gfx/ycbcr/yuv_convert.cpp

   1 // Copyright (c) 2010 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 // YV24 is a full plane of Y and a full height, full width chroma planes
  15 //
  16 // ARGB pixel format is output, which on little endian is stored as BGRA.
  17 // The alpha is set to 255, allowing the application to use RGBA or RGB32.
  18
  19 #include "yuv_convert.h"
  20
  21 // Header for low level row functions.
  22 #include "yuv_row.h"
  23 #include "mozilla/SSE.h"
  24
  25 namespace mozilla {
  26
  27 namespace gfx {
  28
  29 // 16.16 fixed point arithmetic
  30 const int kFractionBits = 16;
  31 const int kFractionMax = 1 << kFractionBits;
  32 const int kFractionMask = ((1 << kFractionBits) - 1);
  33
  34 NS_GFX_(YUVType) TypeFromSize(int ywidth,
  35                               int yheight,
  36                               int cbcrwidth,
  37                               int cbcrheight)
  38 {
  39   if (ywidth == cbcrwidth && yheight == cbcrheight) {
  40     return YV24;
  41   }
  42   else if (ywidth / 2 == cbcrwidth && yheight == cbcrheight) {
  43     return YV16;
  44   }
  45   else {
  46     return YV12;
  47   }
  48 }
  49
  50 // Convert a frame of YUV to 32 bit ARGB.
  51 NS_GFX_(void) ConvertYCbCrToRGB32(const uint8* y_buf,
  52                                   const uint8* u_buf,
  53                                   const uint8* v_buf,
  54                                   uint8* rgb_buf,
  55                                   int pic_x,
  56                                   int pic_y,
  57                                   int pic_width,
  58                                   int pic_height,
  59                                   int y_pitch,
  60                                   int uv_pitch,
  61                                   int rgb_pitch,
  62                                   YUVType yuv_type) {
  63   unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
  64   unsigned int x_shift = yuv_type == YV24 ? 0 : 1;
  65   // Test for SSE because the optimized code uses movntq, which is not part of MMX.
  66   bool has_sse = supports_mmx() && supports_sse();
  67   // There is no optimized YV24 SSE routine so we check for this and
  68   // fall back to the C code.
  69   has_sse &= yuv_type != YV24;
  70   bool odd_pic_x = yuv_type != YV24 && pic_x % 2 != 0;
  71   int x_width = odd_pic_x ? pic_width - 1 : pic_width;
  72
  73   for (int y = pic_y; y < pic_height + pic_y; ++y) {
  74     uint8* rgb_row = rgb_buf + (y - pic_y) * rgb_pitch;
  75     const uint8* y_ptr = y_buf + y * y_pitch + pic_x;
  76     const uint8* u_ptr = u_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
  77     const uint8* v_ptr = v_buf + (y >> y_shift) * uv_pitch + (pic_x >> x_shift);
  78
  79     if (odd_pic_x) {
  80       // Handle the single odd pixel manually and use the
  81       // fast routines for the remaining.
  82       FastConvertYUVToRGB32Row_C(y_ptr++,
  83                                  u_ptr++,
  84                                  v_ptr++,
  85                                  rgb_row,
  86                                  1,
  87                                  x_shift);
  88       rgb_row += 4;
  89     }
  90
  91     if (has_sse) {
  92       FastConvertYUVToRGB32Row(y_ptr,
  93                                u_ptr,
  94                                v_ptr,
  95                                rgb_row,
  96                                x_width);
  97     }
  98     else {
  99       FastConvertYUVToRGB32Row_C(y_ptr,
 100                                  u_ptr,
 101                                  v_ptr,
 102                                  rgb_row,
 103                                  x_width,
 104                                  x_shift);
 105     }
 106   }
 107
 108   // MMX used for FastConvertYUVToRGB32Row requires emms instruction.
 109   if (has_sse)
 110     EMMS();
 111 }
 112
 113 // C version does 8 at a time to mimic MMX code
 114 static void FilterRows_C(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
 115                          int source_width, int source_y_fraction) {
 116   int y1_fraction = source_y_fraction;
 117   int y0_fraction = 256 - y1_fraction;
 118   uint8* end = ybuf + source_width;
 119   do {
 120     ybuf[0] = (y0_ptr[0] * y0_fraction + y1_ptr[0] * y1_fraction) >> 8;
 121     ybuf[1] = (y0_ptr[1] * y0_fraction + y1_ptr[1] * y1_fraction) >> 8;
 122     ybuf[2] = (y0_ptr[2] * y0_fraction + y1_ptr[2] * y1_fraction) >> 8;
 123     ybuf[3] = (y0_ptr[3] * y0_fraction + y1_ptr[3] * y1_fraction) >> 8;
 124     ybuf[4] = (y0_ptr[4] * y0_fraction + y1_ptr[4] * y1_fraction) >> 8;
 125     ybuf[5] = (y0_ptr[5] * y0_fraction + y1_ptr[5] * y1_fraction) >> 8;
 126     ybuf[6] = (y0_ptr[6] * y0_fraction + y1_ptr[6] * y1_fraction) >> 8;
 127     ybuf[7] = (y0_ptr[7] * y0_fraction + y1_ptr[7] * y1_fraction) >> 8;
 128     y0_ptr += 8;
 129     y1_ptr += 8;
 130     ybuf += 8;
 131   } while (ybuf < end);
 132 }
 133
 134 #ifdef MOZILLA_MAY_SUPPORT_MMX
 135 void FilterRows_MMX(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
 136                     int source_width, int source_y_fraction);
 137 #endif
 138
 139 #ifdef MOZILLA_MAY_SUPPORT_SSE2
 140 void FilterRows_SSE2(uint8* ybuf, const uint8* y0_ptr, const uint8* y1_ptr,
 141                      int source_width, int source_y_fraction);
 142 #endif
 143
 144 static inline void FilterRows(uint8* ybuf, const uint8* y0_ptr,
 145                               const uint8* y1_ptr, int source_width,
 146                               int source_y_fraction) {
 147 #ifdef MOZILLA_MAY_SUPPORT_SSE2
 148   if (mozilla::supports_sse2()) {
 149     FilterRows_SSE2(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
 150     return;
 151   }
 152 #endif
 153
 154 #ifdef MOZILLA_MAY_SUPPORT_MMX
 155   if (mozilla::supports_mmx()) {
 156     FilterRows_MMX(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
 157     return;
 158   }
 159 #endif
 160
 161   FilterRows_C(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
 162 }
 163
 164
 165 // Scale a frame of YUV to 32 bit ARGB.
 166 NS_GFX_(void) ScaleYCbCrToRGB32(const uint8* y_buf,
 167                                 const uint8* u_buf,
 168                                 const uint8* v_buf,
 169                                 uint8* rgb_buf,
 170                                 int source_width,
 171                                 int source_height,
 172                                 int width,
 173                                 int height,
 174                                 int y_pitch,
 175                                 int uv_pitch,
 176                                 int rgb_pitch,
 177                                 YUVType yuv_type,
 178                                 Rotate view_rotate,
 179                                 ScaleFilter filter) {
 180   bool has_mmx = supports_mmx();
 181
 182   // 4096 allows 3 buffers to fit in 12k.
 183   // Helps performance on CPU with 16K L1 cache.
 184   // Large enough for 3830x2160 and 30" displays which are 2560x1600.
 185   const int kFilterBufferSize = 4096;
 186   // Disable filtering if the screen is too big (to avoid buffer overflows).
 187   // This should never happen to regular users: they don't have monitors
 188   // wider than 4096 pixels.
 189   // TODO(fbarchard): Allow rotated videos to filter.
 190   if (source_width > kFilterBufferSize || view_rotate)
 191     filter = FILTER_NONE;
 192
 193   unsigned int y_shift = yuv_type == YV12 ? 1 : 0;
 194   // Diagram showing origin and direction of source sampling.
 195   // ->0   4<-
 196   // 7       3
 197   //
 198   // 6       5
 199   // ->1   2<-
 200   // Rotations that start at right side of image.
 201   if ((view_rotate == ROTATE_180) ||
 202       (view_rotate == ROTATE_270) ||
 203       (view_rotate == MIRROR_ROTATE_0) ||
 204       (view_rotate == MIRROR_ROTATE_90)) {
 205     y_buf += source_width - 1;
 206     u_buf += source_width / 2 - 1;
 207     v_buf += source_width / 2 - 1;
 208     source_width = -source_width;
 209   }
 210   // Rotations that start at bottom of image.
 211   if ((view_rotate == ROTATE_90) ||
 212       (view_rotate == ROTATE_180) ||
 213       (view_rotate == MIRROR_ROTATE_90) ||
 214       (view_rotate == MIRROR_ROTATE_180)) {
 215     y_buf += (source_height - 1) * y_pitch;
 216     u_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 217     v_buf += ((source_height >> y_shift) - 1) * uv_pitch;
 218     source_height = -source_height;
 219   }
 220
 221   // Handle zero sized destination.
 222   if (width == 0 || height == 0)
 223     return;
 224   int source_dx = source_width * kFractionMax / width;
 225   int source_dy = source_height * kFractionMax / height;
 226   int source_dx_uv = source_dx;
 227
 228   if ((view_rotate == ROTATE_90) ||
 229       (view_rotate == ROTATE_270)) {
 230     int tmp = height;
 231     height = width;
 232     width = tmp;
 233     tmp = source_height;
 234     source_height = source_width;
 235     source_width = tmp;
 236     int original_dx = source_dx;
 237     int original_dy = source_dy;
 238     source_dx = ((original_dy >> kFractionBits) * y_pitch) << kFractionBits;
 239     source_dx_uv = ((original_dy >> kFractionBits) * uv_pitch) << kFractionBits;
 240     source_dy = original_dx;
 241     if (view_rotate == ROTATE_90) {
 242       y_pitch = -1;
 243       uv_pitch = -1;
 244       source_height = -source_height;
 245     } else {
 246       y_pitch = 1;
 247       uv_pitch = 1;
 248     }
 249   }
 250
 251   // Need padding because FilterRows() will write 1 to 16 extra pixels
 252   // after the end for SSE2 version.
 253   uint8 yuvbuf[16 + kFilterBufferSize * 3 + 16];
 254   uint8* ybuf =
 255       reinterpret_cast<uint8*>(reinterpret_cast<uintptr_t>(yuvbuf + 15) & ~15);
 256   uint8* ubuf = ybuf + kFilterBufferSize;
 257   uint8* vbuf = ubuf + kFilterBufferSize;
 258   // TODO(fbarchard): Fixed point math is off by 1 on negatives.
 259   int yscale_fixed = (source_height << kFractionBits) / height;
 260
 261   // TODO(fbarchard): Split this into separate function for better efficiency.
 262   for (int y = 0; y < height; ++y) {
 263     uint8* dest_pixel = rgb_buf + y * rgb_pitch;
 264     int source_y_subpixel = (y * yscale_fixed);
 265     if (yscale_fixed >= (kFractionMax * 2)) {
 266       source_y_subpixel += kFractionMax / 2;  // For 1/2 or less, center filter.
 267     }
 268     int source_y = source_y_subpixel >> kFractionBits;
 269
 270     const uint8* y0_ptr = y_buf + source_y * y_pitch;
 271     const uint8* y1_ptr = y0_ptr + y_pitch;
 272
 273     const uint8* u0_ptr = u_buf + (source_y >> y_shift) * uv_pitch;
 274     const uint8* u1_ptr = u0_ptr + uv_pitch;
 275     const uint8* v0_ptr = v_buf + (source_y >> y_shift) * uv_pitch;
 276     const uint8* v1_ptr = v0_ptr + uv_pitch;
 277
 278     // vertical scaler uses 16.8 fixed point
 279     int source_y_fraction = (source_y_subpixel & kFractionMask) >> 8;
 280     int source_uv_fraction =
 281         ((source_y_subpixel >> y_shift) & kFractionMask) >> 8;
 282
 283     const uint8* y_ptr = y0_ptr;
 284     const uint8* u_ptr = u0_ptr;
 285     const uint8* v_ptr = v0_ptr;
 286     // Apply vertical filtering if necessary.
 287     // TODO(fbarchard): Remove memcpy when not necessary.
 288     if (filter & mozilla::gfx::FILTER_BILINEAR_V) {
 289       if (yscale_fixed != kFractionMax &&
 290           source_y_fraction && ((source_y + 1) < source_height)) {
 291         FilterRows(ybuf, y0_ptr, y1_ptr, source_width, source_y_fraction);
 292       } else {
 293         memcpy(ybuf, y0_ptr, source_width);
 294       }
 295       y_ptr = ybuf;
 296       ybuf[source_width] = ybuf[source_width-1];
 297       int uv_source_width = (source_width + 1) / 2;
 298       if (yscale_fixed != kFractionMax &&
 299           source_uv_fraction &&
 300           (((source_y >> y_shift) + 1) < (source_height >> y_shift))) {
 301         FilterRows(ubuf, u0_ptr, u1_ptr, uv_source_width, source_uv_fraction);
 302         FilterRows(vbuf, v0_ptr, v1_ptr, uv_source_width, source_uv_fraction);
 303       } else {
 304         memcpy(ubuf, u0_ptr, uv_source_width);
 305         memcpy(vbuf, v0_ptr, uv_source_width);
 306       }
 307       u_ptr = ubuf;
 308       v_ptr = vbuf;
 309       ubuf[uv_source_width] = ubuf[uv_source_width - 1];
 310       vbuf[uv_source_width] = vbuf[uv_source_width - 1];
 311     }
 312     if (source_dx == kFractionMax) {  // Not scaled
 313       FastConvertYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
 314                                dest_pixel, width);
 315     } else if (filter & FILTER_BILINEAR_H) {
 316         LinearScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
 317                                  dest_pixel, width, source_dx);
 318     } else {
 319 // Specialized scalers and rotation.
 320 #if defined(MOZILLA_MAY_SUPPORT_SSE) && defined(_MSC_VER) && defined(_M_IX86)
 321       if(mozilla::supports_sse()) {
 322         if (width == (source_width * 2)) {
 323           DoubleYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
 324                                   dest_pixel, width);
 325         } else if ((source_dx & kFractionMask) == 0) {
 326           // Scaling by integer scale factor. ie half.
 327           ConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
 328                                    dest_pixel, width,
 329                                    source_dx >> kFractionBits);
 330         } else if (source_dx_uv == source_dx) {  // Not rotated.
 331           ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
 332                              dest_pixel, width, source_dx);
 333         } else {
 334           RotateConvertYUVToRGB32Row_SSE(y_ptr, u_ptr, v_ptr,
 335                                          dest_pixel, width,
 336                                          source_dx >> kFractionBits,
 337                                          source_dx_uv >> kFractionBits);
 338         }
 339       }
 340       else {
 341         ScaleYUVToRGB32Row_C(y_ptr, u_ptr, v_ptr,
 342                              dest_pixel, width, source_dx);
 343       }
 344 #else
 345       (void)source_dx_uv;
 346       ScaleYUVToRGB32Row(y_ptr, u_ptr, v_ptr,
 347                          dest_pixel, width, source_dx);
 348 #endif
 349     }
 350   }
 351   // MMX used for FastConvertYUVToRGB32Row and FilterRows requires emms.
 352   if (has_mmx)
 353     EMMS();
 354 }
 355
 356 }  // namespace gfx
 357 }  // namespace mozilla