uint8*,
  63                                       int,
  64                                       int,
  65                                       int,
  66                                       int,
  67                                       int,
  68                                       YUVType);
  69
  70 typedef void (*ConvertYUVAToARGBProc)(const uint8*,
  71                                       const uint8*,
  72                                       const uint8*,
  73                                       const uint8*,
  74                                       uint8*,
  75                                       int,
  76                                       int,
  77                                       int,
  78                                       int,
  79                                       int,
  80                                       int,
  81                                       YUVType);
  82
  83 typedef void (*ConvertYUVToRGB32RowProc)(const uint8*,
  84                                          const uint8*,
  85                                          const uint8*,
  86                                          uint8*,
  87                                          ptrdiff_t,
  88                                          const int16[1024][4]);
  89
  90 typedef void (*ConvertYUVAToARGBRowProc)(const uint8*,
  91                                          const uint8*,
  92                                          const uint8*,
  93                                          const uint8*,
  94                                          uint8*,
  95                                          ptrdiff_t,
  96                                          const int16[1024][4]);
  97
  98 typedef void (*ScaleYUVToRGB32RowProc)(const uint8*,
  99                                        const uint8*,
 100                                        const uint8*,
 101                                        uint8*,
 102                                        ptrdiff_t,
 103                                        ptrdiff_t,
 104                                        const int16[1024][4]);
 105
 106 static FilterYUVRowsProc g_filter_yuv_rows_proc_ = NULL;
 107 static ConvertYUVToRGB32RowProc g_convert_yuv_to_rgb32_row_proc_ = NULL;
 108 static ScaleYUVToRGB32RowProc g_scale_yuv_to_rgb32_row_proc_ = NULL;
 109 static ScaleYUVToRGB32RowProc g_linear_scale_yuv_to_rgb32_row_proc_ = NULL;
 110 static ConvertRGBToYUVProc g_convert_rgb32_to_yuv_proc_ = NULL;
 111 static ConvertRGBToYUVProc g_convert_rgb24_to_yuv_proc_ = NULL;
 112 static ConvertYUVToRGB32Proc g_convert_yuv_to_rgb32_proc_ = NULL;
 113 static ConvertYUVAToARGBProc g_convert_yuva_to_argb_proc_ = NULL;
 114
 115 // Empty SIMD registers state after using them.
 116 void EmptyRegisterStateStub() {}
 117 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 118 void EmptyRegisterStateIntrinsic() { _mm_empty(); }
 119 #endif
 120 typedef void (*EmptyRegisterStateProc)();
 121 static EmptyRegisterStateProc g_empty_register_state_proc_ = NULL;
 122
 123 // Get the appropriate value to bitshift by for vertical indices.
 124 int GetVerticalShift(YUVType type) {
 125   switch (type) {
 126     case YV16:
 127       return 0;
 128     case YV12:
 129     case YV12J:
 130       return 1;
 131   }
 132   NOTREACHED();
 133   return 0;
 134 }
 135
 136 const int16 (&GetLookupTable(YUVType type))[1024][4] {
 137   switch (type) {
 138     case YV12:
 139     case YV16:
 140       return kCoefficientsRgbY;
 141     case YV12J:
 142       return kCoefficientsRgbY_JPEG;
 143   }
 144   NOTREACHED();
 145   return kCoefficientsRgbY;
 146 }
 147
 148 void InitializeCPUSpecificYUVConversions() {
 149   CHECK(!g_filter_yuv_rows_proc_);
 150   CHECK(!g_convert_yuv_to_rgb32_row_proc_);
 151   CHECK(!g_scale_yuv_to_rgb32_row_proc_);
 152   CHECK(!g_linear_scale_yuv_to_rgb32_row_proc_);
 153   CHECK(!g_convert_rgb32_to_yuv_proc_);
 154   CHECK(!g_convert_rgb24_to_yuv_proc_);
 155   CHECK(!g_convert_yuv_to_rgb32_proc_);
 156   CHECK(!g_convert_yuva_to_argb_proc_);
 157   CHECK(!g_empty_register_state_proc_);
 158
 159   g_filter_yuv_rows_proc_ = FilterYUVRows_C;
 160   g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_C;
 161   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_C;
 162   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_C;
 163   g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_C;
 164   g_convert_rgb24_to_yuv_proc_ = ConvertRGB24ToYUV_C;
 165   g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_C;
 166   g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_C;
 167   g_empty_register_state_proc_ = EmptyRegisterStateStub;
 168
 169   // Assembly code confuses MemorySanitizer.
 170 #if defined(ARCH_CPU_X86_FAMILY) && !defined(MEMORY_SANITIZER)
 171   g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_MMX;
 172
 173 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 174   g_empty_register_state_proc_ = EmptyRegisterStateIntrinsic;
 175 #else
 176   g_empty_register_state_proc_ = EmptyRegisterState_MMX;
 177 #endif
 178
 179   g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_SSE;
 180   g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_SSE;
 181
 182   g_filter_yuv_rows_proc_ = FilterYUVRows_SSE2;
 183   g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_SSE2;
 184
 185 #if defined(ARCH_CPU_X86_64)
 186   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE2_X64;
 187
 188   // Technically this should be in the MMX section, but MSVC will optimize out
 189   // the export of LinearScaleYUVToRGB32Row_MMX, which is required by the unit
 190   // tests, if that decision can be made at compile time.  Since all X64 CPUs
 191   // have SSE2, we can hack around this by making the selection here.
 192   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_MMX_X64;
 193 #else
 194   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE;
 195   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_SSE;
 196 #endif
 197
 198   base::CPU cpu;
 199   if (cpu.has_ssse3()) {
 200     g_convert_rgb24_to_yuv_proc_ = &ConvertRGB24ToYUV_SSSE3;
 201
 202     // TODO(hclam): Add ConvertRGB32ToYUV_SSSE3 when the cyan problem is solved.
 203     // See: crbug.com/100462
 204   }
 205 #endif
 206 }
 207
 208 // Empty SIMD registers state after using them.
 209 void EmptyRegisterState() { g_empty_register_state_proc_(); }
 210
 211 // 16.16 fixed point arithmetic
 212 const int kFractionBits = 16;
 213 const int kFractionMax = 1 << kFractionBits;
 214 const int kFractionMask = ((1 << kFractionBits) - 1);
 215
 216 // Scale a frame of YUV to 32 bit ARGB.
 217 void ScaleYUVToRGB32(const uint8* y_buf,
 218                      const uint8* u_buf,
 219                      const uint8* v_buf,
 220                      uint8* rgb_buf,
 221                      int source_width,
 222                      int source_height,
 223                      int width,
 224                      int height,
 225                      int y_pitch,
 226                      int uv_pitch,
 227                      int rgb_pitch,
 228                      YUVType yuv_type,
 229                      Rotate view_rotate,
 230                      ScaleFilter filter) {
 231   // Handle zero sized sources and destinations.
 232   if ((yuv_type == YV12 && (source_width < 2 || source_height < 2)) ||
 233       (yuv_type == YV16 && (source_width < 2 || source_height < 1)) ||
 234       width == 0 || height == 0)
 235     return;
 236
 237   // 4096 allows 3 buffers to fit in 12k.
 238   // Helps performance on CPU with 16K L1 cache.
 239   // Large enough for 3830x2160 and 30" displays which are 2560x1600.
 240   const int kFilterBufferSize = 4096;
 241   // Disable filtering if the screen is too big (to avoid buffer overflows).
 242   // This should never happen to regular users: they don't have monitors
 243   // wider than 4096 pixels.
 244   // TODO(fbarchard): Allow rotated videos to filter.
 245   if (source_width > kFilterBufferSize || view_rotate)
 246     filter = FILTER_NONE;
 247
 248   unsigned int y_shift = GetVerticalShift(yuv_type);
 249   // Diagram showing origin and direction of source sampling.
 250   // ->0   4<-
 251   // 7       3
 252   //
 253   // 6       5
 254   // ->1   2<-
 255   // Rotations that start at right side of image.
 256   if ((view_rotate == ROTATE_180) || (view_rotate == ROTATE_270) ||
 257       (view_rotate == MIRROR_ROTATE_0) || (view_rotate == MIRROR_ROTATE_90)) {
 258     y_buf += source_width - 1;
 259     u_buf += source_width / 2 - 1;
 260     v_buf += source_width / 2 - 1;
 261     source_width = -source_width;
 262   }
 263   // Rotations that start at bottom of image.
 264   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_180) ||
 265       (view_rotate == MIRROR_ROTATE_90) || (view_rotate == MIRROR_ROTATE_180)) {
 266     y_buf += (source_height - 1) * y_pitch;
 267     u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 268     v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 269     source_height = -source_height;
 270   }
 271
 272   int source_dx = source_width * kFractionMax / width;
 273
 274   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_270)) {
 275     int tmp = height;
 276     height = width;
 277     width = tmp;
 278     tmp = source_height;
 279     source_height = source_width;
 280     source_width = tmp;
 281     int source_dy = source_height * kFractionMax / height;
 282     source_dx = ((source_dy >> kFractionBits) * y_pitch) << kFractionBits;
 283     if (view_rotate == ROTATE_90) {
 284       y_pitch = -1;
 285       uv_pitch = -1;
 286       source_height = -source_height;
 287     } else {
 288       y_pitch = 1;
 289       uv_pitch = 1;
 290     }
 291   }
 292
 293   // Need padding because FilterRows() will write 1 to 16 extra pixels
 294   // after the end for SSE2 version.
 295   uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
 296   uint8* ybuf =
 297       reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
 298   uint8* ubuf = ybuf + kFilterBufferSize;
 299   uint8* vbuf = ubuf + kFilterBufferSize;
 300
 301   // TODO(fbarchard): Fixed point math is off by 1 on negatives.
 302
 303   // We take a y-coordinate in [0,1] space in the source image space, and
 304   // transform to a y-coordinate in [0,1] space in the destination image space.
 305   // Note that the coordinate endpoints lie on pixel boundaries, not on pixel
 306   // centers: e.g. a two-pixel-high image will have pixel centers at 0.25 and
 307   // 0.75.  The formula is as follows (in fixed-point arithmetic):
 308   //   y_dst = dst_height * ((y_src + 0.5) / src_height)
 309   //   dst_pixel = clamp([0, dst_height - 1], floor(y_dst - 0.5))
 310   // Implement this here as an accumulator + delta, to avoid expensive math
 311   // in the loop.
 312   int source_y_subpixel_accum =
 313       ((kFractionMax / 2) * source_height) / height - (kFractionMax / 2);
 314   int source_y_subpixel_delta = ((1 << kFractionBits) * source_height) / height;
 315
 316   // TODO(fbarchard): Split this into separate function for better efficiency.
 317   for (int y = 0; y < height; ++y) {
 318     uint8* dest_pixel = rgb_buf + y * rgb_pitch;
 319     int source_y_subpixel = source_y_subpixel_accum;
 320     source_y_subpixel_accum += source_y_subpixel_delta;
 321     if (source_y_subpixel < 0)
 322       source_y_subpixel = 0;
 323     else if (source_y_subpixel > ((source_height - 1) << kFractionBits))
 324       source_y_subpixel = (source_height - 1) << kFractionBits;
 325
 326     const uint8* y_ptr = NULL;
 327     const uint8* u_ptr = NULL;
 328     const uint8* v_ptr = NULL;
 329     // Apply vertical filtering if necessary.
 330     // TODO(fbarchard): Remove memcpy when not necessary.
 331     if (filter & media::FILTER_BILINEAR_V) {
 332       int source_y = source_y_subpixel >> kFractionBits;
 333       y_ptr = y_buf + source_y * y_pitch;
 334       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 335       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 336
 337       // Vertical scaler uses 16.8 fixed point.
 338       uint8 source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
 339       if (source_y_fraction != 0) {
 340         g_filter_yuv_rows_proc_(
 341             ybuf, y_ptr, y_ptr + y_pitch, source_width, source_y_fraction);
 342       } else {
 343         memcpy(ybuf, y_ptr, source_width);
 344       }
 345       y_ptr = ybuf;
 346       ybuf[source_width] = ybuf[source_width - 1];
 347
 348       int uv_source_width = (source_width + 1) / 2;
 349       uint8 source_uv_fraction;
 350
 351       // For formats with half-height UV planes, each even-numbered pixel row
 352       // should not interpolate, since the next row to interpolate from should
 353       // be a duplicate of the current row.
 354       if (y_shift && (source_y & 0x1) == 0)
 355         source_uv_fraction = 0;
 356       else
 357         source_uv_fraction = source_y_fraction;
 358
 359       if (source_uv_fraction != 0) {
 360         g_filter_yuv_rows_proc_(
 361             ubuf, u_ptr, u_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 362         g_filter_yuv_rows_proc_(
 363             vbuf, v_ptr, v_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 364       } else {
 365         memcpy(ubuf, u_ptr, uv_source_width);
 366         memcpy(vbuf, v_ptr, uv_source_width);
 367       }
 368       u_ptr = ubuf;
 369       v_ptr = vbuf;
 370       ubuf[uv_source_width] = ubuf[uv_source_width - 1];
 371       vbuf[uv_source_width] = vbuf[uv_source_width - 1];
 372     } else {
 373       // Offset by 1/2 pixel for center sampling.
 374       int source_y = (source_y_subpixel + (kFractionMax / 2)) >> kFractionBits;
 375       y_ptr = y_buf + source_y * y_pitch;
 376       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 377       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 378     }
 379     if (source_dx == kFractionMax) {  // Not scaled
 380       g_convert_yuv_to_rgb32_row_proc_(
 381           y_ptr, u_ptr, v_ptr, dest_pixel, width, kCoefficientsRgbY);
 382     } else {
 383       if (filter & FILTER_BILINEAR_H) {
 384         g_linear_scale_yuv_to_rgb32_row_proc_(y_ptr,
 385                                               u_ptr,
 386                                               v_ptr,
 387                                               dest_pixel,
 388                                               width,
 389                                               source_dx,
 390                                               kCoefficientsRgbY);
 391       } else {
 392         g_scale_yuv_to_rgb32_row_proc_(y_ptr,
 393                                        u_ptr,
 394                                        v_ptr,
 395                                        dest_pixel,
 396                                        width,
 397                                        source_dx,
 398                                        kCoefficientsRgbY);
 399       }
 400     }
 401   }
 402
 403   g_empty_register_state_proc_();
 404 }
 405
 406 // Scale a frame of YV12 to 32 bit ARGB for a specific rectangle.
 407 void ScaleYUVToRGB32WithRect(const uint8* y_buf,
 408                              const uint8* u_buf,
 409                              const uint8* v_buf,
 410                              uint8* rgb_buf,
 411                              int source_width,
 412                              int source_height,
 413                              int dest_width,
 414                              int dest_height,
 415                              int dest_rect_left,
 416                              int dest_rect_top,
 417                              int dest_rect_right,
 418                              int dest_rect_bottom,
 419                              int y_pitch,
 420                              int uv_pitch,
 421                              int rgb_pitch) {
 422   // This routine doesn't currently support up-scaling.
 423   CHECK_LE(dest_width, source_width);
 424   CHECK_LE(dest_height, source_height);
 425
 426   // Sanity-check the destination rectangle.
 427   DCHECK(dest_rect_left >= 0 && dest_rect_right <= dest_width);
 428   DCHECK(dest_rect_top >= 0 && dest_rect_bottom <= dest_height);
 429   DCHECK(dest_rect_right > dest_rect_left);
 430   DCHECK(dest_rect_bottom > dest_rect_top);
 431
 432   // Fixed-point value of vertical and horizontal scale down factor.
 433   // Values are in the format 16.16.
 434   int y_step = kFractionMax * source_height / dest_height;
 435   int x_step = kFractionMax * source_width / dest_width;
 436
 437   // Determine the coordinates of the rectangle in 16.16 coords.
 438   // NB: Our origin is the *center* of the top/left pixel, NOT its top/left.
 439   // If we're down-scaling by more than a factor of two, we start with a 50%
 440   // fraction to avoid degenerating to point-sampling - we should really just
 441   // fix the fraction at 50% for all pixels in that case.
 442   int source_left = dest_rect_left * x_step;
 443   int source_right = (dest_rect_right - 1) * x_step;
 444   if (x_step < kFractionMax * 2) {
 445     source_left += ((x_step - kFractionMax) / 2);
 446     source_right += ((x_step - kFractionMax) / 2);
 447   } else {
 448     source_left += kFractionMax / 2;
 449     source_right += kFractionMax / 2;
 450   }
 451   int source_top = dest_rect_top * y_step;
 452   if (y_step < kFractionMax * 2) {
 453     source_top += ((y_step - kFractionMax) / 2);
 454   } else {
 455     source_top += kFractionMax / 2;
 456   }
 457
 458   // Determine the parts of the Y, U and V buffers to interpolate.
 459   int source_y_left = source_left >> kFractionBits;
 460   int source_y_right =
 461       std::min((source_right >> kFractionBits) + 2, source_width + 1);
 462
 463   int source_uv_left = source_y_left / 2;
 464   int source_uv_right = std::min((source_right >> (kFractionBits + 1)) + 2,
 465                                  (source_width + 1) / 2);
 466
 467   int source_y_width = source_y_right - source_y_left;
 468   int source_uv_width = source_uv_right - source_uv_left;
 469
 470   // Determine number of pixels in each output row.
 471   int dest_rect_width = dest_rect_right - dest_rect_left;
 472
 473   // Intermediate buffer for vertical interpolation.
 474   // 4096 bytes allows 3 buffers to fit in 12k, which fits in a 16K L1 cache,
 475   // and is bigger than most users will generally need.
 476   // The buffer is 16-byte aligned and padded with 16 extra bytes; some of the
 477   // FilterYUVRowsProcs have alignment requirements, and the SSE version can
 478   // write up to 16 bytes past the end of the buffer.
 479   const int kFilterBufferSize = 4096;
 480   const bool kAvoidUsingOptimizedFilter = source_width > kFilterBufferSize;
 481   uint8 yuv_temp[16 + kFilterBufferSize * 3 + 16];
 482   // memset() yuv_temp to 0 to avoid bogus warnings when running on Valgrind.
 483   if (RunningOnValgrind())
 484     memset(yuv_temp, 0, sizeof(yuv_temp));
 485   uint8* y_temp = reinterpret_cast<uint8*>(
 486       reinterpret_cast<uintptr_t>(yuv_temp + 15) & ~15);
 487   uint8* u_temp = y_temp + kFilterBufferSize;
 488   uint8* v_temp = u_temp + kFilterBufferSize;
 489
 490   // Move to the top-left pixel of output.
 491   rgb_buf += dest_rect_top * rgb_pitch;
 492   rgb_buf += dest_rect_left * 4;
 493
 494   // For each destination row perform interpolation and color space
 495   // conversion to produce the output.
 496   for (int row = dest_rect_top; row < dest_rect_bottom; ++row) {
 497     // Round the fixed-point y position to get the current row.
 498     int source_row = source_top >> kFractionBits;
 499     int source_uv_row = source_row / 2;
 500     DCHECK(source_row < source_height);
 501
 502     // Locate the first row for each plane for interpolation.
 503     const uint8* y0_ptr = y_buf + y_pitch * source_row + source_y_left;
 504     const uint8* u0_ptr = u_buf + uv_pitch * source_uv_row + source_uv_left;
 505     const uint8* v0_ptr = v_buf + uv_pitch * source_uv_row + source_uv_left;
 506     const uint8* y1_ptr = NULL;
 507     const uint8* u1_ptr = NULL;
 508     const uint8* v1_ptr = NULL;
 509
 510     // Locate the second row for interpolation, being careful not to overrun.
 511     if (source_row + 1 >= source_height) {
 512       y1_ptr = y0_ptr;
 513     } else {
 514       y1_ptr = y0_ptr + y_pitch;
 515     }
 516     if (source_uv_row + 1 >= (source_height + 1) / 2) {
 517       u1_ptr = u0_ptr;
 518       v1_ptr = v0_ptr;
 519     } else {
 520       u1_ptr = u0_ptr + uv_pitch;
 521       v1_ptr = v0_ptr + uv_pitch;
 522     }
 523
 524     if (!kAvoidUsingOptimizedFilter) {
 525       // Vertical scaler uses 16.8 fixed point.
 526       uint8 fraction = (source_top & kFractionMask) >> 8;
 527       g_filter_yuv_rows_proc_(
 528           y_temp + source_y_left, y0_ptr, y1_ptr, source_y_width, fraction);
 529       g_filter_yuv_rows_proc_(
 530           u_temp + source_uv_left, u0_ptr, u1_ptr, source_uv_width, fraction);
 531       g_filter_yuv_rows_proc_(
 532           v_temp + source_uv_left, v0_ptr, v1_ptr, source_uv_width, fraction);
 533
 534       // Perform horizontal interpolation and color space conversion.
 535       // TODO(hclam): Use the MMX version after more testing.
 536       LinearScaleYUVToRGB32RowWithRange_C(y_temp,
 537                                           u_temp,
 538                                           v_temp,
 539                                           rgb_buf,
 540                                           dest_rect_width,
 541                                           source_left,
 542                                           x_step,
 543                                           kCoefficientsRgbY);
 544     } else {
 545       // If the frame is too large then we linear scale a single row.
 546       LinearScaleYUVToRGB32RowWithRange_C(y0_ptr,
 547                                           u0_ptr,
 548                                           v0_ptr,
 549                                           rgb_buf,
 550                                           dest_rect_width,
 551                                           source_left,
 552                                           x_step,
 553                                           kCoefficientsRgbY);
 554     }
 555
 556     // Advance vertically in the source and destination image.
 557     source_top += y_step;
 558     rgb_buf += rgb_pitch;
 559   }
 560
 561   g_empty_register_state_proc_();
 562 }
 563
 564 void ConvertRGB32ToYUV(const uint8* rgbframe,
 565                        uint8* yplane,
 566                        uint8* uplane,
 567                        uint8* vplane,
 568                        int width,
 569                        int height,
 570                        int rgbstride,
 571                        int ystride,
 572                        int uvstride) {
 573   g_convert_rgb32_to_yuv_proc_(rgbframe,
 574                                yplane,
 575                                uplane,
 576                                vplane,
 577                                width,
 578                                height,
 579                                rgbstride,
 580                                ystride,
 581                                uvstride);
 582 }
 583
 584 void ConvertRGB24ToYUV(const uint8* rgbframe,
 585                        uint8* yplane,
 586                        uint8* uplane,
 587                        uint8* vplane,
 588                        int width,
 589                        int height,
 590                        int rgbstride,
 591                        int ystride,
 592                        int uvstride) {
 593   g_convert_rgb24_to_yuv_proc_(rgbframe,
 594                                yplane,
 595                                uplane,
 596                                vplane,
 597                                width,
 598                                height,
 599                                rgbstride,
 600                                ystride,
 601                                uvstride);
 602 }
 603
 604 void ConvertYUY2ToYUV(const uint8* src,
 605                       uint8* yplane,
 606                       uint8* uplane,
 607                       uint8* vplane,
 608                       int width,
 609                       int height) {
 610   for (int i = 0; i < height / 2; ++i) {
 611     for (int j = 0; j < (width / 2); ++j) {
 612       yplane[0] = src[0];
 613       *uplane = src[1];
 614       yplane[1] = src[2];
 615       *vplane = src[3];
 616       src += 4;
 617       yplane += 2;
 618       uplane++;
 619       vplane++;
 620     }
 621     for (int j = 0; j < (width / 2); ++j) {
 622       yplane[0] = src[0];
 623       yplane[1] = src[2];
 624       src += 4;
 625       yplane += 2;
 626     }
 627   }
 628 }
 629
 630 void ConvertNV21ToYUV(const uint8* src,
 631                       uint8* yplane,
 632                       uint8* uplane,
 633                       uint8* vplane,
 634                       int width,
 635                       int height) {
 636   int y_plane_size = width * height;
 637   memcpy(yplane, src, y_plane_size);
 638
 639   src += y_plane_size;
 640   int u_plane_size = y_plane_size >> 2;
 641   for (int i = 0; i < u_plane_size; ++i) {
 642     *vplane++ = *src++;
 643     *uplane++ = *src++;
 644   }
 645 }
 646
 647 void ConvertYUVToRGB32(const uint8* yplane,
 648                        const uint8* uplane,
 649                        const uint8* vplane,
 650                        uint8* rgbframe,
 651                        int width,
 652                        int height,
 653                        int ystride,
 654                        int uvstride,
 655                        int rgbstride,
 656                        YUVType yuv_type) {
 657   g_convert_yuv_to_rgb32_proc_(yplane,
 658                                uplane,
 659                                vplane,
 660                                rgbframe,
 661                                width,
 662                                height,
 663                                ystride,
 664                                uvstride,
 665                                rgbstride,
 666                                yuv_type);
 667 }
 668
 669 void ConvertYUVAToARGB(const uint8* yplane,
 670                        const uint8* uplane,
 671                        const uint8* vplane,
 672                        const uint8* aplane,
 673                        uint8* rgbframe,
 674                        int width,
 675                        int height,
 676                        int ystride,
 677                        int uvstride,
 678                        int astride,
 679                        int rgbstride,
 680                        YUVType yuv_type) {
 681   g_convert_yuva_to_argb_proc_(yplane,
 682                                uplane,
 683                                vplane,
 684                                aplane,
 685                                rgbframe,
 686                                width,
 687                                height,
 688                                ystride,
 689                                uvstride,
 690                                astride,
 691                                rgbstride,
 692                                yuv_type);
 693 }
 694
 695 }  // namespace media