git.sesse.net Git - ffmpeg/blob - libavcodec/h264.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavcore/imgutils.h"
  29 #include "internal.h"
  30 #include "dsputil.h"
  31 #include "avcodec.h"
  32 #include "mpegvideo.h"
  33 #include "h264.h"
  34 #include "h264data.h"
  35 #include "h264_mvpred.h"
  36 #include "h264_parser.h"
  37 #include "golomb.h"
  38 #include "mathops.h"
  39 #include "rectangle.h"
  40 #include "vdpau_internal.h"
  41 #include "libavutil/avassert.h"
  42
  43 #include "cabac.h"
  44
  45 //#undef NDEBUG
  46 #include <assert.h>
  47
  48 static const uint8_t rem6[52]={
  49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[52]={
  53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
  54 };
  55
  56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  57     PIX_FMT_DXVA2_VLD,
  58     PIX_FMT_VAAPI_VLD,
  59     PIX_FMT_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 void ff_h264_write_back_intra_pred_mode(H264Context *h){
  64     int8_t *mode= h->intra4x4_pred_mode + h->mb2br_xy[h->mb_xy];
  65
  66     AV_COPY32(mode, h->intra4x4_pred_mode_cache + 4 + 8*4);
  67     mode[4]= h->intra4x4_pred_mode_cache[7+8*3];
  68     mode[5]= h->intra4x4_pred_mode_cache[7+8*2];
  69     mode[6]= h->intra4x4_pred_mode_cache[7+8*1];
  70 }
  71
  72 /**
  73  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  74  */
  75 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  76     MpegEncContext * const s = &h->s;
  77     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  78     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  79     int i;
  80
  81     if(!(h->top_samples_available&0x8000)){
  82         for(i=0; i<4; i++){
  83             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  84             if(status<0){
  85                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  86                 return -1;
  87             } else if(status){
  88                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  89             }
  90         }
  91     }
  92
  93     if((h->left_samples_available&0x8888)!=0x8888){
  94         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  95         for(i=0; i<4; i++){
  96             if(!(h->left_samples_available&mask[i])){
  97                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  98                 if(status<0){
  99                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
 100                     return -1;
 101                 } else if(status){
 102                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
 103                 }
 104             }
 105         }
 106     }
 107
 108     return 0;
 109 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 110
 111 /**
 112  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 113  */
 114 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 115     MpegEncContext * const s = &h->s;
 116     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 117     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 118
 119     if(mode > 6U) {
 120         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 121         return -1;
 122     }
 123
 124     if(!(h->top_samples_available&0x8000)){
 125         mode= top[ mode ];
 126         if(mode<0){
 127             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 128             return -1;
 129         }
 130     }
 131
 132     if((h->left_samples_available&0x8080) != 0x8080){
 133         mode= left[ mode ];
 134         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 135             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 136         }
 137         if(mode<0){
 138             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 139             return -1;
 140         }
 141     }
 142
 143     return mode;
 144 }
 145
 146 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 147     int i, si, di;
 148     uint8_t *dst;
 149     int bufidx;
 150
 151 //    src[0]&0x80;                //forbidden bit
 152     h->nal_ref_idc= src[0]>>5;
 153     h->nal_unit_type= src[0]&0x1F;
 154
 155     src++; length--;
 156 #if 0
 157     for(i=0; i<length; i++)
 158         printf("%2X ", src[i]);
 159 #endif
 160
 161 #if HAVE_FAST_UNALIGNED
 162 # if HAVE_FAST_64BIT
 163 #   define RS 7
 164     for(i=0; i+1<length; i+=9){
 165         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 166 # else
 167 #   define RS 3
 168     for(i=0; i+1<length; i+=5){
 169         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 170 # endif
 171             continue;
 172         if(i>0 && !src[i]) i--;
 173         while(src[i]) i++;
 174 #else
 175 #   define RS 0
 176     for(i=0; i+1<length; i+=2){
 177         if(src[i]) continue;
 178         if(i>0 && src[i-1]==0) i--;
 179 #endif
 180         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 181             if(src[i+2]!=3){
 182                 /* startcode, so we must be past the end */
 183                 length=i;
 184             }
 185             break;
 186         }
 187         i-= RS;
 188     }
 189
 190     if(i>=length-1){ //no escaped 0
 191         *dst_length= length;
 192         *consumed= length+1; //+1 for the header
 193         return src;
 194     }
 195
 196     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 197     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 198     dst= h->rbsp_buffer[bufidx];
 199
 200     if (dst == NULL){
 201         return NULL;
 202     }
 203
 204 //printf("decoding esc\n");
 205     memcpy(dst, src, i);
 206     si=di=i;
 207     while(si+2<length){
 208         //remove escapes (very rare 1:2^22)
 209         if(src[si+2]>3){
 210             dst[di++]= src[si++];
 211             dst[di++]= src[si++];
 212         }else if(src[si]==0 && src[si+1]==0){
 213             if(src[si+2]==3){ //escape
 214                 dst[di++]= 0;
 215                 dst[di++]= 0;
 216                 si+=3;
 217                 continue;
 218             }else //next start code
 219                 goto nsc;
 220         }
 221
 222         dst[di++]= src[si++];
 223     }
 224     while(si<length)
 225         dst[di++]= src[si++];
 226 nsc:
 227
 228     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 229
 230     *dst_length= di;
 231     *consumed= si + 1;//+1 for the header
 232 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 233     return dst;
 234 }
 235
 236 int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 237     int v= *src;
 238     int r;
 239
 240     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 241
 242     for(r=1; r<9; r++){
 243         if(v&1) return r;
 244         v>>=1;
 245     }
 246     return 0;
 247 }
 248
 249 #if 0
 250 /**
 251  * DCT transforms the 16 dc values.
 252  * @param qp quantization parameter ??? FIXME
 253  */
 254 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 255 //    const int qmul= dequant_coeff[qp][0];
 256     int i;
 257     int temp[16]; //FIXME check if this is a good idea
 258     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 259     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 260
 261     for(i=0; i<4; i++){
 262         const int offset= y_offset[i];
 263         const int z0= block[offset+stride*0] + block[offset+stride*4];
 264         const int z1= block[offset+stride*0] - block[offset+stride*4];
 265         const int z2= block[offset+stride*1] - block[offset+stride*5];
 266         const int z3= block[offset+stride*1] + block[offset+stride*5];
 267
 268         temp[4*i+0]= z0+z3;
 269         temp[4*i+1]= z1+z2;
 270         temp[4*i+2]= z1-z2;
 271         temp[4*i+3]= z0-z3;
 272     }
 273
 274     for(i=0; i<4; i++){
 275         const int offset= x_offset[i];
 276         const int z0= temp[4*0+i] + temp[4*2+i];
 277         const int z1= temp[4*0+i] - temp[4*2+i];
 278         const int z2= temp[4*1+i] - temp[4*3+i];
 279         const int z3= temp[4*1+i] + temp[4*3+i];
 280
 281         block[stride*0 +offset]= (z0 + z3)>>1;
 282         block[stride*2 +offset]= (z1 + z2)>>1;
 283         block[stride*8 +offset]= (z1 - z2)>>1;
 284         block[stride*10+offset]= (z0 - z3)>>1;
 285     }
 286 }
 287 #endif
 288
 289 #undef xStride
 290 #undef stride
 291
 292 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qmul){
 293     const int stride= 16*2;
 294     const int xStride= 16;
 295     int a,b,c,d,e;
 296
 297     a= block[stride*0 + xStride*0];
 298     b= block[stride*0 + xStride*1];
 299     c= block[stride*1 + xStride*0];
 300     d= block[stride*1 + xStride*1];
 301
 302     e= a-b;
 303     a= a+b;
 304     b= c-d;
 305     c= c+d;
 306
 307     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
 308     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
 309     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
 310     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
 311 }
 312
 313 #if 0
 314 static void chroma_dc_dct_c(DCTELEM *block){
 315     const int stride= 16*2;
 316     const int xStride= 16;
 317     int a,b,c,d,e;
 318
 319     a= block[stride*0 + xStride*0];
 320     b= block[stride*0 + xStride*1];
 321     c= block[stride*1 + xStride*0];
 322     d= block[stride*1 + xStride*1];
 323
 324     e= a-b;
 325     a= a+b;
 326     b= c-d;
 327     c= c+d;
 328
 329     block[stride*0 + xStride*0]= (a+c);
 330     block[stride*0 + xStride*1]= (e+b);
 331     block[stride*1 + xStride*0]= (a-c);
 332     block[stride*1 + xStride*1]= (e-b);
 333 }
 334 #endif
 335
 336 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 337                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 338                            int src_x_offset, int src_y_offset,
 339                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
 340     MpegEncContext * const s = &h->s;
 341     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 342     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 343     const int luma_xy= (mx&3) + ((my&3)<<2);
 344     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
 345     uint8_t * src_cb, * src_cr;
 346     int extra_width= h->emu_edge_width;
 347     int extra_height= h->emu_edge_height;
 348     int emu=0;
 349     const int full_mx= mx>>2;
 350     const int full_my= my>>2;
 351     const int pic_width  = 16*s->mb_width;
 352     const int pic_height = 16*s->mb_height >> MB_FIELD;
 353
 354     if(mx&7) extra_width -= 3;
 355     if(my&7) extra_height -= 3;
 356
 357     if(   full_mx < 0-extra_width
 358        || full_my < 0-extra_height
 359        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 360        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 361         ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 362             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
 363         emu=1;
 364     }
 365
 366     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 367     if(!square){
 368         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 369     }
 370
 371     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 372
 373     if(MB_FIELD){
 374         // chroma offset when predicting from a field of opposite parity
 375         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
 376         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 377     }
 378     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
 379     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
 380
 381     if(emu){
 382         ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 383             src_cb= s->edge_emu_buffer;
 384     }
 385     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 386
 387     if(emu){
 388         ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 389             src_cr= s->edge_emu_buffer;
 390     }
 391     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 392 }
 393
 394 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 395                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 396                            int x_offset, int y_offset,
 397                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 398                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 399                            int list0, int list1){
 400     MpegEncContext * const s = &h->s;
 401     qpel_mc_func *qpix_op=  qpix_put;
 402     h264_chroma_mc_func chroma_op= chroma_put;
 403
 404     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
 405     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
 406     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
 407     x_offset += 8*s->mb_x;
 408     y_offset += 8*(s->mb_y >> MB_FIELD);
 409
 410     if(list0){
 411         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 412         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 413                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 414                            qpix_op, chroma_op);
 415
 416         qpix_op=  qpix_avg;
 417         chroma_op= chroma_avg;
 418     }
 419
 420     if(list1){
 421         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 422         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 423                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 424                            qpix_op, chroma_op);
 425     }
 426 }
 427
 428 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 429                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 430                            int x_offset, int y_offset,
 431                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 432                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 433                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 434                            int list0, int list1){
 435     MpegEncContext * const s = &h->s;
 436
 437     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
 438     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
 439     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
 440     x_offset += 8*s->mb_x;
 441     y_offset += 8*(s->mb_y >> MB_FIELD);
 442
 443     if(list0 && list1){
 444         /* don't optimize for luma-only case, since B-frames usually
 445          * use implicit weights => chroma too. */
 446         uint8_t *tmp_cb = s->obmc_scratchpad;
 447         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
 448         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
 449         int refn0 = h->ref_cache[0][ scan8[n] ];
 450         int refn1 = h->ref_cache[1][ scan8[n] ];
 451
 452         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 453                     dest_y, dest_cb, dest_cr,
 454                     x_offset, y_offset, qpix_put, chroma_put);
 455         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 456                     tmp_y, tmp_cb, tmp_cr,
 457                     x_offset, y_offset, qpix_put, chroma_put);
 458
 459         if(h->use_weight == 2){
 460             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 461             int weight1 = 64 - weight0;
 462             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 463             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 464             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 465         }else{
 466             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 467                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 468                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 469             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 470                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 471                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 472             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 473                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 474                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 475         }
 476     }else{
 477         int list = list1 ? 1 : 0;
 478         int refn = h->ref_cache[list][ scan8[n] ];
 479         Picture *ref= &h->ref_list[list][refn];
 480         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 481                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 482                     qpix_put, chroma_put);
 483
 484         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 485                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 486         if(h->use_weight_chroma){
 487             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 488                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 489             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 490                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 491         }
 492     }
 493 }
 494
 495 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 496                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 497                            int x_offset, int y_offset,
 498                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 499                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 500                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 501                            int list0, int list1){
 502     if((h->use_weight==2 && list0 && list1
 503         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 504        || h->use_weight==1)
 505         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 506                          x_offset, y_offset, qpix_put, chroma_put,
 507                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
 508     else
 509         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 510                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
 511 }
 512
 513 static inline void prefetch_motion(H264Context *h, int list){
 514     /* fetch pixels for estimated mv 4 macroblocks ahead
 515      * optimized for 64byte cache lines */
 516     MpegEncContext * const s = &h->s;
 517     const int refn = h->ref_cache[list][scan8[0]];
 518     if(refn >= 0){
 519         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 520         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 521         uint8_t **src= h->ref_list[list][refn].data;
 522         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
 523         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 524         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
 525         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 526     }
 527 }
 528
 529 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 530                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 531                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 532                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
 533     MpegEncContext * const s = &h->s;
 534     const int mb_xy= h->mb_xy;
 535     const int mb_type= s->current_picture.mb_type[mb_xy];
 536
 537     assert(IS_INTER(mb_type));
 538
 539     prefetch_motion(h, 0);
 540
 541     if(IS_16X16(mb_type)){
 542         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 543                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 544                 weight_op, weight_avg,
 545                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 546     }else if(IS_16X8(mb_type)){
 547         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
 548                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 549                 &weight_op[1], &weight_avg[1],
 550                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 551         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
 552                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 553                 &weight_op[1], &weight_avg[1],
 554                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
 555     }else if(IS_8X16(mb_type)){
 556         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 557                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 558                 &weight_op[2], &weight_avg[2],
 559                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 560         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 561                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 562                 &weight_op[2], &weight_avg[2],
 563                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
 564     }else{
 565         int i;
 566
 567         assert(IS_8X8(mb_type));
 568
 569         for(i=0; i<4; i++){
 570             const int sub_mb_type= h->sub_mb_type[i];
 571             const int n= 4*i;
 572             int x_offset= (i&1)<<2;
 573             int y_offset= (i&2)<<1;
 574
 575             if(IS_SUB_8X8(sub_mb_type)){
 576                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 577                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 578                     &weight_op[3], &weight_avg[3],
 579                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 580             }else if(IS_SUB_8X4(sub_mb_type)){
 581                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 582                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 583                     &weight_op[4], &weight_avg[4],
 584                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 585                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 586                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 587                     &weight_op[4], &weight_avg[4],
 588                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 589             }else if(IS_SUB_4X8(sub_mb_type)){
 590                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 591                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 592                     &weight_op[5], &weight_avg[5],
 593                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 594                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 595                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 596                     &weight_op[5], &weight_avg[5],
 597                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 598             }else{
 599                 int j;
 600                 assert(IS_SUB_4X4(sub_mb_type));
 601                 for(j=0; j<4; j++){
 602                     int sub_x_offset= x_offset + 2*(j&1);
 603                     int sub_y_offset= y_offset +   (j&2);
 604                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 605                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 606                         &weight_op[6], &weight_avg[6],
 607                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 608                 }
 609             }
 610         }
 611     }
 612
 613     prefetch_motion(h, 1);
 614 }
 615
 616
 617 static void free_tables(H264Context *h){
 618     int i;
 619     H264Context *hx;
 620     av_freep(&h->intra4x4_pred_mode);
 621     av_freep(&h->chroma_pred_mode_table);
 622     av_freep(&h->cbp_table);
 623     av_freep(&h->mvd_table[0]);
 624     av_freep(&h->mvd_table[1]);
 625     av_freep(&h->direct_table);
 626     av_freep(&h->non_zero_count);
 627     av_freep(&h->slice_table_base);
 628     h->slice_table= NULL;
 629     av_freep(&h->list_counts);
 630
 631     av_freep(&h->mb2b_xy);
 632     av_freep(&h->mb2br_xy);
 633
 634     for(i = 0; i < MAX_THREADS; i++) {
 635         hx = h->thread_context[i];
 636         if(!hx) continue;
 637         av_freep(&hx->top_borders[1]);
 638         av_freep(&hx->top_borders[0]);
 639         av_freep(&hx->s.obmc_scratchpad);
 640         av_freep(&hx->rbsp_buffer[1]);
 641         av_freep(&hx->rbsp_buffer[0]);
 642         hx->rbsp_buffer_size[0] = 0;
 643         hx->rbsp_buffer_size[1] = 0;
 644         if (i) av_freep(&h->thread_context[i]);
 645     }
 646 }
 647
 648 static void init_dequant8_coeff_table(H264Context *h){
 649     int i,q,x;
 650     h->dequant8_coeff[0] = h->dequant8_buffer[0];
 651     h->dequant8_coeff[1] = h->dequant8_buffer[1];
 652
 653     for(i=0; i<2; i++ ){
 654         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
 655             h->dequant8_coeff[1] = h->dequant8_buffer[0];
 656             break;
 657         }
 658
 659         for(q=0; q<52; q++){
 660             int shift = div6[q];
 661             int idx = rem6[q];
 662             for(x=0; x<64; x++)
 663                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 664                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 665                     h->pps.scaling_matrix8[i][x]) << shift;
 666         }
 667     }
 668 }
 669
 670 static void init_dequant4_coeff_table(H264Context *h){
 671     int i,j,q,x;
 672     for(i=0; i<6; i++ ){
 673         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 674         for(j=0; j<i; j++){
 675             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 676                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 677                 break;
 678             }
 679         }
 680         if(j<i)
 681             continue;
 682
 683         for(q=0; q<52; q++){
 684             int shift = div6[q] + 2;
 685             int idx = rem6[q];
 686             for(x=0; x<16; x++)
 687                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 688                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 689                     h->pps.scaling_matrix4[i][x]) << shift;
 690         }
 691     }
 692 }
 693
 694 static void init_dequant_tables(H264Context *h){
 695     int i,x;
 696     init_dequant4_coeff_table(h);
 697     if(h->pps.transform_8x8_mode)
 698         init_dequant8_coeff_table(h);
 699     if(h->sps.transform_bypass){
 700         for(i=0; i<6; i++)
 701             for(x=0; x<16; x++)
 702                 h->dequant4_coeff[i][0][x] = 1<<6;
 703         if(h->pps.transform_8x8_mode)
 704             for(i=0; i<2; i++)
 705                 for(x=0; x<64; x++)
 706                     h->dequant8_coeff[i][0][x] = 1<<6;
 707     }
 708 }
 709
 710
 711 int ff_h264_alloc_tables(H264Context *h){
 712     MpegEncContext * const s = &h->s;
 713     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 714     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 715     int x,y;
 716
 717     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 718
 719     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
 720     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 721     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 722
 723     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 724     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 725     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 726     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 727     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 728
 729     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 730     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 731
 732     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 733     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 734     for(y=0; y<s->mb_height; y++){
 735         for(x=0; x<s->mb_width; x++){
 736             const int mb_xy= x + y*s->mb_stride;
 737             const int b_xy = 4*x + 4*y*h->b_stride;
 738
 739             h->mb2b_xy [mb_xy]= b_xy;
 740             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 741         }
 742     }
 743
 744     s->obmc_scratchpad = NULL;
 745
 746     if(!h->dequant4_coeff[0])
 747         init_dequant_tables(h);
 748
 749     return 0;
 750 fail:
 751     free_tables(h);
 752     return -1;
 753 }
 754
 755 /**
 756  * Mimic alloc_tables(), but for every context thread.
 757  */
 758 static void clone_tables(H264Context *dst, H264Context *src, int i){
 759     MpegEncContext * const s = &src->s;
 760     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 761     dst->non_zero_count           = src->non_zero_count;
 762     dst->slice_table              = src->slice_table;
 763     dst->cbp_table                = src->cbp_table;
 764     dst->mb2b_xy                  = src->mb2b_xy;
 765     dst->mb2br_xy                 = src->mb2br_xy;
 766     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 767     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 768     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 769     dst->direct_table             = src->direct_table;
 770     dst->list_counts              = src->list_counts;
 771
 772     dst->s.obmc_scratchpad = NULL;
 773     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
 774 }
 775
 776 /**
 777  * Init context
 778  * Allocate buffers which are not shared amongst multiple threads.
 779  */
 780 static int context_init(H264Context *h){
 781     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)
 782     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)
 783
 784     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 785     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 786
 787     return 0;
 788 fail:
 789     return -1; // free_tables will clean up for us
 790 }
 791
 792 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 793
 794 static av_cold void common_init(H264Context *h){
 795     MpegEncContext * const s = &h->s;
 796
 797     s->width = s->avctx->width;
 798     s->height = s->avctx->height;
 799     s->codec_id= s->avctx->codec->id;
 800
 801     ff_h264dsp_init(&h->h264dsp);
 802     ff_h264_pred_init(&h->hpc, s->codec_id);
 803
 804     h->dequant_coeff_pps= -1;
 805     s->unrestricted_mv=1;
 806     s->decode=1; //FIXME
 807
 808     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 809
 810     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 811     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 812 }
 813
 814 int ff_h264_decode_extradata(H264Context *h)
 815 {
 816     AVCodecContext *avctx = h->s.avctx;
 817
 818     if(*(char *)avctx->extradata == 1){
 819         int i, cnt, nalsize;
 820         unsigned char *p = avctx->extradata;
 821
 822         h->is_avc = 1;
 823
 824         if(avctx->extradata_size < 7) {
 825             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
 826             return -1;
 827         }
 828         /* sps and pps in the avcC always have length coded with 2 bytes,
 829            so put a fake nal_length_size = 2 while parsing them */
 830         h->nal_length_size = 2;
 831         // Decode sps from avcC
 832         cnt = *(p+5) & 0x1f; // Number of sps
 833         p += 6;
 834         for (i = 0; i < cnt; i++) {
 835             nalsize = AV_RB16(p) + 2;
 836             if(decode_nal_units(h, p, nalsize) < 0) {
 837                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
 838                 return -1;
 839             }
 840             p += nalsize;
 841         }
 842         // Decode pps from avcC
 843         cnt = *(p++); // Number of pps
 844         for (i = 0; i < cnt; i++) {
 845             nalsize = AV_RB16(p) + 2;
 846             if(decode_nal_units(h, p, nalsize)  != nalsize) {
 847                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
 848                 return -1;
 849             }
 850             p += nalsize;
 851         }
 852         // Now store right nal length size, that will be use to parse all other nals
 853         h->nal_length_size = ((*(((char*)(avctx->extradata))+4))&0x03)+1;
 854     } else {
 855         h->is_avc = 0;
 856         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
 857             return -1;
 858     }
 859     return 0;
 860 }
 861
 862 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
 863     H264Context *h= avctx->priv_data;
 864     MpegEncContext * const s = &h->s;
 865
 866     MPV_decode_defaults(s);
 867
 868     s->avctx = avctx;
 869     common_init(h);
 870
 871     s->out_format = FMT_H264;
 872     s->workaround_bugs= avctx->workaround_bugs;
 873
 874     // set defaults
 875 //    s->decode_mb= ff_h263_decode_mb;
 876     s->quarter_sample = 1;
 877     if(!avctx->has_b_frames)
 878     s->low_delay= 1;
 879
 880     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
 881
 882     ff_h264_decode_init_vlc();
 883
 884     h->thread_context[0] = h;
 885     h->outputed_poc = INT_MIN;
 886     h->prev_poc_msb= 1<<16;
 887     h->x264_build = -1;
 888     ff_h264_reset_sei(h);
 889     if(avctx->codec_id == CODEC_ID_H264){
 890         if(avctx->ticks_per_frame == 1){
 891             s->avctx->time_base.den *=2;
 892         }
 893         avctx->ticks_per_frame = 2;
 894     }
 895
 896     if(avctx->extradata_size > 0 && avctx->extradata &&
 897         ff_h264_decode_extradata(h))
 898         return -1;
 899
 900     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
 901         s->avctx->has_b_frames = h->sps.num_reorder_frames;
 902         s->low_delay = 0;
 903     }
 904
 905     return 0;
 906 }
 907
 908 int ff_h264_frame_start(H264Context *h){
 909     MpegEncContext * const s = &h->s;
 910     int i;
 911
 912     if(MPV_frame_start(s, s->avctx) < 0)
 913         return -1;
 914     ff_er_frame_start(s);
 915     /*
 916      * MPV_frame_start uses pict_type to derive key_frame.
 917      * This is incorrect for H.264; IDR markings must be used.
 918      * Zero here; IDR markings per slice in frame or fields are ORed in later.
 919      * See decode_nal_units().
 920      */
 921     s->current_picture_ptr->key_frame= 0;
 922     s->current_picture_ptr->mmco_reset= 0;
 923
 924     assert(s->linesize && s->uvlinesize);
 925
 926     for(i=0; i<16; i++){
 927         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
 928         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
 929     }
 930     for(i=0; i<4; i++){
 931         h->block_offset[16+i]=
 932         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
 933         h->block_offset[24+16+i]=
 934         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
 935     }
 936
 937     /* can't be in alloc_tables because linesize isn't known there.
 938      * FIXME: redo bipred weight to not require extra buffer? */
 939     for(i = 0; i < s->avctx->thread_count; i++)
 940         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
 941             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
 942
 943     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
 944     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
 945
 946 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
 947
 948     // We mark the current picture as non-reference after allocating it, so
 949     // that if we break out due to an error it can be released automatically
 950     // in the next MPV_frame_start().
 951     // SVQ3 as well as most other codecs have only last/next/current and thus
 952     // get released even with set reference, besides SVQ3 and others do not
 953     // mark frames as reference later "naturally".
 954     if(s->codec_id != CODEC_ID_SVQ3)
 955         s->current_picture_ptr->reference= 0;
 956
 957     s->current_picture_ptr->field_poc[0]=
 958     s->current_picture_ptr->field_poc[1]= INT_MAX;
 959     assert(s->current_picture_ptr->long_ref==0);
 960
 961     return 0;
 962 }
 963
 964 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
 965     MpegEncContext * const s = &h->s;
 966     uint8_t *top_border;
 967     int top_idx = 1;
 968
 969     src_y  -=   linesize;
 970     src_cb -= uvlinesize;
 971     src_cr -= uvlinesize;
 972
 973     if(!simple && FRAME_MBAFF){
 974         if(s->mb_y&1){
 975             if(!MB_MBAFF){
 976                 top_border = h->top_borders[0][s->mb_x];
 977                 AV_COPY128(top_border, src_y + 15*linesize);
 978                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
 979                     AV_COPY64(top_border+16, src_cb+7*uvlinesize);
 980                     AV_COPY64(top_border+24, src_cr+7*uvlinesize);
 981                 }
 982             }
 983         }else if(MB_MBAFF){
 984             top_idx = 0;
 985         }else
 986             return;
 987     }
 988
 989     top_border = h->top_borders[top_idx][s->mb_x];
 990     // There are two lines saved, the line above the the top macroblock of a pair,
 991     // and the line above the bottom macroblock
 992     AV_COPY128(top_border, src_y + 16*linesize);
 993
 994     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
 995         AV_COPY64(top_border+16, src_cb+8*uvlinesize);
 996         AV_COPY64(top_border+24, src_cr+8*uvlinesize);
 997     }
 998 }
 999
1000 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg, int simple){
1001     MpegEncContext * const s = &h->s;
1002     int deblock_left;
1003     int deblock_top;
1004     int top_idx = 1;
1005     uint8_t *top_border_m1;
1006     uint8_t *top_border;
1007
1008     if(!simple && FRAME_MBAFF){
1009         if(s->mb_y&1){
1010             if(!MB_MBAFF)
1011                 return;
1012         }else{
1013             top_idx = MB_MBAFF ? 0 : 1;
1014         }
1015     }
1016
1017     if(h->deblocking_filter == 2) {
1018         deblock_left = h->left_type[0];
1019         deblock_top  = h->top_type;
1020     } else {
1021         deblock_left = (s->mb_x > 0);
1022         deblock_top =  (s->mb_y > !!MB_FIELD);
1023     }
1024
1025     src_y  -=   linesize + 1;
1026     src_cb -= uvlinesize + 1;
1027     src_cr -= uvlinesize + 1;
1028
1029     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1030     top_border    = h->top_borders[top_idx][s->mb_x];
1031
1032 #define XCHG(a,b,xchg)\
1033 if (xchg) AV_SWAP64(b,a);\
1034 else      AV_COPY64(b,a);
1035
1036     if(deblock_top){
1037         if(deblock_left){
1038             XCHG(top_border_m1+8, src_y -7, 1);
1039         }
1040         XCHG(top_border+0, src_y +1, xchg);
1041         XCHG(top_border+8, src_y +9, 1);
1042         if(s->mb_x+1 < s->mb_width){
1043             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y +17, 1);
1044         }
1045     }
1046
1047     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1048         if(deblock_top){
1049             if(deblock_left){
1050                 XCHG(top_border_m1+16, src_cb -7, 1);
1051                 XCHG(top_border_m1+24, src_cr -7, 1);
1052             }
1053             XCHG(top_border+16, src_cb+1, 1);
1054             XCHG(top_border+24, src_cr+1, 1);
1055         }
1056     }
1057 }
1058
1059 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
1060     MpegEncContext * const s = &h->s;
1061     const int mb_x= s->mb_x;
1062     const int mb_y= s->mb_y;
1063     const int mb_xy= h->mb_xy;
1064     const int mb_type= s->current_picture.mb_type[mb_xy];
1065     uint8_t  *dest_y, *dest_cb, *dest_cr;
1066     int linesize, uvlinesize /*dct_offset*/;
1067     int i;
1068     int *block_offset = &h->block_offset[0];
1069     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1070     /* is_h264 should always be true if SVQ3 is disabled. */
1071     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1072     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1073     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1074
1075     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
1076     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
1077     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
1078
1079     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1080     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
1081
1082     h->list_counts[mb_xy]= h->list_count;
1083
1084     if (!simple && MB_FIELD) {
1085         linesize   = h->mb_linesize   = s->linesize * 2;
1086         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1087         block_offset = &h->block_offset[24];
1088         if(mb_y&1){ //FIXME move out of this function?
1089             dest_y -= s->linesize*15;
1090             dest_cb-= s->uvlinesize*7;
1091             dest_cr-= s->uvlinesize*7;
1092         }
1093         if(FRAME_MBAFF) {
1094             int list;
1095             for(list=0; list<h->list_count; list++){
1096                 if(!USES_LIST(mb_type, list))
1097                     continue;
1098                 if(IS_16X16(mb_type)){
1099                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1100                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1101                 }else{
1102                     for(i=0; i<16; i+=4){
1103                         int ref = h->ref_cache[list][scan8[i]];
1104                         if(ref >= 0)
1105                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1106                     }
1107                 }
1108             }
1109         }
1110     } else {
1111         linesize   = h->mb_linesize   = s->linesize;
1112         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1113 //        dct_offset = s->linesize * 16;
1114     }
1115
1116     if (!simple && IS_INTRA_PCM(mb_type)) {
1117         for (i=0; i<16; i++) {
1118             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1119         }
1120         for (i=0; i<8; i++) {
1121             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1122             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1123         }
1124     } else {
1125         if(IS_INTRA(mb_type)){
1126             if(h->deblocking_filter)
1127                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
1128
1129             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1130                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1131                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1132             }
1133
1134             if(IS_INTRA4x4(mb_type)){
1135                 if(simple || !s->encoding){
1136                     if(IS_8x8DCT(mb_type)){
1137                         if(transform_bypass){
1138                             idct_dc_add =
1139                             idct_add    = s->dsp.add_pixels8;
1140                         }else{
1141                             idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1142                             idct_add    = h->h264dsp.h264_idct8_add;
1143                         }
1144                         for(i=0; i<16; i+=4){
1145                             uint8_t * const ptr= dest_y + block_offset[i];
1146                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1147                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1148                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
1149                             }else{
1150                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
1151                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1152                                                             (h->topright_samples_available<<i)&0x4000, linesize);
1153                                 if(nnz){
1154                                     if(nnz == 1 && h->mb[i*16])
1155                                         idct_dc_add(ptr, h->mb + i*16, linesize);
1156                                     else
1157                                         idct_add   (ptr, h->mb + i*16, linesize);
1158                                 }
1159                             }
1160                         }
1161                     }else{
1162                         if(transform_bypass){
1163                             idct_dc_add =
1164                             idct_add    = s->dsp.add_pixels4;
1165                         }else{
1166                             idct_dc_add = h->h264dsp.h264_idct_dc_add;
1167                             idct_add    = h->h264dsp.h264_idct_add;
1168                         }
1169                         for(i=0; i<16; i++){
1170                             uint8_t * const ptr= dest_y + block_offset[i];
1171                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1172
1173                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1174                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
1175                             }else{
1176                                 uint8_t *topright;
1177                                 int nnz, tr;
1178                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1179                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1180                                     assert(mb_y || linesize <= block_offset[i]);
1181                                     if(!topright_avail){
1182                                         tr= ptr[3 - linesize]*0x01010101;
1183                                         topright= (uint8_t*) &tr;
1184                                     }else
1185                                         topright= ptr + 4 - linesize;
1186                                 }else
1187                                     topright= NULL;
1188
1189                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1190                                 nnz = h->non_zero_count_cache[ scan8[i] ];
1191                                 if(nnz){
1192                                     if(is_h264){
1193                                         if(nnz == 1 && h->mb[i*16])
1194                                             idct_dc_add(ptr, h->mb + i*16, linesize);
1195                                         else
1196                                             idct_add   (ptr, h->mb + i*16, linesize);
1197                                     }else
1198                                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
1199                                 }
1200                             }
1201                         }
1202                     }
1203                 }
1204             }else{
1205                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1206                 if(is_h264){
1207                     if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX] ]){
1208                         if(!transform_bypass)
1209                             h->h264dsp.h264_luma_dc_dequant_idct(h->mb, h->mb_luma_dc, h->dequant4_coeff[0][s->qscale][0]);
1210                         else{
1211                             static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1212                                                                     8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1213                             for(i = 0; i < 16; i++)
1214                                 h->mb[dc_mapping[i]] = h->mb_luma_dc[i];
1215                         }
1216                     }
1217                 }else
1218                     ff_svq3_luma_dc_dequant_idct_c(h->mb, h->mb_luma_dc, s->qscale);
1219             }
1220             if(h->deblocking_filter)
1221                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
1222         }else if(is_h264){
1223             hl_motion(h, dest_y, dest_cb, dest_cr,
1224                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1225                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1226                       h->h264dsp.weight_h264_pixels_tab, h->h264dsp.biweight_h264_pixels_tab);
1227         }
1228
1229
1230         if(!IS_INTRA4x4(mb_type)){
1231             if(is_h264){
1232                 if(IS_INTRA16x16(mb_type)){
1233                     if(transform_bypass){
1234                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1235                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
1236                         }else{
1237                             for(i=0; i<16; i++){
1238                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
1239                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
1240                             }
1241                         }
1242                     }else{
1243                          h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1244                     }
1245                 }else if(h->cbp&15){
1246                     if(transform_bypass){
1247                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1248                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1249                         for(i=0; i<16; i+=di){
1250                             if(h->non_zero_count_cache[ scan8[i] ]){
1251                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
1252                             }
1253                         }
1254                     }else{
1255                         if(IS_8x8DCT(mb_type)){
1256                             h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1257                         }else{
1258                             h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1259                         }
1260                     }
1261                 }
1262             }else{
1263                 for(i=0; i<16; i++){
1264                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
1265                         uint8_t * const ptr= dest_y + block_offset[i];
1266                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1267                     }
1268                 }
1269             }
1270         }
1271
1272         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1273             uint8_t *dest[2] = {dest_cb, dest_cr};
1274             if(transform_bypass){
1275                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1276                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
1277                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
1278                 }else{
1279                     idct_add = s->dsp.add_pixels4;
1280                     for(i=16; i<16+8; i++){
1281                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
1282                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
1283                     }
1284                 }
1285             }else{
1286                 if(is_h264){
1287                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1288                         chroma_dc_dequant_idct_c(h->mb + 16*16     , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1289                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1290                         chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1291                     h->h264dsp.h264_idct_add8(dest, block_offset,
1292                                               h->mb, uvlinesize,
1293                                               h->non_zero_count_cache);
1294                 }else{
1295                     chroma_dc_dequant_idct_c(h->mb + 16*16     , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1296                     chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1297                     for(i=16; i<16+8; i++){
1298                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1299                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
1300                             ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[s->qscale + 12] - 12, 2);
1301                         }
1302                     }
1303                 }
1304             }
1305         }
1306     }
1307     if(h->cbp || IS_INTRA(mb_type))
1308         s->dsp.clear_blocks(h->mb);
1309 }
1310
1311 /**
1312  * Process a macroblock; this case avoids checks for expensive uncommon cases.
1313  */
1314 static void hl_decode_mb_simple(H264Context *h){
1315     hl_decode_mb_internal(h, 1);
1316 }
1317
1318 /**
1319  * Process a macroblock; this handles edge cases, such as interlacing.
1320  */
1321 static void av_noinline hl_decode_mb_complex(H264Context *h){
1322     hl_decode_mb_internal(h, 0);
1323 }
1324
1325 void ff_h264_hl_decode_mb(H264Context *h){
1326     MpegEncContext * const s = &h->s;
1327     const int mb_xy= h->mb_xy;
1328     const int mb_type= s->current_picture.mb_type[mb_xy];
1329     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1330
1331     if (is_complex)
1332         hl_decode_mb_complex(h);
1333     else hl_decode_mb_simple(h);
1334 }
1335
1336 static int pred_weight_table(H264Context *h){
1337     MpegEncContext * const s = &h->s;
1338     int list, i;
1339     int luma_def, chroma_def;
1340
1341     h->use_weight= 0;
1342     h->use_weight_chroma= 0;
1343     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
1344     if(CHROMA)
1345         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
1346     luma_def = 1<<h->luma_log2_weight_denom;
1347     chroma_def = 1<<h->chroma_log2_weight_denom;
1348
1349     for(list=0; list<2; list++){
1350         h->luma_weight_flag[list]   = 0;
1351         h->chroma_weight_flag[list] = 0;
1352         for(i=0; i<h->ref_count[list]; i++){
1353             int luma_weight_flag, chroma_weight_flag;
1354
1355             luma_weight_flag= get_bits1(&s->gb);
1356             if(luma_weight_flag){
1357                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
1358                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
1359                 if(   h->luma_weight[i][list][0] != luma_def
1360                    || h->luma_weight[i][list][1] != 0) {
1361                     h->use_weight= 1;
1362                     h->luma_weight_flag[list]= 1;
1363                 }
1364             }else{
1365                 h->luma_weight[i][list][0]= luma_def;
1366                 h->luma_weight[i][list][1]= 0;
1367             }
1368
1369             if(CHROMA){
1370                 chroma_weight_flag= get_bits1(&s->gb);
1371                 if(chroma_weight_flag){
1372                     int j;
1373                     for(j=0; j<2; j++){
1374                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
1375                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
1376                         if(   h->chroma_weight[i][list][j][0] != chroma_def
1377                            || h->chroma_weight[i][list][j][1] != 0) {
1378                             h->use_weight_chroma= 1;
1379                             h->chroma_weight_flag[list]= 1;
1380                         }
1381                     }
1382                 }else{
1383                     int j;
1384                     for(j=0; j<2; j++){
1385                         h->chroma_weight[i][list][j][0]= chroma_def;
1386                         h->chroma_weight[i][list][j][1]= 0;
1387                     }
1388                 }
1389             }
1390         }
1391         if(h->slice_type_nos != FF_B_TYPE) break;
1392     }
1393     h->use_weight= h->use_weight || h->use_weight_chroma;
1394     return 0;
1395 }
1396
1397 /**
1398  * Initialize implicit_weight table.
1399  * @param field  0/1 initialize the weight for interlaced MBAFF
1400  *                -1 initializes the rest
1401  */
1402 static void implicit_weight_table(H264Context *h, int field){
1403     MpegEncContext * const s = &h->s;
1404     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
1405
1406     for (i = 0; i < 2; i++) {
1407         h->luma_weight_flag[i]   = 0;
1408         h->chroma_weight_flag[i] = 0;
1409     }
1410
1411     if(field < 0){
1412         cur_poc = s->current_picture_ptr->poc;
1413     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
1414        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
1415         h->use_weight= 0;
1416         h->use_weight_chroma= 0;
1417         return;
1418     }
1419         ref_start= 0;
1420         ref_count0= h->ref_count[0];
1421         ref_count1= h->ref_count[1];
1422     }else{
1423         cur_poc = s->current_picture_ptr->field_poc[field];
1424         ref_start= 16;
1425         ref_count0= 16+2*h->ref_count[0];
1426         ref_count1= 16+2*h->ref_count[1];
1427     }
1428
1429     h->use_weight= 2;
1430     h->use_weight_chroma= 2;
1431     h->luma_log2_weight_denom= 5;
1432     h->chroma_log2_weight_denom= 5;
1433
1434     for(ref0=ref_start; ref0 < ref_count0; ref0++){
1435         int poc0 = h->ref_list[0][ref0].poc;
1436         for(ref1=ref_start; ref1 < ref_count1; ref1++){
1437             int poc1 = h->ref_list[1][ref1].poc;
1438             int td = av_clip(poc1 - poc0, -128, 127);
1439             int w= 32;
1440             if(td){
1441                 int tb = av_clip(cur_poc - poc0, -128, 127);
1442                 int tx = (16384 + (FFABS(td) >> 1)) / td;
1443                 int dist_scale_factor = (tb*tx + 32) >> 8;
1444                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
1445                     w = 64 - dist_scale_factor;
1446             }
1447             if(field<0){
1448                 h->implicit_weight[ref0][ref1][0]=
1449                 h->implicit_weight[ref0][ref1][1]= w;
1450             }else{
1451                 h->implicit_weight[ref0][ref1][field]=w;
1452             }
1453         }
1454     }
1455 }
1456
1457 /**
1458  * instantaneous decoder refresh.
1459  */
1460 static void idr(H264Context *h){
1461     ff_h264_remove_all_refs(h);
1462     h->prev_frame_num= 0;
1463     h->prev_frame_num_offset= 0;
1464     h->prev_poc_msb=
1465     h->prev_poc_lsb= 0;
1466 }
1467
1468 /* forget old pics after a seek */
1469 static void flush_dpb(AVCodecContext *avctx){
1470     H264Context *h= avctx->priv_data;
1471     int i;
1472     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
1473         if(h->delayed_pic[i])
1474             h->delayed_pic[i]->reference= 0;
1475         h->delayed_pic[i]= NULL;
1476     }
1477     h->outputed_poc= INT_MIN;
1478     h->prev_interlaced_frame = 1;
1479     idr(h);
1480     if(h->s.current_picture_ptr)
1481         h->s.current_picture_ptr->reference= 0;
1482     h->s.first_field= 0;
1483     ff_h264_reset_sei(h);
1484     ff_mpeg_flush(avctx);
1485 }
1486
1487 static int init_poc(H264Context *h){
1488     MpegEncContext * const s = &h->s;
1489     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
1490     int field_poc[2];
1491     Picture *cur = s->current_picture_ptr;
1492
1493     h->frame_num_offset= h->prev_frame_num_offset;
1494     if(h->frame_num < h->prev_frame_num)
1495         h->frame_num_offset += max_frame_num;
1496
1497     if(h->sps.poc_type==0){
1498         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
1499
1500         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
1501             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
1502         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
1503             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
1504         else
1505             h->poc_msb = h->prev_poc_msb;
1506 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
1507         field_poc[0] =
1508         field_poc[1] = h->poc_msb + h->poc_lsb;
1509         if(s->picture_structure == PICT_FRAME)
1510             field_poc[1] += h->delta_poc_bottom;
1511     }else if(h->sps.poc_type==1){
1512         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
1513         int i;
1514
1515         if(h->sps.poc_cycle_length != 0)
1516             abs_frame_num = h->frame_num_offset + h->frame_num;
1517         else
1518             abs_frame_num = 0;
1519
1520         if(h->nal_ref_idc==0 && abs_frame_num > 0)
1521             abs_frame_num--;
1522
1523         expected_delta_per_poc_cycle = 0;
1524         for(i=0; i < h->sps.poc_cycle_length; i++)
1525             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
1526
1527         if(abs_frame_num > 0){
1528             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
1529             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
1530
1531             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
1532             for(i = 0; i <= frame_num_in_poc_cycle; i++)
1533                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
1534         } else
1535             expectedpoc = 0;
1536
1537         if(h->nal_ref_idc == 0)
1538             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
1539
1540         field_poc[0] = expectedpoc + h->delta_poc[0];
1541         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
1542
1543         if(s->picture_structure == PICT_FRAME)
1544             field_poc[1] += h->delta_poc[1];
1545     }else{
1546         int poc= 2*(h->frame_num_offset + h->frame_num);
1547
1548         if(!h->nal_ref_idc)
1549             poc--;
1550
1551         field_poc[0]= poc;
1552         field_poc[1]= poc;
1553     }
1554
1555     if(s->picture_structure != PICT_BOTTOM_FIELD)
1556         s->current_picture_ptr->field_poc[0]= field_poc[0];
1557     if(s->picture_structure != PICT_TOP_FIELD)
1558         s->current_picture_ptr->field_poc[1]= field_poc[1];
1559     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
1560
1561     return 0;
1562 }
1563
1564
1565 /**
1566  * initialize scan tables
1567  */
1568 static void init_scan_tables(H264Context *h){
1569     int i;
1570     for(i=0; i<16; i++){
1571 #define T(x) (x>>2) | ((x<<2) & 0xF)
1572         h->zigzag_scan[i] = T(zigzag_scan[i]);
1573         h-> field_scan[i] = T( field_scan[i]);
1574 #undef T
1575     }
1576     for(i=0; i<64; i++){
1577 #define T(x) (x>>3) | ((x&7)<<3)
1578         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
1579         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
1580         h->field_scan8x8[i]        = T(field_scan8x8[i]);
1581         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
1582 #undef T
1583     }
1584     if(h->sps.transform_bypass){ //FIXME same ugly
1585         h->zigzag_scan_q0          = zigzag_scan;
1586         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
1587         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
1588         h->field_scan_q0           = field_scan;
1589         h->field_scan8x8_q0        = field_scan8x8;
1590         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
1591     }else{
1592         h->zigzag_scan_q0          = h->zigzag_scan;
1593         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
1594         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
1595         h->field_scan_q0           = h->field_scan;
1596         h->field_scan8x8_q0        = h->field_scan8x8;
1597         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
1598     }
1599 }
1600
1601 static void field_end(H264Context *h){
1602     MpegEncContext * const s = &h->s;
1603     AVCodecContext * const avctx= s->avctx;
1604     s->mb_y= 0;
1605
1606     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1607     s->current_picture_ptr->pict_type= s->pict_type;
1608
1609     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1610         ff_vdpau_h264_set_reference_frames(s);
1611
1612     if(!s->dropable) {
1613         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1614         h->prev_poc_msb= h->poc_msb;
1615         h->prev_poc_lsb= h->poc_lsb;
1616     }
1617     h->prev_frame_num_offset= h->frame_num_offset;
1618     h->prev_frame_num= h->frame_num;
1619
1620     if (avctx->hwaccel) {
1621         if (avctx->hwaccel->end_frame(avctx) < 0)
1622             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
1623     }
1624
1625     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1626         ff_vdpau_h264_picture_complete(s);
1627
1628     /*
1629      * FIXME: Error handling code does not seem to support interlaced
1630      * when slices span multiple rows
1631      * The ff_er_add_slice calls don't work right for bottom
1632      * fields; they cause massive erroneous error concealing
1633      * Error marking covers both fields (top and bottom).
1634      * This causes a mismatched s->error_count
1635      * and a bad error table. Further, the error count goes to
1636      * INT_MAX when called for bottom field, because mb_y is
1637      * past end by one (callers fault) and resync_mb_y != 0
1638      * causes problems for the first MB line, too.
1639      */
1640     if (!FIELD_PICTURE)
1641         ff_er_frame_end(s);
1642
1643     MPV_frame_end(s);
1644
1645     h->current_slice=0;
1646 }
1647
1648 /**
1649  * Replicate H264 "master" context to thread contexts.
1650  */
1651 static void clone_slice(H264Context *dst, H264Context *src)
1652 {
1653     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
1654     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
1655     dst->s.current_picture      = src->s.current_picture;
1656     dst->s.linesize             = src->s.linesize;
1657     dst->s.uvlinesize           = src->s.uvlinesize;
1658     dst->s.first_field          = src->s.first_field;
1659
1660     dst->prev_poc_msb           = src->prev_poc_msb;
1661     dst->prev_poc_lsb           = src->prev_poc_lsb;
1662     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
1663     dst->prev_frame_num         = src->prev_frame_num;
1664     dst->short_ref_count        = src->short_ref_count;
1665
1666     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
1667     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
1668     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
1669     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
1670
1671     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
1672     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
1673 }
1674
1675 /**
1676  * decodes a slice header.
1677  * This will also call MPV_common_init() and frame_start() as needed.
1678  *
1679  * @param h h264context
1680  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
1681  *
1682  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
1683  */
1684 static int decode_slice_header(H264Context *h, H264Context *h0){
1685     MpegEncContext * const s = &h->s;
1686     MpegEncContext * const s0 = &h0->s;
1687     unsigned int first_mb_in_slice;
1688     unsigned int pps_id;
1689     int num_ref_idx_active_override_flag;
1690     unsigned int slice_type, tmp, i, j;
1691     int default_ref_list_done = 0;
1692     int last_pic_structure;
1693
1694     s->dropable= h->nal_ref_idc == 0;
1695
1696     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
1697         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
1698         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
1699     }else{
1700         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
1701         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
1702     }
1703
1704     first_mb_in_slice= get_ue_golomb(&s->gb);
1705
1706     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
1707         if(h0->current_slice && FIELD_PICTURE){
1708             field_end(h);
1709         }
1710
1711         h0->current_slice = 0;
1712         if (!s0->first_field)
1713             s->current_picture_ptr= NULL;
1714     }
1715
1716     slice_type= get_ue_golomb_31(&s->gb);
1717     if(slice_type > 9){
1718         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
1719         return -1;
1720     }
1721     if(slice_type > 4){
1722         slice_type -= 5;
1723         h->slice_type_fixed=1;
1724     }else
1725         h->slice_type_fixed=0;
1726
1727     slice_type= golomb_to_pict_type[ slice_type ];
1728     if (slice_type == FF_I_TYPE
1729         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
1730         default_ref_list_done = 1;
1731     }
1732     h->slice_type= slice_type;
1733     h->slice_type_nos= slice_type & 3;
1734
1735     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
1736
1737     pps_id= get_ue_golomb(&s->gb);
1738     if(pps_id>=MAX_PPS_COUNT){
1739         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
1740         return -1;
1741     }
1742     if(!h0->pps_buffers[pps_id]) {
1743         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
1744         return -1;
1745     }
1746     h->pps= *h0->pps_buffers[pps_id];
1747
1748     if(!h0->sps_buffers[h->pps.sps_id]) {
1749         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
1750         return -1;
1751     }
1752     h->sps = *h0->sps_buffers[h->pps.sps_id];
1753
1754     s->avctx->profile = h->sps.profile_idc;
1755     s->avctx->level   = h->sps.level_idc;
1756     s->avctx->refs    = h->sps.ref_frame_count;
1757
1758     if(h == h0 && h->dequant_coeff_pps != pps_id){
1759         h->dequant_coeff_pps = pps_id;
1760         init_dequant_tables(h);
1761     }
1762
1763     s->mb_width= h->sps.mb_width;
1764     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
1765
1766     h->b_stride=  s->mb_width*4;
1767
1768     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
1769     if(h->sps.frame_mbs_only_flag)
1770         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
1771     else
1772         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 7);
1773
1774     if (s->context_initialized
1775         && (   s->width != s->avctx->width || s->height != s->avctx->height
1776             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
1777         if(h != h0)
1778             return -1;   // width / height changed during parallelized decoding
1779         free_tables(h);
1780         flush_dpb(s->avctx);
1781         MPV_common_end(s);
1782     }
1783     if (!s->context_initialized) {
1784         if(h != h0)
1785             return -1;  // we cant (re-)initialize context during parallel decoding
1786
1787         avcodec_set_dimensions(s->avctx, s->width, s->height);
1788         s->avctx->sample_aspect_ratio= h->sps.sar;
1789         av_assert0(s->avctx->sample_aspect_ratio.den);
1790
1791         if(h->sps.video_signal_type_present_flag){
1792             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
1793             if(h->sps.colour_description_present_flag){
1794                 s->avctx->color_primaries = h->sps.color_primaries;
1795                 s->avctx->color_trc       = h->sps.color_trc;
1796                 s->avctx->colorspace      = h->sps.colorspace;
1797             }
1798         }
1799
1800         if(h->sps.timing_info_present_flag){
1801             int64_t den= h->sps.time_scale;
1802             if(h->x264_build < 44U)
1803                 den *= 2;
1804             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
1805                       h->sps.num_units_in_tick, den, 1<<30);
1806         }
1807         s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
1808                                                  s->avctx->codec->pix_fmts ?
1809                                                  s->avctx->codec->pix_fmts :
1810                                                  s->avctx->color_range == AVCOL_RANGE_JPEG ?
1811                                                  hwaccel_pixfmt_list_h264_jpeg_420 :
1812                                                  ff_hwaccel_pixfmt_list_420);
1813         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
1814
1815         if (MPV_common_init(s) < 0)
1816             return -1;
1817         s->first_field = 0;
1818         h->prev_interlaced_frame = 1;
1819
1820         init_scan_tables(h);
1821         ff_h264_alloc_tables(h);
1822
1823         for(i = 1; i < s->avctx->thread_count; i++) {
1824             H264Context *c;
1825             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
1826             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
1827             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
1828             c->h264dsp = h->h264dsp;
1829             c->sps = h->sps;
1830             c->pps = h->pps;
1831             init_scan_tables(c);
1832             clone_tables(c, h, i);
1833         }
1834
1835         for(i = 0; i < s->avctx->thread_count; i++)
1836             if(context_init(h->thread_context[i]) < 0)
1837                 return -1;
1838     }
1839
1840     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
1841
1842     h->mb_mbaff = 0;
1843     h->mb_aff_frame = 0;
1844     last_pic_structure = s0->picture_structure;
1845     if(h->sps.frame_mbs_only_flag){
1846         s->picture_structure= PICT_FRAME;
1847     }else{
1848         if(get_bits1(&s->gb)) { //field_pic_flag
1849             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
1850         } else {
1851             s->picture_structure= PICT_FRAME;
1852             h->mb_aff_frame = h->sps.mb_aff;
1853         }
1854     }
1855     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
1856
1857     if(h0->current_slice == 0){
1858         while(h->frame_num !=  h->prev_frame_num &&
1859               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
1860             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
1861             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
1862             if (ff_h264_frame_start(h) < 0)
1863                 return -1;
1864             h->prev_frame_num++;
1865             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
1866             s->current_picture_ptr->frame_num= h->prev_frame_num;
1867             ff_generate_sliding_window_mmcos(h);
1868             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1869             /* Error concealment: if a ref is missing, copy the previous ref in its place.
1870              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
1871              * about there being no actual duplicates.
1872              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
1873              * concealing a lost frame, this probably isn't noticable by comparison, but it should
1874              * be fixed. */
1875             if (h->short_ref_count) {
1876                 if (prev) {
1877                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
1878                                   (const uint8_t**)prev->data, prev->linesize,
1879                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
1880                     h->short_ref[0]->poc = prev->poc+2;
1881                 }
1882                 h->short_ref[0]->frame_num = h->prev_frame_num;
1883             }
1884         }
1885
1886         /* See if we have a decoded first field looking for a pair... */
1887         if (s0->first_field) {
1888             assert(s0->current_picture_ptr);
1889             assert(s0->current_picture_ptr->data[0]);
1890             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
1891
1892             /* figure out if we have a complementary field pair */
1893             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
1894                 /*
1895                  * Previous field is unmatched. Don't display it, but let it
1896                  * remain for reference if marked as such.
1897                  */
1898                 s0->current_picture_ptr = NULL;
1899                 s0->first_field = FIELD_PICTURE;
1900
1901             } else {
1902                 if (h->nal_ref_idc &&
1903                         s0->current_picture_ptr->reference &&
1904                         s0->current_picture_ptr->frame_num != h->frame_num) {
1905                     /*
1906                      * This and previous field were reference, but had
1907                      * different frame_nums. Consider this field first in
1908                      * pair. Throw away previous field except for reference
1909                      * purposes.
1910                      */
1911                     s0->first_field = 1;
1912                     s0->current_picture_ptr = NULL;
1913
1914                 } else {
1915                     /* Second field in complementary pair */
1916                     s0->first_field = 0;
1917                 }
1918             }
1919
1920         } else {
1921             /* Frame or first field in a potentially complementary pair */
1922             assert(!s0->current_picture_ptr);
1923             s0->first_field = FIELD_PICTURE;
1924         }
1925
1926         if((!FIELD_PICTURE || s0->first_field) && ff_h264_frame_start(h) < 0) {
1927             s0->first_field = 0;
1928             return -1;
1929         }
1930     }
1931     if(h != h0)
1932         clone_slice(h, h0);
1933
1934     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
1935
1936     assert(s->mb_num == s->mb_width * s->mb_height);
1937     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
1938        first_mb_in_slice                    >= s->mb_num){
1939         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
1940         return -1;
1941     }
1942     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
1943     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
1944     if (s->picture_structure == PICT_BOTTOM_FIELD)
1945         s->resync_mb_y = s->mb_y = s->mb_y + 1;
1946     assert(s->mb_y < s->mb_height);
1947
1948     if(s->picture_structure==PICT_FRAME){
1949         h->curr_pic_num=   h->frame_num;
1950         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
1951     }else{
1952         h->curr_pic_num= 2*h->frame_num + 1;
1953         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
1954     }
1955
1956     if(h->nal_unit_type == NAL_IDR_SLICE){
1957         get_ue_golomb(&s->gb); /* idr_pic_id */
1958     }
1959
1960     if(h->sps.poc_type==0){
1961         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
1962
1963         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
1964             h->delta_poc_bottom= get_se_golomb(&s->gb);
1965         }
1966     }
1967
1968     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
1969         h->delta_poc[0]= get_se_golomb(&s->gb);
1970
1971         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
1972             h->delta_poc[1]= get_se_golomb(&s->gb);
1973     }
1974
1975     init_poc(h);
1976
1977     if(h->pps.redundant_pic_cnt_present){
1978         h->redundant_pic_count= get_ue_golomb(&s->gb);
1979     }
1980
1981     //set defaults, might be overridden a few lines later
1982     h->ref_count[0]= h->pps.ref_count[0];
1983     h->ref_count[1]= h->pps.ref_count[1];
1984
1985     if(h->slice_type_nos != FF_I_TYPE){
1986         if(h->slice_type_nos == FF_B_TYPE){
1987             h->direct_spatial_mv_pred= get_bits1(&s->gb);
1988         }
1989         num_ref_idx_active_override_flag= get_bits1(&s->gb);
1990
1991         if(num_ref_idx_active_override_flag){
1992             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
1993             if(h->slice_type_nos==FF_B_TYPE)
1994                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
1995
1996             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
1997                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
1998                 h->ref_count[0]= h->ref_count[1]= 1;
1999                 return -1;
2000             }
2001         }
2002         if(h->slice_type_nos == FF_B_TYPE)
2003             h->list_count= 2;
2004         else
2005             h->list_count= 1;
2006     }else
2007         h->list_count= 0;
2008
2009     if(!default_ref_list_done){
2010         ff_h264_fill_default_ref_list(h);
2011     }
2012
2013     if(h->slice_type_nos!=FF_I_TYPE && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2014         return -1;
2015
2016     if(h->slice_type_nos!=FF_I_TYPE){
2017         s->last_picture_ptr= &h->ref_list[0][0];
2018         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2019     }
2020     if(h->slice_type_nos==FF_B_TYPE){
2021         s->next_picture_ptr= &h->ref_list[1][0];
2022         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2023     }
2024
2025     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
2026        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
2027         pred_weight_table(h);
2028     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE){
2029         implicit_weight_table(h, -1);
2030     }else {
2031         h->use_weight = 0;
2032         for (i = 0; i < 2; i++) {
2033             h->luma_weight_flag[i]   = 0;
2034             h->chroma_weight_flag[i] = 0;
2035         }
2036     }
2037
2038     if(h->nal_ref_idc)
2039         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2040
2041     if(FRAME_MBAFF){
2042         ff_h264_fill_mbaff_ref_list(h);
2043
2044         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE){
2045             implicit_weight_table(h, 0);
2046             implicit_weight_table(h, 1);
2047         }
2048     }
2049
2050     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
2051         ff_h264_direct_dist_scale_factor(h);
2052     ff_h264_direct_ref_list_init(h);
2053
2054     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
2055         tmp = get_ue_golomb_31(&s->gb);
2056         if(tmp > 2){
2057             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2058             return -1;
2059         }
2060         h->cabac_init_idc= tmp;
2061     }
2062
2063     h->last_qscale_diff = 0;
2064     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2065     if(tmp>51){
2066         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2067         return -1;
2068     }
2069     s->qscale= tmp;
2070     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2071     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2072     //FIXME qscale / qp ... stuff
2073     if(h->slice_type == FF_SP_TYPE){
2074         get_bits1(&s->gb); /* sp_for_switch_flag */
2075     }
2076     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
2077         get_se_golomb(&s->gb); /* slice_qs_delta */
2078     }
2079
2080     h->deblocking_filter = 1;
2081     h->slice_alpha_c0_offset = 52;
2082     h->slice_beta_offset = 52;
2083     if( h->pps.deblocking_filter_parameters_present ) {
2084         tmp= get_ue_golomb_31(&s->gb);
2085         if(tmp > 2){
2086             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2087             return -1;
2088         }
2089         h->deblocking_filter= tmp;
2090         if(h->deblocking_filter < 2)
2091             h->deblocking_filter^= 1; // 1<->0
2092
2093         if( h->deblocking_filter ) {
2094             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2095             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2096             if(   h->slice_alpha_c0_offset > 104U
2097                || h->slice_beta_offset     > 104U){
2098                 av_log(s->avctx, AV_LOG_ERROR, "deblocking filter parameters %d %d out of range\n", h->slice_alpha_c0_offset, h->slice_beta_offset);
2099                 return -1;
2100             }
2101         }
2102     }
2103
2104     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2105        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != FF_I_TYPE)
2106        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == FF_B_TYPE)
2107        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2108         h->deblocking_filter= 0;
2109
2110     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2111         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2112             /* Cheat slightly for speed:
2113                Do not bother to deblock across slices. */
2114             h->deblocking_filter = 2;
2115         } else {
2116             h0->max_contexts = 1;
2117             if(!h0->single_decode_warning) {
2118                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2119                 h0->single_decode_warning = 1;
2120             }
2121             if(h != h0)
2122                 return 1; // deblocking switched inside frame
2123         }
2124     }
2125     h->qp_thresh= 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1]);
2126
2127 #if 0 //FMO
2128     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2129         slice_group_change_cycle= get_bits(&s->gb, ?);
2130 #endif
2131
2132     h0->last_slice_type = slice_type;
2133     h->slice_num = ++h0->current_slice;
2134     if(h->slice_num >= MAX_SLICES){
2135         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2136     }
2137
2138     for(j=0; j<2; j++){
2139         int id_list[16];
2140         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2141         for(i=0; i<16; i++){
2142             id_list[i]= 60;
2143             if(h->ref_list[j][i].data[0]){
2144                 int k;
2145                 uint8_t *base= h->ref_list[j][i].base[0];
2146                 for(k=0; k<h->short_ref_count; k++)
2147                     if(h->short_ref[k]->base[0] == base){
2148                         id_list[i]= k;
2149                         break;
2150                     }
2151                 for(k=0; k<h->long_ref_count; k++)
2152                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2153                         id_list[i]= h->short_ref_count + k;
2154                         break;
2155                     }
2156             }
2157         }
2158
2159         ref2frm[0]=
2160         ref2frm[1]= -1;
2161         for(i=0; i<16; i++)
2162             ref2frm[i+2]= 4*id_list[i]
2163                           +(h->ref_list[j][i].reference&3);
2164         ref2frm[18+0]=
2165         ref2frm[18+1]= -1;
2166         for(i=16; i<48; i++)
2167             ref2frm[i+4]= 4*id_list[(i-16)>>1]
2168                           +(h->ref_list[j][i].reference&3);
2169     }
2170
2171     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE) ? 0 : 16;
2172     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
2173
2174     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
2175         av_log(h->s.avctx, AV_LOG_DEBUG, "slice:%d %s mb:%d %c%s%s pps:%u frame:%d poc:%d/%d ref:%d/%d qp:%d loop:%d:%d:%d weight:%d%s %s\n",
2176                h->slice_num,
2177                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
2178                first_mb_in_slice,
2179                av_get_pict_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
2180                pps_id, h->frame_num,
2181                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
2182                h->ref_count[0], h->ref_count[1],
2183                s->qscale,
2184                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
2185                h->use_weight,
2186                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
2187                h->slice_type == FF_B_TYPE ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
2188                );
2189     }
2190
2191     return 0;
2192 }
2193
2194 int ff_h264_get_slice_type(const H264Context *h)
2195 {
2196     switch (h->slice_type) {
2197     case FF_P_TYPE:  return 0;
2198     case FF_B_TYPE:  return 1;
2199     case FF_I_TYPE:  return 2;
2200     case FF_SP_TYPE: return 3;
2201     case FF_SI_TYPE: return 4;
2202     default:         return -1;
2203     }
2204 }
2205
2206 /**
2207  *
2208  * @return non zero if the loop filter can be skiped
2209  */
2210 static int fill_filter_caches(H264Context *h, int mb_type){
2211     MpegEncContext * const s = &h->s;
2212     const int mb_xy= h->mb_xy;
2213     int top_xy, left_xy[2];
2214     int top_type, left_type[2];
2215
2216     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
2217
2218     //FIXME deblocking could skip the intra and nnz parts.
2219
2220     /* Wow, what a mess, why didn't they simplify the interlacing & intra
2221      * stuff, I can't imagine that these complex rules are worth it. */
2222
2223     left_xy[1] = left_xy[0] = mb_xy-1;
2224     if(FRAME_MBAFF){
2225         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
2226         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
2227         if(s->mb_y&1){
2228             if (left_mb_field_flag != curr_mb_field_flag) {
2229                 left_xy[0] -= s->mb_stride;
2230             }
2231         }else{
2232             if(curr_mb_field_flag){
2233                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
2234             }
2235             if (left_mb_field_flag != curr_mb_field_flag) {
2236                 left_xy[1] += s->mb_stride;
2237             }
2238         }
2239     }
2240
2241     h->top_mb_xy = top_xy;
2242     h->left_mb_xy[0] = left_xy[0];
2243     h->left_mb_xy[1] = left_xy[1];
2244     {
2245         //for sufficiently low qp, filtering wouldn't do anything
2246         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
2247         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
2248         int qp = s->current_picture.qscale_table[mb_xy];
2249         if(qp <= qp_thresh
2250            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
2251            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
2252             if(!FRAME_MBAFF)
2253                 return 1;
2254             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
2255                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
2256                 return 1;
2257         }
2258     }
2259
2260     top_type     = s->current_picture.mb_type[top_xy]    ;
2261     left_type[0] = s->current_picture.mb_type[left_xy[0]];
2262     left_type[1] = s->current_picture.mb_type[left_xy[1]];
2263     if(h->deblocking_filter == 2){
2264         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
2265         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
2266     }else{
2267         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
2268         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
2269     }
2270     h->top_type    = top_type    ;
2271     h->left_type[0]= left_type[0];
2272     h->left_type[1]= left_type[1];
2273
2274     if(IS_INTRA(mb_type))
2275         return 0;
2276
2277     AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
2278     AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
2279     AV_COPY32(&h->non_zero_count_cache[0+8*5], &h->non_zero_count[mb_xy][16]);
2280     AV_COPY32(&h->non_zero_count_cache[4+8*3], &h->non_zero_count[mb_xy][20]);
2281     AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
2282
2283     h->cbp= h->cbp_table[mb_xy];
2284
2285     {
2286         int list;
2287         for(list=0; list<h->list_count; list++){
2288             int8_t *ref;
2289             int y, b_stride;
2290             int16_t (*mv_dst)[2];
2291             int16_t (*mv_src)[2];
2292
2293             if(!USES_LIST(mb_type, list)){
2294                 fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
2295                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2296                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2297                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2298                 AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2299                 continue;
2300             }
2301
2302             ref = &s->current_picture.ref_index[list][4*mb_xy];
2303             {
2304                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2305                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2306                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2307                 ref += 2;
2308                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2309                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2310             }
2311
2312             b_stride = h->b_stride;
2313             mv_dst   = &h->mv_cache[list][scan8[0]];
2314             mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
2315             for(y=0; y<4; y++){
2316                 AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
2317             }
2318
2319         }
2320     }
2321
2322
2323 /*
2324 0 . T T. T T T T
2325 1 L . .L . . . .
2326 2 L . .L . . . .
2327 3 . T TL . . . .
2328 4 L . .L . . . .
2329 5 L . .. . . . .
2330 */
2331 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
2332     if(top_type){
2333         AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
2334     }
2335
2336     if(left_type[0]){
2337         h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][7+0*8];
2338         h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][7+1*8];
2339         h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][7+2*8];
2340         h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][7+3*8];
2341     }
2342
2343     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
2344     if(!CABAC && h->pps.transform_8x8_mode){
2345         if(IS_8x8DCT(top_type)){
2346             h->non_zero_count_cache[4+8*0]=
2347             h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
2348             h->non_zero_count_cache[6+8*0]=
2349             h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
2350         }
2351         if(IS_8x8DCT(left_type[0])){
2352             h->non_zero_count_cache[3+8*1]=
2353             h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
2354         }
2355         if(IS_8x8DCT(left_type[1])){
2356             h->non_zero_count_cache[3+8*3]=
2357             h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
2358         }
2359
2360         if(IS_8x8DCT(mb_type)){
2361             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
2362             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
2363
2364             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
2365             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
2366
2367             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
2368             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
2369
2370             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
2371             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
2372         }
2373     }
2374
2375     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
2376         int list;
2377         for(list=0; list<h->list_count; list++){
2378             if(USES_LIST(top_type, list)){
2379                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
2380                 const int b8_xy= 4*top_xy + 2;
2381                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2382                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
2383                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
2384                 h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
2385                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
2386                 h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
2387             }else{
2388                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
2389                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2390             }
2391
2392             if(!IS_INTERLACED(mb_type^left_type[0])){
2393                 if(USES_LIST(left_type[0], list)){
2394                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
2395                     const int b8_xy= 4*left_xy[0] + 1;
2396                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2397                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
2398                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
2399                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
2400                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
2401                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
2402                     h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
2403                     h->ref_cache[list][scan8[0] - 1 +16 ]=
2404                     h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
2405                 }else{
2406                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
2407                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
2408                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
2409                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
2410                     h->ref_cache[list][scan8[0] - 1 + 0  ]=
2411                     h->ref_cache[list][scan8[0] - 1 + 8  ]=
2412                     h->ref_cache[list][scan8[0] - 1 + 16 ]=
2413                     h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
2414                 }
2415             }
2416         }
2417     }
2418
2419     return 0;
2420 }
2421
2422 static void loop_filter(H264Context *h){
2423     MpegEncContext * const s = &h->s;
2424     uint8_t  *dest_y, *dest_cb, *dest_cr;
2425     int linesize, uvlinesize, mb_x, mb_y;
2426     const int end_mb_y= s->mb_y + FRAME_MBAFF;
2427     const int old_slice_type= h->slice_type;
2428
2429     if(h->deblocking_filter) {
2430         for(mb_x= 0; mb_x<s->mb_width; mb_x++){
2431             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
2432                 int mb_xy, mb_type;
2433                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
2434                 h->slice_num= h->slice_table[mb_xy];
2435                 mb_type= s->current_picture.mb_type[mb_xy];
2436                 h->list_count= h->list_counts[mb_xy];
2437
2438                 if(FRAME_MBAFF)
2439                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
2440
2441                 s->mb_x= mb_x;
2442                 s->mb_y= mb_y;
2443                 dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2444                 dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2445                 dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2446                     //FIXME simplify above
2447
2448                 if (MB_FIELD) {
2449                     linesize   = h->mb_linesize   = s->linesize * 2;
2450                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2451                     if(mb_y&1){ //FIXME move out of this function?
2452                         dest_y -= s->linesize*15;
2453                         dest_cb-= s->uvlinesize*7;
2454                         dest_cr-= s->uvlinesize*7;
2455                     }
2456                 } else {
2457                     linesize   = h->mb_linesize   = s->linesize;
2458                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2459                 }
2460                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
2461                 if(fill_filter_caches(h, mb_type))
2462                     continue;
2463                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2464                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2465
2466                 if (FRAME_MBAFF) {
2467                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2468                 } else {
2469                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2470                 }
2471             }
2472         }
2473     }
2474     h->slice_type= old_slice_type;
2475     s->mb_x= 0;
2476     s->mb_y= end_mb_y - FRAME_MBAFF;
2477     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2478     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2479 }
2480
2481 static void predict_field_decoding_flag(H264Context *h){
2482     MpegEncContext * const s = &h->s;
2483     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
2484     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
2485                 ? s->current_picture.mb_type[mb_xy-1]
2486                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
2487                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
2488                 : 0;
2489     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
2490 }
2491
2492 static int decode_slice(struct AVCodecContext *avctx, void *arg){
2493     H264Context *h = *(void**)arg;
2494     MpegEncContext * const s = &h->s;
2495     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
2496
2497     s->mb_skip_run= -1;
2498
2499     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
2500                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
2501
2502     if( h->pps.cabac ) {
2503         /* realign */
2504         align_get_bits( &s->gb );
2505
2506         /* init cabac */
2507         ff_init_cabac_states( &h->cabac);
2508         ff_init_cabac_decoder( &h->cabac,
2509                                s->gb.buffer + get_bits_count(&s->gb)/8,
2510                                (get_bits_left(&s->gb) + 7)/8);
2511
2512         ff_h264_init_cabac_states(h);
2513
2514         for(;;){
2515 //START_TIMER
2516             int ret = ff_h264_decode_mb_cabac(h);
2517             int eos;
2518 //STOP_TIMER("decode_mb_cabac")
2519
2520             if(ret>=0) ff_h264_hl_decode_mb(h);
2521
2522             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
2523                 s->mb_y++;
2524
2525                 ret = ff_h264_decode_mb_cabac(h);
2526
2527                 if(ret>=0) ff_h264_hl_decode_mb(h);
2528                 s->mb_y--;
2529             }
2530             eos = get_cabac_terminate( &h->cabac );
2531
2532             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
2533                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2534                 return 0;
2535             }
2536             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2537                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d, bytestream (%td)\n", s->mb_x, s->mb_y, h->cabac.bytestream_end - h->cabac.bytestream);
2538                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
2539                 return -1;
2540             }
2541
2542             if( ++s->mb_x >= s->mb_width ) {
2543                 s->mb_x = 0;
2544                 loop_filter(h);
2545                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2546                 ++s->mb_y;
2547                 if(FIELD_OR_MBAFF_PICTURE) {
2548                     ++s->mb_y;
2549                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2550                         predict_field_decoding_flag(h);
2551                 }
2552             }
2553
2554             if( eos || s->mb_y >= s->mb_height ) {
2555                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2556                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2557                 return 0;
2558             }
2559         }
2560
2561     } else {
2562         for(;;){
2563             int ret = ff_h264_decode_mb_cavlc(h);
2564
2565             if(ret>=0) ff_h264_hl_decode_mb(h);
2566
2567             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
2568                 s->mb_y++;
2569                 ret = ff_h264_decode_mb_cavlc(h);
2570
2571                 if(ret>=0) ff_h264_hl_decode_mb(h);
2572                 s->mb_y--;
2573             }
2574
2575             if(ret<0){
2576                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2577                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
2578
2579                 return -1;
2580             }
2581
2582             if(++s->mb_x >= s->mb_width){
2583                 s->mb_x=0;
2584                 loop_filter(h);
2585                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2586                 ++s->mb_y;
2587                 if(FIELD_OR_MBAFF_PICTURE) {
2588                     ++s->mb_y;
2589                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2590                         predict_field_decoding_flag(h);
2591                 }
2592                 if(s->mb_y >= s->mb_height){
2593                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2594
2595                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
2596                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2597
2598                         return 0;
2599                     }else{
2600                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2601
2602                         return -1;
2603                     }
2604                 }
2605             }
2606
2607             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
2608                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2609                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
2610                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2611
2612                     return 0;
2613                 }else{
2614                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
2615
2616                     return -1;
2617                 }
2618             }
2619         }
2620     }
2621
2622 #if 0
2623     for(;s->mb_y < s->mb_height; s->mb_y++){
2624         for(;s->mb_x < s->mb_width; s->mb_x++){
2625             int ret= decode_mb(h);
2626
2627             ff_h264_hl_decode_mb(h);
2628
2629             if(ret<0){
2630                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2631                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
2632
2633                 return -1;
2634             }
2635
2636             if(++s->mb_x >= s->mb_width){
2637                 s->mb_x=0;
2638                 if(++s->mb_y >= s->mb_height){
2639                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
2640                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2641
2642                         return 0;
2643                     }else{
2644                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2645
2646                         return -1;
2647                     }
2648                 }
2649             }
2650
2651             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
2652                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
2653                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
2654
2655                     return 0;
2656                 }else{
2657                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
2658
2659                     return -1;
2660                 }
2661             }
2662         }
2663         s->mb_x=0;
2664         ff_draw_horiz_band(s, 16*s->mb_y, 16);
2665     }
2666 #endif
2667     return -1; //not reached
2668 }
2669
2670 /**
2671  * Call decode_slice() for each context.
2672  *
2673  * @param h h264 master context
2674  * @param context_count number of contexts to execute
2675  */
2676 static void execute_decode_slices(H264Context *h, int context_count){
2677     MpegEncContext * const s = &h->s;
2678     AVCodecContext * const avctx= s->avctx;
2679     H264Context *hx;
2680     int i;
2681
2682     if (s->avctx->hwaccel)
2683         return;
2684     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2685         return;
2686     if(context_count == 1) {
2687         decode_slice(avctx, &h);
2688     } else {
2689         for(i = 1; i < context_count; i++) {
2690             hx = h->thread_context[i];
2691             hx->s.error_recognition = avctx->error_recognition;
2692             hx->s.error_count = 0;
2693         }
2694
2695         avctx->execute(avctx, (void *)decode_slice,
2696                        h->thread_context, NULL, context_count, sizeof(void*));
2697
2698         /* pull back stuff from slices to master context */
2699         hx = h->thread_context[context_count - 1];
2700         s->mb_x = hx->s.mb_x;
2701         s->mb_y = hx->s.mb_y;
2702         s->dropable = hx->s.dropable;
2703         s->picture_structure = hx->s.picture_structure;
2704         for(i = 1; i < context_count; i++)
2705             h->s.error_count += h->thread_context[i]->s.error_count;
2706     }
2707 }
2708
2709
2710 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
2711     MpegEncContext * const s = &h->s;
2712     AVCodecContext * const avctx= s->avctx;
2713     int buf_index=0;
2714     H264Context *hx; ///< thread context
2715     int context_count = 0;
2716     int next_avc= h->is_avc ? 0 : buf_size;
2717
2718     h->max_contexts = avctx->thread_count;
2719 #if 0
2720     int i;
2721     for(i=0; i<50; i++){
2722         av_log(NULL, AV_LOG_ERROR,"%02X ", buf[i]);
2723     }
2724 #endif
2725     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
2726         h->current_slice = 0;
2727         if (!s->first_field)
2728             s->current_picture_ptr= NULL;
2729         ff_h264_reset_sei(h);
2730     }
2731
2732     for(;;){
2733         int consumed;
2734         int dst_length;
2735         int bit_length;
2736         const uint8_t *ptr;
2737         int i, nalsize = 0;
2738         int err;
2739
2740         if(buf_index >= next_avc) {
2741             if(buf_index >= buf_size) break;
2742             nalsize = 0;
2743             for(i = 0; i < h->nal_length_size; i++)
2744                 nalsize = (nalsize << 8) | buf[buf_index++];
2745             if(nalsize <= 0 || nalsize > buf_size - buf_index){
2746                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
2747                 break;
2748             }
2749             next_avc= buf_index + nalsize;
2750         } else {
2751             // start code prefix search
2752             for(; buf_index + 3 < next_avc; buf_index++){
2753                 // This should always succeed in the first iteration.
2754                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
2755                     break;
2756             }
2757
2758             if(buf_index+3 >= buf_size) break;
2759
2760             buf_index+=3;
2761             if(buf_index >= next_avc) continue;
2762         }
2763
2764         hx = h->thread_context[context_count];
2765
2766         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
2767         if (ptr==NULL || dst_length < 0){
2768             return -1;
2769         }
2770         i= buf_index + consumed;
2771         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
2772            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
2773             s->workaround_bugs |= FF_BUG_TRUNCATED;
2774
2775         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
2776         while(ptr[dst_length - 1] == 0 && dst_length > 0)
2777             dst_length--;
2778         }
2779         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
2780
2781         if(s->avctx->debug&FF_DEBUG_STARTCODE){
2782             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
2783         }
2784
2785         if (h->is_avc && (nalsize != consumed) && nalsize){
2786             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
2787         }
2788
2789         buf_index += consumed;
2790
2791         if(  (s->hurry_up == 1 && h->nal_ref_idc  == 0) //FIXME do not discard SEI id
2792            ||(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0))
2793             continue;
2794
2795       again:
2796         err = 0;
2797         switch(hx->nal_unit_type){
2798         case NAL_IDR_SLICE:
2799             if (h->nal_unit_type != NAL_IDR_SLICE) {
2800                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
2801                 return -1;
2802             }
2803             idr(h); //FIXME ensure we don't loose some frames if there is reordering
2804         case NAL_SLICE:
2805             init_get_bits(&hx->s.gb, ptr, bit_length);
2806             hx->intra_gb_ptr=
2807             hx->inter_gb_ptr= &hx->s.gb;
2808             hx->s.data_partitioning = 0;
2809
2810             if((err = decode_slice_header(hx, h)))
2811                break;
2812
2813             if (h->current_slice == 1) {
2814                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
2815                     return -1;
2816                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2817                     ff_vdpau_h264_picture_start(s);
2818             }
2819
2820             s->current_picture_ptr->key_frame |=
2821                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
2822                     (h->sei_recovery_frame_cnt >= 0);
2823             if(hx->redundant_pic_count==0 && hx->s.hurry_up < 5
2824                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2825                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=FF_B_TYPE)
2826                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==FF_I_TYPE)
2827                && avctx->skip_frame < AVDISCARD_ALL){
2828                 if(avctx->hwaccel) {
2829                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
2830                         return -1;
2831                 }else
2832                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
2833                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
2834                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
2835                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
2836                 }else
2837                     context_count++;
2838             }
2839             break;
2840         case NAL_DPA:
2841             init_get_bits(&hx->s.gb, ptr, bit_length);
2842             hx->intra_gb_ptr=
2843             hx->inter_gb_ptr= NULL;
2844
2845             if ((err = decode_slice_header(hx, h)) < 0)
2846                 break;
2847
2848             hx->s.data_partitioning = 1;
2849
2850             break;
2851         case NAL_DPB:
2852             init_get_bits(&hx->intra_gb, ptr, bit_length);
2853             hx->intra_gb_ptr= &hx->intra_gb;
2854             break;
2855         case NAL_DPC:
2856             init_get_bits(&hx->inter_gb, ptr, bit_length);
2857             hx->inter_gb_ptr= &hx->inter_gb;
2858
2859             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
2860                && s->context_initialized
2861                && s->hurry_up < 5
2862                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2863                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=FF_B_TYPE)
2864                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==FF_I_TYPE)
2865                && avctx->skip_frame < AVDISCARD_ALL)
2866                 context_count++;
2867             break;
2868         case NAL_SEI:
2869             init_get_bits(&s->gb, ptr, bit_length);
2870             ff_h264_decode_sei(h);
2871             break;
2872         case NAL_SPS:
2873             init_get_bits(&s->gb, ptr, bit_length);
2874             ff_h264_decode_seq_parameter_set(h);
2875
2876             if(s->flags& CODEC_FLAG_LOW_DELAY)
2877                 s->low_delay=1;
2878
2879             if(avctx->has_b_frames < 2)
2880                 avctx->has_b_frames= !s->low_delay;
2881             break;
2882         case NAL_PPS:
2883             init_get_bits(&s->gb, ptr, bit_length);
2884
2885             ff_h264_decode_picture_parameter_set(h, bit_length);
2886
2887             break;
2888         case NAL_AUD:
2889         case NAL_END_SEQUENCE:
2890         case NAL_END_STREAM:
2891         case NAL_FILLER_DATA:
2892         case NAL_SPS_EXT:
2893         case NAL_AUXILIARY_SLICE:
2894             break;
2895         default:
2896             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
2897         }
2898
2899         if(context_count == h->max_contexts) {
2900             execute_decode_slices(h, context_count);
2901             context_count = 0;
2902         }
2903
2904         if (err < 0)
2905             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
2906         else if(err == 1) {
2907             /* Slice could not be decoded in parallel mode, copy down
2908              * NAL unit stuff to context 0 and restart. Note that
2909              * rbsp_buffer is not transferred, but since we no longer
2910              * run in parallel mode this should not be an issue. */
2911             h->nal_unit_type = hx->nal_unit_type;
2912             h->nal_ref_idc   = hx->nal_ref_idc;
2913             hx = h;
2914             goto again;
2915         }
2916     }
2917     if(context_count)
2918         execute_decode_slices(h, context_count);
2919     return buf_index;
2920 }
2921
2922 /**
2923  * returns the number of bytes consumed for building the current frame
2924  */
2925 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
2926         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
2927         if(pos+10>buf_size) pos=buf_size; // oops ;)
2928
2929         return pos;
2930 }
2931
2932 static int decode_frame(AVCodecContext *avctx,
2933                              void *data, int *data_size,
2934                              AVPacket *avpkt)
2935 {
2936     const uint8_t *buf = avpkt->data;
2937     int buf_size = avpkt->size;
2938     H264Context *h = avctx->priv_data;
2939     MpegEncContext *s = &h->s;
2940     AVFrame *pict = data;
2941     int buf_index;
2942
2943     s->flags= avctx->flags;
2944     s->flags2= avctx->flags2;
2945
2946    /* end of stream, output what is still in the buffers */
2947  out:
2948     if (buf_size == 0) {
2949         Picture *out;
2950         int i, out_idx;
2951
2952 //FIXME factorize this with the output code below
2953         out = h->delayed_pic[0];
2954         out_idx = 0;
2955         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
2956             if(h->delayed_pic[i]->poc < out->poc){
2957                 out = h->delayed_pic[i];
2958                 out_idx = i;
2959             }
2960
2961         for(i=out_idx; h->delayed_pic[i]; i++)
2962             h->delayed_pic[i] = h->delayed_pic[i+1];
2963
2964         if(out){
2965             *data_size = sizeof(AVFrame);
2966             *pict= *(AVFrame*)out;
2967         }
2968
2969         return 0;
2970     }
2971
2972     buf_index=decode_nal_units(h, buf, buf_size);
2973     if(buf_index < 0)
2974         return -1;
2975
2976     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
2977         buf_size = 0;
2978         goto out;
2979     }
2980
2981     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
2982         if (avctx->skip_frame >= AVDISCARD_NONREF || s->hurry_up) return 0;
2983         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
2984         return -1;
2985     }
2986
2987     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
2988         Picture *out = s->current_picture_ptr;
2989         Picture *cur = s->current_picture_ptr;
2990         int i, pics, out_of_order, out_idx;
2991
2992         field_end(h);
2993
2994         if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
2995             /* Wait for second field. */
2996             *data_size = 0;
2997
2998         } else {
2999             cur->interlaced_frame = 0;
3000             cur->repeat_pict = 0;
3001
3002             /* Signal interlacing information externally. */
3003             /* Prioritize picture timing SEI information over used decoding process if it exists. */
3004
3005             if(h->sps.pic_struct_present_flag){
3006                 switch (h->sei_pic_struct)
3007                 {
3008                 case SEI_PIC_STRUCT_FRAME:
3009                     break;
3010                 case SEI_PIC_STRUCT_TOP_FIELD:
3011                 case SEI_PIC_STRUCT_BOTTOM_FIELD:
3012                     cur->interlaced_frame = 1;
3013                     break;
3014                 case SEI_PIC_STRUCT_TOP_BOTTOM:
3015                 case SEI_PIC_STRUCT_BOTTOM_TOP:
3016                     if (FIELD_OR_MBAFF_PICTURE)
3017                         cur->interlaced_frame = 1;
3018                     else
3019                         // try to flag soft telecine progressive
3020                         cur->interlaced_frame = h->prev_interlaced_frame;
3021                     break;
3022                 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
3023                 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
3024                     // Signal the possibility of telecined film externally (pic_struct 5,6)
3025                     // From these hints, let the applications decide if they apply deinterlacing.
3026                     cur->repeat_pict = 1;
3027                     break;
3028                 case SEI_PIC_STRUCT_FRAME_DOUBLING:
3029                     // Force progressive here, as doubling interlaced frame is a bad idea.
3030                     cur->repeat_pict = 2;
3031                     break;
3032                 case SEI_PIC_STRUCT_FRAME_TRIPLING:
3033                     cur->repeat_pict = 4;
3034                     break;
3035                 }
3036
3037                 if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
3038                     cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
3039             }else{
3040                 /* Derive interlacing flag from used decoding process. */
3041                 cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
3042             }
3043             h->prev_interlaced_frame = cur->interlaced_frame;
3044
3045             if (cur->field_poc[0] != cur->field_poc[1]){
3046                 /* Derive top_field_first from field pocs. */
3047                 cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
3048             }else{
3049                 if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
3050                     /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
3051                     if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
3052                       || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
3053                         cur->top_field_first = 1;
3054                     else
3055                         cur->top_field_first = 0;
3056                 }else{
3057                     /* Most likely progressive */
3058                     cur->top_field_first = 0;
3059                 }
3060             }
3061
3062         //FIXME do something with unavailable reference frames
3063
3064             /* Sort B-frames into display order */
3065
3066             if(h->sps.bitstream_restriction_flag
3067                && s->avctx->has_b_frames < h->sps.num_reorder_frames){
3068                 s->avctx->has_b_frames = h->sps.num_reorder_frames;
3069                 s->low_delay = 0;
3070             }
3071
3072             if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
3073                && !h->sps.bitstream_restriction_flag){
3074                 s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
3075                 s->low_delay= 0;
3076             }
3077
3078             pics = 0;
3079             while(h->delayed_pic[pics]) pics++;
3080
3081             assert(pics <= MAX_DELAYED_PIC_COUNT);
3082
3083             h->delayed_pic[pics++] = cur;
3084             if(cur->reference == 0)
3085                 cur->reference = DELAYED_PIC_REF;
3086
3087             out = h->delayed_pic[0];
3088             out_idx = 0;
3089             for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3090                 if(h->delayed_pic[i]->poc < out->poc){
3091                     out = h->delayed_pic[i];
3092                     out_idx = i;
3093                 }
3094             if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
3095                 h->outputed_poc= INT_MIN;
3096             out_of_order = out->poc < h->outputed_poc;
3097
3098             if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
3099                 { }
3100             else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
3101                || (s->low_delay &&
3102                 ((h->outputed_poc != INT_MIN && out->poc > h->outputed_poc + 2)
3103                  || cur->pict_type == FF_B_TYPE)))
3104             {
3105                 s->low_delay = 0;
3106                 s->avctx->has_b_frames++;
3107             }
3108
3109             if(out_of_order || pics > s->avctx->has_b_frames){
3110                 out->reference &= ~DELAYED_PIC_REF;
3111                 for(i=out_idx; h->delayed_pic[i]; i++)
3112                     h->delayed_pic[i] = h->delayed_pic[i+1];
3113             }
3114             if(!out_of_order && pics > s->avctx->has_b_frames){
3115                 *data_size = sizeof(AVFrame);
3116
3117                 if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
3118                     h->outputed_poc = INT_MIN;
3119                 } else
3120                     h->outputed_poc = out->poc;
3121                 *pict= *(AVFrame*)out;
3122             }else{
3123                 av_log(avctx, AV_LOG_DEBUG, "no picture\n");
3124             }
3125         }
3126     }
3127
3128     assert(pict->data[0] || !*data_size);
3129     ff_print_debug_info(s, pict);
3130 //printf("out %d\n", (int)pict->data[0]);
3131
3132     return get_consumed_bytes(s, buf_index, buf_size);
3133 }
3134 #if 0
3135 static inline void fill_mb_avail(H264Context *h){
3136     MpegEncContext * const s = &h->s;
3137     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3138
3139     if(s->mb_y){
3140         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3141         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3142         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3143     }else{
3144         h->mb_avail[0]=
3145         h->mb_avail[1]=
3146         h->mb_avail[2]= 0;
3147     }
3148     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3149     h->mb_avail[4]= 1; //FIXME move out
3150     h->mb_avail[5]= 0; //FIXME move out
3151 }
3152 #endif
3153
3154 #ifdef TEST
3155 #undef printf
3156 #undef random
3157 #define COUNT 8000
3158 #define SIZE (COUNT*40)
3159 int main(void){
3160     int i;
3161     uint8_t temp[SIZE];
3162     PutBitContext pb;
3163     GetBitContext gb;
3164 //    int int_temp[10000];
3165     DSPContext dsp;
3166     AVCodecContext avctx;
3167
3168     dsputil_init(&dsp, &avctx);
3169
3170     init_put_bits(&pb, temp, SIZE);
3171     printf("testing unsigned exp golomb\n");
3172     for(i=0; i<COUNT; i++){
3173         START_TIMER
3174         set_ue_golomb(&pb, i);
3175         STOP_TIMER("set_ue_golomb");
3176     }
3177     flush_put_bits(&pb);
3178
3179     init_get_bits(&gb, temp, 8*SIZE);
3180     for(i=0; i<COUNT; i++){
3181         int j, s;
3182
3183         s= show_bits(&gb, 24);
3184
3185         START_TIMER
3186         j= get_ue_golomb(&gb);
3187         if(j != i){
3188             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3189 //            return -1;
3190         }
3191         STOP_TIMER("get_ue_golomb");
3192     }
3193
3194
3195     init_put_bits(&pb, temp, SIZE);
3196     printf("testing signed exp golomb\n");
3197     for(i=0; i<COUNT; i++){
3198         START_TIMER
3199         set_se_golomb(&pb, i - COUNT/2);
3200         STOP_TIMER("set_se_golomb");
3201     }
3202     flush_put_bits(&pb);
3203
3204     init_get_bits(&gb, temp, 8*SIZE);
3205     for(i=0; i<COUNT; i++){
3206         int j, s;
3207
3208         s= show_bits(&gb, 24);
3209
3210         START_TIMER
3211         j= get_se_golomb(&gb);
3212         if(j != i - COUNT/2){
3213             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3214 //            return -1;
3215         }
3216         STOP_TIMER("get_se_golomb");
3217     }
3218
3219 #if 0
3220     printf("testing 4x4 (I)DCT\n");
3221
3222     DCTELEM block[16];
3223     uint8_t src[16], ref[16];
3224     uint64_t error= 0, max_error=0;
3225
3226     for(i=0; i<COUNT; i++){
3227         int j;
3228 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
3229         for(j=0; j<16; j++){
3230             ref[j]= random()%255;
3231             src[j]= random()%255;
3232         }
3233
3234         h264_diff_dct_c(block, src, ref, 4);
3235
3236         //normalize
3237         for(j=0; j<16; j++){
3238 //            printf("%d ", block[j]);
3239             block[j]= block[j]*4;
3240             if(j&1) block[j]= (block[j]*4 + 2)/5;
3241             if(j&4) block[j]= (block[j]*4 + 2)/5;
3242         }
3243 //        printf("\n");
3244
3245         h->h264dsp.h264_idct_add(ref, block, 4);
3246 /*        for(j=0; j<16; j++){
3247             printf("%d ", ref[j]);
3248         }
3249         printf("\n");*/
3250
3251         for(j=0; j<16; j++){
3252             int diff= FFABS(src[j] - ref[j]);
3253
3254             error+= diff*diff;
3255             max_error= FFMAX(max_error, diff);
3256         }
3257     }
3258     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
3259     printf("testing quantizer\n");
3260     for(qp=0; qp<52; qp++){
3261         for(i=0; i<16; i++)
3262             src1_block[i]= src2_block[i]= random()%255;
3263
3264     }
3265     printf("Testing NAL layer\n");
3266
3267     uint8_t bitstream[COUNT];
3268     uint8_t nal[COUNT*2];
3269     H264Context h;
3270     memset(&h, 0, sizeof(H264Context));
3271
3272     for(i=0; i<COUNT; i++){
3273         int zeros= i;
3274         int nal_length;
3275         int consumed;
3276         int out_length;
3277         uint8_t *out;
3278         int j;
3279
3280         for(j=0; j<COUNT; j++){
3281             bitstream[j]= (random() % 255) + 1;
3282         }
3283
3284         for(j=0; j<zeros; j++){
3285             int pos= random() % COUNT;
3286             while(bitstream[pos] == 0){
3287                 pos++;
3288                 pos %= COUNT;
3289             }
3290             bitstream[pos]=0;
3291         }
3292
3293         START_TIMER
3294
3295         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
3296         if(nal_length<0){
3297             printf("encoding failed\n");
3298             return -1;
3299         }
3300
3301         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
3302
3303         STOP_TIMER("NAL")
3304
3305         if(out_length != COUNT){
3306             printf("incorrect length %d %d\n", out_length, COUNT);
3307             return -1;
3308         }
3309
3310         if(consumed != nal_length){
3311             printf("incorrect consumed length %d %d\n", nal_length, consumed);
3312             return -1;
3313         }
3314
3315         if(memcmp(bitstream, out, COUNT)){
3316             printf("mismatch\n");
3317             return -1;
3318         }
3319     }
3320 #endif
3321
3322     printf("Testing RBSP\n");
3323
3324
3325     return 0;
3326 }
3327 #endif /* TEST */
3328
3329
3330 av_cold void ff_h264_free_context(H264Context *h)
3331 {
3332     int i;
3333
3334     free_tables(h); //FIXME cleanup init stuff perhaps
3335
3336     for(i = 0; i < MAX_SPS_COUNT; i++)
3337         av_freep(h->sps_buffers + i);
3338
3339     for(i = 0; i < MAX_PPS_COUNT; i++)
3340         av_freep(h->pps_buffers + i);
3341 }
3342
3343 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
3344 {
3345     H264Context *h = avctx->priv_data;
3346     MpegEncContext *s = &h->s;
3347
3348     ff_h264_free_context(h);
3349
3350     MPV_common_end(s);
3351
3352 //    memset(h, 0, sizeof(H264Context));
3353
3354     return 0;
3355 }
3356
3357
3358 AVCodec h264_decoder = {
3359     "h264",
3360     AVMEDIA_TYPE_VIDEO,
3361     CODEC_ID_H264,
3362     sizeof(H264Context),
3363     ff_h264_decode_init,
3364     NULL,
3365     ff_h264_decode_end,
3366     decode_frame,
3367     /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY,
3368     .flush= flush_dpb,
3369     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
3370 };
3371
3372 #if CONFIG_H264_VDPAU_DECODER
3373 AVCodec h264_vdpau_decoder = {
3374     "h264_vdpau",
3375     AVMEDIA_TYPE_VIDEO,
3376     CODEC_ID_H264,
3377     sizeof(H264Context),
3378     ff_h264_decode_init,
3379     NULL,
3380     ff_h264_decode_end,
3381     decode_frame,
3382     CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
3383     .flush= flush_dpb,
3384     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
3385     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
3386 };
3387 #endif