git.sesse.net Git - ffmpeg/blob - libavcodec/huffyuv.c

   1 /*
   2  * huffyuv codec for libavcodec
   3  *
   4  * Copyright (c) 2002-2003 Michael Niedermayer <michaelni@gmx.at>
   5  *
   6  * see http://www.pcisys.net/~melanson/codecs/huffyuv.txt for a description of
   7  * the algorithm used
   8  *
   9  * This file is part of Libav.
  10  *
  11  * Libav 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  * Libav 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 Libav; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  24  */
  25
  26 /**
  27  * @file
  28  * huffyuv codec for libavcodec.
  29  */
  30
  31 #include "avcodec.h"
  32 #include "get_bits.h"
  33 #include "put_bits.h"
  34 #include "dsputil.h"
  35 #include "thread.h"
  36
  37 #define VLC_BITS 11
  38
  39 #if HAVE_BIGENDIAN
  40 #define B 3
  41 #define G 2
  42 #define R 1
  43 #define A 0
  44 #else
  45 #define B 0
  46 #define G 1
  47 #define R 2
  48 #define A 3
  49 #endif
  50
  51 typedef enum Predictor{
  52     LEFT= 0,
  53     PLANE,
  54     MEDIAN,
  55 } Predictor;
  56
  57 typedef struct HYuvContext{
  58     AVCodecContext *avctx;
  59     Predictor predictor;
  60     GetBitContext gb;
  61     PutBitContext pb;
  62     int interlaced;
  63     int decorrelate;
  64     int bitstream_bpp;
  65     int version;
  66     int yuy2;                               //use yuy2 instead of 422P
  67     int bgr32;                              //use bgr32 instead of bgr24
  68     int width, height;
  69     int flags;
  70     int context;
  71     int picture_number;
  72     int last_slice_end;
  73     uint8_t *temp[3];
  74     uint64_t stats[3][256];
  75     uint8_t len[3][256];
  76     uint32_t bits[3][256];
  77     uint32_t pix_bgr_map[1<<VLC_BITS];
  78     VLC vlc[6];                             //Y,U,V,YY,YU,YV
  79     AVFrame picture;
  80     uint8_t *bitstream_buffer;
  81     unsigned int bitstream_buffer_size;
  82     DSPContext dsp;
  83 }HYuvContext;
  84
  85 #define classic_shift_luma_table_size 42
  86 static const unsigned char classic_shift_luma[classic_shift_luma_table_size + FF_INPUT_BUFFER_PADDING_SIZE] = {
  87   34,36,35,69,135,232,9,16,10,24,11,23,12,16,13,10,14,8,15,8,
  88   16,8,17,20,16,10,207,206,205,236,11,8,10,21,9,23,8,8,199,70,
  89   69,68, 0
  90 };
  91
  92 #define classic_shift_chroma_table_size 59
  93 static const unsigned char classic_shift_chroma[classic_shift_chroma_table_size + FF_INPUT_BUFFER_PADDING_SIZE] = {
  94   66,36,37,38,39,40,41,75,76,77,110,239,144,81,82,83,84,85,118,183,
  95   56,57,88,89,56,89,154,57,58,57,26,141,57,56,58,57,58,57,184,119,
  96   214,245,116,83,82,49,80,79,78,77,44,75,41,40,39,38,37,36,34, 0
  97 };
  98
  99 static const unsigned char classic_add_luma[256] = {
 100     3,  9,  5, 12, 10, 35, 32, 29, 27, 50, 48, 45, 44, 41, 39, 37,
 101    73, 70, 68, 65, 64, 61, 58, 56, 53, 50, 49, 46, 44, 41, 38, 36,
 102    68, 65, 63, 61, 58, 55, 53, 51, 48, 46, 45, 43, 41, 39, 38, 36,
 103    35, 33, 32, 30, 29, 27, 26, 25, 48, 47, 46, 44, 43, 41, 40, 39,
 104    37, 36, 35, 34, 32, 31, 30, 28, 27, 26, 24, 23, 22, 20, 19, 37,
 105    35, 34, 33, 31, 30, 29, 27, 26, 24, 23, 21, 20, 18, 17, 15, 29,
 106    27, 26, 24, 22, 21, 19, 17, 16, 14, 26, 25, 23, 21, 19, 18, 16,
 107    15, 27, 25, 23, 21, 19, 17, 16, 14, 26, 25, 23, 21, 18, 17, 14,
 108    12, 17, 19, 13,  4,  9,  2, 11,  1,  7,  8,  0, 16,  3, 14,  6,
 109    12, 10,  5, 15, 18, 11, 10, 13, 15, 16, 19, 20, 22, 24, 27, 15,
 110    18, 20, 22, 24, 26, 14, 17, 20, 22, 24, 27, 15, 18, 20, 23, 25,
 111    28, 16, 19, 22, 25, 28, 32, 36, 21, 25, 29, 33, 38, 42, 45, 49,
 112    28, 31, 34, 37, 40, 42, 44, 47, 49, 50, 52, 54, 56, 57, 59, 60,
 113    62, 64, 66, 67, 69, 35, 37, 39, 40, 42, 43, 45, 47, 48, 51, 52,
 114    54, 55, 57, 59, 60, 62, 63, 66, 67, 69, 71, 72, 38, 40, 42, 43,
 115    46, 47, 49, 51, 26, 28, 30, 31, 33, 34, 18, 19, 11, 13,  7,  8,
 116 };
 117
 118 static const unsigned char classic_add_chroma[256] = {
 119     3,  1,  2,  2,  2,  2,  3,  3,  7,  5,  7,  5,  8,  6, 11,  9,
 120     7, 13, 11, 10,  9,  8,  7,  5,  9,  7,  6,  4,  7,  5,  8,  7,
 121    11,  8, 13, 11, 19, 15, 22, 23, 20, 33, 32, 28, 27, 29, 51, 77,
 122    43, 45, 76, 81, 46, 82, 75, 55, 56,144, 58, 80, 60, 74,147, 63,
 123   143, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79,
 124    80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 27, 30, 21, 22,
 125    17, 14,  5,  6,100, 54, 47, 50, 51, 53,106,107,108,109,110,111,
 126   112,113,114,115,  4,117,118, 92, 94,121,122,  3,124,103,  2,  1,
 127     0,129,130,131,120,119,126,125,136,137,138,139,140,141,142,134,
 128   135,132,133,104, 64,101, 62, 57,102, 95, 93, 59, 61, 28, 97, 96,
 129    52, 49, 48, 29, 32, 25, 24, 46, 23, 98, 45, 44, 43, 20, 42, 41,
 130    19, 18, 99, 40, 15, 39, 38, 16, 13, 12, 11, 37, 10,  9,  8, 36,
 131     7,128,127,105,123,116, 35, 34, 33,145, 31, 79, 42,146, 78, 26,
 132    83, 48, 49, 50, 44, 47, 26, 31, 30, 18, 17, 19, 21, 24, 25, 13,
 133    14, 16, 17, 18, 20, 21, 12, 14, 15,  9, 10,  6,  9,  6,  5,  8,
 134     6, 12,  8, 10,  7,  9,  6,  4,  6,  2,  2,  3,  3,  3,  3,  2,
 135 };
 136
 137 static inline int sub_left_prediction(HYuvContext *s, uint8_t *dst, uint8_t *src, int w, int left){
 138     int i;
 139     if(w<32){
 140         for(i=0; i<w; i++){
 141             const int temp= src[i];
 142             dst[i]= temp - left;
 143             left= temp;
 144         }
 145         return left;
 146     }else{
 147         for(i=0; i<16; i++){
 148             const int temp= src[i];
 149             dst[i]= temp - left;
 150             left= temp;
 151         }
 152         s->dsp.diff_bytes(dst+16, src+16, src+15, w-16);
 153         return src[w-1];
 154     }
 155 }
 156
 157 static inline void sub_left_prediction_bgr32(HYuvContext *s, uint8_t *dst, uint8_t *src, int w, int *red, int *green, int *blue){
 158     int i;
 159     int r,g,b;
 160     r= *red;
 161     g= *green;
 162     b= *blue;
 163     for(i=0; i<FFMIN(w,4); i++){
 164         const int rt= src[i*4+R];
 165         const int gt= src[i*4+G];
 166         const int bt= src[i*4+B];
 167         dst[i*4+R]= rt - r;
 168         dst[i*4+G]= gt - g;
 169         dst[i*4+B]= bt - b;
 170         r = rt;
 171         g = gt;
 172         b = bt;
 173     }
 174     s->dsp.diff_bytes(dst+16, src+16, src+12, w*4-16);
 175     *red=   src[(w-1)*4+R];
 176     *green= src[(w-1)*4+G];
 177     *blue=  src[(w-1)*4+B];
 178 }
 179
 180 static int read_len_table(uint8_t *dst, GetBitContext *gb){
 181     int i, val, repeat;
 182
 183     for(i=0; i<256;){
 184         repeat= get_bits(gb, 3);
 185         val   = get_bits(gb, 5);
 186         if(repeat==0)
 187             repeat= get_bits(gb, 8);
 188 //printf("%d %d\n", val, repeat);
 189         if(i+repeat > 256 || get_bits_left(gb) < 0) {
 190             av_log(NULL, AV_LOG_ERROR, "Error reading huffman table\n");
 191             return -1;
 192         }
 193         while (repeat--)
 194             dst[i++] = val;
 195     }
 196     return 0;
 197 }
 198
 199 static int generate_bits_table(uint32_t *dst, const uint8_t *len_table){
 200     int len, index;
 201     uint32_t bits=0;
 202
 203     for(len=32; len>0; len--){
 204         for(index=0; index<256; index++){
 205             if(len_table[index]==len)
 206                 dst[index]= bits++;
 207         }
 208         if(bits & 1){
 209             av_log(NULL, AV_LOG_ERROR, "Error generating huffman table\n");
 210             return -1;
 211         }
 212         bits >>= 1;
 213     }
 214     return 0;
 215 }
 216
 217 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER
 218 typedef struct {
 219     uint64_t val;
 220     int name;
 221 } HeapElem;
 222
 223 static void heap_sift(HeapElem *h, int root, int size)
 224 {
 225     while(root*2+1 < size) {
 226         int child = root*2+1;
 227         if(child < size-1 && h[child].val > h[child+1].val)
 228             child++;
 229         if(h[root].val > h[child].val) {
 230             FFSWAP(HeapElem, h[root], h[child]);
 231             root = child;
 232         } else
 233             break;
 234     }
 235 }
 236
 237 static void generate_len_table(uint8_t *dst, const uint64_t *stats){
 238     HeapElem h[256];
 239     int up[2*256];
 240     int len[2*256];
 241     int offset, i, next;
 242     int size = 256;
 243
 244     for(offset=1; ; offset<<=1){
 245         for(i=0; i<size; i++){
 246             h[i].name = i;
 247             h[i].val = (stats[i] << 8) + offset;
 248         }
 249         for(i=size/2-1; i>=0; i--)
 250             heap_sift(h, i, size);
 251
 252         for(next=size; next<size*2-1; next++){
 253             // merge the two smallest entries, and put it back in the heap
 254             uint64_t min1v = h[0].val;
 255             up[h[0].name] = next;
 256             h[0].val = INT64_MAX;
 257             heap_sift(h, 0, size);
 258             up[h[0].name] = next;
 259             h[0].name = next;
 260             h[0].val += min1v;
 261             heap_sift(h, 0, size);
 262         }
 263
 264         len[2*size-2] = 0;
 265         for(i=2*size-3; i>=size; i--)
 266             len[i] = len[up[i]] + 1;
 267         for(i=0; i<size; i++) {
 268             dst[i] = len[up[i]] + 1;
 269             if(dst[i] >= 32) break;
 270         }
 271         if(i==size) break;
 272     }
 273 }
 274 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */
 275
 276 static void generate_joint_tables(HYuvContext *s){
 277     uint16_t symbols[1<<VLC_BITS];
 278     uint16_t bits[1<<VLC_BITS];
 279     uint8_t len[1<<VLC_BITS];
 280     if(s->bitstream_bpp < 24){
 281         int p, i, y, u;
 282         for(p=0; p<3; p++){
 283             for(i=y=0; y<256; y++){
 284                 int len0 = s->len[0][y];
 285                 int limit = VLC_BITS - len0;
 286                 if(limit <= 0)
 287                     continue;
 288                 for(u=0; u<256; u++){
 289                     int len1 = s->len[p][u];
 290                     if(len1 > limit)
 291                         continue;
 292                     len[i] = len0 + len1;
 293                     bits[i] = (s->bits[0][y] << len1) + s->bits[p][u];
 294                     symbols[i] = (y<<8) + u;
 295                     if(symbols[i] != 0xffff) // reserved to mean "invalid"
 296                         i++;
 297                 }
 298             }
 299             ff_free_vlc(&s->vlc[3+p]);
 300             ff_init_vlc_sparse(&s->vlc[3+p], VLC_BITS, i, len, 1, 1, bits, 2, 2, symbols, 2, 2, 0);
 301         }
 302     }else{
 303         uint8_t (*map)[4] = (uint8_t(*)[4])s->pix_bgr_map;
 304         int i, b, g, r, code;
 305         int p0 = s->decorrelate;
 306         int p1 = !s->decorrelate;
 307         // restrict the range to +/-16 becaues that's pretty much guaranteed to
 308         // cover all the combinations that fit in 11 bits total, and it doesn't
 309         // matter if we miss a few rare codes.
 310         for(i=0, g=-16; g<16; g++){
 311             int len0 = s->len[p0][g&255];
 312             int limit0 = VLC_BITS - len0;
 313             if(limit0 < 2)
 314                 continue;
 315             for(b=-16; b<16; b++){
 316                 int len1 = s->len[p1][b&255];
 317                 int limit1 = limit0 - len1;
 318                 if(limit1 < 1)
 319                     continue;
 320                 code = (s->bits[p0][g&255] << len1) + s->bits[p1][b&255];
 321                 for(r=-16; r<16; r++){
 322                     int len2 = s->len[2][r&255];
 323                     if(len2 > limit1)
 324                         continue;
 325                     len[i] = len0 + len1 + len2;
 326                     bits[i] = (code << len2) + s->bits[2][r&255];
 327                     if(s->decorrelate){
 328                         map[i][G] = g;
 329                         map[i][B] = g+b;
 330                         map[i][R] = g+r;
 331                     }else{
 332                         map[i][B] = g;
 333                         map[i][G] = b;
 334                         map[i][R] = r;
 335                     }
 336                     i++;
 337                 }
 338             }
 339         }
 340         ff_free_vlc(&s->vlc[3]);
 341         init_vlc(&s->vlc[3], VLC_BITS, i, len, 1, 1, bits, 2, 2, 0);
 342     }
 343 }
 344
 345 static int read_huffman_tables(HYuvContext *s, const uint8_t *src, int length){
 346     GetBitContext gb;
 347     int i;
 348
 349     init_get_bits(&gb, src, length*8);
 350
 351     for(i=0; i<3; i++){
 352         if(read_len_table(s->len[i], &gb)<0)
 353             return -1;
 354         if(generate_bits_table(s->bits[i], s->len[i])<0){
 355             return -1;
 356         }
 357         ff_free_vlc(&s->vlc[i]);
 358         init_vlc(&s->vlc[i], VLC_BITS, 256, s->len[i], 1, 1, s->bits[i], 4, 4, 0);
 359     }
 360
 361     generate_joint_tables(s);
 362
 363     return (get_bits_count(&gb)+7)/8;
 364 }
 365
 366 static int read_old_huffman_tables(HYuvContext *s){
 367 #if 1
 368     GetBitContext gb;
 369     int i;
 370
 371     init_get_bits(&gb, classic_shift_luma, classic_shift_luma_table_size*8);
 372     if(read_len_table(s->len[0], &gb)<0)
 373         return -1;
 374     init_get_bits(&gb, classic_shift_chroma, classic_shift_chroma_table_size*8);
 375     if(read_len_table(s->len[1], &gb)<0)
 376         return -1;
 377
 378     for(i=0; i<256; i++) s->bits[0][i] = classic_add_luma  [i];
 379     for(i=0; i<256; i++) s->bits[1][i] = classic_add_chroma[i];
 380
 381     if(s->bitstream_bpp >= 24){
 382         memcpy(s->bits[1], s->bits[0], 256*sizeof(uint32_t));
 383         memcpy(s->len[1] , s->len [0], 256*sizeof(uint8_t));
 384     }
 385     memcpy(s->bits[2], s->bits[1], 256*sizeof(uint32_t));
 386     memcpy(s->len[2] , s->len [1], 256*sizeof(uint8_t));
 387
 388     for(i=0; i<3; i++){
 389         ff_free_vlc(&s->vlc[i]);
 390         init_vlc(&s->vlc[i], VLC_BITS, 256, s->len[i], 1, 1, s->bits[i], 4, 4, 0);
 391     }
 392
 393     generate_joint_tables(s);
 394
 395     return 0;
 396 #else
 397     av_log(s->avctx, AV_LOG_DEBUG, "v1 huffyuv is not supported \n");
 398     return -1;
 399 #endif
 400 }
 401
 402 static av_cold void alloc_temp(HYuvContext *s){
 403     int i;
 404
 405     if(s->bitstream_bpp<24){
 406         for(i=0; i<3; i++){
 407             s->temp[i]= av_malloc(s->width + 16);
 408         }
 409     }else{
 410         s->temp[0]= av_mallocz(4*s->width + 16);
 411     }
 412 }
 413
 414 static av_cold int common_init(AVCodecContext *avctx){
 415     HYuvContext *s = avctx->priv_data;
 416
 417     s->avctx= avctx;
 418     s->flags= avctx->flags;
 419
 420     ff_dsputil_init(&s->dsp, avctx);
 421
 422     s->width= avctx->width;
 423     s->height= avctx->height;
 424     assert(s->width>0 && s->height>0);
 425
 426     return 0;
 427 }
 428
 429 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER
 430 static av_cold int decode_init(AVCodecContext *avctx)
 431 {
 432     HYuvContext *s = avctx->priv_data;
 433
 434     common_init(avctx);
 435     memset(s->vlc, 0, 3*sizeof(VLC));
 436
 437     avctx->coded_frame= &s->picture;
 438     s->interlaced= s->height > 288;
 439
 440 s->bgr32=1;
 441 //if(avctx->extradata)
 442 //  printf("extradata:%X, extradata_size:%d\n", *(uint32_t*)avctx->extradata, avctx->extradata_size);
 443     if(avctx->extradata_size){
 444         if((avctx->bits_per_coded_sample&7) && avctx->bits_per_coded_sample != 12)
 445             s->version=1; // do such files exist at all?
 446         else
 447             s->version=2;
 448     }else
 449         s->version=0;
 450
 451     if(s->version==2){
 452         int method, interlace;
 453
 454         if (avctx->extradata_size < 4)
 455             return -1;
 456
 457         method= ((uint8_t*)avctx->extradata)[0];
 458         s->decorrelate= method&64 ? 1 : 0;
 459         s->predictor= method&63;
 460         s->bitstream_bpp= ((uint8_t*)avctx->extradata)[1];
 461         if(s->bitstream_bpp==0)
 462             s->bitstream_bpp= avctx->bits_per_coded_sample&~7;
 463         interlace= (((uint8_t*)avctx->extradata)[2] & 0x30) >> 4;
 464         s->interlaced= (interlace==1) ? 1 : (interlace==2) ? 0 : s->interlaced;
 465         s->context= ((uint8_t*)avctx->extradata)[2] & 0x40 ? 1 : 0;
 466
 467         if(read_huffman_tables(s, ((uint8_t*)avctx->extradata)+4, avctx->extradata_size-4) < 0)
 468             return -1;
 469     }else{
 470         switch(avctx->bits_per_coded_sample&7){
 471         case 1:
 472             s->predictor= LEFT;
 473             s->decorrelate= 0;
 474             break;
 475         case 2:
 476             s->predictor= LEFT;
 477             s->decorrelate= 1;
 478             break;
 479         case 3:
 480             s->predictor= PLANE;
 481             s->decorrelate= avctx->bits_per_coded_sample >= 24;
 482             break;
 483         case 4:
 484             s->predictor= MEDIAN;
 485             s->decorrelate= 0;
 486             break;
 487         default:
 488             s->predictor= LEFT; //OLD
 489             s->decorrelate= 0;
 490             break;
 491         }
 492         s->bitstream_bpp= avctx->bits_per_coded_sample & ~7;
 493         s->context= 0;
 494
 495         if(read_old_huffman_tables(s) < 0)
 496             return -1;
 497     }
 498
 499     switch(s->bitstream_bpp){
 500     case 12:
 501         avctx->pix_fmt = PIX_FMT_YUV420P;
 502         break;
 503     case 16:
 504         if(s->yuy2){
 505             avctx->pix_fmt = PIX_FMT_YUYV422;
 506         }else{
 507             avctx->pix_fmt = PIX_FMT_YUV422P;
 508         }
 509         break;
 510     case 24:
 511     case 32:
 512         if(s->bgr32){
 513             avctx->pix_fmt = PIX_FMT_RGB32;
 514         }else{
 515             avctx->pix_fmt = PIX_FMT_BGR24;
 516         }
 517         break;
 518     default:
 519         return AVERROR_INVALIDDATA;
 520     }
 521
 522     alloc_temp(s);
 523
 524 //    av_log(NULL, AV_LOG_DEBUG, "pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_coded_sample, s->interlaced);
 525
 526     return 0;
 527 }
 528
 529 static av_cold int decode_init_thread_copy(AVCodecContext *avctx)
 530 {
 531     HYuvContext *s = avctx->priv_data;
 532     int i;
 533
 534     avctx->coded_frame= &s->picture;
 535     alloc_temp(s);
 536
 537     for (i = 0; i < 6; i++)
 538         s->vlc[i].table = NULL;
 539
 540     if(s->version==2){
 541         if(read_huffman_tables(s, ((uint8_t*)avctx->extradata)+4, avctx->extradata_size) < 0)
 542             return -1;
 543     }else{
 544         if(read_old_huffman_tables(s) < 0)
 545             return -1;
 546     }
 547
 548     return 0;
 549 }
 550 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */
 551
 552 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER
 553 static int store_table(HYuvContext *s, const uint8_t *len, uint8_t *buf){
 554     int i;
 555     int index= 0;
 556
 557     for(i=0; i<256;){
 558         int val= len[i];
 559         int repeat=0;
 560
 561         for(; i<256 && len[i]==val && repeat<255; i++)
 562             repeat++;
 563
 564         assert(val < 32 && val >0 && repeat<256 && repeat>0);
 565         if(repeat>7){
 566             buf[index++]= val;
 567             buf[index++]= repeat;
 568         }else{
 569             buf[index++]= val | (repeat<<5);
 570         }
 571     }
 572
 573     return index;
 574 }
 575
 576 static av_cold int encode_init(AVCodecContext *avctx)
 577 {
 578     HYuvContext *s = avctx->priv_data;
 579     int i, j;
 580
 581     common_init(avctx);
 582
 583     avctx->extradata= av_mallocz(1024*30); // 256*3+4 == 772
 584     avctx->stats_out= av_mallocz(1024*30); // 21*256*3(%llu ) + 3(\n) + 1(0) = 16132
 585     s->version=2;
 586
 587     avctx->coded_frame= &s->picture;
 588
 589     switch(avctx->pix_fmt){
 590     case PIX_FMT_YUV420P:
 591         s->bitstream_bpp= 12;
 592         break;
 593     case PIX_FMT_YUV422P:
 594         s->bitstream_bpp= 16;
 595         break;
 596     case PIX_FMT_RGB32:
 597         s->bitstream_bpp= 24;
 598         break;
 599     default:
 600         av_log(avctx, AV_LOG_ERROR, "format not supported\n");
 601         return -1;
 602     }
 603     avctx->bits_per_coded_sample= s->bitstream_bpp;
 604     s->decorrelate= s->bitstream_bpp >= 24;
 605     s->predictor= avctx->prediction_method;
 606     s->interlaced= avctx->flags&CODEC_FLAG_INTERLACED_ME ? 1 : 0;
 607     if(avctx->context_model==1){
 608         s->context= avctx->context_model;
 609         if(s->flags & (CODEC_FLAG_PASS1|CODEC_FLAG_PASS2)){
 610             av_log(avctx, AV_LOG_ERROR, "context=1 is not compatible with 2 pass huffyuv encoding\n");
 611             return -1;
 612         }
 613     }else s->context= 0;
 614
 615     if(avctx->codec->id==AV_CODEC_ID_HUFFYUV){
 616         if(avctx->pix_fmt==PIX_FMT_YUV420P){
 617             av_log(avctx, AV_LOG_ERROR, "Error: YV12 is not supported by huffyuv; use vcodec=ffvhuff or format=422p\n");
 618             return -1;
 619         }
 620         if(avctx->context_model){
 621             av_log(avctx, AV_LOG_ERROR, "Error: per-frame huffman tables are not supported by huffyuv; use vcodec=ffvhuff\n");
 622             return -1;
 623         }
 624         if(s->interlaced != ( s->height > 288 ))
 625             av_log(avctx, AV_LOG_INFO, "using huffyuv 2.2.0 or newer interlacing flag\n");
 626     }
 627
 628     if(s->bitstream_bpp>=24 && s->predictor==MEDIAN){
 629         av_log(avctx, AV_LOG_ERROR, "Error: RGB is incompatible with median predictor\n");
 630         return -1;
 631     }
 632
 633     ((uint8_t*)avctx->extradata)[0]= s->predictor | (s->decorrelate << 6);
 634     ((uint8_t*)avctx->extradata)[1]= s->bitstream_bpp;
 635     ((uint8_t*)avctx->extradata)[2]= s->interlaced ? 0x10 : 0x20;
 636     if(s->context)
 637         ((uint8_t*)avctx->extradata)[2]|= 0x40;
 638     ((uint8_t*)avctx->extradata)[3]= 0;
 639     s->avctx->extradata_size= 4;
 640
 641     if(avctx->stats_in){
 642         char *p= avctx->stats_in;
 643
 644         for(i=0; i<3; i++)
 645             for(j=0; j<256; j++)
 646                 s->stats[i][j]= 1;
 647
 648         for(;;){
 649             for(i=0; i<3; i++){
 650                 char *next;
 651
 652                 for(j=0; j<256; j++){
 653                     s->stats[i][j]+= strtol(p, &next, 0);
 654                     if(next==p) return -1;
 655                     p=next;
 656                 }
 657             }
 658             if(p[0]==0 || p[1]==0 || p[2]==0) break;
 659         }
 660     }else{
 661         for(i=0; i<3; i++)
 662             for(j=0; j<256; j++){
 663                 int d= FFMIN(j, 256-j);
 664
 665                 s->stats[i][j]= 100000000/(d+1);
 666             }
 667     }
 668
 669     for(i=0; i<3; i++){
 670         generate_len_table(s->len[i], s->stats[i]);
 671
 672         if(generate_bits_table(s->bits[i], s->len[i])<0){
 673             return -1;
 674         }
 675
 676         s->avctx->extradata_size+=
 677         store_table(s, s->len[i], &((uint8_t*)s->avctx->extradata)[s->avctx->extradata_size]);
 678     }
 679
 680     if(s->context){
 681         for(i=0; i<3; i++){
 682             int pels = s->width*s->height / (i?40:10);
 683             for(j=0; j<256; j++){
 684                 int d= FFMIN(j, 256-j);
 685                 s->stats[i][j]= pels/(d+1);
 686             }
 687         }
 688     }else{
 689         for(i=0; i<3; i++)
 690             for(j=0; j<256; j++)
 691                 s->stats[i][j]= 0;
 692     }
 693
 694 //    printf("pred:%d bpp:%d hbpp:%d il:%d\n", s->predictor, s->bitstream_bpp, avctx->bits_per_coded_sample, s->interlaced);
 695
 696     alloc_temp(s);
 697
 698     s->picture_number=0;
 699
 700     return 0;
 701 }
 702 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */
 703
 704 /* TODO instead of restarting the read when the code isn't in the first level
 705  * of the joint table, jump into the 2nd level of the individual table. */
 706 #define READ_2PIX(dst0, dst1, plane1){\
 707     uint16_t code = get_vlc2(&s->gb, s->vlc[3+plane1].table, VLC_BITS, 1);\
 708     if(code != 0xffff){\
 709         dst0 = code>>8;\
 710         dst1 = code;\
 711     }else{\
 712         dst0 = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3);\
 713         dst1 = get_vlc2(&s->gb, s->vlc[plane1].table, VLC_BITS, 3);\
 714     }\
 715 }
 716
 717 static void decode_422_bitstream(HYuvContext *s, int count){
 718     int i;
 719
 720     count/=2;
 721
 722     if(count >= (get_bits_left(&s->gb))/(31*4)){
 723         for (i = 0; i < count && get_bits_left(&s->gb) > 0; i++) {
 724             READ_2PIX(s->temp[0][2*i  ], s->temp[1][i], 1);
 725             READ_2PIX(s->temp[0][2*i+1], s->temp[2][i], 2);
 726         }
 727     }else{
 728         for(i=0; i<count; i++){
 729             READ_2PIX(s->temp[0][2*i  ], s->temp[1][i], 1);
 730             READ_2PIX(s->temp[0][2*i+1], s->temp[2][i], 2);
 731         }
 732     }
 733 }
 734
 735 static void decode_gray_bitstream(HYuvContext *s, int count){
 736     int i;
 737
 738     count/=2;
 739
 740     if(count >= (get_bits_left(&s->gb))/(31*2)){
 741         for (i = 0; i < count && get_bits_left(&s->gb) > 0; i++) {
 742             READ_2PIX(s->temp[0][2*i  ], s->temp[0][2*i+1], 0);
 743         }
 744     }else{
 745         for(i=0; i<count; i++){
 746             READ_2PIX(s->temp[0][2*i  ], s->temp[0][2*i+1], 0);
 747         }
 748     }
 749 }
 750
 751 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER
 752 static int encode_422_bitstream(HYuvContext *s, int offset, int count){
 753     int i;
 754     const uint8_t *y = s->temp[0] + offset;
 755     const uint8_t *u = s->temp[1] + offset/2;
 756     const uint8_t *v = s->temp[2] + offset/2;
 757
 758     if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 2*4*count){
 759         av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 760         return -1;
 761     }
 762
 763 #define LOAD4\
 764             int y0 = y[2*i];\
 765             int y1 = y[2*i+1];\
 766             int u0 = u[i];\
 767             int v0 = v[i];
 768
 769     count/=2;
 770     if(s->flags&CODEC_FLAG_PASS1){
 771         for(i=0; i<count; i++){
 772             LOAD4;
 773             s->stats[0][y0]++;
 774             s->stats[1][u0]++;
 775             s->stats[0][y1]++;
 776             s->stats[2][v0]++;
 777         }
 778     }
 779     if(s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT)
 780         return 0;
 781     if(s->context){
 782         for(i=0; i<count; i++){
 783             LOAD4;
 784             s->stats[0][y0]++;
 785             put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]);
 786             s->stats[1][u0]++;
 787             put_bits(&s->pb, s->len[1][u0], s->bits[1][u0]);
 788             s->stats[0][y1]++;
 789             put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]);
 790             s->stats[2][v0]++;
 791             put_bits(&s->pb, s->len[2][v0], s->bits[2][v0]);
 792         }
 793     }else{
 794         for(i=0; i<count; i++){
 795             LOAD4;
 796             put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]);
 797             put_bits(&s->pb, s->len[1][u0], s->bits[1][u0]);
 798             put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]);
 799             put_bits(&s->pb, s->len[2][v0], s->bits[2][v0]);
 800         }
 801     }
 802     return 0;
 803 }
 804
 805 static int encode_gray_bitstream(HYuvContext *s, int count){
 806     int i;
 807
 808     if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 4*count){
 809         av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 810         return -1;
 811     }
 812
 813 #define LOAD2\
 814             int y0 = s->temp[0][2*i];\
 815             int y1 = s->temp[0][2*i+1];
 816 #define STAT2\
 817             s->stats[0][y0]++;\
 818             s->stats[0][y1]++;
 819 #define WRITE2\
 820             put_bits(&s->pb, s->len[0][y0], s->bits[0][y0]);\
 821             put_bits(&s->pb, s->len[0][y1], s->bits[0][y1]);
 822
 823     count/=2;
 824     if(s->flags&CODEC_FLAG_PASS1){
 825         for(i=0; i<count; i++){
 826             LOAD2;
 827             STAT2;
 828         }
 829     }
 830     if(s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT)
 831         return 0;
 832
 833     if(s->context){
 834         for(i=0; i<count; i++){
 835             LOAD2;
 836             STAT2;
 837             WRITE2;
 838         }
 839     }else{
 840         for(i=0; i<count; i++){
 841             LOAD2;
 842             WRITE2;
 843         }
 844     }
 845     return 0;
 846 }
 847 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */
 848
 849 static av_always_inline void decode_bgr_1(HYuvContext *s, int count, int decorrelate, int alpha){
 850     int i;
 851     for(i=0; i<count; i++){
 852         int code = get_vlc2(&s->gb, s->vlc[3].table, VLC_BITS, 1);
 853         if(code != -1){
 854             *(uint32_t*)&s->temp[0][4*i] = s->pix_bgr_map[code];
 855         }else if(decorrelate){
 856             s->temp[0][4*i+G] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3);
 857             s->temp[0][4*i+B] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3) + s->temp[0][4*i+G];
 858             s->temp[0][4*i+R] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3) + s->temp[0][4*i+G];
 859         }else{
 860             s->temp[0][4*i+B] = get_vlc2(&s->gb, s->vlc[0].table, VLC_BITS, 3);
 861             s->temp[0][4*i+G] = get_vlc2(&s->gb, s->vlc[1].table, VLC_BITS, 3);
 862             s->temp[0][4*i+R] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3);
 863         }
 864         if(alpha)
 865             s->temp[0][4*i+A] = get_vlc2(&s->gb, s->vlc[2].table, VLC_BITS, 3);
 866     }
 867 }
 868
 869 static void decode_bgr_bitstream(HYuvContext *s, int count){
 870     if(s->decorrelate){
 871         if(s->bitstream_bpp==24)
 872             decode_bgr_1(s, count, 1, 0);
 873         else
 874             decode_bgr_1(s, count, 1, 1);
 875     }else{
 876         if(s->bitstream_bpp==24)
 877             decode_bgr_1(s, count, 0, 0);
 878         else
 879             decode_bgr_1(s, count, 0, 1);
 880     }
 881 }
 882
 883 static int encode_bgr_bitstream(HYuvContext *s, int count){
 884     int i;
 885
 886     if(s->pb.buf_end - s->pb.buf - (put_bits_count(&s->pb)>>3) < 3*4*count){
 887         av_log(s->avctx, AV_LOG_ERROR, "encoded frame too large\n");
 888         return -1;
 889     }
 890
 891 #define LOAD3\
 892             int g= s->temp[0][4*i+G];\
 893             int b= (s->temp[0][4*i+B] - g) & 0xff;\
 894             int r= (s->temp[0][4*i+R] - g) & 0xff;
 895 #define STAT3\
 896             s->stats[0][b]++;\
 897             s->stats[1][g]++;\
 898             s->stats[2][r]++;
 899 #define WRITE3\
 900             put_bits(&s->pb, s->len[1][g], s->bits[1][g]);\
 901             put_bits(&s->pb, s->len[0][b], s->bits[0][b]);\
 902             put_bits(&s->pb, s->len[2][r], s->bits[2][r]);
 903
 904     if((s->flags&CODEC_FLAG_PASS1) && (s->avctx->flags2&CODEC_FLAG2_NO_OUTPUT)){
 905         for(i=0; i<count; i++){
 906             LOAD3;
 907             STAT3;
 908         }
 909     }else if(s->context || (s->flags&CODEC_FLAG_PASS1)){
 910         for(i=0; i<count; i++){
 911             LOAD3;
 912             STAT3;
 913             WRITE3;
 914         }
 915     }else{
 916         for(i=0; i<count; i++){
 917             LOAD3;
 918             WRITE3;
 919         }
 920     }
 921     return 0;
 922 }
 923
 924 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER
 925 static void draw_slice(HYuvContext *s, int y){
 926     int h, cy, i;
 927     int offset[AV_NUM_DATA_POINTERS];
 928
 929     if(s->avctx->draw_horiz_band==NULL)
 930         return;
 931
 932     h= y - s->last_slice_end;
 933     y -= h;
 934
 935     if(s->bitstream_bpp==12){
 936         cy= y>>1;
 937     }else{
 938         cy= y;
 939     }
 940
 941     offset[0] = s->picture.linesize[0]*y;
 942     offset[1] = s->picture.linesize[1]*cy;
 943     offset[2] = s->picture.linesize[2]*cy;
 944     for (i = 3; i < AV_NUM_DATA_POINTERS; i++)
 945         offset[i] = 0;
 946     emms_c();
 947
 948     s->avctx->draw_horiz_band(s->avctx, &s->picture, offset, y, 3, h);
 949
 950     s->last_slice_end= y + h;
 951 }
 952
 953 static int decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt){
 954     const uint8_t *buf = avpkt->data;
 955     int buf_size = avpkt->size;
 956     HYuvContext *s = avctx->priv_data;
 957     const int width= s->width;
 958     const int width2= s->width>>1;
 959     const int height= s->height;
 960     int fake_ystride, fake_ustride, fake_vstride;
 961     AVFrame * const p= &s->picture;
 962     int table_size= 0;
 963
 964     AVFrame *picture = data;
 965
 966     av_fast_malloc(&s->bitstream_buffer, &s->bitstream_buffer_size, buf_size + FF_INPUT_BUFFER_PADDING_SIZE);
 967     if (!s->bitstream_buffer)
 968         return AVERROR(ENOMEM);
 969
 970     memset(s->bitstream_buffer + buf_size, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 971     s->dsp.bswap_buf((uint32_t*)s->bitstream_buffer, (const uint32_t*)buf, buf_size/4);
 972
 973     if(p->data[0])
 974         ff_thread_release_buffer(avctx, p);
 975
 976     p->reference= 0;
 977     if(ff_thread_get_buffer(avctx, p) < 0){
 978         av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
 979         return -1;
 980     }
 981
 982     if(s->context){
 983         table_size = read_huffman_tables(s, s->bitstream_buffer, buf_size);
 984         if(table_size < 0)
 985             return -1;
 986     }
 987
 988     if((unsigned)(buf_size-table_size) >= INT_MAX/8)
 989         return -1;
 990
 991     init_get_bits(&s->gb, s->bitstream_buffer+table_size, (buf_size-table_size)*8);
 992
 993     fake_ystride= s->interlaced ? p->linesize[0]*2  : p->linesize[0];
 994     fake_ustride= s->interlaced ? p->linesize[1]*2  : p->linesize[1];
 995     fake_vstride= s->interlaced ? p->linesize[2]*2  : p->linesize[2];
 996
 997     s->last_slice_end= 0;
 998
 999     if(s->bitstream_bpp<24){
1000         int y, cy;
1001         int lefty, leftu, leftv;
1002         int lefttopy, lefttopu, lefttopv;
1003
1004         if(s->yuy2){
1005             p->data[0][3]= get_bits(&s->gb, 8);
1006             p->data[0][2]= get_bits(&s->gb, 8);
1007             p->data[0][1]= get_bits(&s->gb, 8);
1008             p->data[0][0]= get_bits(&s->gb, 8);
1009
1010             av_log(avctx, AV_LOG_ERROR, "YUY2 output is not implemented yet\n");
1011             return -1;
1012         }else{
1013
1014             leftv= p->data[2][0]= get_bits(&s->gb, 8);
1015             lefty= p->data[0][1]= get_bits(&s->gb, 8);
1016             leftu= p->data[1][0]= get_bits(&s->gb, 8);
1017                    p->data[0][0]= get_bits(&s->gb, 8);
1018
1019             switch(s->predictor){
1020             case LEFT:
1021             case PLANE:
1022                 decode_422_bitstream(s, width-2);
1023                 lefty= s->dsp.add_hfyu_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty);
1024                 if(!(s->flags&CODEC_FLAG_GRAY)){
1025                     leftu= s->dsp.add_hfyu_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu);
1026                     leftv= s->dsp.add_hfyu_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv);
1027                 }
1028
1029                 for(cy=y=1; y<s->height; y++,cy++){
1030                     uint8_t *ydst, *udst, *vdst;
1031
1032                     if(s->bitstream_bpp==12){
1033                         decode_gray_bitstream(s, width);
1034
1035                         ydst= p->data[0] + p->linesize[0]*y;
1036
1037                         lefty= s->dsp.add_hfyu_left_prediction(ydst, s->temp[0], width, lefty);
1038                         if(s->predictor == PLANE){
1039                             if(y>s->interlaced)
1040                                 s->dsp.add_bytes(ydst, ydst - fake_ystride, width);
1041                         }
1042                         y++;
1043                         if(y>=s->height) break;
1044                     }
1045
1046                     draw_slice(s, y);
1047
1048                     ydst= p->data[0] + p->linesize[0]*y;
1049                     udst= p->data[1] + p->linesize[1]*cy;
1050                     vdst= p->data[2] + p->linesize[2]*cy;
1051
1052                     decode_422_bitstream(s, width);
1053                     lefty= s->dsp.add_hfyu_left_prediction(ydst, s->temp[0], width, lefty);
1054                     if(!(s->flags&CODEC_FLAG_GRAY)){
1055                         leftu= s->dsp.add_hfyu_left_prediction(udst, s->temp[1], width2, leftu);
1056                         leftv= s->dsp.add_hfyu_left_prediction(vdst, s->temp[2], width2, leftv);
1057                     }
1058                     if(s->predictor == PLANE){
1059                         if(cy>s->interlaced){
1060                             s->dsp.add_bytes(ydst, ydst - fake_ystride, width);
1061                             if(!(s->flags&CODEC_FLAG_GRAY)){
1062                                 s->dsp.add_bytes(udst, udst - fake_ustride, width2);
1063                                 s->dsp.add_bytes(vdst, vdst - fake_vstride, width2);
1064                             }
1065                         }
1066                     }
1067                 }
1068                 draw_slice(s, height);
1069
1070                 break;
1071             case MEDIAN:
1072                 /* first line except first 2 pixels is left predicted */
1073                 decode_422_bitstream(s, width-2);
1074                 lefty= s->dsp.add_hfyu_left_prediction(p->data[0] + 2, s->temp[0], width-2, lefty);
1075                 if(!(s->flags&CODEC_FLAG_GRAY)){
1076                     leftu= s->dsp.add_hfyu_left_prediction(p->data[1] + 1, s->temp[1], width2-1, leftu);
1077                     leftv= s->dsp.add_hfyu_left_prediction(p->data[2] + 1, s->temp[2], width2-1, leftv);
1078                 }
1079
1080                 cy=y=1;
1081
1082                 /* second line is left predicted for interlaced case */
1083                 if(s->interlaced){
1084                     decode_422_bitstream(s, width);
1085                     lefty= s->dsp.add_hfyu_left_prediction(p->data[0] + p->linesize[0], s->temp[0], width, lefty);
1086                     if(!(s->flags&CODEC_FLAG_GRAY)){
1087                         leftu= s->dsp.add_hfyu_left_prediction(p->data[1] + p->linesize[2], s->temp[1], width2, leftu);
1088                         leftv= s->dsp.add_hfyu_left_prediction(p->data[2] + p->linesize[1], s->temp[2], width2, leftv);
1089                     }
1090                     y++; cy++;
1091                 }
1092
1093                 /* next 4 pixels are left predicted too */
1094                 decode_422_bitstream(s, 4);
1095                 lefty= s->dsp.add_hfyu_left_prediction(p->data[0] + fake_ystride, s->temp[0], 4, lefty);
1096                 if(!(s->flags&CODEC_FLAG_GRAY)){
1097                     leftu= s->dsp.add_hfyu_left_prediction(p->data[1] + fake_ustride, s->temp[1], 2, leftu);
1098                     leftv= s->dsp.add_hfyu_left_prediction(p->data[2] + fake_vstride, s->temp[2], 2, leftv);
1099                 }
1100
1101                 /* next line except the first 4 pixels is median predicted */
1102                 lefttopy= p->data[0][3];
1103                 decode_422_bitstream(s, width-4);
1104                 s->dsp.add_hfyu_median_prediction(p->data[0] + fake_ystride+4, p->data[0]+4, s->temp[0], width-4, &lefty, &lefttopy);
1105                 if(!(s->flags&CODEC_FLAG_GRAY)){
1106                     lefttopu= p->data[1][1];
1107                     lefttopv= p->data[2][1];
1108                     s->dsp.add_hfyu_median_prediction(p->data[1] + fake_ustride+2, p->data[1]+2, s->temp[1], width2-2, &leftu, &lefttopu);
1109                     s->dsp.add_hfyu_median_prediction(p->data[2] + fake_vstride+2, p->data[2]+2, s->temp[2], width2-2, &leftv, &lefttopv);
1110                 }
1111                 y++; cy++;
1112
1113                 for(; y<height; y++,cy++){
1114                     uint8_t *ydst, *udst, *vdst;
1115
1116                     if(s->bitstream_bpp==12){
1117                         while(2*cy > y){
1118                             decode_gray_bitstream(s, width);
1119                             ydst= p->data[0] + p->linesize[0]*y;
1120                             s->dsp.add_hfyu_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
1121                             y++;
1122                         }
1123                         if(y>=height) break;
1124                     }
1125                     draw_slice(s, y);
1126
1127                     decode_422_bitstream(s, width);
1128
1129                     ydst= p->data[0] + p->linesize[0]*y;
1130                     udst= p->data[1] + p->linesize[1]*cy;
1131                     vdst= p->data[2] + p->linesize[2]*cy;
1132
1133                     s->dsp.add_hfyu_median_prediction(ydst, ydst - fake_ystride, s->temp[0], width, &lefty, &lefttopy);
1134                     if(!(s->flags&CODEC_FLAG_GRAY)){
1135                         s->dsp.add_hfyu_median_prediction(udst, udst - fake_ustride, s->temp[1], width2, &leftu, &lefttopu);
1136                         s->dsp.add_hfyu_median_prediction(vdst, vdst - fake_vstride, s->temp[2], width2, &leftv, &lefttopv);
1137                     }
1138                 }
1139
1140                 draw_slice(s, height);
1141                 break;
1142             }
1143         }
1144     }else{
1145         int y;
1146         int leftr, leftg, leftb, lefta;
1147         const int last_line= (height-1)*p->linesize[0];
1148
1149         if(s->bitstream_bpp==32){
1150             lefta= p->data[0][last_line+A]= get_bits(&s->gb, 8);
1151             leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8);
1152             leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8);
1153             leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8);
1154         }else{
1155             leftr= p->data[0][last_line+R]= get_bits(&s->gb, 8);
1156             leftg= p->data[0][last_line+G]= get_bits(&s->gb, 8);
1157             leftb= p->data[0][last_line+B]= get_bits(&s->gb, 8);
1158             lefta= p->data[0][last_line+A]= 255;
1159             skip_bits(&s->gb, 8);
1160         }
1161
1162         if(s->bgr32){
1163             switch(s->predictor){
1164             case LEFT:
1165             case PLANE:
1166                 decode_bgr_bitstream(s, width-1);
1167                 s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + last_line+4, s->temp[0], width-1, &leftr, &leftg, &leftb, &lefta);
1168
1169                 for(y=s->height-2; y>=0; y--){ //Yes it is stored upside down.
1170                     decode_bgr_bitstream(s, width);
1171
1172                     s->dsp.add_hfyu_left_prediction_bgr32(p->data[0] + p->linesize[0]*y, s->temp[0], width, &leftr, &leftg, &leftb, &lefta);
1173                     if(s->predictor == PLANE){
1174                         if(s->bitstream_bpp!=32) lefta=0;
1175                         if((y&s->interlaced)==0 && y<s->height-1-s->interlaced){
1176                             s->dsp.add_bytes(p->data[0] + p->linesize[0]*y,
1177                                              p->data[0] + p->linesize[0]*y + fake_ystride, fake_ystride);
1178                         }
1179                     }
1180                 }
1181                 draw_slice(s, height); // just 1 large slice as this is not possible in reverse order
1182                 break;
1183             default:
1184                 av_log(avctx, AV_LOG_ERROR, "prediction type not supported!\n");
1185             }
1186         }else{
1187
1188             av_log(avctx, AV_LOG_ERROR, "BGR24 output is not implemented yet\n");
1189             return -1;
1190         }
1191     }
1192     emms_c();
1193
1194     *picture= *p;
1195     *data_size = sizeof(AVFrame);
1196
1197     return (get_bits_count(&s->gb)+31)/32*4 + table_size;
1198 }
1199 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */
1200
1201 static int common_end(HYuvContext *s){
1202     int i;
1203
1204     for(i=0; i<3; i++){
1205         av_freep(&s->temp[i]);
1206     }
1207     return 0;
1208 }
1209
1210 #if CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER
1211 static av_cold int decode_end(AVCodecContext *avctx)
1212 {
1213     HYuvContext *s = avctx->priv_data;
1214     int i;
1215
1216     if (s->picture.data[0])
1217         avctx->release_buffer(avctx, &s->picture);
1218
1219     common_end(s);
1220     av_freep(&s->bitstream_buffer);
1221
1222     for(i=0; i<6; i++){
1223         ff_free_vlc(&s->vlc[i]);
1224     }
1225
1226     return 0;
1227 }
1228 #endif /* CONFIG_HUFFYUV_DECODER || CONFIG_FFVHUFF_DECODER */
1229
1230 #if CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER
1231 static int encode_frame(AVCodecContext *avctx, AVPacket *pkt,
1232                         const AVFrame *pict, int *got_packet)
1233 {
1234     HYuvContext *s = avctx->priv_data;
1235     const int width= s->width;
1236     const int width2= s->width>>1;
1237     const int height= s->height;
1238     const int fake_ystride= s->interlaced ? pict->linesize[0]*2  : pict->linesize[0];
1239     const int fake_ustride= s->interlaced ? pict->linesize[1]*2  : pict->linesize[1];
1240     const int fake_vstride= s->interlaced ? pict->linesize[2]*2  : pict->linesize[2];
1241     AVFrame * const p= &s->picture;
1242     int i, j, size = 0, ret;
1243
1244     if (!pkt->data &&
1245         (ret = av_new_packet(pkt, width * height * 3 * 4 + FF_MIN_BUFFER_SIZE)) < 0) {
1246         av_log(avctx, AV_LOG_ERROR, "Error allocating output packet.\n");
1247         return ret;
1248     }
1249
1250     *p = *pict;
1251     p->pict_type= AV_PICTURE_TYPE_I;
1252     p->key_frame= 1;
1253
1254     if(s->context){
1255         for(i=0; i<3; i++){
1256             generate_len_table(s->len[i], s->stats[i]);
1257             if(generate_bits_table(s->bits[i], s->len[i])<0)
1258                 return -1;
1259             size += store_table(s, s->len[i], &pkt->data[size]);
1260         }
1261
1262         for(i=0; i<3; i++)
1263             for(j=0; j<256; j++)
1264                 s->stats[i][j] >>= 1;
1265     }
1266
1267     init_put_bits(&s->pb, pkt->data + size, pkt->size - size);
1268
1269     if(avctx->pix_fmt == PIX_FMT_YUV422P || avctx->pix_fmt == PIX_FMT_YUV420P){
1270         int lefty, leftu, leftv, y, cy;
1271
1272         put_bits(&s->pb, 8, leftv= p->data[2][0]);
1273         put_bits(&s->pb, 8, lefty= p->data[0][1]);
1274         put_bits(&s->pb, 8, leftu= p->data[1][0]);
1275         put_bits(&s->pb, 8,        p->data[0][0]);
1276
1277         lefty= sub_left_prediction(s, s->temp[0], p->data[0], width , 0);
1278         leftu= sub_left_prediction(s, s->temp[1], p->data[1], width2, 0);
1279         leftv= sub_left_prediction(s, s->temp[2], p->data[2], width2, 0);
1280
1281         encode_422_bitstream(s, 2, width-2);
1282
1283         if(s->predictor==MEDIAN){
1284             int lefttopy, lefttopu, lefttopv;
1285             cy=y=1;
1286             if(s->interlaced){
1287                 lefty= sub_left_prediction(s, s->temp[0], p->data[0]+p->linesize[0], width , lefty);
1288                 leftu= sub_left_prediction(s, s->temp[1], p->data[1]+p->linesize[1], width2, leftu);
1289                 leftv= sub_left_prediction(s, s->temp[2], p->data[2]+p->linesize[2], width2, leftv);
1290
1291                 encode_422_bitstream(s, 0, width);
1292                 y++; cy++;
1293             }
1294
1295             lefty= sub_left_prediction(s, s->temp[0], p->data[0]+fake_ystride, 4, lefty);
1296             leftu= sub_left_prediction(s, s->temp[1], p->data[1]+fake_ustride, 2, leftu);
1297             leftv= sub_left_prediction(s, s->temp[2], p->data[2]+fake_vstride, 2, leftv);
1298
1299             encode_422_bitstream(s, 0, 4);
1300
1301             lefttopy= p->data[0][3];
1302             lefttopu= p->data[1][1];
1303             lefttopv= p->data[2][1];
1304             s->dsp.sub_hfyu_median_prediction(s->temp[0], p->data[0]+4, p->data[0] + fake_ystride+4, width-4 , &lefty, &lefttopy);
1305             s->dsp.sub_hfyu_median_prediction(s->temp[1], p->data[1]+2, p->data[1] + fake_ustride+2, width2-2, &leftu, &lefttopu);
1306             s->dsp.sub_hfyu_median_prediction(s->temp[2], p->data[2]+2, p->data[2] + fake_vstride+2, width2-2, &leftv, &lefttopv);
1307             encode_422_bitstream(s, 0, width-4);
1308             y++; cy++;
1309
1310             for(; y<height; y++,cy++){
1311                 uint8_t *ydst, *udst, *vdst;
1312
1313                 if(s->bitstream_bpp==12){
1314                     while(2*cy > y){
1315                         ydst= p->data[0] + p->linesize[0]*y;
1316                         s->dsp.sub_hfyu_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
1317                         encode_gray_bitstream(s, width);
1318                         y++;
1319                     }
1320                     if(y>=height) break;
1321                 }
1322                 ydst= p->data[0] + p->linesize[0]*y;
1323                 udst= p->data[1] + p->linesize[1]*cy;
1324                 vdst= p->data[2] + p->linesize[2]*cy;
1325
1326                 s->dsp.sub_hfyu_median_prediction(s->temp[0], ydst - fake_ystride, ydst, width , &lefty, &lefttopy);
1327                 s->dsp.sub_hfyu_median_prediction(s->temp[1], udst - fake_ustride, udst, width2, &leftu, &lefttopu);
1328                 s->dsp.sub_hfyu_median_prediction(s->temp[2], vdst - fake_vstride, vdst, width2, &leftv, &lefttopv);
1329
1330                 encode_422_bitstream(s, 0, width);
1331             }
1332         }else{
1333             for(cy=y=1; y<height; y++,cy++){
1334                 uint8_t *ydst, *udst, *vdst;
1335
1336                 /* encode a luma only line & y++ */
1337                 if(s->bitstream_bpp==12){
1338                     ydst= p->data[0] + p->linesize[0]*y;
1339
1340                     if(s->predictor == PLANE && s->interlaced < y){
1341                         s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);
1342
1343                         lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty);
1344                     }else{
1345                         lefty= sub_left_prediction(s, s->temp[0], ydst, width , lefty);
1346                     }
1347                     encode_gray_bitstream(s, width);
1348                     y++;
1349                     if(y>=height) break;
1350                 }
1351
1352                 ydst= p->data[0] + p->linesize[0]*y;
1353                 udst= p->data[1] + p->linesize[1]*cy;
1354                 vdst= p->data[2] + p->linesize[2]*cy;
1355
1356                 if(s->predictor == PLANE && s->interlaced < cy){
1357                     s->dsp.diff_bytes(s->temp[1], ydst, ydst - fake_ystride, width);
1358                     s->dsp.diff_bytes(s->temp[2], udst, udst - fake_ustride, width2);
1359                     s->dsp.diff_bytes(s->temp[2] + width2, vdst, vdst - fake_vstride, width2);
1360
1361                     lefty= sub_left_prediction(s, s->temp[0], s->temp[1], width , lefty);
1362                     leftu= sub_left_prediction(s, s->temp[1], s->temp[2], width2, leftu);
1363                     leftv= sub_left_prediction(s, s->temp[2], s->temp[2] + width2, width2, leftv);
1364                 }else{
1365                     lefty= sub_left_prediction(s, s->temp[0], ydst, width , lefty);
1366                     leftu= sub_left_prediction(s, s->temp[1], udst, width2, leftu);
1367                     leftv= sub_left_prediction(s, s->temp[2], vdst, width2, leftv);
1368                 }
1369
1370                 encode_422_bitstream(s, 0, width);
1371             }
1372         }
1373     }else if(avctx->pix_fmt == PIX_FMT_RGB32){
1374         uint8_t *data = p->data[0] + (height-1)*p->linesize[0];
1375         const int stride = -p->linesize[0];
1376         const int fake_stride = -fake_ystride;
1377         int y;
1378         int leftr, leftg, leftb;
1379
1380         put_bits(&s->pb, 8, leftr= data[R]);
1381         put_bits(&s->pb, 8, leftg= data[G]);
1382         put_bits(&s->pb, 8, leftb= data[B]);
1383         put_bits(&s->pb, 8, 0);
1384
1385         sub_left_prediction_bgr32(s, s->temp[0], data+4, width-1, &leftr, &leftg, &leftb);
1386         encode_bgr_bitstream(s, width-1);
1387
1388         for(y=1; y<s->height; y++){
1389             uint8_t *dst = data + y*stride;
1390             if(s->predictor == PLANE && s->interlaced < y){
1391                 s->dsp.diff_bytes(s->temp[1], dst, dst - fake_stride, width*4);
1392                 sub_left_prediction_bgr32(s, s->temp[0], s->temp[1], width, &leftr, &leftg, &leftb);
1393             }else{
1394                 sub_left_prediction_bgr32(s, s->temp[0], dst, width, &leftr, &leftg, &leftb);
1395             }
1396             encode_bgr_bitstream(s, width);
1397         }
1398     }else{
1399         av_log(avctx, AV_LOG_ERROR, "Format not supported!\n");
1400     }
1401     emms_c();
1402
1403     size+= (put_bits_count(&s->pb)+31)/8;
1404     put_bits(&s->pb, 16, 0);
1405     put_bits(&s->pb, 15, 0);
1406     size/= 4;
1407
1408     if((s->flags&CODEC_FLAG_PASS1) && (s->picture_number&31)==0){
1409         int j;
1410         char *p= avctx->stats_out;
1411         char *end= p + 1024*30;
1412         for(i=0; i<3; i++){
1413             for(j=0; j<256; j++){
1414                 snprintf(p, end-p, "%"PRIu64" ", s->stats[i][j]);
1415                 p+= strlen(p);
1416                 s->stats[i][j]= 0;
1417             }
1418             snprintf(p, end-p, "\n");
1419             p++;
1420         }
1421     } else
1422         avctx->stats_out[0] = '\0';
1423     if(!(s->avctx->flags2 & CODEC_FLAG2_NO_OUTPUT)){
1424         flush_put_bits(&s->pb);
1425         s->dsp.bswap_buf((uint32_t*)pkt->data, (uint32_t*)pkt->data, size);
1426     }
1427
1428     s->picture_number++;
1429
1430     pkt->size   = size*4;
1431     pkt->flags |= AV_PKT_FLAG_KEY;
1432     *got_packet = 1;
1433
1434     return 0;
1435 }
1436
1437 static av_cold int encode_end(AVCodecContext *avctx)
1438 {
1439     HYuvContext *s = avctx->priv_data;
1440
1441     common_end(s);
1442
1443     av_freep(&avctx->extradata);
1444     av_freep(&avctx->stats_out);
1445
1446     return 0;
1447 }
1448 #endif /* CONFIG_HUFFYUV_ENCODER || CONFIG_FFVHUFF_ENCODER */
1449
1450 #if CONFIG_HUFFYUV_DECODER
1451 AVCodec ff_huffyuv_decoder = {
1452     .name             = "huffyuv",
1453     .type             = AVMEDIA_TYPE_VIDEO,
1454     .id               = AV_CODEC_ID_HUFFYUV,
1455     .priv_data_size   = sizeof(HYuvContext),
1456     .init             = decode_init,
1457     .close            = decode_end,
1458     .decode           = decode_frame,
1459     .capabilities     = CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND |
1460                         CODEC_CAP_FRAME_THREADS,
1461     .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
1462     .long_name        = NULL_IF_CONFIG_SMALL("Huffyuv / HuffYUV"),
1463 };
1464 #endif
1465
1466 #if CONFIG_FFVHUFF_DECODER
1467 AVCodec ff_ffvhuff_decoder = {
1468     .name             = "ffvhuff",
1469     .type             = AVMEDIA_TYPE_VIDEO,
1470     .id               = AV_CODEC_ID_FFVHUFF,
1471     .priv_data_size   = sizeof(HYuvContext),
1472     .init             = decode_init,
1473     .close            = decode_end,
1474     .decode           = decode_frame,
1475     .capabilities     = CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND |
1476                         CODEC_CAP_FRAME_THREADS,
1477     .init_thread_copy = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
1478     .long_name        = NULL_IF_CONFIG_SMALL("Huffyuv FFmpeg variant"),
1479 };
1480 #endif
1481
1482 #if CONFIG_HUFFYUV_ENCODER
1483 AVCodec ff_huffyuv_encoder = {
1484     .name           = "huffyuv",
1485     .type           = AVMEDIA_TYPE_VIDEO,
1486     .id             = AV_CODEC_ID_HUFFYUV,
1487     .priv_data_size = sizeof(HYuvContext),
1488     .init           = encode_init,
1489     .encode2        = encode_frame,
1490     .close          = encode_end,
1491     .pix_fmts       = (const enum PixelFormat[]){
1492         PIX_FMT_YUV422P, PIX_FMT_RGB32, PIX_FMT_NONE
1493     },
1494     .long_name      = NULL_IF_CONFIG_SMALL("Huffyuv / HuffYUV"),
1495 };
1496 #endif
1497
1498 #if CONFIG_FFVHUFF_ENCODER
1499 AVCodec ff_ffvhuff_encoder = {
1500     .name           = "ffvhuff",
1501     .type           = AVMEDIA_TYPE_VIDEO,
1502     .id             = AV_CODEC_ID_FFVHUFF,
1503     .priv_data_size = sizeof(HYuvContext),
1504     .init           = encode_init,
1505     .encode2        = encode_frame,
1506     .close          = encode_end,
1507     .pix_fmts       = (const enum PixelFormat[]){
1508         PIX_FMT_YUV420P, PIX_FMT_YUV422P, PIX_FMT_RGB32, PIX_FMT_NONE
1509     },
1510     .long_name      = NULL_IF_CONFIG_SMALL("Huffyuv FFmpeg variant"),
1511 };
1512 #endif