libswscale/yuv2rgb.c

   1 /*
   2  * software YUV to RGB converter
   3  *
   4  * Copyright (C) 2009 Konstantin Shishkov
   5  *
   6  * 1,4,8bpp support and context / deglobalize stuff
   7  * by Michael Niedermayer (michaelni@gmx.at)
   8  *
   9  * This file is part of FFmpeg.
  10  *
  11  * FFmpeg is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; either
  14  * version 2.1 of the License, or (at your option) any later version.
  15  *
  16  * FFmpeg is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with FFmpeg; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  24  */
  25
  26 #include <stdio.h>
  27 #include <stdlib.h>
  28 #include <inttypes.h>
  29 #include <assert.h>
  30
  31 #include "config.h"
  32 #include "rgb2rgb.h"
  33 #include "swscale.h"
  34 #include "swscale_internal.h"
  35 #include "libavutil/x86_cpu.h"
  36
  37 extern const uint8_t dither_8x8_32[8][8];
  38 extern const uint8_t dither_8x8_73[8][8];
  39 extern const uint8_t dither_8x8_220[8][8];
  40
  41 const int32_t ff_yuv2rgb_coeffs[8][4] = {
  42     {117504, 138453, 13954, 34903}, /* no sequence_display_extension */
  43     {117504, 138453, 13954, 34903}, /* ITU-R Rec. 709 (1990) */
  44     {104597, 132201, 25675, 53279}, /* unspecified */
  45     {104597, 132201, 25675, 53279}, /* reserved */
  46     {104448, 132798, 24759, 53109}, /* FCC */
  47     {104597, 132201, 25675, 53279}, /* ITU-R Rec. 624-4 System B, G */
  48     {104597, 132201, 25675, 53279}, /* SMPTE 170M */
  49     {117579, 136230, 16907, 35559}  /* SMPTE 240M (1987) */
  50 };
  51
  52 #define LOADCHROMA(i)                               \
  53     U = pu[i];                                      \
  54     V = pv[i];                                      \
  55     r = (void *)c->table_rV[V];                     \
  56     g = (void *)(c->table_gU[U] + c->table_gV[V]);  \
  57     b = (void *)c->table_bU[U];
  58
  59 #define PUTRGB(dst,src,i,o)          \
  60     Y = src[2*i+o];                  \
  61     dst[2*i  ] = r[Y] + g[Y] + b[Y]; \
  62     Y = src[2*i+1-o];                \
  63     dst[2*i+1] = r[Y] + g[Y] + b[Y];
  64
  65 #define PUTRGB24(dst,src,i)                                  \
  66     Y = src[2*i];                                            \
  67     dst[6*i+0] = r[Y]; dst[6*i+1] = g[Y]; dst[6*i+2] = b[Y]; \
  68     Y = src[2*i+1];                                          \
  69     dst[6*i+3] = r[Y]; dst[6*i+4] = g[Y]; dst[6*i+5] = b[Y];
  70
  71 #define PUTBGR24(dst,src,i)                                  \
  72     Y = src[2*i];                                            \
  73     dst[6*i+0] = b[Y]; dst[6*i+1] = g[Y]; dst[6*i+2] = r[Y]; \
  74     Y = src[2*i+1];                                          \
  75     dst[6*i+3] = b[Y]; dst[6*i+4] = g[Y]; dst[6*i+5] = r[Y];
  76
  77 #define PUTRGBA(dst,ysrc,asrc,i,o,s)                    \
  78     Y = ysrc[2*i+o];                                    \
  79     dst[2*i  ] = r[Y] + g[Y] + b[Y] + (asrc[2*i  ]<<s); \
  80     Y = ysrc[2*i+1-o];                                  \
  81     dst[2*i+1] = r[Y] + g[Y] + b[Y] + (asrc[2*i+1]<<s);
  82
  83 #define YUV2RGBFUNC(func_name, dst_type, alpha) \
  84 static int func_name(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY, \
  85                      int srcSliceH, uint8_t* dst[], int dstStride[]){\
  86     int y;\
  87 \
  88     if (!alpha && c->srcFormat == PIX_FMT_YUV422P) {\
  89         srcStride[1] *= 2;\
  90         srcStride[2] *= 2;\
  91     }\
  92     for (y=0; y<srcSliceH; y+=2) {\
  93         dst_type *dst_1 = (dst_type*)(dst[0] + (y+srcSliceY  )*dstStride[0]);\
  94         dst_type *dst_2 = (dst_type*)(dst[0] + (y+srcSliceY+1)*dstStride[0]);\
  95         dst_type av_unused *r, *b;\
  96         dst_type *g;\
  97         uint8_t *py_1 = src[0] + y*srcStride[0];\
  98         uint8_t *py_2 = py_1 + srcStride[0];\
  99         uint8_t *pu = src[1] + (y>>1)*srcStride[1];\
 100         uint8_t *pv = src[2] + (y>>1)*srcStride[2];\
 101         uint8_t av_unused *pa_1, *pa_2;\
 102         unsigned int h_size = c->dstW>>3;\
 103         if (alpha){\
 104             pa_1 = src[3] + y*srcStride[3];\
 105             pa_2 = pa_1 + srcStride[3];\
 106         }\
 107         while (h_size--) {\
 108             int av_unused U, V;\
 109             int Y;\
 110
 111 #define ENDYUV2RGBLINE(dst_delta)\
 112             pu += 4;\
 113             pv += 4;\
 114             py_1 += 8;\
 115             py_2 += 8;\
 116             dst_1 += dst_delta;\
 117             dst_2 += dst_delta;\
 118         }\
 119         if (c->dstW & 4) {\
 120             int av_unused Y, U, V;\
 121
 122 #define ENDYUV2RGBFUNC()\
 123         }\
 124     }\
 125     return srcSliceH;\
 126 }
 127
 128 #define CLOSEYUV2RGBFUNC(dst_delta)\
 129     ENDYUV2RGBLINE(dst_delta)\
 130     ENDYUV2RGBFUNC()
 131
 132 YUV2RGBFUNC(yuv2rgb_c_32, uint32_t, 0)
 133     LOADCHROMA(0);
 134     PUTRGB(dst_1,py_1,0,0);
 135     PUTRGB(dst_2,py_2,0,1);
 136
 137     LOADCHROMA(1);
 138     PUTRGB(dst_2,py_2,1,1);
 139     PUTRGB(dst_1,py_1,1,0);
 140
 141     LOADCHROMA(2);
 142     PUTRGB(dst_1,py_1,2,0);
 143     PUTRGB(dst_2,py_2,2,1);
 144
 145     LOADCHROMA(3);
 146     PUTRGB(dst_2,py_2,3,1);
 147     PUTRGB(dst_1,py_1,3,0);
 148 ENDYUV2RGBLINE(8)
 149     LOADCHROMA(0);
 150     PUTRGB(dst_1,py_1,0,0);
 151     PUTRGB(dst_2,py_2,0,1);
 152
 153     LOADCHROMA(1);
 154     PUTRGB(dst_2,py_2,1,1);
 155     PUTRGB(dst_1,py_1,1,0);
 156 ENDYUV2RGBFUNC()
 157
 158 YUV2RGBFUNC(yuva2rgba_c, uint32_t, 1)
 159     LOADCHROMA(0);
 160     PUTRGBA(dst_1,py_1,pa_1,0,0,24);
 161     PUTRGBA(dst_2,py_2,pa_2,0,1,24);
 162
 163     LOADCHROMA(1);
 164     PUTRGBA(dst_2,py_2,pa_1,1,1,24);
 165     PUTRGBA(dst_1,py_1,pa_2,1,0,24);
 166
 167     LOADCHROMA(2);
 168     PUTRGBA(dst_1,py_1,pa_1,2,0,24);
 169     PUTRGBA(dst_2,py_2,pa_2,2,1,24);
 170
 171     LOADCHROMA(3);
 172     PUTRGBA(dst_2,py_2,pa_1,3,1,24);
 173     PUTRGBA(dst_1,py_1,pa_2,3,0,24);
 174     pa_1 += 8;\
 175     pa_2 += 8;\
 176 ENDYUV2RGBLINE(8)
 177     LOADCHROMA(0);
 178     PUTRGBA(dst_1,py_1,pa_1,0,0,24);
 179     PUTRGBA(dst_2,py_2,pa_2,0,1,24);
 180
 181     LOADCHROMA(1);
 182     PUTRGBA(dst_2,py_2,pa_1,1,1,24);
 183     PUTRGBA(dst_1,py_1,pa_2,1,0,24);
 184 ENDYUV2RGBFUNC()
 185
 186 YUV2RGBFUNC(yuva2argb_c, uint32_t, 1)
 187     LOADCHROMA(0);
 188     PUTRGBA(dst_1,py_1,pa_1,0,0,0);
 189     PUTRGBA(dst_2,py_2,pa_2,0,1,0);
 190
 191     LOADCHROMA(1);
 192     PUTRGBA(dst_2,py_2,pa_2,1,1,0);
 193     PUTRGBA(dst_1,py_1,pa_1,1,0,0);
 194
 195     LOADCHROMA(2);
 196     PUTRGBA(dst_1,py_1,pa_1,2,0,0);
 197     PUTRGBA(dst_2,py_2,pa_2,2,1,0);
 198
 199     LOADCHROMA(3);
 200     PUTRGBA(dst_2,py_2,pa_2,3,1,0);
 201     PUTRGBA(dst_1,py_1,pa_1,3,0,0);
 202     pa_1 += 8;\
 203     pa_2 += 8;\
 204 ENDYUV2RGBLINE(8)
 205     LOADCHROMA(0);
 206     PUTRGBA(dst_1,py_1,pa_1,0,0,0);
 207     PUTRGBA(dst_2,py_2,pa_2,0,1,0);
 208
 209     LOADCHROMA(1);
 210     PUTRGBA(dst_2,py_2,pa_2,1,1,0);
 211     PUTRGBA(dst_1,py_1,pa_1,1,0,0);
 212 ENDYUV2RGBFUNC()
 213
 214 YUV2RGBFUNC(yuv2rgb_c_24_rgb, uint8_t, 0)
 215     LOADCHROMA(0);
 216     PUTRGB24(dst_1,py_1,0);
 217     PUTRGB24(dst_2,py_2,0);
 218
 219     LOADCHROMA(1);
 220     PUTRGB24(dst_2,py_2,1);
 221     PUTRGB24(dst_1,py_1,1);
 222
 223     LOADCHROMA(2);
 224     PUTRGB24(dst_1,py_1,2);
 225     PUTRGB24(dst_2,py_2,2);
 226
 227     LOADCHROMA(3);
 228     PUTRGB24(dst_2,py_2,3);
 229     PUTRGB24(dst_1,py_1,3);
 230 ENDYUV2RGBLINE(24)
 231     LOADCHROMA(0);
 232     PUTRGB24(dst_1,py_1,0);
 233     PUTRGB24(dst_2,py_2,0);
 234
 235     LOADCHROMA(1);
 236     PUTRGB24(dst_2,py_2,1);
 237     PUTRGB24(dst_1,py_1,1);
 238 ENDYUV2RGBFUNC()
 239
 240 // only trivial mods from yuv2rgb_c_24_rgb
 241 YUV2RGBFUNC(yuv2rgb_c_24_bgr, uint8_t, 0)
 242     LOADCHROMA(0);
 243     PUTBGR24(dst_1,py_1,0);
 244     PUTBGR24(dst_2,py_2,0);
 245
 246     LOADCHROMA(1);
 247     PUTBGR24(dst_2,py_2,1);
 248     PUTBGR24(dst_1,py_1,1);
 249
 250     LOADCHROMA(2);
 251     PUTBGR24(dst_1,py_1,2);
 252     PUTBGR24(dst_2,py_2,2);
 253
 254     LOADCHROMA(3);
 255     PUTBGR24(dst_2,py_2,3);
 256     PUTBGR24(dst_1,py_1,3);
 257 ENDYUV2RGBLINE(24)
 258     LOADCHROMA(0);
 259     PUTBGR24(dst_1,py_1,0);
 260     PUTBGR24(dst_2,py_2,0);
 261
 262     LOADCHROMA(1);
 263     PUTBGR24(dst_2,py_2,1);
 264     PUTBGR24(dst_1,py_1,1);
 265 ENDYUV2RGBFUNC()
 266
 267 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 268 // r, g, b, dst_1, dst_2
 269 YUV2RGBFUNC(yuv2rgb_c_16, uint16_t, 0)
 270     LOADCHROMA(0);
 271     PUTRGB(dst_1,py_1,0,0);
 272     PUTRGB(dst_2,py_2,0,1);
 273
 274     LOADCHROMA(1);
 275     PUTRGB(dst_2,py_2,1,1);
 276     PUTRGB(dst_1,py_1,1,0);
 277
 278     LOADCHROMA(2);
 279     PUTRGB(dst_1,py_1,2,0);
 280     PUTRGB(dst_2,py_2,2,1);
 281
 282     LOADCHROMA(3);
 283     PUTRGB(dst_2,py_2,3,1);
 284     PUTRGB(dst_1,py_1,3,0);
 285 CLOSEYUV2RGBFUNC(8)
 286
 287 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 288 // r, g, b, dst_1, dst_2
 289 YUV2RGBFUNC(yuv2rgb_c_8, uint8_t, 0)
 290     LOADCHROMA(0);
 291     PUTRGB(dst_1,py_1,0,0);
 292     PUTRGB(dst_2,py_2,0,1);
 293
 294     LOADCHROMA(1);
 295     PUTRGB(dst_2,py_2,1,1);
 296     PUTRGB(dst_1,py_1,1,0);
 297
 298     LOADCHROMA(2);
 299     PUTRGB(dst_1,py_1,2,0);
 300     PUTRGB(dst_2,py_2,2,1);
 301
 302     LOADCHROMA(3);
 303     PUTRGB(dst_2,py_2,3,1);
 304     PUTRGB(dst_1,py_1,3,0);
 305 CLOSEYUV2RGBFUNC(8)
 306
 307 // r, g, b, dst_1, dst_2
 308 YUV2RGBFUNC(yuv2rgb_c_8_ordered_dither, uint8_t, 0)
 309     const uint8_t *d32 = dither_8x8_32[y&7];
 310     const uint8_t *d64 = dither_8x8_73[y&7];
 311 #define PUTRGB8(dst,src,i,o)                                    \
 312     Y = src[2*i];                                               \
 313     dst[2*i]   = r[Y+d32[0+o]] + g[Y+d32[0+o]] + b[Y+d64[0+o]]; \
 314     Y = src[2*i+1];                                             \
 315     dst[2*i+1] = r[Y+d32[1+o]] + g[Y+d32[1+o]] + b[Y+d64[1+o]];
 316
 317     LOADCHROMA(0);
 318     PUTRGB8(dst_1,py_1,0,0);
 319     PUTRGB8(dst_2,py_2,0,0+8);
 320
 321     LOADCHROMA(1);
 322     PUTRGB8(dst_2,py_2,1,2+8);
 323     PUTRGB8(dst_1,py_1,1,2);
 324
 325     LOADCHROMA(2);
 326     PUTRGB8(dst_1,py_1,2,4);
 327     PUTRGB8(dst_2,py_2,2,4+8);
 328
 329     LOADCHROMA(3);
 330     PUTRGB8(dst_2,py_2,3,6+8);
 331     PUTRGB8(dst_1,py_1,3,6);
 332 CLOSEYUV2RGBFUNC(8)
 333
 334
 335 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 336 // r, g, b, dst_1, dst_2
 337 YUV2RGBFUNC(yuv2rgb_c_4, uint8_t, 0)
 338     int acc;
 339 #define PUTRGB4(dst,src,i)          \
 340     Y = src[2*i];                   \
 341     acc = r[Y] + g[Y] + b[Y];       \
 342     Y = src[2*i+1];                 \
 343     acc |= (r[Y] + g[Y] + b[Y])<<4; \
 344     dst[i] = acc;
 345
 346     LOADCHROMA(0);
 347     PUTRGB4(dst_1,py_1,0);
 348     PUTRGB4(dst_2,py_2,0);
 349
 350     LOADCHROMA(1);
 351     PUTRGB4(dst_2,py_2,1);
 352     PUTRGB4(dst_1,py_1,1);
 353
 354     LOADCHROMA(2);
 355     PUTRGB4(dst_1,py_1,2);
 356     PUTRGB4(dst_2,py_2,2);
 357
 358     LOADCHROMA(3);
 359     PUTRGB4(dst_2,py_2,3);
 360     PUTRGB4(dst_1,py_1,3);
 361 CLOSEYUV2RGBFUNC(4)
 362
 363 YUV2RGBFUNC(yuv2rgb_c_4_ordered_dither, uint8_t, 0)
 364     const uint8_t *d64 =  dither_8x8_73[y&7];
 365     const uint8_t *d128 = dither_8x8_220[y&7];
 366     int acc;
 367
 368 #define PUTRGB4D(dst,src,i,o)                                     \
 369     Y = src[2*i];                                                 \
 370     acc = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]];        \
 371     Y = src[2*i+1];                                               \
 372     acc |= (r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]])<<4;  \
 373     dst[i]= acc;
 374
 375     LOADCHROMA(0);
 376     PUTRGB4D(dst_1,py_1,0,0);
 377     PUTRGB4D(dst_2,py_2,0,0+8);
 378
 379     LOADCHROMA(1);
 380     PUTRGB4D(dst_2,py_2,1,2+8);
 381     PUTRGB4D(dst_1,py_1,1,2);
 382
 383     LOADCHROMA(2);
 384     PUTRGB4D(dst_1,py_1,2,4);
 385     PUTRGB4D(dst_2,py_2,2,4+8);
 386
 387     LOADCHROMA(3);
 388     PUTRGB4D(dst_2,py_2,3,6+8);
 389     PUTRGB4D(dst_1,py_1,3,6);
 390 CLOSEYUV2RGBFUNC(4)
 391
 392 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 393 // r, g, b, dst_1, dst_2
 394 YUV2RGBFUNC(yuv2rgb_c_4b, uint8_t, 0)
 395     LOADCHROMA(0);
 396     PUTRGB(dst_1,py_1,0,0);
 397     PUTRGB(dst_2,py_2,0,1);
 398
 399     LOADCHROMA(1);
 400     PUTRGB(dst_2,py_2,1,1);
 401     PUTRGB(dst_1,py_1,1,0);
 402
 403     LOADCHROMA(2);
 404     PUTRGB(dst_1,py_1,2,0);
 405     PUTRGB(dst_2,py_2,2,1);
 406
 407     LOADCHROMA(3);
 408     PUTRGB(dst_2,py_2,3,1);
 409     PUTRGB(dst_1,py_1,3,0);
 410 CLOSEYUV2RGBFUNC(8)
 411
 412 YUV2RGBFUNC(yuv2rgb_c_4b_ordered_dither, uint8_t, 0)
 413     const uint8_t *d64 =  dither_8x8_73[y&7];
 414     const uint8_t *d128 = dither_8x8_220[y&7];
 415
 416 #define PUTRGB4DB(dst,src,i,o)                                    \
 417     Y = src[2*i];                                                 \
 418     dst[2*i]   = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]]; \
 419     Y = src[2*i+1];                                               \
 420     dst[2*i+1] = r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]];
 421
 422     LOADCHROMA(0);
 423     PUTRGB4DB(dst_1,py_1,0,0);
 424     PUTRGB4DB(dst_2,py_2,0,0+8);
 425
 426     LOADCHROMA(1);
 427     PUTRGB4DB(dst_2,py_2,1,2+8);
 428     PUTRGB4DB(dst_1,py_1,1,2);
 429
 430     LOADCHROMA(2);
 431     PUTRGB4DB(dst_1,py_1,2,4);
 432     PUTRGB4DB(dst_2,py_2,2,4+8);
 433
 434     LOADCHROMA(3);
 435     PUTRGB4DB(dst_2,py_2,3,6+8);
 436     PUTRGB4DB(dst_1,py_1,3,6);
 437 CLOSEYUV2RGBFUNC(8)
 438
 439 YUV2RGBFUNC(yuv2rgb_c_1_ordered_dither, uint8_t, 0)
 440         const uint8_t *d128 = dither_8x8_220[y&7];
 441         char out_1 = 0, out_2 = 0;
 442         g= c->table_gU[128] + c->table_gV[128];
 443
 444 #define PUTRGB1(out,src,i,o)    \
 445     Y = src[2*i];               \
 446     out+= out + g[Y+d128[0+o]]; \
 447     Y = src[2*i+1];             \
 448     out+= out + g[Y+d128[1+o]];
 449
 450     PUTRGB1(out_1,py_1,0,0);
 451     PUTRGB1(out_2,py_2,0,0+8);
 452
 453     PUTRGB1(out_2,py_2,1,2+8);
 454     PUTRGB1(out_1,py_1,1,2);
 455
 456     PUTRGB1(out_1,py_1,2,4);
 457     PUTRGB1(out_2,py_2,2,4+8);
 458
 459     PUTRGB1(out_2,py_2,3,6+8);
 460     PUTRGB1(out_1,py_1,3,6);
 461
 462     dst_1[0]= out_1;
 463     dst_2[0]= out_2;
 464 CLOSEYUV2RGBFUNC(1)
 465
 466 SwsFunc ff_yuv2rgb_get_func_ptr(SwsContext *c)
 467 {
 468     SwsFunc t = NULL;
 469 #if (HAVE_MMX2 || HAVE_MMX) && CONFIG_GPL
 470      t = ff_yuv2rgb_init_mmx(c);
 471 #endif
 472 #if HAVE_VIS
 473     t = ff_yuv2rgb_init_vis(c);
 474 #endif
 475 #if CONFIG_MLIB
 476     t = ff_yuv2rgb_init_mlib(c);
 477 #endif
 478 #if HAVE_ALTIVEC && CONFIG_GPL
 479     if (c->flags & SWS_CPU_CAPS_ALTIVEC)
 480         t = ff_yuv2rgb_init_altivec(c);
 481 #endif
 482
 483 #if ARCH_BFIN
 484     if (c->flags & SWS_CPU_CAPS_BFIN)
 485         t = ff_yuv2rgb_get_func_ptr_bfin(c);
 486 #endif
 487
 488     if (t)
 489         return t;
 490
 491     av_log(c, AV_LOG_WARNING, "No accelerated colorspace conversion found.\n");
 492
 493     switch (c->dstFormat) {
 494     case PIX_FMT_ARGB:
 495     case PIX_FMT_ABGR:       if (CONFIG_SWSCALE_ALPHA && c->srcFormat == PIX_FMT_YUVA420P) return yuva2argb_c;
 496     case PIX_FMT_RGBA:
 497     case PIX_FMT_BGRA:       return (CONFIG_SWSCALE_ALPHA && c->srcFormat == PIX_FMT_YUVA420P) ? yuva2rgba_c : yuv2rgb_c_32;
 498     case PIX_FMT_RGB24:      return yuv2rgb_c_24_rgb;
 499     case PIX_FMT_BGR24:      return yuv2rgb_c_24_bgr;
 500     case PIX_FMT_RGB565:
 501     case PIX_FMT_BGR565:
 502     case PIX_FMT_RGB555:
 503     case PIX_FMT_BGR555:     return yuv2rgb_c_16;
 504     case PIX_FMT_RGB8:
 505     case PIX_FMT_BGR8:       return yuv2rgb_c_8_ordered_dither;
 506     case PIX_FMT_RGB4:
 507     case PIX_FMT_BGR4:       return yuv2rgb_c_4_ordered_dither;
 508     case PIX_FMT_RGB4_BYTE:
 509     case PIX_FMT_BGR4_BYTE:  return yuv2rgb_c_4b_ordered_dither;
 510     case PIX_FMT_MONOBLACK:  return yuv2rgb_c_1_ordered_dither;
 511     default:
 512         assert(0);
 513     }
 514     return NULL;
 515 }
 516
 517 static void fill_table(uint8_t* table[256], const int elemsize, const int inc, uint8_t *y_table)
 518 {
 519     int i;
 520     int64_t cb = 0;
 521
 522     y_table -= elemsize * (inc >> 9);
 523
 524     for (i = 0; i < 256; i++) {
 525         table[i] = y_table + elemsize * (cb >> 16);
 526         cb += inc;
 527     }
 528 }
 529
 530 static void fill_gv_table(int table[256], const int elemsize, const int inc)
 531 {
 532     int i;
 533     int64_t cb = 0;
 534     int off = -(inc >> 9);
 535
 536     for (i = 0; i < 256; i++) {
 537         table[i] = elemsize * (off + (cb >> 16));
 538         cb += inc;
 539     }
 540 }
 541
 542 av_cold int ff_yuv2rgb_c_init_tables(SwsContext *c, const int inv_table[4], int fullRange,
 543                                      int brightness, int contrast, int saturation)
 544 {
 545     const int isRgb =      c->dstFormat==PIX_FMT_RGB32
 546                         || c->dstFormat==PIX_FMT_RGB32_1
 547                         || c->dstFormat==PIX_FMT_BGR24
 548                         || c->dstFormat==PIX_FMT_RGB565
 549                         || c->dstFormat==PIX_FMT_RGB555
 550                         || c->dstFormat==PIX_FMT_RGB8
 551                         || c->dstFormat==PIX_FMT_RGB4
 552                         || c->dstFormat==PIX_FMT_RGB4_BYTE
 553                         || c->dstFormat==PIX_FMT_MONOBLACK;
 554     const int bpp = fmt_depth(c->dstFormat);
 555     uint8_t *y_table;
 556     uint16_t *y_table16;
 557     uint32_t *y_table32;
 558     int i, base, rbase, gbase, bbase, abase, needAlpha;
 559     const int yoffs = fullRange ? 384 : 326;
 560
 561     int64_t crv =  inv_table[0];
 562     int64_t cbu =  inv_table[1];
 563     int64_t cgu = -inv_table[2];
 564     int64_t cgv = -inv_table[3];
 565     int64_t cy  = 1<<16;
 566     int64_t oy  = 0;
 567
 568     int64_t yb = 0;
 569
 570     if (!fullRange) {
 571         cy = (cy*255) / 219;
 572         oy = 16<<16;
 573     } else {
 574         crv = (crv*224) / 255;
 575         cbu = (cbu*224) / 255;
 576         cgu = (cgu*224) / 255;
 577         cgv = (cgv*224) / 255;
 578     }
 579
 580     cy  = (cy *contrast             ) >> 16;
 581     crv = (crv*contrast * saturation) >> 32;
 582     cbu = (cbu*contrast * saturation) >> 32;
 583     cgu = (cgu*contrast * saturation) >> 32;
 584     cgv = (cgv*contrast * saturation) >> 32;
 585     oy -= 256*brightness;
 586
 587     //scale coefficients by cy
 588     crv = ((crv << 16) + 0x8000) / cy;
 589     cbu = ((cbu << 16) + 0x8000) / cy;
 590     cgu = ((cgu << 16) + 0x8000) / cy;
 591     cgv = ((cgv << 16) + 0x8000) / cy;
 592
 593     av_free(c->yuvTable);
 594
 595     switch (bpp) {
 596     case 1:
 597         c->yuvTable = av_malloc(1024);
 598         y_table = c->yuvTable;
 599         yb = -(384<<16) - oy;
 600         for (i = 0; i < 1024-110; i++) {
 601             y_table[i+110] = av_clip_uint8((yb + 0x8000) >> 16) >> 7;
 602             yb += cy;
 603         }
 604         fill_table(c->table_gU, 1, cgu, y_table + yoffs);
 605         fill_gv_table(c->table_gV, 1, cgv);
 606         break;
 607     case 4:
 608     case 4|128:
 609         rbase = isRgb ? 3 : 0;
 610         gbase = 1;
 611         bbase = isRgb ? 0 : 3;
 612         c->yuvTable = av_malloc(1024*3);
 613         y_table = c->yuvTable;
 614         yb = -(384<<16) - oy;
 615         for (i = 0; i < 1024-110; i++) {
 616             int yval = av_clip_uint8((yb + 0x8000) >> 16);
 617             y_table[i+110     ] =  (yval >> 7)       << rbase;
 618             y_table[i+ 37+1024] = ((yval + 43) / 85) << gbase;
 619             y_table[i+110+2048] =  (yval >> 7)       << bbase;
 620             yb += cy;
 621         }
 622         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 623         fill_table(c->table_gU, 1, cgu, y_table + yoffs + 1024);
 624         fill_table(c->table_bU, 1, cbu, y_table + yoffs + 2048);
 625         fill_gv_table(c->table_gV, 1, cgv);
 626         break;
 627     case 8:
 628         rbase = isRgb ? 5 : 0;
 629         gbase = isRgb ? 2 : 3;
 630         bbase = isRgb ? 0 : 6;
 631         c->yuvTable = av_malloc(1024*3);
 632         y_table = c->yuvTable;
 633         yb = -(384<<16) - oy;
 634         for (i = 0; i < 1024-38; i++) {
 635             int yval = av_clip_uint8((yb + 0x8000) >> 16);
 636             y_table[i+16     ] = ((yval + 18) / 36) << rbase;
 637             y_table[i+16+1024] = ((yval + 18) / 36) << gbase;
 638             y_table[i+37+2048] = ((yval + 43) / 85) << bbase;
 639             yb += cy;
 640         }
 641         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 642         fill_table(c->table_gU, 1, cgu, y_table + yoffs + 1024);
 643         fill_table(c->table_bU, 1, cbu, y_table + yoffs + 2048);
 644         fill_gv_table(c->table_gV, 1, cgv);
 645         break;
 646     case 15:
 647     case 16:
 648         rbase = isRgb ? bpp - 5 : 0;
 649         gbase = 5;
 650         bbase = isRgb ? 0 : (bpp - 5);
 651         c->yuvTable = av_malloc(1024*3*2);
 652         y_table16 = c->yuvTable;
 653         yb = -(384<<16) - oy;
 654         for (i = 0; i < 1024; i++) {
 655             uint8_t yval = av_clip_uint8((yb + 0x8000) >> 16);
 656             y_table16[i     ] = (yval >> 3)          << rbase;
 657             y_table16[i+1024] = (yval >> (18 - bpp)) << gbase;
 658             y_table16[i+2048] = (yval >> 3)          << bbase;
 659             yb += cy;
 660         }
 661         fill_table(c->table_rV, 2, crv, y_table16 + yoffs);
 662         fill_table(c->table_gU, 2, cgu, y_table16 + yoffs + 1024);
 663         fill_table(c->table_bU, 2, cbu, y_table16 + yoffs + 2048);
 664         fill_gv_table(c->table_gV, 2, cgv);
 665         break;
 666     case 24:
 667         c->yuvTable = av_malloc(1024);
 668         y_table = c->yuvTable;
 669         yb = -(384<<16) - oy;
 670         for (i = 0; i < 1024; i++) {
 671             y_table[i] = av_clip_uint8((yb + 0x8000) >> 16);
 672             yb += cy;
 673         }
 674         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 675         fill_table(c->table_gU, 1, cgu, y_table + yoffs);
 676         fill_table(c->table_bU, 1, cbu, y_table + yoffs);
 677         fill_gv_table(c->table_gV, 1, cgv);
 678         break;
 679     case 32:
 680         base = (c->dstFormat == PIX_FMT_RGB32_1 || c->dstFormat == PIX_FMT_BGR32_1) ? 8 : 0;
 681         rbase = base + (isRgb ? 16 : 0);
 682         gbase = base + 8;
 683         bbase = base + (isRgb ? 0 : 16);
 684         needAlpha = CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat);
 685         if (!needAlpha)
 686             abase = (base + 24) & 31;
 687         c->yuvTable = av_malloc(1024*3*4);
 688         y_table32 = c->yuvTable;
 689         yb = -(384<<16) - oy;
 690         for (i = 0; i < 1024; i++) {
 691             uint8_t yval = av_clip_uint8((yb + 0x8000) >> 16);
 692             y_table32[i     ] = (yval << rbase) + (needAlpha ? 0 : (255 << abase));
 693             y_table32[i+1024] = yval << gbase;
 694             y_table32[i+2048] = yval << bbase;
 695             yb += cy;
 696         }
 697         fill_table(c->table_rV, 4, crv, y_table32 + yoffs);
 698         fill_table(c->table_gU, 4, cgu, y_table32 + yoffs + 1024);
 699         fill_table(c->table_bU, 4, cbu, y_table32 + yoffs + 2048);
 700         fill_gv_table(c->table_gV, 4, cgv);
 701         break;
 702     default:
 703         c->yuvTable = NULL;
 704         av_log(c, AV_LOG_ERROR, "%ibpp not supported by yuv2rgb\n", bpp);
 705         return -1;
 706     }
 707     return 0;
 708 }