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 const int *sws_getCoefficients(int colorspace)
  53 {
  54     if (colorspace > 7 || colorspace < 0)
  55         colorspace = SWS_CS_DEFAULT;
  56     return ff_yuv2rgb_coeffs[colorspace];
  57 }
  58
  59 #define LOADCHROMA(i)                               \
  60     U = pu[i];                                      \
  61     V = pv[i];                                      \
  62     r = (void *)c->table_rV[V];                     \
  63     g = (void *)(c->table_gU[U] + c->table_gV[V]);  \
  64     b = (void *)c->table_bU[U];
  65
  66 #define PUTRGB(dst,src,i)            \
  67     Y = src[2*i];                    \
  68     dst[2*i  ] = r[Y] + g[Y] + b[Y]; \
  69     Y = src[2*i+1];                  \
  70     dst[2*i+1] = r[Y] + g[Y] + b[Y];
  71
  72 #define PUTRGB24(dst,src,i)                                  \
  73     Y = src[2*i];                                            \
  74     dst[6*i+0] = r[Y]; dst[6*i+1] = g[Y]; dst[6*i+2] = b[Y]; \
  75     Y = src[2*i+1];                                          \
  76     dst[6*i+3] = r[Y]; dst[6*i+4] = g[Y]; dst[6*i+5] = b[Y];
  77
  78 #define PUTBGR24(dst,src,i)                                  \
  79     Y = src[2*i];                                            \
  80     dst[6*i+0] = b[Y]; dst[6*i+1] = g[Y]; dst[6*i+2] = r[Y]; \
  81     Y = src[2*i+1];                                          \
  82     dst[6*i+3] = b[Y]; dst[6*i+4] = g[Y]; dst[6*i+5] = r[Y];
  83
  84 #define PUTRGBA(dst,ysrc,asrc,i,s)                      \
  85     Y = ysrc[2*i];                                      \
  86     dst[2*i  ] = r[Y] + g[Y] + b[Y] + (asrc[2*i  ]<<s); \
  87     Y = ysrc[2*i+1];                                    \
  88     dst[2*i+1] = r[Y] + g[Y] + b[Y] + (asrc[2*i+1]<<s);
  89
  90 #define PUTRGB48(dst,src,i)             \
  91     Y = src[2*i];                       \
  92     dst[12*i+ 0] = dst[12*i+ 1] = r[Y]; \
  93     dst[12*i+ 2] = dst[12*i+ 3] = g[Y]; \
  94     dst[12*i+ 4] = dst[12*i+ 5] = b[Y]; \
  95     Y = src[2*i+1];                     \
  96     dst[12*i+ 6] = dst[12*i+ 7] = r[Y]; \
  97     dst[12*i+ 8] = dst[12*i+ 9] = g[Y]; \
  98     dst[12*i+10] = dst[12*i+11] = b[Y];
  99
 100 #define YUV2RGBFUNC(func_name, dst_type, alpha) \
 101 static int func_name(SwsContext *c, const uint8_t* src[], int srcStride[], int srcSliceY, \
 102                      int srcSliceH, uint8_t* dst[], int dstStride[]) \
 103 {\
 104     int y;\
 105 \
 106     if (!alpha && c->srcFormat == PIX_FMT_YUV422P) {\
 107         srcStride[1] *= 2;\
 108         srcStride[2] *= 2;\
 109     }\
 110     for (y=0; y<srcSliceH; y+=2) {\
 111         dst_type *dst_1 = (dst_type*)(dst[0] + (y+srcSliceY  )*dstStride[0]);\
 112         dst_type *dst_2 = (dst_type*)(dst[0] + (y+srcSliceY+1)*dstStride[0]);\
 113         dst_type av_unused *r, *b;\
 114         dst_type *g;\
 115         const uint8_t *py_1 = src[0] + y*srcStride[0];\
 116         const uint8_t *py_2 = py_1 + srcStride[0];\
 117         const uint8_t *pu = src[1] + (y>>1)*srcStride[1];\
 118         const uint8_t *pv = src[2] + (y>>1)*srcStride[2];\
 119         const uint8_t av_unused *pa_1, *pa_2;\
 120         unsigned int h_size = c->dstW>>3;\
 121         if (alpha) {\
 122             pa_1 = src[3] + y*srcStride[3];\
 123             pa_2 = pa_1 + srcStride[3];\
 124         }\
 125         while (h_size--) {\
 126             int av_unused U, V;\
 127             int Y;\
 128
 129 #define ENDYUV2RGBLINE(dst_delta)\
 130             pu += 4;\
 131             pv += 4;\
 132             py_1 += 8;\
 133             py_2 += 8;\
 134             dst_1 += dst_delta;\
 135             dst_2 += dst_delta;\
 136         }\
 137         if (c->dstW & 4) {\
 138             int av_unused Y, U, V;\
 139
 140 #define ENDYUV2RGBFUNC()\
 141         }\
 142     }\
 143     return srcSliceH;\
 144 }
 145
 146 #define CLOSEYUV2RGBFUNC(dst_delta)\
 147     ENDYUV2RGBLINE(dst_delta)\
 148     ENDYUV2RGBFUNC()
 149
 150 YUV2RGBFUNC(yuv2rgb_c_48, uint8_t, 0)
 151     LOADCHROMA(0);
 152     PUTRGB48(dst_1,py_1,0);
 153     PUTRGB48(dst_2,py_2,0);
 154
 155     LOADCHROMA(1);
 156     PUTRGB48(dst_2,py_2,1);
 157     PUTRGB48(dst_1,py_1,1);
 158
 159     LOADCHROMA(2);
 160     PUTRGB48(dst_1,py_1,2);
 161     PUTRGB48(dst_2,py_2,2);
 162
 163     LOADCHROMA(3);
 164     PUTRGB48(dst_2,py_2,3);
 165     PUTRGB48(dst_1,py_1,3);
 166 ENDYUV2RGBLINE(48)
 167     LOADCHROMA(0);
 168     PUTRGB48(dst_1,py_1,0);
 169     PUTRGB48(dst_2,py_2,0);
 170
 171     LOADCHROMA(1);
 172     PUTRGB48(dst_2,py_2,1);
 173     PUTRGB48(dst_1,py_1,1);
 174 ENDYUV2RGBFUNC()
 175
 176 YUV2RGBFUNC(yuv2rgb_c_32, uint32_t, 0)
 177     LOADCHROMA(0);
 178     PUTRGB(dst_1,py_1,0);
 179     PUTRGB(dst_2,py_2,0);
 180
 181     LOADCHROMA(1);
 182     PUTRGB(dst_2,py_2,1);
 183     PUTRGB(dst_1,py_1,1);
 184
 185     LOADCHROMA(2);
 186     PUTRGB(dst_1,py_1,2);
 187     PUTRGB(dst_2,py_2,2);
 188
 189     LOADCHROMA(3);
 190     PUTRGB(dst_2,py_2,3);
 191     PUTRGB(dst_1,py_1,3);
 192 ENDYUV2RGBLINE(8)
 193     LOADCHROMA(0);
 194     PUTRGB(dst_1,py_1,0);
 195     PUTRGB(dst_2,py_2,0);
 196
 197     LOADCHROMA(1);
 198     PUTRGB(dst_2,py_2,1);
 199     PUTRGB(dst_1,py_1,1);
 200 ENDYUV2RGBFUNC()
 201
 202 YUV2RGBFUNC(yuva2rgba_c, uint32_t, 1)
 203     LOADCHROMA(0);
 204     PUTRGBA(dst_1,py_1,pa_1,0,24);
 205     PUTRGBA(dst_2,py_2,pa_2,0,24);
 206
 207     LOADCHROMA(1);
 208     PUTRGBA(dst_2,py_2,pa_1,1,24);
 209     PUTRGBA(dst_1,py_1,pa_2,1,24);
 210
 211     LOADCHROMA(2);
 212     PUTRGBA(dst_1,py_1,pa_1,2,24);
 213     PUTRGBA(dst_2,py_2,pa_2,2,24);
 214
 215     LOADCHROMA(3);
 216     PUTRGBA(dst_2,py_2,pa_1,3,24);
 217     PUTRGBA(dst_1,py_1,pa_2,3,24);
 218     pa_1 += 8;\
 219     pa_2 += 8;\
 220 ENDYUV2RGBLINE(8)
 221     LOADCHROMA(0);
 222     PUTRGBA(dst_1,py_1,pa_1,0,24);
 223     PUTRGBA(dst_2,py_2,pa_2,0,24);
 224
 225     LOADCHROMA(1);
 226     PUTRGBA(dst_2,py_2,pa_1,1,24);
 227     PUTRGBA(dst_1,py_1,pa_2,1,24);
 228 ENDYUV2RGBFUNC()
 229
 230 YUV2RGBFUNC(yuva2argb_c, uint32_t, 1)
 231     LOADCHROMA(0);
 232     PUTRGBA(dst_1,py_1,pa_1,0,0);
 233     PUTRGBA(dst_2,py_2,pa_2,0,0);
 234
 235     LOADCHROMA(1);
 236     PUTRGBA(dst_2,py_2,pa_2,1,0);
 237     PUTRGBA(dst_1,py_1,pa_1,1,0);
 238
 239     LOADCHROMA(2);
 240     PUTRGBA(dst_1,py_1,pa_1,2,0);
 241     PUTRGBA(dst_2,py_2,pa_2,2,0);
 242
 243     LOADCHROMA(3);
 244     PUTRGBA(dst_2,py_2,pa_2,3,0);
 245     PUTRGBA(dst_1,py_1,pa_1,3,0);
 246     pa_1 += 8;\
 247     pa_2 += 8;\
 248 ENDYUV2RGBLINE(8)
 249     LOADCHROMA(0);
 250     PUTRGBA(dst_1,py_1,pa_1,0,0);
 251     PUTRGBA(dst_2,py_2,pa_2,0,0);
 252
 253     LOADCHROMA(1);
 254     PUTRGBA(dst_2,py_2,pa_2,1,0);
 255     PUTRGBA(dst_1,py_1,pa_1,1,0);
 256 ENDYUV2RGBFUNC()
 257
 258 YUV2RGBFUNC(yuv2rgb_c_24_rgb, uint8_t, 0)
 259     LOADCHROMA(0);
 260     PUTRGB24(dst_1,py_1,0);
 261     PUTRGB24(dst_2,py_2,0);
 262
 263     LOADCHROMA(1);
 264     PUTRGB24(dst_2,py_2,1);
 265     PUTRGB24(dst_1,py_1,1);
 266
 267     LOADCHROMA(2);
 268     PUTRGB24(dst_1,py_1,2);
 269     PUTRGB24(dst_2,py_2,2);
 270
 271     LOADCHROMA(3);
 272     PUTRGB24(dst_2,py_2,3);
 273     PUTRGB24(dst_1,py_1,3);
 274 ENDYUV2RGBLINE(24)
 275     LOADCHROMA(0);
 276     PUTRGB24(dst_1,py_1,0);
 277     PUTRGB24(dst_2,py_2,0);
 278
 279     LOADCHROMA(1);
 280     PUTRGB24(dst_2,py_2,1);
 281     PUTRGB24(dst_1,py_1,1);
 282 ENDYUV2RGBFUNC()
 283
 284 // only trivial mods from yuv2rgb_c_24_rgb
 285 YUV2RGBFUNC(yuv2rgb_c_24_bgr, uint8_t, 0)
 286     LOADCHROMA(0);
 287     PUTBGR24(dst_1,py_1,0);
 288     PUTBGR24(dst_2,py_2,0);
 289
 290     LOADCHROMA(1);
 291     PUTBGR24(dst_2,py_2,1);
 292     PUTBGR24(dst_1,py_1,1);
 293
 294     LOADCHROMA(2);
 295     PUTBGR24(dst_1,py_1,2);
 296     PUTBGR24(dst_2,py_2,2);
 297
 298     LOADCHROMA(3);
 299     PUTBGR24(dst_2,py_2,3);
 300     PUTBGR24(dst_1,py_1,3);
 301 ENDYUV2RGBLINE(24)
 302     LOADCHROMA(0);
 303     PUTBGR24(dst_1,py_1,0);
 304     PUTBGR24(dst_2,py_2,0);
 305
 306     LOADCHROMA(1);
 307     PUTBGR24(dst_2,py_2,1);
 308     PUTBGR24(dst_1,py_1,1);
 309 ENDYUV2RGBFUNC()
 310
 311 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 312 // r, g, b, dst_1, dst_2
 313 YUV2RGBFUNC(yuv2rgb_c_16, uint16_t, 0)
 314     LOADCHROMA(0);
 315     PUTRGB(dst_1,py_1,0);
 316     PUTRGB(dst_2,py_2,0);
 317
 318     LOADCHROMA(1);
 319     PUTRGB(dst_2,py_2,1);
 320     PUTRGB(dst_1,py_1,1);
 321
 322     LOADCHROMA(2);
 323     PUTRGB(dst_1,py_1,2);
 324     PUTRGB(dst_2,py_2,2);
 325
 326     LOADCHROMA(3);
 327     PUTRGB(dst_2,py_2,3);
 328     PUTRGB(dst_1,py_1,3);
 329 CLOSEYUV2RGBFUNC(8)
 330
 331 #if 0 // Currently unused
 332 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 333 // r, g, b, dst_1, dst_2
 334 YUV2RGBFUNC(yuv2rgb_c_8, uint8_t, 0)
 335     LOADCHROMA(0);
 336     PUTRGB(dst_1,py_1,0);
 337     PUTRGB(dst_2,py_2,0);
 338
 339     LOADCHROMA(1);
 340     PUTRGB(dst_2,py_2,1);
 341     PUTRGB(dst_1,py_1,1);
 342
 343     LOADCHROMA(2);
 344     PUTRGB(dst_1,py_1,2);
 345     PUTRGB(dst_2,py_2,2);
 346
 347     LOADCHROMA(3);
 348     PUTRGB(dst_2,py_2,3);
 349     PUTRGB(dst_1,py_1,3);
 350 CLOSEYUV2RGBFUNC(8)
 351 #endif
 352
 353 // r, g, b, dst_1, dst_2
 354 YUV2RGBFUNC(yuv2rgb_c_8_ordered_dither, uint8_t, 0)
 355     const uint8_t *d32 = dither_8x8_32[y&7];
 356     const uint8_t *d64 = dither_8x8_73[y&7];
 357 #define PUTRGB8(dst,src,i,o)                                    \
 358     Y = src[2*i];                                               \
 359     dst[2*i]   = r[Y+d32[0+o]] + g[Y+d32[0+o]] + b[Y+d64[0+o]]; \
 360     Y = src[2*i+1];                                             \
 361     dst[2*i+1] = r[Y+d32[1+o]] + g[Y+d32[1+o]] + b[Y+d64[1+o]];
 362
 363     LOADCHROMA(0);
 364     PUTRGB8(dst_1,py_1,0,0);
 365     PUTRGB8(dst_2,py_2,0,0+8);
 366
 367     LOADCHROMA(1);
 368     PUTRGB8(dst_2,py_2,1,2+8);
 369     PUTRGB8(dst_1,py_1,1,2);
 370
 371     LOADCHROMA(2);
 372     PUTRGB8(dst_1,py_1,2,4);
 373     PUTRGB8(dst_2,py_2,2,4+8);
 374
 375     LOADCHROMA(3);
 376     PUTRGB8(dst_2,py_2,3,6+8);
 377     PUTRGB8(dst_1,py_1,3,6);
 378 CLOSEYUV2RGBFUNC(8)
 379
 380 #if 0 // Currently unused
 381 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 382 // r, g, b, dst_1, dst_2
 383 YUV2RGBFUNC(yuv2rgb_c_4, uint8_t, 0)
 384     int acc;
 385 #define PUTRGB4(dst,src,i)          \
 386     Y = src[2*i];                   \
 387     acc = r[Y] + g[Y] + b[Y];       \
 388     Y = src[2*i+1];                 \
 389     acc |= (r[Y] + g[Y] + b[Y])<<4; \
 390     dst[i] = acc;
 391
 392     LOADCHROMA(0);
 393     PUTRGB4(dst_1,py_1,0);
 394     PUTRGB4(dst_2,py_2,0);
 395
 396     LOADCHROMA(1);
 397     PUTRGB4(dst_2,py_2,1);
 398     PUTRGB4(dst_1,py_1,1);
 399
 400     LOADCHROMA(2);
 401     PUTRGB4(dst_1,py_1,2);
 402     PUTRGB4(dst_2,py_2,2);
 403
 404     LOADCHROMA(3);
 405     PUTRGB4(dst_2,py_2,3);
 406     PUTRGB4(dst_1,py_1,3);
 407 CLOSEYUV2RGBFUNC(4)
 408 #endif
 409
 410 YUV2RGBFUNC(yuv2rgb_c_4_ordered_dither, uint8_t, 0)
 411     const uint8_t *d64 =  dither_8x8_73[y&7];
 412     const uint8_t *d128 = dither_8x8_220[y&7];
 413     int acc;
 414
 415 #define PUTRGB4D(dst,src,i,o)                                     \
 416     Y = src[2*i];                                                 \
 417     acc = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]];        \
 418     Y = src[2*i+1];                                               \
 419     acc |= (r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]])<<4;  \
 420     dst[i]= acc;
 421
 422     LOADCHROMA(0);
 423     PUTRGB4D(dst_1,py_1,0,0);
 424     PUTRGB4D(dst_2,py_2,0,0+8);
 425
 426     LOADCHROMA(1);
 427     PUTRGB4D(dst_2,py_2,1,2+8);
 428     PUTRGB4D(dst_1,py_1,1,2);
 429
 430     LOADCHROMA(2);
 431     PUTRGB4D(dst_1,py_1,2,4);
 432     PUTRGB4D(dst_2,py_2,2,4+8);
 433
 434     LOADCHROMA(3);
 435     PUTRGB4D(dst_2,py_2,3,6+8);
 436     PUTRGB4D(dst_1,py_1,3,6);
 437 CLOSEYUV2RGBFUNC(4)
 438
 439 #if 0 // Currently unused
 440 // This is exactly the same code as yuv2rgb_c_32 except for the types of
 441 // r, g, b, dst_1, dst_2
 442 YUV2RGBFUNC(yuv2rgb_c_4b, uint8_t, 0)
 443     LOADCHROMA(0);
 444     PUTRGB(dst_1,py_1,0);
 445     PUTRGB(dst_2,py_2,0);
 446
 447     LOADCHROMA(1);
 448     PUTRGB(dst_2,py_2,1);
 449     PUTRGB(dst_1,py_1,1);
 450
 451     LOADCHROMA(2);
 452     PUTRGB(dst_1,py_1,2);
 453     PUTRGB(dst_2,py_2,2);
 454
 455     LOADCHROMA(3);
 456     PUTRGB(dst_2,py_2,3);
 457     PUTRGB(dst_1,py_1,3);
 458 CLOSEYUV2RGBFUNC(8)
 459 #endif
 460
 461 YUV2RGBFUNC(yuv2rgb_c_4b_ordered_dither, uint8_t, 0)
 462     const uint8_t *d64 =  dither_8x8_73[y&7];
 463     const uint8_t *d128 = dither_8x8_220[y&7];
 464
 465 #define PUTRGB4DB(dst,src,i,o)                                    \
 466     Y = src[2*i];                                                 \
 467     dst[2*i]   = r[Y+d128[0+o]] + g[Y+d64[0+o]] + b[Y+d128[0+o]]; \
 468     Y = src[2*i+1];                                               \
 469     dst[2*i+1] = r[Y+d128[1+o]] + g[Y+d64[1+o]] + b[Y+d128[1+o]];
 470
 471     LOADCHROMA(0);
 472     PUTRGB4DB(dst_1,py_1,0,0);
 473     PUTRGB4DB(dst_2,py_2,0,0+8);
 474
 475     LOADCHROMA(1);
 476     PUTRGB4DB(dst_2,py_2,1,2+8);
 477     PUTRGB4DB(dst_1,py_1,1,2);
 478
 479     LOADCHROMA(2);
 480     PUTRGB4DB(dst_1,py_1,2,4);
 481     PUTRGB4DB(dst_2,py_2,2,4+8);
 482
 483     LOADCHROMA(3);
 484     PUTRGB4DB(dst_2,py_2,3,6+8);
 485     PUTRGB4DB(dst_1,py_1,3,6);
 486 CLOSEYUV2RGBFUNC(8)
 487
 488 YUV2RGBFUNC(yuv2rgb_c_1_ordered_dither, uint8_t, 0)
 489         const uint8_t *d128 = dither_8x8_220[y&7];
 490         char out_1 = 0, out_2 = 0;
 491         g= c->table_gU[128] + c->table_gV[128];
 492
 493 #define PUTRGB1(out,src,i,o)    \
 494     Y = src[2*i];               \
 495     out+= out + g[Y+d128[0+o]]; \
 496     Y = src[2*i+1];             \
 497     out+= out + g[Y+d128[1+o]];
 498
 499     PUTRGB1(out_1,py_1,0,0);
 500     PUTRGB1(out_2,py_2,0,0+8);
 501
 502     PUTRGB1(out_2,py_2,1,2+8);
 503     PUTRGB1(out_1,py_1,1,2);
 504
 505     PUTRGB1(out_1,py_1,2,4);
 506     PUTRGB1(out_2,py_2,2,4+8);
 507
 508     PUTRGB1(out_2,py_2,3,6+8);
 509     PUTRGB1(out_1,py_1,3,6);
 510
 511     dst_1[0]= out_1;
 512     dst_2[0]= out_2;
 513 CLOSEYUV2RGBFUNC(1)
 514
 515 SwsFunc ff_yuv2rgb_get_func_ptr(SwsContext *c)
 516 {
 517     SwsFunc t = NULL;
 518 #if (HAVE_MMX2 || HAVE_MMX) && CONFIG_GPL
 519      t = ff_yuv2rgb_init_mmx(c);
 520 #endif
 521 #if HAVE_VIS
 522     t = ff_yuv2rgb_init_vis(c);
 523 #endif
 524 #if CONFIG_MLIB
 525     t = ff_yuv2rgb_init_mlib(c);
 526 #endif
 527 #if HAVE_ALTIVEC
 528     if (c->flags & SWS_CPU_CAPS_ALTIVEC)
 529         t = ff_yuv2rgb_init_altivec(c);
 530 #endif
 531
 532 #if ARCH_BFIN
 533     if (c->flags & SWS_CPU_CAPS_BFIN)
 534         t = ff_yuv2rgb_get_func_ptr_bfin(c);
 535 #endif
 536
 537     if (t)
 538         return t;
 539
 540     av_log(c, AV_LOG_WARNING, "No accelerated colorspace conversion found from %s to %s.\n", sws_format_name(c->srcFormat), sws_format_name(c->dstFormat));
 541
 542     switch (c->dstFormat) {
 543     case PIX_FMT_RGB48BE:
 544     case PIX_FMT_RGB48LE:    return yuv2rgb_c_48;
 545     case PIX_FMT_ARGB:
 546     case PIX_FMT_ABGR:       if (CONFIG_SWSCALE_ALPHA && c->srcFormat == PIX_FMT_YUVA420P) return yuva2argb_c;
 547     case PIX_FMT_RGBA:
 548     case PIX_FMT_BGRA:       return (CONFIG_SWSCALE_ALPHA && c->srcFormat == PIX_FMT_YUVA420P) ? yuva2rgba_c : yuv2rgb_c_32;
 549     case PIX_FMT_RGB24:      return yuv2rgb_c_24_rgb;
 550     case PIX_FMT_BGR24:      return yuv2rgb_c_24_bgr;
 551     case PIX_FMT_RGB565:
 552     case PIX_FMT_BGR565:
 553     case PIX_FMT_RGB555:
 554     case PIX_FMT_BGR555:     return yuv2rgb_c_16;
 555     case PIX_FMT_RGB8:
 556     case PIX_FMT_BGR8:       return yuv2rgb_c_8_ordered_dither;
 557     case PIX_FMT_RGB4:
 558     case PIX_FMT_BGR4:       return yuv2rgb_c_4_ordered_dither;
 559     case PIX_FMT_RGB4_BYTE:
 560     case PIX_FMT_BGR4_BYTE:  return yuv2rgb_c_4b_ordered_dither;
 561     case PIX_FMT_MONOBLACK:  return yuv2rgb_c_1_ordered_dither;
 562     default:
 563         assert(0);
 564     }
 565     return NULL;
 566 }
 567
 568 static void fill_table(uint8_t* table[256], const int elemsize, const int inc, uint8_t *y_table)
 569 {
 570     int i;
 571     int64_t cb = 0;
 572
 573     y_table -= elemsize * (inc >> 9);
 574
 575     for (i = 0; i < 256; i++) {
 576         table[i] = y_table + elemsize * (cb >> 16);
 577         cb += inc;
 578     }
 579 }
 580
 581 static void fill_gv_table(int table[256], const int elemsize, const int inc)
 582 {
 583     int i;
 584     int64_t cb = 0;
 585     int off = -(inc >> 9);
 586
 587     for (i = 0; i < 256; i++) {
 588         table[i] = elemsize * (off + (cb >> 16));
 589         cb += inc;
 590     }
 591 }
 592
 593 av_cold int ff_yuv2rgb_c_init_tables(SwsContext *c, const int inv_table[4], int fullRange,
 594                                      int brightness, int contrast, int saturation)
 595 {
 596     const int isRgb =      c->dstFormat==PIX_FMT_RGB32
 597                         || c->dstFormat==PIX_FMT_RGB32_1
 598                         || c->dstFormat==PIX_FMT_BGR24
 599                         || c->dstFormat==PIX_FMT_RGB565
 600                         || c->dstFormat==PIX_FMT_RGB555
 601                         || c->dstFormat==PIX_FMT_RGB8
 602                         || c->dstFormat==PIX_FMT_RGB4
 603                         || c->dstFormat==PIX_FMT_RGB4_BYTE
 604                         || c->dstFormat==PIX_FMT_MONOBLACK;
 605     const int bpp = c->dstFormatBpp;
 606     uint8_t *y_table;
 607     uint16_t *y_table16;
 608     uint32_t *y_table32;
 609     int i, base, rbase, gbase, bbase, abase, needAlpha;
 610     const int yoffs = fullRange ? 384 : 326;
 611
 612     int64_t crv =  inv_table[0];
 613     int64_t cbu =  inv_table[1];
 614     int64_t cgu = -inv_table[2];
 615     int64_t cgv = -inv_table[3];
 616     int64_t cy  = 1<<16;
 617     int64_t oy  = 0;
 618
 619     int64_t yb = 0;
 620
 621     if (!fullRange) {
 622         cy = (cy*255) / 219;
 623         oy = 16<<16;
 624     } else {
 625         crv = (crv*224) / 255;
 626         cbu = (cbu*224) / 255;
 627         cgu = (cgu*224) / 255;
 628         cgv = (cgv*224) / 255;
 629     }
 630
 631     cy  = (cy *contrast             ) >> 16;
 632     crv = (crv*contrast * saturation) >> 32;
 633     cbu = (cbu*contrast * saturation) >> 32;
 634     cgu = (cgu*contrast * saturation) >> 32;
 635     cgv = (cgv*contrast * saturation) >> 32;
 636     oy -= 256*brightness;
 637
 638     //scale coefficients by cy
 639     crv = ((crv << 16) + 0x8000) / cy;
 640     cbu = ((cbu << 16) + 0x8000) / cy;
 641     cgu = ((cgu << 16) + 0x8000) / cy;
 642     cgv = ((cgv << 16) + 0x8000) / cy;
 643
 644     av_free(c->yuvTable);
 645
 646     switch (bpp) {
 647     case 1:
 648         c->yuvTable = av_malloc(1024);
 649         y_table = c->yuvTable;
 650         yb = -(384<<16) - oy;
 651         for (i = 0; i < 1024-110; i++) {
 652             y_table[i+110] = av_clip_uint8((yb + 0x8000) >> 16) >> 7;
 653             yb += cy;
 654         }
 655         fill_table(c->table_gU, 1, cgu, y_table + yoffs);
 656         fill_gv_table(c->table_gV, 1, cgv);
 657         break;
 658     case 4:
 659     case 4|128:
 660         rbase = isRgb ? 3 : 0;
 661         gbase = 1;
 662         bbase = isRgb ? 0 : 3;
 663         c->yuvTable = av_malloc(1024*3);
 664         y_table = c->yuvTable;
 665         yb = -(384<<16) - oy;
 666         for (i = 0; i < 1024-110; i++) {
 667             int yval = av_clip_uint8((yb + 0x8000) >> 16);
 668             y_table[i+110     ] =  (yval >> 7)       << rbase;
 669             y_table[i+ 37+1024] = ((yval + 43) / 85) << gbase;
 670             y_table[i+110+2048] =  (yval >> 7)       << bbase;
 671             yb += cy;
 672         }
 673         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 674         fill_table(c->table_gU, 1, cgu, y_table + yoffs + 1024);
 675         fill_table(c->table_bU, 1, cbu, y_table + yoffs + 2048);
 676         fill_gv_table(c->table_gV, 1, cgv);
 677         break;
 678     case 8:
 679         rbase = isRgb ? 5 : 0;
 680         gbase = isRgb ? 2 : 3;
 681         bbase = isRgb ? 0 : 6;
 682         c->yuvTable = av_malloc(1024*3);
 683         y_table = c->yuvTable;
 684         yb = -(384<<16) - oy;
 685         for (i = 0; i < 1024-38; i++) {
 686             int yval = av_clip_uint8((yb + 0x8000) >> 16);
 687             y_table[i+16     ] = ((yval + 18) / 36) << rbase;
 688             y_table[i+16+1024] = ((yval + 18) / 36) << gbase;
 689             y_table[i+37+2048] = ((yval + 43) / 85) << bbase;
 690             yb += cy;
 691         }
 692         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 693         fill_table(c->table_gU, 1, cgu, y_table + yoffs + 1024);
 694         fill_table(c->table_bU, 1, cbu, y_table + yoffs + 2048);
 695         fill_gv_table(c->table_gV, 1, cgv);
 696         break;
 697     case 15:
 698     case 16:
 699         rbase = isRgb ? bpp - 5 : 0;
 700         gbase = 5;
 701         bbase = isRgb ? 0 : (bpp - 5);
 702         c->yuvTable = av_malloc(1024*3*2);
 703         y_table16 = c->yuvTable;
 704         yb = -(384<<16) - oy;
 705         for (i = 0; i < 1024; i++) {
 706             uint8_t yval = av_clip_uint8((yb + 0x8000) >> 16);
 707             y_table16[i     ] = (yval >> 3)          << rbase;
 708             y_table16[i+1024] = (yval >> (18 - bpp)) << gbase;
 709             y_table16[i+2048] = (yval >> 3)          << bbase;
 710             yb += cy;
 711         }
 712         fill_table(c->table_rV, 2, crv, y_table16 + yoffs);
 713         fill_table(c->table_gU, 2, cgu, y_table16 + yoffs + 1024);
 714         fill_table(c->table_bU, 2, cbu, y_table16 + yoffs + 2048);
 715         fill_gv_table(c->table_gV, 2, cgv);
 716         break;
 717     case 24:
 718     case 48:
 719         c->yuvTable = av_malloc(1024);
 720         y_table = c->yuvTable;
 721         yb = -(384<<16) - oy;
 722         for (i = 0; i < 1024; i++) {
 723             y_table[i] = av_clip_uint8((yb + 0x8000) >> 16);
 724             yb += cy;
 725         }
 726         fill_table(c->table_rV, 1, crv, y_table + yoffs);
 727         fill_table(c->table_gU, 1, cgu, y_table + yoffs);
 728         fill_table(c->table_bU, 1, cbu, y_table + yoffs);
 729         fill_gv_table(c->table_gV, 1, cgv);
 730         break;
 731     case 32:
 732         base = (c->dstFormat == PIX_FMT_RGB32_1 || c->dstFormat == PIX_FMT_BGR32_1) ? 8 : 0;
 733         rbase = base + (isRgb ? 16 : 0);
 734         gbase = base + 8;
 735         bbase = base + (isRgb ? 0 : 16);
 736         needAlpha = CONFIG_SWSCALE_ALPHA && isALPHA(c->srcFormat);
 737         if (!needAlpha)
 738             abase = (base + 24) & 31;
 739         c->yuvTable = av_malloc(1024*3*4);
 740         y_table32 = c->yuvTable;
 741         yb = -(384<<16) - oy;
 742         for (i = 0; i < 1024; i++) {
 743             uint8_t yval = av_clip_uint8((yb + 0x8000) >> 16);
 744             y_table32[i     ] = (yval << rbase) + (needAlpha ? 0 : (255 << abase));
 745             y_table32[i+1024] = yval << gbase;
 746             y_table32[i+2048] = yval << bbase;
 747             yb += cy;
 748         }
 749         fill_table(c->table_rV, 4, crv, y_table32 + yoffs);
 750         fill_table(c->table_gU, 4, cgu, y_table32 + yoffs + 1024);
 751         fill_table(c->table_bU, 4, cbu, y_table32 + yoffs + 2048);
 752         fill_gv_table(c->table_gV, 4, cgv);
 753         break;
 754     default:
 755         c->yuvTable = NULL;
 756         av_log(c, AV_LOG_ERROR, "%ibpp not supported by yuv2rgb\n", bpp);
 757         return -1;
 758     }
 759     return 0;
 760 }