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
  29 #if defined(ARCH_CPU_X86_FAMILY)
  30 #if defined(COMPILER_MSVC)
  31 #include <intrin.h>
  32 #else
  33 #include <mmintrin.h>
  34 #endif
  35 #endif
  36
  37 // Assembly functions are declared without namespace.
  38 extern "C" { void EmptyRegisterState_MMX(); }  // extern "C"
  39
  40 namespace media {
  41
  42 typedef void (*FilterYUVRowsProc)(uint8*, const uint8*, const uint8*, int, int);
  43
  44 typedef void (*ConvertRGBToYUVProc)(const uint8*,
  45                                     uint8*,
  46                                     uint8*,
  47                                     uint8*,
  48                                     int,
  49                                     int,
  50                                     int,
  51                                     int,
  52                                     int);
  53
  54 typedef void (*ConvertYUVToRGB32Proc)(const uint8*,
  55                                       const uint8*,
  56                                       const uint8*,
  57                                       uint8*,
  58                                       int,
  59                                       int,
  60                                       int,
  61                                       int,
  62                                       int,
  63                                       YUVType);
  64
  65 typedef void (*ConvertYUVAToARGBProc)(const uint8*,
  66                                       const uint8*,
  67                                       const uint8*,
  68                                       const uint8*,
  69                                       uint8*,
  70                                       int,
  71                                       int,
  72                                       int,
  73                                       int,
  74                                       int,
  75                                       int,
  76                                       YUVType);
  77
  78 typedef void (*ConvertYUVToRGB32RowProc)(const uint8*,
  79                                          const uint8*,
  80                                          const uint8*,
  81                                          uint8*,
  82                                          ptrdiff_t);
  83
  84 typedef void (*ConvertYUVAToARGBRowProc)(const uint8*,
  85                                          const uint8*,
  86                                          const uint8*,
  87                                          const uint8*,
  88                                          uint8*,
  89                                          ptrdiff_t);
  90
  91 typedef void (*ScaleYUVToRGB32RowProc)(const uint8*,
  92                                        const uint8*,
  93                                        const uint8*,
  94                                        uint8*,
  95                                        ptrdiff_t,
  96                                        ptrdiff_t);
  97
  98 static FilterYUVRowsProc g_filter_yuv_rows_proc_ = NULL;
  99 static ConvertYUVToRGB32RowProc g_convert_yuv_to_rgb32_row_proc_ = NULL;
 100 static ScaleYUVToRGB32RowProc g_scale_yuv_to_rgb32_row_proc_ = NULL;
 101 static ScaleYUVToRGB32RowProc g_linear_scale_yuv_to_rgb32_row_proc_ = NULL;
 102 static ConvertRGBToYUVProc g_convert_rgb32_to_yuv_proc_ = NULL;
 103 static ConvertRGBToYUVProc g_convert_rgb24_to_yuv_proc_ = NULL;
 104 static ConvertYUVToRGB32Proc g_convert_yuv_to_rgb32_proc_ = NULL;
 105 static ConvertYUVAToARGBProc g_convert_yuva_to_argb_proc_ = NULL;
 106
 107 // Empty SIMD registers state after using them.
 108 void EmptyRegisterStateStub() {}
 109 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 110 void EmptyRegisterStateIntrinsic() { _mm_empty(); }
 111 #endif
 112 typedef void (*EmptyRegisterStateProc)();
 113 static EmptyRegisterStateProc g_empty_register_state_proc_ = NULL;
 114
 115 void InitializeCPUSpecificYUVConversions() {
 116   CHECK(!g_filter_yuv_rows_proc_);
 117   CHECK(!g_convert_yuv_to_rgb32_row_proc_);
 118   CHECK(!g_scale_yuv_to_rgb32_row_proc_);
 119   CHECK(!g_linear_scale_yuv_to_rgb32_row_proc_);
 120   CHECK(!g_convert_rgb32_to_yuv_proc_);
 121   CHECK(!g_convert_rgb24_to_yuv_proc_);
 122   CHECK(!g_convert_yuv_to_rgb32_proc_);
 123   CHECK(!g_convert_yuva_to_argb_proc_);
 124   CHECK(!g_empty_register_state_proc_);
 125
 126   g_filter_yuv_rows_proc_ = FilterYUVRows_C;
 127   g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_C;
 128   g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_C;
 129   g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_C;
 130   g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_C;
 131   g_convert_rgb24_to_yuv_proc_ = ConvertRGB24ToYUV_C;
 132   g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_C;
 133   g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_C;
 134   g_empty_register_state_proc_ = EmptyRegisterStateStub;
 135
 136   // Assembly code confuses MemorySanitizer.
 137 #if defined(ARCH_CPU_X86_FAMILY) && !defined(MEMORY_SANITIZER)
 138   base::CPU cpu;
 139   if (cpu.has_mmx()) {
 140     g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_MMX;
 141     g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_MMX;
 142     g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_MMX;
 143     g_convert_yuva_to_argb_proc_ = ConvertYUVAToARGB_MMX;
 144     g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_MMX;
 145
 146 #if defined(MEDIA_MMX_INTRINSICS_AVAILABLE)
 147     g_filter_yuv_rows_proc_ = FilterYUVRows_MMX;
 148     g_empty_register_state_proc_ = EmptyRegisterStateIntrinsic;
 149 #else
 150     g_empty_register_state_proc_ = EmptyRegisterState_MMX;
 151 #endif
 152   }
 153
 154   if (cpu.has_sse()) {
 155     g_convert_yuv_to_rgb32_row_proc_ = ConvertYUVToRGB32Row_SSE;
 156     g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE;
 157     g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_SSE;
 158     g_convert_yuv_to_rgb32_proc_ = ConvertYUVToRGB32_SSE;
 159   }
 160
 161   if (cpu.has_sse2()) {
 162     g_filter_yuv_rows_proc_ = FilterYUVRows_SSE2;
 163     g_convert_rgb32_to_yuv_proc_ = ConvertRGB32ToYUV_SSE2;
 164
 165 #if defined(ARCH_CPU_X86_64)
 166     g_scale_yuv_to_rgb32_row_proc_ = ScaleYUVToRGB32Row_SSE2_X64;
 167
 168     // Technically this should be in the MMX section, but MSVC will optimize out
 169     // the export of LinearScaleYUVToRGB32Row_MMX, which is required by the unit
 170     // tests, if that decision can be made at compile time.  Since all X64 CPUs
 171     // have SSE2, we can hack around this by making the selection here.
 172     g_linear_scale_yuv_to_rgb32_row_proc_ = LinearScaleYUVToRGB32Row_MMX_X64;
 173 #endif
 174   }
 175
 176   if (cpu.has_ssse3()) {
 177     g_convert_rgb24_to_yuv_proc_ = &ConvertRGB24ToYUV_SSSE3;
 178
 179     // TODO(hclam): Add ConvertRGB32ToYUV_SSSE3 when the cyan problem is solved.
 180     // See: crbug.com/100462
 181   }
 182 #endif
 183 }
 184
 185 // Empty SIMD registers state after using them.
 186 void EmptyRegisterState() { g_empty_register_state_proc_(); }
 187
 188 // 16.16 fixed point arithmetic
 189 const int kFractionBits = 16;
 190 const int kFractionMax = 1 << kFractionBits;
 191 const int kFractionMask = ((1 << kFractionBits) - 1);
 192
 193 // Scale a frame of YUV to 32 bit ARGB.
 194 void ScaleYUVToRGB32(const uint8* y_buf,
 195                      const uint8* u_buf,
 196                      const uint8* v_buf,
 197                      uint8* rgb_buf,
 198                      int source_width,
 199                      int source_height,
 200                      int width,
 201                      int height,
 202                      int y_pitch,
 203                      int uv_pitch,
 204                      int rgb_pitch,
 205                      YUVType yuv_type,
 206                      Rotate view_rotate,
 207                      ScaleFilter filter) {
 208   // Handle zero sized sources and destinations.
 209   if ((yuv_type == YV12 && (source_width < 2 || source_height < 2)) ||
 210       (yuv_type == YV16 && (source_width < 2 || source_height < 1)) ||
 211       width == 0 || height == 0)
 212     return;
 213
 214   // 4096 allows 3 buffers to fit in 12k.
 215   // Helps performance on CPU with 16K L1 cache.
 216   // Large enough for 3830x2160 and 30" displays which are 2560x1600.
 217   const int kFilterBufferSize = 4096;
 218   // Disable filtering if the screen is too big (to avoid buffer overflows).
 219   // This should never happen to regular users: they don't have monitors
 220   // wider than 4096 pixels.
 221   // TODO(fbarchard): Allow rotated videos to filter.
 222   if (source_width > kFilterBufferSize || view_rotate)
 223     filter = FILTER_NONE;
 224
 225   unsigned int y_shift = yuv_type;
 226   // Diagram showing origin and direction of source sampling.
 227   // ->0   4<-
 228   // 7       3
 229   //
 230   // 6       5
 231   // ->1   2<-
 232   // Rotations that start at right side of image.
 233   if ((view_rotate == ROTATE_180) || (view_rotate == ROTATE_270) ||
 234       (view_rotate == MIRROR_ROTATE_0) || (view_rotate == MIRROR_ROTATE_90)) {
 235     y_buf += source_width - 1;
 236     u_buf += source_width / 2 - 1;
 237     v_buf += source_width / 2 - 1;
 238     source_width = -source_width;
 239   }
 240   // Rotations that start at bottom of image.
 241   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_180) ||
 242       (view_rotate == MIRROR_ROTATE_90) || (view_rotate == MIRROR_ROTATE_180)) {
 243     y_buf += (source_height - 1) * y_pitch;
 244     u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 245     v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 246     source_height = -source_height;
 247   }
 248
 249   int source_dx = source_width * kFractionMax / width;
 250
 251   if ((view_rotate == ROTATE_90) || (view_rotate == ROTATE_270)) {
 252     int tmp = height;
 253     height = width;
 254     width = tmp;
 255     tmp = source_height;
 256     source_height = source_width;
 257     source_width = tmp;
 258     int source_dy = source_height * kFractionMax / height;
 259     source_dx = ((source_dy >> kFractionBits) * y_pitch) << kFractionBits;
 260     if (view_rotate == ROTATE_90) {
 261       y_pitch = -1;
 262       uv_pitch = -1;
 263       source_height = -source_height;
 264     } else {
 265       y_pitch = 1;
 266       uv_pitch = 1;
 267     }
 268   }
 269
 270   // Need padding because FilterRows() will write 1 to 16 extra pixels
 271   // after the end for SSE2 version.
 272   uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
 273   uint8* ybuf =
 274       reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
 275   uint8* ubuf = ybuf + kFilterBufferSize;
 276   uint8* vbuf = ubuf + kFilterBufferSize;
 277
 278   // TODO(fbarchard): Fixed point math is off by 1 on negatives.
 279
 280   // We take a y-coordinate in [0,1] space in the source image space, and
 281   // transform to a y-coordinate in [0,1] space in the destination image space.
 282   // Note that the coordinate endpoints lie on pixel boundaries, not on pixel
 283   // centers: e.g. a two-pixel-high image will have pixel centers at 0.25 and
 284   // 0.75.  The formula is as follows (in fixed-point arithmetic):
 285   //   y_dst = dst_height * ((y_src + 0.5) / src_height)
 286   //   dst_pixel = clamp([0, dst_height - 1], floor(y_dst - 0.5))
 287   // Implement this here as an accumulator + delta, to avoid expensive math
 288   // in the loop.
 289   int source_y_subpixel_accum =
 290       ((kFractionMax / 2) * source_height) / height - (kFractionMax / 2);
 291   int source_y_subpixel_delta = ((1 << kFractionBits) * source_height) / height;
 292
 293   // TODO(fbarchard): Split this into separate function for better efficiency.
 294   for (int y = 0; y < height; ++y) {
 295     uint8* dest_pixel = rgb_buf + y * rgb_pitch;
 296     int source_y_subpixel = source_y_subpixel_accum;
 297     source_y_subpixel_accum += source_y_subpixel_delta;
 298     if (source_y_subpixel < 0)
 299       source_y_subpixel = 0;
 300     else if (source_y_subpixel > ((source_height - 1) << kFractionBits))
 301       source_y_subpixel = (source_height - 1) << kFractionBits;
 302
 303     const uint8* y_ptr = NULL;
 304     const uint8* u_ptr = NULL;
 305     const uint8* v_ptr = NULL;
 306     // Apply vertical filtering if necessary.
 307     // TODO(fbarchard): Remove memcpy when not necessary.
 308     if (filter & media::FILTER_BILINEAR_V) {
 309       int source_y = source_y_subpixel >> kFractionBits;
 310       y_ptr = y_buf + source_y * y_pitch;
 311       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 312       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 313
 314       // Vertical scaler uses 16.8 fixed point.
 315       int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
 316       if (source_y_fraction != 0) {
 317         g_filter_yuv_rows_proc_(
 318             ybuf, y_ptr, y_ptr + y_pitch, source_width, source_y_fraction);
 319       } else {
 320         memcpy(ybuf, y_ptr, source_width);
 321       }
 322       y_ptr = ybuf;
 323       ybuf[source_width] = ybuf[source_width - 1];
 324
 325       int uv_source_width = (source_width + 1) / 2;
 326       int source_uv_fraction;
 327
 328       // For formats with half-height UV planes, each even-numbered pixel row
 329       // should not interpolate, since the next row to interpolate from should
 330       // be a duplicate of the current row.
 331       if (y_shift && (source_y & 0x1) == 0)
 332         source_uv_fraction = 0;
 333       else
 334         source_uv_fraction = source_y_fraction;
 335
 336       if (source_uv_fraction != 0) {
 337         g_filter_yuv_rows_proc_(
 338             ubuf, u_ptr, u_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 339         g_filter_yuv_rows_proc_(
 340             vbuf, v_ptr, v_ptr + uv_pitch, uv_source_width, source_uv_fraction);
 341       } else {
 342         memcpy(ubuf, u_ptr, uv_source_width);
 343         memcpy(vbuf, v_ptr, uv_source_width);
 344       }
 345       u_ptr = ubuf;
 346       v_ptr = vbuf;
 347       ubuf[uv_source_width] = ubuf[uv_source_width - 1];
 348       vbuf[uv_source_width] = vbuf[uv_source_width - 1];
 349     } else {
 350       // Offset by 1/2 pixel for center sampling.
 351       int source_y = (source_y_subpixel + (kFractionMax / 2)) >> kFractionBits;
 352       y_ptr = y_buf + source_y * y_pitch;
 353       u_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 354       v_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 355     }
 356     if (source_dx == kFractionMax) {  // Not scaled
 357       g_convert_yuv_to_rgb32_row_proc_(y_ptr, u_ptr, v_ptr, dest_pixel, width);
 358     } else {
 359       if (filter & FILTER_BILINEAR_H) {
 360         g_linear_scale_yuv_to_rgb32_row_proc_(
 361             y_ptr, u_ptr, v_ptr, dest_pixel, width, source_dx);
 362       } else {
 363         g_scale_yuv_to_rgb32_row_proc_(
 364             y_ptr, u_ptr, v_ptr, dest_pixel, width, source_dx);
 365       }
 366     }
 367   }
 368
 369   g_empty_register_state_proc_();
 370 }
 371
 372 // Scale a frame of YV12 to 32 bit ARGB for a specific rectangle.
 373 void ScaleYUVToRGB32WithRect(const uint8* y_buf,
 374                              const uint8* u_buf,
 375                              const uint8* v_buf,
 376                              uint8* rgb_buf,
 377                              int source_width,
 378                              int source_height,
 379                              int dest_width,
 380                              int dest_height,
 381                              int dest_rect_left,
 382                              int dest_rect_top,
 383                              int dest_rect_right,
 384                              int dest_rect_bottom,
 385                              int y_pitch,
 386                              int uv_pitch,
 387                              int rgb_pitch) {
 388   // This routine doesn't currently support up-scaling.
 389   CHECK_LE(dest_width, source_width);
 390   CHECK_LE(dest_height, source_height);
 391
 392   // Sanity-check the destination rectangle.
 393   DCHECK(dest_rect_left >= 0 && dest_rect_right <= dest_width);
 394   DCHECK(dest_rect_top >= 0 && dest_rect_bottom <= dest_height);
 395   DCHECK(dest_rect_right > dest_rect_left);
 396   DCHECK(dest_rect_bottom > dest_rect_top);
 397
 398   // Fixed-point value of vertical and horizontal scale down factor.
 399   // Values are in the format 16.16.
 400   int y_step = kFractionMax * source_height / dest_height;
 401   int x_step = kFractionMax * source_width / dest_width;
 402
 403   // Determine the coordinates of the rectangle in 16.16 coords.
 404   // NB: Our origin is the *center* of the top/left pixel, NOT its top/left.
 405   // If we're down-scaling by more than a factor of two, we start with a 50%
 406   // fraction to avoid degenerating to point-sampling - we should really just
 407   // fix the fraction at 50% for all pixels in that case.
 408   int source_left = dest_rect_left * x_step;
 409   int source_right = (dest_rect_right - 1) * x_step;
 410   if (x_step < kFractionMax * 2) {
 411     source_left += ((x_step - kFractionMax) / 2);
 412     source_right += ((x_step - kFractionMax) / 2);
 413   } else {
 414     source_left += kFractionMax / 2;
 415     source_right += kFractionMax / 2;
 416   }
 417   int source_top = dest_rect_top * y_step;
 418   if (y_step < kFractionMax * 2) {
 419     source_top += ((y_step - kFractionMax) / 2);
 420   } else {
 421     source_top += kFractionMax / 2;
 422   }
 423
 424   // Determine the parts of the Y, U and V buffers to interpolate.
 425   int source_y_left = source_left >> kFractionBits;
 426   int source_y_right =
 427       std::min((source_right >> kFractionBits) + 2, source_width + 1);
 428
 429   int source_uv_left = source_y_left / 2;
 430   int source_uv_right = std::min((source_right >> (kFractionBits + 1)) + 2,
 431                                  (source_width + 1) / 2);
 432
 433   int source_y_width = source_y_right - source_y_left;
 434   int source_uv_width = source_uv_right - source_uv_left;
 435
 436   // Determine number of pixels in each output row.
 437   int dest_rect_width = dest_rect_right - dest_rect_left;
 438
 439   // Intermediate buffer for vertical interpolation.
 440   // 4096 bytes allows 3 buffers to fit in 12k, which fits in a 16K L1 cache,
 441   // and is bigger than most users will generally need.
 442   // The buffer is 16-byte aligned and padded with 16 extra bytes; some of the
 443   // FilterYUVRowProcs have alignment requirements, and the SSE version can
 444   // write up to 16 bytes past the end of the buffer.
 445   const int kFilterBufferSize = 4096;
 446   const bool kAvoidUsingOptimizedFilter = source_width > kFilterBufferSize;
 447   uint8 yuv_temp[16 + kFilterBufferSize * 3 + 16];
 448   // memset() yuv_temp to 0 to avoid bogus warnings when running on Valgrind.
 449   if (RunningOnValgrind())
 450     memset(yuv_temp, 0, sizeof(yuv_temp));
 451   uint8* y_temp = reinterpret_cast<uint8*>(
 452       reinterpret_cast<uintptr_t>(yuv_temp + 15) & ~15);
 453   uint8* u_temp = y_temp + kFilterBufferSize;
 454   uint8* v_temp = u_temp + kFilterBufferSize;
 455
 456   // Move to the top-left pixel of output.
 457   rgb_buf += dest_rect_top * rgb_pitch;
 458   rgb_buf += dest_rect_left * 4;
 459
 460   // For each destination row perform interpolation and color space
 461   // conversion to produce the output.
 462   for (int row = dest_rect_top; row < dest_rect_bottom; ++row) {
 463     // Round the fixed-point y position to get the current row.
 464     int source_row = source_top >> kFractionBits;
 465     int source_uv_row = source_row / 2;
 466     DCHECK(source_row < source_height);
 467
 468     // Locate the first row for each plane for interpolation.
 469     const uint8* y0_ptr = y_buf + y_pitch * source_row + source_y_left;
 470     const uint8* u0_ptr = u_buf + uv_pitch * source_uv_row + source_uv_left;
 471     const uint8* v0_ptr = v_buf + uv_pitch * source_uv_row + source_uv_left;
 472     const uint8* y1_ptr = NULL;
 473     const uint8* u1_ptr = NULL;
 474     const uint8* v1_ptr = NULL;
 475
 476     // Locate the second row for interpolation, being careful not to overrun.
 477     if (source_row + 1 >= source_height) {
 478       y1_ptr = y0_ptr;
 479     } else {
 480       y1_ptr = y0_ptr + y_pitch;
 481     }
 482     if (source_uv_row + 1 >= (source_height + 1) / 2) {
 483       u1_ptr = u0_ptr;
 484       v1_ptr = v0_ptr;
 485     } else {
 486       u1_ptr = u0_ptr + uv_pitch;
 487       v1_ptr = v0_ptr + uv_pitch;
 488     }
 489
 490     if (!kAvoidUsingOptimizedFilter) {
 491       // Vertical scaler uses 16.8 fixed point.
 492       int fraction = (source_top & kFractionMask) >> 8;
 493       g_filter_yuv_rows_proc_(
 494           y_temp + source_y_left, y0_ptr, y1_ptr, source_y_width, fraction);
 495       g_filter_yuv_rows_proc_(
 496           u_temp + source_uv_left, u0_ptr, u1_ptr, source_uv_width, fraction);
 497       g_filter_yuv_rows_proc_(
 498           v_temp + source_uv_left, v0_ptr, v1_ptr, source_uv_width, fraction);
 499
 500       // Perform horizontal interpolation and color space conversion.
 501       // TODO(hclam): Use the MMX version after more testing.
 502       LinearScaleYUVToRGB32RowWithRange_C(y_temp,
 503                                           u_temp,
 504                                           v_temp,
 505                                           rgb_buf,
 506                                           dest_rect_width,
 507                                           source_left,
 508                                           x_step);
 509     } else {
 510       // If the frame is too large then we linear scale a single row.
 511       LinearScaleYUVToRGB32RowWithRange_C(y0_ptr,
 512                                           u0_ptr,
 513                                           v0_ptr,
 514                                           rgb_buf,
 515                                           dest_rect_width,
 516                                           source_left,
 517                                           x_step);
 518     }
 519
 520     // Advance vertically in the source and destination image.
 521     source_top += y_step;
 522     rgb_buf += rgb_pitch;
 523   }
 524
 525   g_empty_register_state_proc_();
 526 }
 527
 528 void ConvertRGB32ToYUV(const uint8* rgbframe,
 529                        uint8* yplane,
 530                        uint8* uplane,
 531                        uint8* vplane,
 532                        int width,
 533                        int height,
 534                        int rgbstride,
 535                        int ystride,
 536                        int uvstride) {
 537   g_convert_rgb32_to_yuv_proc_(rgbframe,
 538                                yplane,
 539                                uplane,
 540                                vplane,
 541                                width,
 542                                height,
 543                                rgbstride,
 544                                ystride,
 545                                uvstride);
 546 }
 547
 548 void ConvertRGB24ToYUV(const uint8* rgbframe,
 549                        uint8* yplane,
 550                        uint8* uplane,
 551                        uint8* vplane,
 552                        int width,
 553                        int height,
 554                        int rgbstride,
 555                        int ystride,
 556                        int uvstride) {
 557   g_convert_rgb24_to_yuv_proc_(rgbframe,
 558                                yplane,
 559                                uplane,
 560                                vplane,
 561                                width,
 562                                height,
 563                                rgbstride,
 564                                ystride,
 565                                uvstride);
 566 }
 567
 568 void ConvertYUY2ToYUV(const uint8* src,
 569                       uint8* yplane,
 570                       uint8* uplane,
 571                       uint8* vplane,
 572                       int width,
 573                       int height) {
 574   for (int i = 0; i < height / 2; ++i) {
 575     for (int j = 0; j < (width / 2); ++j) {
 576       yplane[0] = src[0];
 577       *uplane = src[1];
 578       yplane[1] = src[2];
 579       *vplane = src[3];
 580       src += 4;
 581       yplane += 2;
 582       uplane++;
 583       vplane++;
 584     }
 585     for (int j = 0; j < (width / 2); ++j) {
 586       yplane[0] = src[0];
 587       yplane[1] = src[2];
 588       src += 4;
 589       yplane += 2;
 590     }
 591   }
 592 }
 593
 594 void ConvertNV21ToYUV(const uint8* src,
 595                       uint8* yplane,
 596                       uint8* uplane,
 597                       uint8* vplane,
 598                       int width,
 599                       int height) {
 600   int y_plane_size = width * height;
 601   memcpy(yplane, src, y_plane_size);
 602
 603   src += y_plane_size;
 604   int u_plane_size = y_plane_size >> 2;
 605   for (int i = 0; i < u_plane_size; ++i) {
 606     *vplane++ = *src++;
 607     *uplane++ = *src++;
 608   }
 609 }
 610
 611 void ConvertYUVToRGB32(const uint8* yplane,
 612                        const uint8* uplane,
 613                        const uint8* vplane,
 614                        uint8* rgbframe,
 615                        int width,
 616                        int height,
 617                        int ystride,
 618                        int uvstride,
 619                        int rgbstride,
 620                        YUVType yuv_type) {
 621   g_convert_yuv_to_rgb32_proc_(yplane,
 622                                uplane,
 623                                vplane,
 624                                rgbframe,
 625                                width,
 626                                height,
 627                                ystride,
 628                                uvstride,
 629                                rgbstride,
 630                                yuv_type);
 631 }
 632
 633 void ConvertYUVAToARGB(const uint8* yplane,
 634                        const uint8* uplane,
 635                        const uint8* vplane,
 636                        const uint8* aplane,
 637                        uint8* rgbframe,
 638                        int width,
 639                        int height,
 640                        int ystride,
 641                        int uvstride,
 642                        int astride,
 643                        int rgbstride,
 644                        YUVType yuv_type) {
 645   g_convert_yuva_to_argb_proc_(yplane,
 646                                uplane,
 647                                vplane,
 648                                aplane,
 649                                rgbframe,
 650                                width,
 651                                height,
 652                                ystride,
 653                                uvstride,
 654                                astride,
 655                                rgbstride,
 656                                yuv_type);
 657 }
 658
 659 }  // namespace media