libswscale/yuv2rgb.c

   1 /*
   2  * yuv2rgb.c, Software YUV to RGB converter
   3  *
   4  *  Copyright (C) 1999, Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
   5  *
   6  *  Functions broken out from display_x11.c and several new modes
   7  *  added by Håkan Hjort <d95hjort@dtek.chalmers.se>
   8  *
   9  *  15 & 16 bpp support by Franck Sicard <Franck.Sicard@solsoft.fr>
  10  *
  11  *  MMX/MMX2 template stuff (needed for fast movntq support),
  12  *  1,4,8bpp support and context / deglobalize stuff
  13  *  by Michael Niedermayer (michaelni@gmx.at)
  14  *
  15  *  This file is part of mpeg2dec, a free MPEG-2 video decoder
  16  *
  17  *  mpeg2dec is free software; you can redistribute it and/or modify
  18  *  it under the terms of the GNU General Public License as published by
  19  *  the Free Software Foundation; either version 2, or (at your option)
  20  *  any later version.
  21  *
  22  *  mpeg2dec is distributed in the hope that it will be useful,
  23  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  24  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  25  *  GNU General Public License for more details.
  26  *
  27  *  You should have received a copy of the GNU General Public License
  28  *  along with mpeg2dec; if not, write to the Free Software
  29  *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  30  */
  31
  32 #include <stdio.h>
  33 #include <stdlib.h>
  34 #include <inttypes.h>
  35 #include <assert.h>
  36
  37 #include "config.h"
  38 #include "rgb2rgb.h"
  39 #include "swscale.h"
  40 #include "swscale_internal.h"
  41
  42 #define DITHER1XBPP // only for MMX
  43
  44 extern const uint8_t dither_8x8_32[8][8];
  45 extern const uint8_t dither_8x8_73[8][8];
  46 extern const uint8_t dither_8x8_220[8][8];
  47
  48 #ifdef HAVE_MMX
  49
  50 /* hope these constant values are cache line aligned */
  51 DECLARE_ASM_CONST(8, uint64_t, mmx_00ffw)   = 0x00ff00ff00ff00ffULL;
  52 DECLARE_ASM_CONST(8, uint64_t, mmx_redmask) = 0xf8f8f8f8f8f8f8f8ULL;
  53 DECLARE_ASM_CONST(8, uint64_t, mmx_grnmask) = 0xfcfcfcfcfcfcfcfcULL;
  54
  55 #undef HAVE_MMX
  56
  57 //MMX versions
  58 #undef RENAME
  59 #define HAVE_MMX
  60 #undef HAVE_MMX2
  61 #undef HAVE_3DNOW
  62 #define RENAME(a) a ## _MMX
  63 #include "yuv2rgb_template.c"
  64
  65 //MMX2 versions
  66 #undef RENAME
  67 #define HAVE_MMX
  68 #define HAVE_MMX2
  69 #undef HAVE_3DNOW
  70 #define RENAME(a) a ## _MMX2
  71 #include "yuv2rgb_template.c"
  72
  73 #endif /* HAVE_MMX */
  74
  75 const int32_t Inverse_Table_6_9[8][4] = {
  76     {117504, 138453, 13954, 34903}, /* no sequence_display_extension */
  77     {117504, 138453, 13954, 34903}, /* ITU-R Rec. 709 (1990) */
  78     {104597, 132201, 25675, 53279}, /* unspecified */
  79     {104597, 132201, 25675, 53279}, /* reserved */
  80     {104448, 132798, 24759, 53109}, /* FCC */
  81     {104597, 132201, 25675, 53279}, /* ITU-R Rec. 624-4 System B, G */
  82     {104597, 132201, 25675, 53279}, /* SMPTE 170M */
  83     {117579, 136230, 16907, 35559}  /* SMPTE 240M (1987) */
  84 };
  85
  86 #define RGB(i)                                      \
  87     U = pu[i];                                      \
  88     V = pv[i];                                      \
  89     r = (void *)c->table_rV[V];                     \
  90     g = (void *)(c->table_gU[U] + c->table_gV[V]);  \
  91     b = (void *)c->table_bU[U];
  92
  93 #define DST1(i)                         \
  94     Y = py_1[2*i];                      \
  95     dst_1[2*i] = r[Y] + g[Y] + b[Y];    \
  96     Y = py_1[2*i+1];                    \
  97     dst_1[2*i+1] = r[Y] + g[Y] + b[Y];
  98
  99 #define DST2(i)                         \
 100     Y = py_2[2*i];                      \
 101     dst_2[2*i] = r[Y] + g[Y] + b[Y];    \
 102     Y = py_2[2*i+1];                    \
 103     dst_2[2*i+1] = r[Y] + g[Y] + b[Y];
 104
 105 #define DST1RGB(i)                                                \
 106     Y = py_1[2*i];                                                \
 107     dst_1[6*i] = r[Y]; dst_1[6*i+1] = g[Y]; dst_1[6*i+2] = b[Y];  \
 108     Y = py_1[2*i+1];                                              \
 109     dst_1[6*i+3] = r[Y]; dst_1[6*i+4] = g[Y]; dst_1[6*i+5] = b[Y];
 110
 111 #define DST2RGB(i)                                                \
 112     Y = py_2[2*i];                                                \
 113     dst_2[6*i] = r[Y]; dst_2[6*i+1] = g[Y]; dst_2[6*i+2] = b[Y];  \
 114     Y = py_2[2*i+1];                                              \
 115     dst_2[6*i+3] = r[Y]; dst_2[6*i+4] = g[Y]; dst_2[6*i+5] = b[Y];
 116
 117 #define DST1BGR(i)                                                \
 118     Y = py_1[2*i];                                                \
 119     dst_1[6*i] = b[Y]; dst_1[6*i+1] = g[Y]; dst_1[6*i+2] = r[Y];  \
 120     Y = py_1[2*i+1];                                              \
 121     dst_1[6*i+3] = b[Y]; dst_1[6*i+4] = g[Y]; dst_1[6*i+5] = r[Y];
 122
 123 #define DST2BGR(i)                                                \
 124     Y = py_2[2*i];                                                \
 125     dst_2[6*i] = b[Y]; dst_2[6*i+1] = g[Y]; dst_2[6*i+2] = r[Y];  \
 126     Y = py_2[2*i+1];                                              \
 127     dst_2[6*i+3] = b[Y]; dst_2[6*i+4] = g[Y]; dst_2[6*i+5] = r[Y];
 128
 129 #define PROLOG(func_name, dst_type) \
 130 static int func_name(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY, \
 131                      int srcSliceH, uint8_t* dst[], int dstStride[]){\
 132     int y;\
 133 \
 134     if (c->srcFormat == PIX_FMT_YUV422P){\
 135         srcStride[1] *= 2;\
 136         srcStride[2] *= 2;\
 137     }\
 138     for (y=0; y<srcSliceH; y+=2){\
 139         dst_type *dst_1= (dst_type*)(dst[0] + (y+srcSliceY  )*dstStride[0]);\
 140         dst_type *dst_2= (dst_type*)(dst[0] + (y+srcSliceY+1)*dstStride[0]);\
 141         dst_type av_unused *r, *b;\
 142         dst_type *g;\
 143         uint8_t *py_1= src[0] + y*srcStride[0];\
 144         uint8_t *py_2= py_1 + srcStride[0];\
 145         uint8_t *pu= src[1] + (y>>1)*srcStride[1];\
 146         uint8_t *pv= src[2] + (y>>1)*srcStride[2];\
 147         unsigned int h_size= c->dstW>>3;\
 148         while (h_size--) {\
 149             int av_unused U, V;\
 150             int Y;\
 151
 152 #define EPILOG1(dst_delta)\
 153             pu += 4;\
 154             pv += 4;\
 155             py_1 += 8;\
 156             py_2 += 8;\
 157             dst_1 += dst_delta;\
 158             dst_2 += dst_delta;\
 159         }\
 160         if (c->dstW & 4) {\
 161             int av_unused Y, U, V;\
 162
 163 #define EPILOG2()\
 164         }\
 165     }\
 166     return srcSliceH;\
 167 }
 168
 169 #define EPILOG(dst_delta)\
 170     EPILOG1(dst_delta)\
 171     EPILOG2()
 172
 173 PROLOG(yuv2rgb_c_32, uint32_t)
 174     RGB(0);
 175     DST1(0);
 176     DST2(0);
 177
 178     RGB(1);
 179     DST2(1);
 180     DST1(1);
 181
 182     RGB(2);
 183     DST1(2);
 184     DST2(2);
 185
 186     RGB(3);
 187     DST2(3);
 188     DST1(3);
 189 EPILOG1(8)
 190     RGB(0);
 191     DST1(0);
 192     DST2(0);
 193
 194     RGB(1);
 195     DST2(1);
 196     DST1(1);
 197 EPILOG2()
 198
 199 PROLOG(yuv2rgb_c_24_rgb, uint8_t)
 200     RGB(0);
 201     DST1RGB(0);
 202     DST2RGB(0);
 203
 204     RGB(1);
 205     DST2RGB(1);
 206     DST1RGB(1);
 207
 208     RGB(2);
 209     DST1RGB(2);
 210     DST2RGB(2);
 211
 212     RGB(3);
 213     DST2RGB(3);
 214     DST1RGB(3);
 215 EPILOG1(24)
 216     RGB(0);
 217     DST1RGB(0);
 218     DST2RGB(0);
 219
 220     RGB(1);
 221     DST2RGB(1);
 222     DST1RGB(1);
 223 EPILOG2()
 224
 225 // only trivial mods from yuv2rgb_c_24_rgb
 226 PROLOG(yuv2rgb_c_24_bgr, uint8_t)
 227     RGB(0);
 228     DST1BGR(0);
 229     DST2BGR(0);
 230
 231     RGB(1);
 232     DST2BGR(1);
 233     DST1BGR(1);
 234
 235     RGB(2);
 236     DST1BGR(2);
 237     DST2BGR(2);
 238
 239     RGB(3);
 240     DST2BGR(3);
 241     DST1BGR(3);
 242 EPILOG1(24)
 243     RGB(0);
 244     DST1BGR(0);
 245     DST2BGR(0);
 246
 247     RGB(1);
 248     DST2BGR(1);
 249     DST1BGR(1);
 250 EPILOG2()
 251
 252 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 253 // r, g, b, dst_1, dst_2
 254 PROLOG(yuv2rgb_c_16, uint16_t)
 255     RGB(0);
 256     DST1(0);
 257     DST2(0);
 258
 259     RGB(1);
 260     DST2(1);
 261     DST1(1);
 262
 263     RGB(2);
 264     DST1(2);
 265     DST2(2);
 266
 267     RGB(3);
 268     DST2(3);
 269     DST1(3);
 270 EPILOG(8)
 271
 272 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 273 // r, g, b, dst_1, dst_2
 274 PROLOG(yuv2rgb_c_8, uint8_t)
 275     RGB(0);
 276     DST1(0);
 277     DST2(0);
 278
 279     RGB(1);
 280     DST2(1);
 281     DST1(1);
 282
 283     RGB(2);
 284     DST1(2);
 285     DST2(2);
 286
 287     RGB(3);
 288     DST2(3);
 289     DST1(3);
 290 EPILOG(8)
 291
 292 // r, g, b, dst_1, dst_2
 293 PROLOG(yuv2rgb_c_8_ordered_dither, uint8_t)
 294     const uint8_t *d32= dither_8x8_32[y&7];
 295     const uint8_t *d64= dither_8x8_73[y&7];
 296 #define DST1bpp8(i,o)                                               \
 297     Y = py_1[2*i];                                                  \
 298     dst_1[2*i]   = r[Y+d32[0+o]] + g[Y+d32[0+o]] + b[Y+d64[0+o]];   \
 299     Y = py_1[2*i+1];                                                \
 300     dst_1[2*i+1] = r[Y+d32[1+o]] + g[Y+d32[1+o]] + b[Y+d64[1+o]];
 301
 302 #define DST2bpp8(i,o)                                               \
 303     Y = py_2[2*i];                                                  \
 304     dst_2[2*i]   =  r[Y+d32[8+o]] + g[Y+d32[8+o]] + b[Y+d64[8+o]];  \
 305     Y = py_2[2*i+1];                                                \
 306     dst_2[2*i+1] =  r[Y+d32[9+o]] + g[Y+d32[9+o]] + b[Y+d64[9+o]];
 307
 308
 309     RGB(0);
 310     DST1bpp8(0,0);
 311     DST2bpp8(0,0);
 312
 313     RGB(1);
 314     DST2bpp8(1,2);
 315     DST1bpp8(1,2);
 316
 317     RGB(2);
 318     DST1bpp8(2,4);
 319     DST2bpp8(2,4);
 320
 321     RGB(3);
 322     DST2bpp8(3,6);
 323     DST1bpp8(3,6);
 324 EPILOG(8)
 325
 326
 327 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 328 // r, g, b, dst_1, dst_2
 329 PROLOG(yuv2rgb_c_4, uint8_t)
 330     int acc;
 331 #define DST1_4(i)                   \
 332     Y = py_1[2*i];                  \
 333     acc = r[Y] + g[Y] + b[Y];       \
 334     Y = py_1[2*i+1];                \
 335     acc |= (r[Y] + g[Y] + b[Y])<<4; \
 336     dst_1[i] = acc;
 337
 338 #define DST2_4(i)                   \
 339     Y = py_2[2*i];                  \
 340     acc = r[Y] + g[Y] + b[Y];       \
 341     Y = py_2[2*i+1];                \
 342     acc |= (r[Y] + g[Y] + b[Y])<<4; \
 343     dst_2[i] = acc;
 344
 345     RGB(0);
 346     DST1_4(0);
 347     DST2_4(0);
 348
 349     RGB(1);
 350     DST2_4(1);
 351     DST1_4(1);
 352
 353     RGB(2);
 354     DST1_4(2);
 355     DST2_4(2);
 356
 357     RGB(3);
 358     DST2_4(3);
 359     DST1_4(3);
 360 EPILOG(4)
 361
 362 PROLOG(yuv2rgb_c_4_ordered_dither, uint8_t)
 363     const uint8_t *d64= dither_8x8_73[y&7];
 364     const uint8_t *d128=dither_8x8_220[y&7];
 365     int acc;
 366
 367 #define DST1bpp4(i,o)                                             \
 368     Y = py_1[2*i];                                                \
 369     acc = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]];        \
 370     Y = py_1[2*i+1];                                              \
 371     acc |= (r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]])<<4;  \
 372     dst_1[i]= acc;
 373
 374 #define DST2bpp4(i,o)                                             \
 375     Y = py_2[2*i];                                                \
 376     acc =  r[Y+d128[8+o]] + g[Y+d64[8+o]] + b[Y+d128[8+o]];       \
 377     Y = py_2[2*i+1];                                              \
 378     acc |=  (r[Y+d128[9+o]] + g[Y+d64[9+o]] + b[Y+d128[9+o]])<<4; \
 379     dst_2[i]= acc;
 380
 381
 382     RGB(0);
 383     DST1bpp4(0,0);
 384     DST2bpp4(0,0);
 385
 386     RGB(1);
 387     DST2bpp4(1,2);
 388     DST1bpp4(1,2);
 389
 390     RGB(2);
 391     DST1bpp4(2,4);
 392     DST2bpp4(2,4);
 393
 394     RGB(3);
 395     DST2bpp4(3,6);
 396     DST1bpp4(3,6);
 397 EPILOG(4)
 398
 399 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 400 // r, g, b, dst_1, dst_2
 401 PROLOG(yuv2rgb_c_4b, uint8_t)
 402     RGB(0);
 403     DST1(0);
 404     DST2(0);
 405
 406     RGB(1);
 407     DST2(1);
 408     DST1(1);
 409
 410     RGB(2);
 411     DST1(2);
 412     DST2(2);
 413
 414     RGB(3);
 415     DST2(3);
 416     DST1(3);
 417 EPILOG(8)
 418
 419 PROLOG(yuv2rgb_c_4b_ordered_dither, uint8_t)
 420     const uint8_t *d64= dither_8x8_73[y&7];
 421     const uint8_t *d128=dither_8x8_220[y&7];
 422
 423 #define DST1bpp4b(i,o)                                                \
 424     Y = py_1[2*i];                                                    \
 425     dst_1[2*i]   = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]];   \
 426     Y = py_1[2*i+1];                                                  \
 427     dst_1[2*i+1] = r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]];
 428
 429 #define DST2bpp4b(i,o)                                                \
 430     Y = py_2[2*i];                                                    \
 431     dst_2[2*i]   =  r[Y+d128[8+o]] + g[Y+d64[8+o]] + b[Y+d128[8+o]];  \
 432     Y = py_2[2*i+1];                                                  \
 433     dst_2[2*i+1] =  r[Y+d128[9+o]] + g[Y+d64[9+o]] + b[Y+d128[9+o]];
 434
 435
 436     RGB(0);
 437     DST1bpp4b(0,0);
 438     DST2bpp4b(0,0);
 439
 440     RGB(1);
 441     DST2bpp4b(1,2);
 442     DST1bpp4b(1,2);
 443
 444     RGB(2);
 445     DST1bpp4b(2,4);
 446     DST2bpp4b(2,4);
 447
 448     RGB(3);
 449     DST2bpp4b(3,6);
 450     DST1bpp4b(3,6);
 451 EPILOG(8)
 452
 453 PROLOG(yuv2rgb_c_1_ordered_dither, uint8_t)
 454         const uint8_t *d128=dither_8x8_220[y&7];
 455         char out_1=0, out_2=0;
 456         g= c->table_gU[128] + c->table_gV[128];
 457
 458 #define DST1bpp1(i,o)               \
 459     Y = py_1[2*i];                  \
 460     out_1+= out_1 + g[Y+d128[0+o]]; \
 461     Y = py_1[2*i+1];                \
 462     out_1+= out_1 + g[Y+d128[1+o]];
 463
 464 #define DST2bpp1(i,o)               \
 465     Y = py_2[2*i];                  \
 466     out_2+= out_2 + g[Y+d128[8+o]]; \
 467     Y = py_2[2*i+1];                \
 468     out_2+= out_2 + g[Y+d128[9+o]];
 469
 470     DST1bpp1(0,0);
 471     DST2bpp1(0,0);
 472
 473     DST2bpp1(1,2);
 474     DST1bpp1(1,2);
 475
 476     DST1bpp1(2,4);
 477     DST2bpp1(2,4);
 478
 479     DST2bpp1(3,6);
 480     DST1bpp1(3,6);
 481
 482     dst_1[0]= out_1;
 483     dst_2[0]= out_2;
 484 EPILOG(1)
 485
 486 SwsFunc yuv2rgb_get_func_ptr (SwsContext *c)
 487 {
 488 #if defined(HAVE_MMX2) || defined(HAVE_MMX)
 489     if (c->flags & SWS_CPU_CAPS_MMX2){
 490         switch(c->dstFormat){
 491         case PIX_FMT_RGB32:  return yuv420_rgb32_MMX2;
 492         case PIX_FMT_BGR24:  return yuv420_rgb24_MMX2;
 493         case PIX_FMT_RGB565: return yuv420_rgb16_MMX2;
 494         case PIX_FMT_RGB555: return yuv420_rgb15_MMX2;
 495         }
 496     }
 497     if (c->flags & SWS_CPU_CAPS_MMX){
 498         switch(c->dstFormat){
 499         case PIX_FMT_RGB32:  return yuv420_rgb32_MMX;
 500         case PIX_FMT_BGR24:  return yuv420_rgb24_MMX;
 501         case PIX_FMT_RGB565: return yuv420_rgb16_MMX;
 502         case PIX_FMT_RGB555: return yuv420_rgb15_MMX;
 503         }
 504     }
 505 #endif
 506 #ifdef HAVE_VIS
 507     {
 508         SwsFunc t= yuv2rgb_init_vis(c);
 509         if (t) return t;
 510     }
 511 #endif
 512 #ifdef CONFIG_MLIB
 513     {
 514         SwsFunc t= yuv2rgb_init_mlib(c);
 515         if (t) return t;
 516     }
 517 #endif
 518 #ifdef HAVE_ALTIVEC
 519     if (c->flags & SWS_CPU_CAPS_ALTIVEC)
 520     {
 521         SwsFunc t = yuv2rgb_init_altivec(c);
 522         if (t) return t;
 523     }
 524 #endif
 525
 526 #ifdef ARCH_BFIN
 527     if (c->flags & SWS_CPU_CAPS_BFIN)
 528     {
 529         SwsFunc t = ff_bfin_yuv2rgb_get_func_ptr (c);
 530         if (t) return t;
 531     }
 532 #endif
 533
 534     av_log(c, AV_LOG_WARNING, "No accelerated colorspace conversion found.\n");
 535
 536     switch(c->dstFormat){
 537     case PIX_FMT_BGR32_1:
 538     case PIX_FMT_RGB32_1:
 539     case PIX_FMT_BGR32:
 540     case PIX_FMT_RGB32: return yuv2rgb_c_32;
 541     case PIX_FMT_RGB24: return yuv2rgb_c_24_rgb;
 542     case PIX_FMT_BGR24: return yuv2rgb_c_24_bgr;
 543     case PIX_FMT_RGB565:
 544     case PIX_FMT_BGR565:
 545     case PIX_FMT_RGB555:
 546     case PIX_FMT_BGR555: return yuv2rgb_c_16;
 547     case PIX_FMT_RGB8:
 548     case PIX_FMT_BGR8:  return yuv2rgb_c_8_ordered_dither;
 549     case PIX_FMT_RGB4:
 550     case PIX_FMT_BGR4:  return yuv2rgb_c_4_ordered_dither;
 551     case PIX_FMT_RGB4_BYTE:
 552     case PIX_FMT_BGR4_BYTE:  return yuv2rgb_c_4b_ordered_dither;
 553     case PIX_FMT_MONOBLACK:  return yuv2rgb_c_1_ordered_dither;
 554     default:
 555         assert(0);
 556     }
 557     return NULL;
 558 }
 559
 560 static int div_round (int dividend, int divisor)
 561 {
 562     if (dividend > 0)
 563         return (dividend + (divisor>>1)) / divisor;
 564     else
 565         return -((-dividend + (divisor>>1)) / divisor);
 566 }
 567
 568 int yuv2rgb_c_init_tables (SwsContext *c, const int inv_table[4], int fullRange, int brightness, int contrast, int saturation)
 569 {
 570     const int isRgb =      c->dstFormat==PIX_FMT_RGB32
 571                         || c->dstFormat==PIX_FMT_RGB32_1
 572                         || c->dstFormat==PIX_FMT_BGR24
 573                         || c->dstFormat==PIX_FMT_RGB565
 574                         || c->dstFormat==PIX_FMT_RGB555
 575                         || c->dstFormat==PIX_FMT_RGB8
 576                         || c->dstFormat==PIX_FMT_RGB4
 577                         || c->dstFormat==PIX_FMT_RGB4_BYTE
 578                         || c->dstFormat==PIX_FMT_MONOBLACK;
 579     const int bpp = fmt_depth(c->dstFormat);
 580     int i, base;
 581     uint8_t table_Y[1024];
 582     uint32_t *table_32 = 0;
 583     uint16_t *table_16 = 0;
 584     uint8_t *table_8 = 0;
 585     uint8_t *table_332 = 0;
 586     uint8_t *table_121 = 0;
 587     uint8_t *table_1 = 0;
 588     int entry_size = 0;
 589     void *table_r = 0, *table_g = 0, *table_b = 0;
 590     void *table_start;
 591
 592     int64_t crv =  inv_table[0];
 593     int64_t cbu =  inv_table[1];
 594     int64_t cgu = -inv_table[2];
 595     int64_t cgv = -inv_table[3];
 596     int64_t cy  = 1<<16;
 597     int64_t oy  = 0;
 598
 599 //printf("%lld %lld %lld %lld %lld\n", cy, crv, cbu, cgu, cgv);
 600     if (!fullRange){
 601         cy= (cy*255) / 219;
 602         oy= 16<<16;
 603     }else{
 604         crv= (crv*224) / 255;
 605         cbu= (cbu*224) / 255;
 606         cgu= (cgu*224) / 255;
 607         cgv= (cgv*224) / 255;
 608     }
 609
 610     cy = (cy *contrast             )>>16;
 611     crv= (crv*contrast * saturation)>>32;
 612     cbu= (cbu*contrast * saturation)>>32;
 613     cgu= (cgu*contrast * saturation)>>32;
 614     cgv= (cgv*contrast * saturation)>>32;
 615 //printf("%lld %lld %lld %lld %lld\n", cy, crv, cbu, cgu, cgv);
 616     oy -= 256*brightness;
 617
 618     for (i = 0; i < 1024; i++) {
 619         int j;
 620
 621         j= (cy*(((i - 384)<<16) - oy) + (1<<31))>>32;
 622         j = (j < 0) ? 0 : ((j > 255) ? 255 : j);
 623         table_Y[i] = j;
 624     }
 625
 626     switch (bpp) {
 627     case 32:
 628         table_start= table_32 = av_malloc ((197 + 2*682 + 256 + 132) * sizeof (uint32_t));
 629         base= (c->dstFormat == PIX_FMT_RGB32_1 || c->dstFormat == PIX_FMT_BGR32_1) ? 8 : 0;
 630
 631         entry_size = sizeof (uint32_t);
 632         table_r = table_32 + 197;
 633         table_b = table_32 + 197 + 685;
 634         table_g = table_32 + 197 + 2*682;
 635
 636         for (i = -197; i < 256+197; i++)
 637             ((uint32_t *)table_r)[i] = table_Y[i+384] << ((isRgb ? 16 : 0) + base);
 638         for (i = -132; i < 256+132; i++)
 639             ((uint32_t *)table_g)[i] = table_Y[i+384] << (8                + base);
 640         for (i = -232; i < 256+232; i++)
 641             ((uint32_t *)table_b)[i] = table_Y[i+384] << ((isRgb ? 0 : 16) + base);
 642         break;
 643
 644     case 24:
 645         table_start= table_8 = av_malloc ((256 + 2*232) * sizeof (uint8_t));
 646
 647         entry_size = sizeof (uint8_t);
 648         table_r = table_g = table_b = table_8 + 232;
 649
 650         for (i = -232; i < 256+232; i++)
 651             ((uint8_t * )table_b)[i] = table_Y[i+384];
 652         break;
 653
 654     case 15:
 655     case 16:
 656         table_start= table_16 = av_malloc ((197 + 2*682 + 256 + 132) * sizeof (uint16_t));
 657
 658         entry_size = sizeof (uint16_t);
 659         table_r = table_16 + 197;
 660         table_b = table_16 + 197 + 685;
 661         table_g = table_16 + 197 + 2*682;
 662
 663         for (i = -197; i < 256+197; i++) {
 664             int j = table_Y[i+384] >> 3;
 665
 666             if (isRgb)
 667                 j <<= ((bpp==16) ? 11 : 10);
 668
 669             ((uint16_t *)table_r)[i] = j;
 670         }
 671         for (i = -132; i < 256+132; i++) {
 672             int j = table_Y[i+384] >> ((bpp==16) ? 2 : 3);
 673
 674             ((uint16_t *)table_g)[i] = j << 5;
 675         }
 676         for (i = -232; i < 256+232; i++) {
 677             int j = table_Y[i+384] >> 3;
 678
 679             if (!isRgb)
 680                 j <<= ((bpp==16) ? 11 : 10);
 681
 682             ((uint16_t *)table_b)[i] = j;
 683         }
 684         break;
 685
 686     case 8:
 687         table_start= table_332 = av_malloc ((197 + 2*682 + 256 + 132) * sizeof (uint8_t));
 688
 689         entry_size = sizeof (uint8_t);
 690         table_r = table_332 + 197;
 691         table_b = table_332 + 197 + 685;
 692         table_g = table_332 + 197 + 2*682;
 693
 694         for (i = -197; i < 256+197; i++) {
 695             int j = (table_Y[i+384 - 16] + 18)/36;
 696
 697             if (isRgb)
 698                 j <<= 5;
 699
 700             ((uint8_t *)table_r)[i] = j;
 701         }
 702         for (i = -132; i < 256+132; i++) {
 703             int j = (table_Y[i+384 - 16] + 18)/36;
 704
 705             if (!isRgb)
 706                 j <<= 1;
 707
 708             ((uint8_t *)table_g)[i] = j << 2;
 709         }
 710         for (i = -232; i < 256+232; i++) {
 711             int j = (table_Y[i+384 - 37] + 43)/85;
 712
 713             if (!isRgb)
 714                 j <<= 6;
 715
 716             ((uint8_t *)table_b)[i] = j;
 717         }
 718         break;
 719     case 4:
 720     case 4|128:
 721         table_start= table_121 = av_malloc ((197 + 2*682 + 256 + 132) * sizeof (uint8_t));
 722
 723         entry_size = sizeof (uint8_t);
 724         table_r = table_121 + 197;
 725         table_b = table_121 + 197 + 685;
 726         table_g = table_121 + 197 + 2*682;
 727
 728         for (i = -197; i < 256+197; i++) {
 729             int j = table_Y[i+384 - 110] >> 7;
 730
 731             if (isRgb)
 732                 j <<= 3;
 733
 734             ((uint8_t *)table_r)[i] = j;
 735         }
 736         for (i = -132; i < 256+132; i++) {
 737             int j = (table_Y[i+384 - 37]+ 43)/85;
 738
 739             ((uint8_t *)table_g)[i] = j << 1;
 740         }
 741         for (i = -232; i < 256+232; i++) {
 742             int j =table_Y[i+384 - 110] >> 7;
 743
 744             if (!isRgb)
 745                 j <<= 3;
 746
 747             ((uint8_t *)table_b)[i] = j;
 748         }
 749         break;
 750
 751     case 1:
 752         table_start= table_1 = av_malloc (256*2 * sizeof (uint8_t));
 753
 754         entry_size = sizeof (uint8_t);
 755         table_g = table_1;
 756         table_r = table_b = NULL;
 757
 758         for (i = 0; i < 256+256; i++) {
 759             int j = table_Y[i + 384 - 110]>>7;
 760
 761             ((uint8_t *)table_g)[i] = j;
 762         }
 763         break;
 764
 765     default:
 766         table_start= NULL;
 767         av_log(c, AV_LOG_ERROR, "%ibpp not supported by yuv2rgb\n", bpp);
 768         //free mem?
 769         return -1;
 770     }
 771
 772     for (i = 0; i < 256; i++) {
 773         c->table_rV[i] = (uint8_t *)table_r + entry_size * div_round (crv * (i-128), cy);
 774         c->table_gU[i] = (uint8_t *)table_g + entry_size * div_round (cgu * (i-128), cy);
 775         c->table_gV[i] = entry_size * div_round (cgv * (i-128), cy);
 776         c->table_bU[i] = (uint8_t *)table_b + entry_size * div_round (cbu * (i-128), cy);
 777     }
 778
 779     av_free(c->yuvTable);
 780     c->yuvTable= table_start;
 781     return 0;
 782 }