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 "libavutil/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 "golomb.h"
  37 #include "mathops.h"
  38 #include "rectangle.h"
  39 #include "thread.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[QP_MAX_NUM+1]={
  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, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[QP_MAX_NUM+1]={
  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, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
  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 /**
  64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  65  */
  66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  67     MpegEncContext * const s = &h->s;
  68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  70     int i;
  71
  72     if(!(h->top_samples_available&0x8000)){
  73         for(i=0; i<4; i++){
  74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  75             if(status<0){
  76                 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);
  77                 return -1;
  78             } else if(status){
  79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  80             }
  81         }
  82     }
  83
  84     if((h->left_samples_available&0x8888)!=0x8888){
  85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  86         for(i=0; i<4; i++){
  87             if(!(h->left_samples_available&mask[i])){
  88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  89                 if(status<0){
  90                     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);
  91                     return -1;
  92                 } else if(status){
  93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  94                 }
  95             }
  96         }
  97     }
  98
  99     return 0;
 100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 101
 102 /**
 103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 104  */
 105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 106     MpegEncContext * const s = &h->s;
 107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 109
 110     if(mode > 6U) {
 111         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);
 112         return -1;
 113     }
 114
 115     if(!(h->top_samples_available&0x8000)){
 116         mode= top[ mode ];
 117         if(mode<0){
 118             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);
 119             return -1;
 120         }
 121     }
 122
 123     if((h->left_samples_available&0x8080) != 0x8080){
 124         mode= left[ mode ];
 125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 127         }
 128         if(mode<0){
 129             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);
 130             return -1;
 131         }
 132     }
 133
 134     return mode;
 135 }
 136
 137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 138     int i, si, di;
 139     uint8_t *dst;
 140     int bufidx;
 141
 142 //    src[0]&0x80;                //forbidden bit
 143     h->nal_ref_idc= src[0]>>5;
 144     h->nal_unit_type= src[0]&0x1F;
 145
 146     src++; length--;
 147
 148 #if HAVE_FAST_UNALIGNED
 149 # if HAVE_FAST_64BIT
 150 #   define RS 7
 151     for(i=0; i+1<length; i+=9){
 152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 153 # else
 154 #   define RS 3
 155     for(i=0; i+1<length; i+=5){
 156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 157 # endif
 158             continue;
 159         if(i>0 && !src[i]) i--;
 160         while(src[i]) i++;
 161 #else
 162 #   define RS 0
 163     for(i=0; i+1<length; i+=2){
 164         if(src[i]) continue;
 165         if(i>0 && src[i-1]==0) i--;
 166 #endif
 167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 168             if(src[i+2]!=3){
 169                 /* startcode, so we must be past the end */
 170                 length=i;
 171             }
 172             break;
 173         }
 174         i-= RS;
 175     }
 176
 177     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 178     si=h->rbsp_buffer_size[bufidx];
 179     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 180     dst= h->rbsp_buffer[bufidx];
 181     if(si != h->rbsp_buffer_size[bufidx])
 182         memset(dst + length, 0, FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 183
 184     if (dst == NULL){
 185         return NULL;
 186     }
 187
 188     if(i>=length-1){ //no escaped 0
 189         *dst_length= length;
 190         *consumed= length+1; //+1 for the header
 191         memcpy(dst, src, length);
 192         return dst;
 193     }
 194
 195 //printf("decoding esc\n");
 196     memcpy(dst, src, i);
 197     si=di=i;
 198     while(si+2<length){
 199         //remove escapes (very rare 1:2^22)
 200         if(src[si+2]>3){
 201             dst[di++]= src[si++];
 202             dst[di++]= src[si++];
 203         }else if(src[si]==0 && src[si+1]==0){
 204             if(src[si+2]==3){ //escape
 205                 dst[di++]= 0;
 206                 dst[di++]= 0;
 207                 si+=3;
 208                 continue;
 209             }else //next start code
 210                 goto nsc;
 211         }
 212
 213         dst[di++]= src[si++];
 214     }
 215     while(si<length)
 216         dst[di++]= src[si++];
 217 nsc:
 218
 219     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 220
 221     *dst_length= di;
 222     *consumed= si + 1;//+1 for the header
 223 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 224     return dst;
 225 }
 226
 227 /**
 228  * Identify the exact end of the bitstream
 229  * @return the length of the trailing, or 0 if damaged
 230  */
 231 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 232     int v= *src;
 233     int r;
 234
 235     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 236
 237     for(r=1; r<9; r++){
 238         if(v&1) return r;
 239         v>>=1;
 240     }
 241     return 0;
 242 }
 243
 244 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 245                                  int y_offset, int list){
 246     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 247     int filter_height= (raw_my&3) ? 2 : 0;
 248     int full_my= (raw_my>>2) + y_offset;
 249     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 250
 251     return FFMAX(abs(top), bottom);
 252 }
 253
 254 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 255                                int y_offset, int list0, int list1, int *nrefs){
 256     MpegEncContext * const s = &h->s;
 257     int my;
 258
 259     y_offset += 16*(s->mb_y >> MB_FIELD);
 260
 261     if(list0){
 262         int ref_n = h->ref_cache[0][ scan8[n] ];
 263         Picture *ref= &h->ref_list[0][ref_n];
 264
 265         // Error resilience puts the current picture in the ref list.
 266         // Don't try to wait on these as it will cause a deadlock.
 267         // Fields can wait on each other, though.
 268         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 269            (ref->f.reference & 3) != s->picture_structure) {
 270             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 271             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 272             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 273         }
 274     }
 275
 276     if(list1){
 277         int ref_n = h->ref_cache[1][ scan8[n] ];
 278         Picture *ref= &h->ref_list[1][ref_n];
 279
 280         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 281            (ref->f.reference & 3) != s->picture_structure) {
 282             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 283             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 284             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 285         }
 286     }
 287 }
 288
 289 /**
 290  * Wait until all reference frames are available for MC operations.
 291  *
 292  * @param h the H264 context
 293  */
 294 static void await_references(H264Context *h){
 295     MpegEncContext * const s = &h->s;
 296     const int mb_xy= h->mb_xy;
 297     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 298     int refs[2][48];
 299     int nrefs[2] = {0};
 300     int ref, list;
 301
 302     memset(refs, -1, sizeof(refs));
 303
 304     if(IS_16X16(mb_type)){
 305         get_lowest_part_y(h, refs, 0, 16, 0,
 306                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 307     }else if(IS_16X8(mb_type)){
 308         get_lowest_part_y(h, refs, 0, 8, 0,
 309                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 310         get_lowest_part_y(h, refs, 8, 8, 8,
 311                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 312     }else if(IS_8X16(mb_type)){
 313         get_lowest_part_y(h, refs, 0, 16, 0,
 314                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 315         get_lowest_part_y(h, refs, 4, 16, 0,
 316                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 317     }else{
 318         int i;
 319
 320         assert(IS_8X8(mb_type));
 321
 322         for(i=0; i<4; i++){
 323             const int sub_mb_type= h->sub_mb_type[i];
 324             const int n= 4*i;
 325             int y_offset= (i&2)<<2;
 326
 327             if(IS_SUB_8X8(sub_mb_type)){
 328                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 329                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 330             }else if(IS_SUB_8X4(sub_mb_type)){
 331                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 332                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 333                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 334                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 335             }else if(IS_SUB_4X8(sub_mb_type)){
 336                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 337                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 338                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 339                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 340             }else{
 341                 int j;
 342                 assert(IS_SUB_4X4(sub_mb_type));
 343                 for(j=0; j<4; j++){
 344                     int sub_y_offset= y_offset + 2*(j&2);
 345                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 346                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 347                 }
 348             }
 349         }
 350     }
 351
 352     for(list=h->list_count-1; list>=0; list--){
 353         for(ref=0; ref<48 && nrefs[list]; ref++){
 354             int row = refs[list][ref];
 355             if(row >= 0){
 356                 Picture *ref_pic = &h->ref_list[list][ref];
 357                 int ref_field = ref_pic->f.reference - 1;
 358                 int ref_field_picture = ref_pic->field_picture;
 359                 int pic_height = 16*s->mb_height >> ref_field_picture;
 360
 361                 row <<= MB_MBAFF;
 362                 nrefs[list]--;
 363
 364                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 365                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 367                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 369                 }else if(FIELD_PICTURE){
 370                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 371                 }else{
 372                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 373                 }
 374             }
 375         }
 376     }
 377 }
 378
 379 #if 0
 380 /**
 381  * DCT transforms the 16 dc values.
 382  * @param qp quantization parameter ??? FIXME
 383  */
 384 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 385 //    const int qmul= dequant_coeff[qp][0];
 386     int i;
 387     int temp[16]; //FIXME check if this is a good idea
 388     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 389     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 390
 391     for(i=0; i<4; i++){
 392         const int offset= y_offset[i];
 393         const int z0= block[offset+stride*0] + block[offset+stride*4];
 394         const int z1= block[offset+stride*0] - block[offset+stride*4];
 395         const int z2= block[offset+stride*1] - block[offset+stride*5];
 396         const int z3= block[offset+stride*1] + block[offset+stride*5];
 397
 398         temp[4*i+0]= z0+z3;
 399         temp[4*i+1]= z1+z2;
 400         temp[4*i+2]= z1-z2;
 401         temp[4*i+3]= z0-z3;
 402     }
 403
 404     for(i=0; i<4; i++){
 405         const int offset= x_offset[i];
 406         const int z0= temp[4*0+i] + temp[4*2+i];
 407         const int z1= temp[4*0+i] - temp[4*2+i];
 408         const int z2= temp[4*1+i] - temp[4*3+i];
 409         const int z3= temp[4*1+i] + temp[4*3+i];
 410
 411         block[stride*0 +offset]= (z0 + z3)>>1;
 412         block[stride*2 +offset]= (z1 + z2)>>1;
 413         block[stride*8 +offset]= (z1 - z2)>>1;
 414         block[stride*10+offset]= (z0 - z3)>>1;
 415     }
 416 }
 417 #endif
 418
 419 #undef xStride
 420 #undef stride
 421
 422 #if 0
 423 static void chroma_dc_dct_c(DCTELEM *block){
 424     const int stride= 16*2;
 425     const int xStride= 16;
 426     int a,b,c,d,e;
 427
 428     a= block[stride*0 + xStride*0];
 429     b= block[stride*0 + xStride*1];
 430     c= block[stride*1 + xStride*0];
 431     d= block[stride*1 + xStride*1];
 432
 433     e= a-b;
 434     a= a+b;
 435     b= c-d;
 436     c= c+d;
 437
 438     block[stride*0 + xStride*0]= (a+c);
 439     block[stride*0 + xStride*1]= (e+b);
 440     block[stride*1 + xStride*0]= (a-c);
 441     block[stride*1 + xStride*1]= (e-b);
 442 }
 443 #endif
 444
 445 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 446                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 447                            int src_x_offset, int src_y_offset,
 448                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 449                            int pixel_shift, int chroma444){
 450     MpegEncContext * const s = &h->s;
 451     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 452     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 453     const int luma_xy= (mx&3) + ((my&3)<<2);
 454     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 455     uint8_t * src_y = pic->f.data[0] + offset;
 456     uint8_t * src_cb, * src_cr;
 457     int extra_width= h->emu_edge_width;
 458     int extra_height= h->emu_edge_height;
 459     int emu=0;
 460     const int full_mx= mx>>2;
 461     const int full_my= my>>2;
 462     const int pic_width  = 16*s->mb_width;
 463     const int pic_height = 16*s->mb_height >> MB_FIELD;
 464
 465     if(mx&7) extra_width -= 3;
 466     if(my&7) extra_height -= 3;
 467
 468     if(   full_mx < 0-extra_width
 469        || full_my < 0-extra_height
 470        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 471        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 472         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 473             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 474         emu=1;
 475     }
 476
 477     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 478     if(!square){
 479         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 480     }
 481
 482     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 483
 484     if(chroma444){
 485         src_cb = pic->f.data[1] + offset;
 486         if(emu){
 487             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 488                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 489             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 490         }
 491         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 492         if(!square){
 493             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 494         }
 495
 496         src_cr = pic->f.data[2] + offset;
 497         if(emu){
 498             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 499                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 500             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 501         }
 502         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 503         if(!square){
 504             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 505         }
 506         return;
 507     }
 508
 509     if(MB_FIELD){
 510         // chroma offset when predicting from a field of opposite parity
 511         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 512         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 513     }
 514     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 515     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 516
 517     if(emu){
 518         s->dsp.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);
 519             src_cb= s->edge_emu_buffer;
 520     }
 521     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 522
 523     if(emu){
 524         s->dsp.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);
 525             src_cr= s->edge_emu_buffer;
 526     }
 527     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 528 }
 529
 530 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 531                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 532                            int x_offset, int y_offset,
 533                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 534                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 535                            int list0, int list1, int pixel_shift, int chroma444){
 536     MpegEncContext * const s = &h->s;
 537     qpel_mc_func *qpix_op=  qpix_put;
 538     h264_chroma_mc_func chroma_op= chroma_put;
 539
 540     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 541     if(chroma444){
 542         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 543         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 544     }else{
 545         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 546         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 547     }
 548     x_offset += 8*s->mb_x;
 549     y_offset += 8*(s->mb_y >> MB_FIELD);
 550
 551     if(list0){
 552         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 553         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 554                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 555                            qpix_op, chroma_op, pixel_shift, chroma444);
 556
 557         qpix_op=  qpix_avg;
 558         chroma_op= chroma_avg;
 559     }
 560
 561     if(list1){
 562         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 563         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 564                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 565                            qpix_op, chroma_op, pixel_shift, chroma444);
 566     }
 567 }
 568
 569 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 570                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 571                            int x_offset, int y_offset,
 572                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 573                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 574                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 575                            int list0, int list1, int pixel_shift, int chroma444){
 576     MpegEncContext * const s = &h->s;
 577
 578     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 579     if(chroma444){
 580         chroma_weight_avg = luma_weight_avg;
 581         chroma_weight_op = luma_weight_op;
 582         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 583         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 584     }else{
 585         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 586         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 587     }
 588     x_offset += 8*s->mb_x;
 589     y_offset += 8*(s->mb_y >> MB_FIELD);
 590
 591     if(list0 && list1){
 592         /* don't optimize for luma-only case, since B-frames usually
 593          * use implicit weights => chroma too. */
 594         uint8_t *tmp_cb = s->obmc_scratchpad;
 595         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 596         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 597         int refn0 = h->ref_cache[0][ scan8[n] ];
 598         int refn1 = h->ref_cache[1][ scan8[n] ];
 599
 600         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 601                     dest_y, dest_cb, dest_cr,
 602                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 603         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 604                     tmp_y, tmp_cb, tmp_cr,
 605                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 606
 607         if(h->use_weight == 2){
 608             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 609             int weight1 = 64 - weight0;
 610             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 611             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 612             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 613         }else{
 614             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 615                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 616                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 617             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 618                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 619                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 620             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 621                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 622                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 623         }
 624     }else{
 625         int list = list1 ? 1 : 0;
 626         int refn = h->ref_cache[list][ scan8[n] ];
 627         Picture *ref= &h->ref_list[list][refn];
 628         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 629                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 630                     qpix_put, chroma_put, pixel_shift, chroma444);
 631
 632         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 633                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 634         if(h->use_weight_chroma){
 635             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 636                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 637             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 638                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 639         }
 640     }
 641 }
 642
 643 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 644                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 645                            int x_offset, int y_offset,
 646                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 647                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 648                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 649                            int list0, int list1, int pixel_shift, int chroma444){
 650     if((h->use_weight==2 && list0 && list1
 651         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 652        || h->use_weight==1)
 653         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 654                          x_offset, y_offset, qpix_put, chroma_put,
 655                          weight_op[0], weight_op[3], weight_avg[0],
 656                          weight_avg[3], list0, list1, pixel_shift, chroma444);
 657     else
 658         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 659                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 660                     chroma_avg, list0, list1, pixel_shift, chroma444);
 661 }
 662
 663 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
 664     /* fetch pixels for estimated mv 4 macroblocks ahead
 665      * optimized for 64byte cache lines */
 666     MpegEncContext * const s = &h->s;
 667     const int refn = h->ref_cache[list][scan8[0]];
 668     if(refn >= 0){
 669         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 670         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 671         uint8_t **src = h->ref_list[list][refn].f.data;
 672         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 673         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 674         if(chroma444){
 675             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 676             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 677         }else{
 678             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 679             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 680         }
 681     }
 682 }
 683
 684 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 685                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 686                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 687                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 688                       int pixel_shift, int chroma444){
 689     MpegEncContext * const s = &h->s;
 690     const int mb_xy= h->mb_xy;
 691     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 692
 693     assert(IS_INTER(mb_type));
 694
 695     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 696         await_references(h);
 697     prefetch_motion(h, 0, pixel_shift, chroma444);
 698
 699     if(IS_16X16(mb_type)){
 700         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 701                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 702                 weight_op, weight_avg,
 703                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 704                 pixel_shift, chroma444);
 705     }else if(IS_16X8(mb_type)){
 706         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 707                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 708                 &weight_op[1], &weight_avg[1],
 709                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 710                 pixel_shift, chroma444);
 711         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 712                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 713                 &weight_op[1], &weight_avg[1],
 714                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 715                 pixel_shift, chroma444);
 716     }else if(IS_8X16(mb_type)){
 717         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 718                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 719                 &weight_op[2], &weight_avg[2],
 720                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 721                 pixel_shift, chroma444);
 722         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 723                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 724                 &weight_op[2], &weight_avg[2],
 725                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 726                 pixel_shift, chroma444);
 727     }else{
 728         int i;
 729
 730         assert(IS_8X8(mb_type));
 731
 732         for(i=0; i<4; i++){
 733             const int sub_mb_type= h->sub_mb_type[i];
 734             const int n= 4*i;
 735             int x_offset= (i&1)<<2;
 736             int y_offset= (i&2)<<1;
 737
 738             if(IS_SUB_8X8(sub_mb_type)){
 739                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 740                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 741                     &weight_op[3], &weight_avg[3],
 742                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 743                     pixel_shift, chroma444);
 744             }else if(IS_SUB_8X4(sub_mb_type)){
 745                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 746                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 747                     &weight_op[4], &weight_avg[4],
 748                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 749                     pixel_shift, chroma444);
 750                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 751                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 752                     &weight_op[4], &weight_avg[4],
 753                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 754                     pixel_shift, chroma444);
 755             }else if(IS_SUB_4X8(sub_mb_type)){
 756                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 757                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 758                     &weight_op[5], &weight_avg[5],
 759                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 760                     pixel_shift, chroma444);
 761                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 762                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 763                     &weight_op[5], &weight_avg[5],
 764                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 765                     pixel_shift, chroma444);
 766             }else{
 767                 int j;
 768                 assert(IS_SUB_4X4(sub_mb_type));
 769                 for(j=0; j<4; j++){
 770                     int sub_x_offset= x_offset + 2*(j&1);
 771                     int sub_y_offset= y_offset +   (j&2);
 772                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 773                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 774                         &weight_op[6], &weight_avg[6],
 775                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 776                         pixel_shift, chroma444);
 777                 }
 778             }
 779         }
 780     }
 781
 782     prefetch_motion(h, 1, pixel_shift, chroma444);
 783 }
 784
 785 static void free_tables(H264Context *h, int free_rbsp){
 786     int i;
 787     H264Context *hx;
 788
 789     av_freep(&h->intra4x4_pred_mode);
 790     av_freep(&h->chroma_pred_mode_table);
 791     av_freep(&h->cbp_table);
 792     av_freep(&h->mvd_table[0]);
 793     av_freep(&h->mvd_table[1]);
 794     av_freep(&h->direct_table);
 795     av_freep(&h->non_zero_count);
 796     av_freep(&h->slice_table_base);
 797     h->slice_table= NULL;
 798     av_freep(&h->list_counts);
 799
 800     av_freep(&h->mb2b_xy);
 801     av_freep(&h->mb2br_xy);
 802
 803     for(i = 0; i < MAX_THREADS; i++) {
 804         hx = h->thread_context[i];
 805         if(!hx) continue;
 806         av_freep(&hx->top_borders[1]);
 807         av_freep(&hx->top_borders[0]);
 808         av_freep(&hx->s.obmc_scratchpad);
 809         if (free_rbsp){
 810             av_freep(&hx->rbsp_buffer[1]);
 811             av_freep(&hx->rbsp_buffer[0]);
 812             hx->rbsp_buffer_size[0] = 0;
 813             hx->rbsp_buffer_size[1] = 0;
 814         }
 815         if (i) av_freep(&h->thread_context[i]);
 816     }
 817 }
 818
 819 static void init_dequant8_coeff_table(H264Context *h){
 820     int i,j,q,x;
 821     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 822
 823     for(i=0; i<6; i++ ){
 824         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 825         for(j=0; j<i; j++){
 826             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 827                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 828                 break;
 829             }
 830         }
 831         if(j<i)
 832             continue;
 833
 834         for(q=0; q<max_qp+1; q++){
 835             int shift = div6[q];
 836             int idx = rem6[q];
 837             for(x=0; x<64; x++)
 838                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 839                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 840                     h->pps.scaling_matrix8[i][x]) << shift;
 841         }
 842     }
 843 }
 844
 845 static void init_dequant4_coeff_table(H264Context *h){
 846     int i,j,q,x;
 847     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 848     for(i=0; i<6; i++ ){
 849         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 850         for(j=0; j<i; j++){
 851             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 852                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 853                 break;
 854             }
 855         }
 856         if(j<i)
 857             continue;
 858
 859         for(q=0; q<max_qp+1; q++){
 860             int shift = div6[q] + 2;
 861             int idx = rem6[q];
 862             for(x=0; x<16; x++)
 863                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 864                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 865                     h->pps.scaling_matrix4[i][x]) << shift;
 866         }
 867     }
 868 }
 869
 870 static void init_dequant_tables(H264Context *h){
 871     int i,x;
 872     init_dequant4_coeff_table(h);
 873     if(h->pps.transform_8x8_mode)
 874         init_dequant8_coeff_table(h);
 875     if(h->sps.transform_bypass){
 876         for(i=0; i<6; i++)
 877             for(x=0; x<16; x++)
 878                 h->dequant4_coeff[i][0][x] = 1<<6;
 879         if(h->pps.transform_8x8_mode)
 880             for(i=0; i<6; i++)
 881                 for(x=0; x<64; x++)
 882                     h->dequant8_coeff[i][0][x] = 1<<6;
 883     }
 884 }
 885
 886
 887 int ff_h264_alloc_tables(H264Context *h){
 888     MpegEncContext * const s = &h->s;
 889     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 890     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 891     int x,y;
 892
 893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 894
 895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 896     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 898
 899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 900     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 901     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 902     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 903     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 904
 905     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 906     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 907
 908     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 909     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 910     for(y=0; y<s->mb_height; y++){
 911         for(x=0; x<s->mb_width; x++){
 912             const int mb_xy= x + y*s->mb_stride;
 913             const int b_xy = 4*x + 4*y*h->b_stride;
 914
 915             h->mb2b_xy [mb_xy]= b_xy;
 916             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 917         }
 918     }
 919
 920     s->obmc_scratchpad = NULL;
 921
 922     if(!h->dequant4_coeff[0])
 923         init_dequant_tables(h);
 924
 925     return 0;
 926 fail:
 927     free_tables(h, 1);
 928     return -1;
 929 }
 930
 931 /**
 932  * Mimic alloc_tables(), but for every context thread.
 933  */
 934 static void clone_tables(H264Context *dst, H264Context *src, int i){
 935     MpegEncContext * const s = &src->s;
 936     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 937     dst->non_zero_count           = src->non_zero_count;
 938     dst->slice_table              = src->slice_table;
 939     dst->cbp_table                = src->cbp_table;
 940     dst->mb2b_xy                  = src->mb2b_xy;
 941     dst->mb2br_xy                 = src->mb2br_xy;
 942     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 943     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 944     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 945     dst->direct_table             = src->direct_table;
 946     dst->list_counts              = src->list_counts;
 947
 948     dst->s.obmc_scratchpad = NULL;
 949     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
 950 }
 951
 952 /**
 953  * Init context
 954  * Allocate buffers which are not shared amongst multiple threads.
 955  */
 956 static int context_init(H264Context *h){
 957     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 958     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 959
 960     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 961     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 962
 963     return 0;
 964 fail:
 965     return -1; // free_tables will clean up for us
 966 }
 967
 968 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 969
 970 static av_cold void common_init(H264Context *h){
 971     MpegEncContext * const s = &h->s;
 972
 973     s->width = s->avctx->width;
 974     s->height = s->avctx->height;
 975     s->codec_id= s->avctx->codec->id;
 976
 977     ff_h264dsp_init(&h->h264dsp, 8, 1);
 978     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
 979
 980     h->dequant_coeff_pps= -1;
 981     s->unrestricted_mv=1;
 982     s->decode=1; //FIXME
 983
 984     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 985
 986     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 987     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 988 }
 989
 990 int ff_h264_decode_extradata(H264Context *h)
 991 {
 992     AVCodecContext *avctx = h->s.avctx;
 993
 994     if(avctx->extradata[0] == 1){
 995         int i, cnt, nalsize;
 996         unsigned char *p = avctx->extradata;
 997
 998         h->is_avc = 1;
 999
1000         if(avctx->extradata_size < 7) {
1001             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1002             return -1;
1003         }
1004         /* sps and pps in the avcC always have length coded with 2 bytes,
1005            so put a fake nal_length_size = 2 while parsing them */
1006         h->nal_length_size = 2;
1007         // Decode sps from avcC
1008         cnt = *(p+5) & 0x1f; // Number of sps
1009         p += 6;
1010         for (i = 0; i < cnt; i++) {
1011             nalsize = AV_RB16(p) + 2;
1012             if(decode_nal_units(h, p, nalsize) < 0) {
1013                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1014                 return -1;
1015             }
1016             p += nalsize;
1017         }
1018         // Decode pps from avcC
1019         cnt = *(p++); // Number of pps
1020         for (i = 0; i < cnt; i++) {
1021             nalsize = AV_RB16(p) + 2;
1022             if (decode_nal_units(h, p, nalsize) < 0) {
1023                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1024                 return -1;
1025             }
1026             p += nalsize;
1027         }
1028         // Now store right nal length size, that will be use to parse all other nals
1029         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1030     } else {
1031         h->is_avc = 0;
1032         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1033             return -1;
1034     }
1035     return 0;
1036 }
1037
1038 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1039     H264Context *h= avctx->priv_data;
1040     MpegEncContext * const s = &h->s;
1041
1042     MPV_decode_defaults(s);
1043
1044     s->avctx = avctx;
1045     common_init(h);
1046
1047     s->out_format = FMT_H264;
1048     s->workaround_bugs= avctx->workaround_bugs;
1049
1050     // set defaults
1051 //    s->decode_mb= ff_h263_decode_mb;
1052     s->quarter_sample = 1;
1053     if(!avctx->has_b_frames)
1054     s->low_delay= 1;
1055
1056     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1057
1058     ff_h264_decode_init_vlc();
1059
1060     h->pixel_shift = 0;
1061     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1062
1063     h->thread_context[0] = h;
1064     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1065     h->prev_poc_msb= 1<<16;
1066     h->x264_build = -1;
1067     ff_h264_reset_sei(h);
1068     if(avctx->codec_id == CODEC_ID_H264){
1069         if(avctx->ticks_per_frame == 1){
1070             s->avctx->time_base.den *=2;
1071         }
1072         avctx->ticks_per_frame = 2;
1073     }
1074
1075     if(avctx->extradata_size > 0 && avctx->extradata &&
1076         ff_h264_decode_extradata(h))
1077         return -1;
1078
1079     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1080         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1081         s->low_delay = 0;
1082     }
1083
1084     return 0;
1085 }
1086
1087 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1088 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1089 {
1090     int i;
1091
1092     for (i=0; i<count; i++){
1093         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1094                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1095                 !from[i]));
1096         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1097     }
1098 }
1099
1100 static void copy_parameter_set(void **to, void **from, int count, int size)
1101 {
1102     int i;
1103
1104     for (i=0; i<count; i++){
1105         if (to[i] && !from[i]) av_freep(&to[i]);
1106         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1107
1108         if (from[i]) memcpy(to[i], from[i], size);
1109     }
1110 }
1111
1112 static int decode_init_thread_copy(AVCodecContext *avctx){
1113     H264Context *h= avctx->priv_data;
1114
1115     if (!avctx->is_copy) return 0;
1116     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1117     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1118
1119     return 0;
1120 }
1121
1122 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1123 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1124     H264Context *h= dst->priv_data, *h1= src->priv_data;
1125     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1126     int inited = s->context_initialized, err;
1127     int i;
1128
1129     if(dst == src || !s1->context_initialized) return 0;
1130
1131     err = ff_mpeg_update_thread_context(dst, src);
1132     if(err) return err;
1133
1134     //FIXME handle width/height changing
1135     if(!inited){
1136         for(i = 0; i < MAX_SPS_COUNT; i++)
1137             av_freep(h->sps_buffers + i);
1138
1139         for(i = 0; i < MAX_PPS_COUNT; i++)
1140             av_freep(h->pps_buffers + i);
1141
1142         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1143         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1144         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1145         if (ff_h264_alloc_tables(h) < 0) {
1146             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1147             return AVERROR(ENOMEM);
1148         }
1149         context_init(h);
1150
1151         for(i=0; i<2; i++){
1152             h->rbsp_buffer[i] = NULL;
1153             h->rbsp_buffer_size[i] = 0;
1154         }
1155
1156         h->thread_context[0] = h;
1157
1158         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1159         // so this has to be allocated here
1160         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1161
1162         s->dsp.clear_blocks(h->mb);
1163         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1164     }
1165
1166     //extradata/NAL handling
1167     h->is_avc          = h1->is_avc;
1168
1169     //SPS/PPS
1170     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1171     h->sps             = h1->sps;
1172     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1173     h->pps             = h1->pps;
1174
1175     //Dequantization matrices
1176     //FIXME these are big - can they be only copied when PPS changes?
1177     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1178
1179     for(i=0; i<6; i++)
1180         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1181
1182     for(i=0; i<6; i++)
1183         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1184
1185     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1186
1187     //POC timing
1188     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1189
1190     //reference lists
1191     copy_fields(h, h1, ref_count, list_count);
1192     copy_fields(h, h1, ref_list,  intra_gb);
1193     copy_fields(h, h1, short_ref, cabac_init_idc);
1194
1195     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1196     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1197     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1198
1199     h->last_slice_type = h1->last_slice_type;
1200
1201     if(!s->current_picture_ptr) return 0;
1202
1203     if(!s->dropable) {
1204         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1205         h->prev_poc_msb     = h->poc_msb;
1206         h->prev_poc_lsb     = h->poc_lsb;
1207     }
1208     h->prev_frame_num_offset= h->frame_num_offset;
1209     h->prev_frame_num       = h->frame_num;
1210     h->outputed_poc         = h->next_outputed_poc;
1211
1212     return err;
1213 }
1214
1215 int ff_h264_frame_start(H264Context *h){
1216     MpegEncContext * const s = &h->s;
1217     int i;
1218     const int pixel_shift = h->pixel_shift;
1219     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1220
1221     if(MPV_frame_start(s, s->avctx) < 0)
1222         return -1;
1223     ff_er_frame_start(s);
1224     /*
1225      * MPV_frame_start uses pict_type to derive key_frame.
1226      * This is incorrect for H.264; IDR markings must be used.
1227      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1228      * See decode_nal_units().
1229      */
1230     s->current_picture_ptr->f.key_frame = 0;
1231     s->current_picture_ptr->mmco_reset= 0;
1232
1233     assert(s->linesize && s->uvlinesize);
1234
1235     for(i=0; i<16; i++){
1236         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1237         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1238     }
1239     for(i=0; i<16; i++){
1240         h->block_offset[16+i]=
1241         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1242         h->block_offset[48+16+i]=
1243         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1244     }
1245
1246     /* can't be in alloc_tables because linesize isn't known there.
1247      * FIXME: redo bipred weight to not require extra buffer? */
1248     for(i = 0; i < thread_count; i++)
1249         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1250             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1251
1252     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1253     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1254
1255 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1256
1257     // We mark the current picture as non-reference after allocating it, so
1258     // that if we break out due to an error it can be released automatically
1259     // in the next MPV_frame_start().
1260     // SVQ3 as well as most other codecs have only last/next/current and thus
1261     // get released even with set reference, besides SVQ3 and others do not
1262     // mark frames as reference later "naturally".
1263     if(s->codec_id != CODEC_ID_SVQ3)
1264         s->current_picture_ptr->f.reference = 0;
1265
1266     s->current_picture_ptr->field_poc[0]=
1267     s->current_picture_ptr->field_poc[1]= INT_MAX;
1268
1269     h->next_output_pic = NULL;
1270
1271     assert(s->current_picture_ptr->long_ref==0);
1272
1273     return 0;
1274 }
1275
1276 /**
1277   * Run setup operations that must be run after slice header decoding.
1278   * This includes finding the next displayed frame.
1279   *
1280   * @param h h264 master context
1281   * @param setup_finished enough NALs have been read that we can call
1282   * ff_thread_finish_setup()
1283   */
1284 static void decode_postinit(H264Context *h, int setup_finished){
1285     MpegEncContext * const s = &h->s;
1286     Picture *out = s->current_picture_ptr;
1287     Picture *cur = s->current_picture_ptr;
1288     int i, pics, out_of_order, out_idx;
1289
1290     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1291     s->current_picture_ptr->f.pict_type   = s->pict_type;
1292
1293     if (h->next_output_pic) return;
1294
1295     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1296         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1297         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1298         //to find this yet, so we assume the worst for now.
1299         //if (setup_finished)
1300         //    ff_thread_finish_setup(s->avctx);
1301         return;
1302     }
1303
1304     cur->f.interlaced_frame = 0;
1305     cur->f.repeat_pict      = 0;
1306
1307     /* Signal interlacing information externally. */
1308     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1309
1310     if(h->sps.pic_struct_present_flag){
1311         switch (h->sei_pic_struct)
1312         {
1313         case SEI_PIC_STRUCT_FRAME:
1314             break;
1315         case SEI_PIC_STRUCT_TOP_FIELD:
1316         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1317             cur->f.interlaced_frame = 1;
1318             break;
1319         case SEI_PIC_STRUCT_TOP_BOTTOM:
1320         case SEI_PIC_STRUCT_BOTTOM_TOP:
1321             if (FIELD_OR_MBAFF_PICTURE)
1322                 cur->f.interlaced_frame = 1;
1323             else
1324                 // try to flag soft telecine progressive
1325                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1326             break;
1327         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1328         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1329             // Signal the possibility of telecined film externally (pic_struct 5,6)
1330             // From these hints, let the applications decide if they apply deinterlacing.
1331             cur->f.repeat_pict = 1;
1332             break;
1333         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1334             // Force progressive here, as doubling interlaced frame is a bad idea.
1335             cur->f.repeat_pict = 2;
1336             break;
1337         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1338             cur->f.repeat_pict = 4;
1339             break;
1340         }
1341
1342         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1343             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1344     }else{
1345         /* Derive interlacing flag from used decoding process. */
1346         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1347     }
1348     h->prev_interlaced_frame = cur->f.interlaced_frame;
1349
1350     if (cur->field_poc[0] != cur->field_poc[1]){
1351         /* Derive top_field_first from field pocs. */
1352         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1353     }else{
1354         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1355             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1356             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1357               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1358                 cur->f.top_field_first = 1;
1359             else
1360                 cur->f.top_field_first = 0;
1361         }else{
1362             /* Most likely progressive */
1363             cur->f.top_field_first = 0;
1364         }
1365     }
1366
1367     //FIXME do something with unavailable reference frames
1368
1369     /* Sort B-frames into display order */
1370
1371     if(h->sps.bitstream_restriction_flag
1372        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1373         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1374         s->low_delay = 0;
1375     }
1376
1377     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1378        && !h->sps.bitstream_restriction_flag){
1379         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1380         s->low_delay= 0;
1381     }
1382
1383     pics = 0;
1384     while(h->delayed_pic[pics]) pics++;
1385
1386     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1387
1388     h->delayed_pic[pics++] = cur;
1389     if (cur->f.reference == 0)
1390         cur->f.reference = DELAYED_PIC_REF;
1391
1392     out = h->delayed_pic[0];
1393     out_idx = 0;
1394     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1395         if(h->delayed_pic[i]->poc < out->poc){
1396             out = h->delayed_pic[i];
1397             out_idx = i;
1398         }
1399     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1400         h->next_outputed_poc= INT_MIN;
1401     out_of_order = out->poc < h->next_outputed_poc;
1402
1403     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1404         { }
1405     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1406        || (s->low_delay &&
1407         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1408          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1409     {
1410         s->low_delay = 0;
1411         s->avctx->has_b_frames++;
1412     }
1413
1414     if(out_of_order || pics > s->avctx->has_b_frames){
1415         out->f.reference &= ~DELAYED_PIC_REF;
1416         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1417                          // or else the first thread can release the picture and reuse it unsafely
1418         for(i=out_idx; h->delayed_pic[i]; i++)
1419             h->delayed_pic[i] = h->delayed_pic[i+1];
1420     }
1421     if(!out_of_order && pics > s->avctx->has_b_frames){
1422         h->next_output_pic = out;
1423         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1424             h->next_outputed_poc = INT_MIN;
1425         } else
1426             h->next_outputed_poc = out->poc;
1427     }else{
1428         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1429     }
1430
1431     if (setup_finished)
1432         ff_thread_finish_setup(s->avctx);
1433 }
1434
1435 static av_always_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){
1436     MpegEncContext * const s = &h->s;
1437     uint8_t *top_border;
1438     int top_idx = 1;
1439     const int pixel_shift = h->pixel_shift;
1440     int chroma444 = CHROMA444;
1441     int chroma422 = CHROMA422;
1442
1443     src_y  -=   linesize;
1444     src_cb -= uvlinesize;
1445     src_cr -= uvlinesize;
1446
1447     if(!simple && FRAME_MBAFF){
1448         if(s->mb_y&1){
1449             if(!MB_MBAFF){
1450                 top_border = h->top_borders[0][s->mb_x];
1451                 AV_COPY128(top_border, src_y + 15*linesize);
1452                 if (pixel_shift)
1453                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1454                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1455                     if(chroma444){
1456                         if (pixel_shift){
1457                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1458                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1459                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1460                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1461                         } else {
1462                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1463                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1464                         }
1465                     } else if(chroma422){
1466                         if (pixel_shift) {
1467                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1468                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1469                         } else {
1470                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1471                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1472                         }
1473                     } else {
1474                         if (pixel_shift) {
1475                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1476                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1477                         } else {
1478                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1479                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1480                         }
1481                     }
1482                 }
1483             }
1484         }else if(MB_MBAFF){
1485             top_idx = 0;
1486         }else
1487             return;
1488     }
1489
1490     top_border = h->top_borders[top_idx][s->mb_x];
1491     // There are two lines saved, the line above the the top macroblock of a pair,
1492     // and the line above the bottom macroblock
1493     AV_COPY128(top_border, src_y + 16*linesize);
1494     if (pixel_shift)
1495         AV_COPY128(top_border+16, src_y+16*linesize+16);
1496
1497     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1498         if(chroma444){
1499             if (pixel_shift){
1500                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1501                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1502                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1503                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1504             } else {
1505                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1506                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1507             }
1508         } else if(chroma422) {
1509             if (pixel_shift) {
1510                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1511                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1512             } else {
1513                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1514                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1515             }
1516         } else {
1517             if (pixel_shift) {
1518                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1519                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1520             } else {
1521                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1522                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1523             }
1524         }
1525     }
1526 }
1527
1528 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1529                                   uint8_t *src_cb, uint8_t *src_cr,
1530                                   int linesize, int uvlinesize,
1531                                   int xchg, int chroma444,
1532                                   int simple, int pixel_shift){
1533     MpegEncContext * const s = &h->s;
1534     int deblock_topleft;
1535     int deblock_top;
1536     int top_idx = 1;
1537     uint8_t *top_border_m1;
1538     uint8_t *top_border;
1539
1540     if(!simple && FRAME_MBAFF){
1541         if(s->mb_y&1){
1542             if(!MB_MBAFF)
1543                 return;
1544         }else{
1545             top_idx = MB_MBAFF ? 0 : 1;
1546         }
1547     }
1548
1549     if(h->deblocking_filter == 2) {
1550         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1551         deblock_top     = h->top_type;
1552     } else {
1553         deblock_topleft = (s->mb_x > 0);
1554         deblock_top     = (s->mb_y > !!MB_FIELD);
1555     }
1556
1557     src_y  -=   linesize + 1 + pixel_shift;
1558     src_cb -= uvlinesize + 1 + pixel_shift;
1559     src_cr -= uvlinesize + 1 + pixel_shift;
1560
1561     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1562     top_border    = h->top_borders[top_idx][s->mb_x];
1563
1564 #define XCHG(a,b,xchg)\
1565     if (pixel_shift) {\
1566         if (xchg) {\
1567             AV_SWAP64(b+0,a+0);\
1568             AV_SWAP64(b+8,a+8);\
1569         } else {\
1570             AV_COPY128(b,a); \
1571         }\
1572     } else \
1573 if (xchg) AV_SWAP64(b,a);\
1574 else      AV_COPY64(b,a);
1575
1576     if(deblock_top){
1577         if(deblock_topleft){
1578             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1579         }
1580         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1581         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1582         if(s->mb_x+1 < s->mb_width){
1583             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1584         }
1585     }
1586     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1587         if(chroma444){
1588             if(deblock_topleft){
1589                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1590                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1591             }
1592             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1593             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1594             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1595             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1596             if(s->mb_x+1 < s->mb_width){
1597                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1598                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1599             }
1600         } else {
1601             if(deblock_top){
1602                 if(deblock_topleft){
1603                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1604                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1605                 }
1606                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1607                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1608             }
1609         }
1610     }
1611 }
1612
1613 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1614     if (high_bit_depth) {
1615         return AV_RN32A(((int32_t*)mb) + index);
1616     } else
1617         return AV_RN16A(mb + index);
1618 }
1619
1620 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1621     if (high_bit_depth) {
1622         AV_WN32A(((int32_t*)mb) + index, value);
1623     } else
1624         AV_WN16A(mb + index, value);
1625 }
1626
1627 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1628                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1629 {
1630     MpegEncContext * const s = &h->s;
1631     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1632     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1633     int i;
1634     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1635     block_offset += 16*p;
1636     if(IS_INTRA4x4(mb_type)){
1637         if(simple || !s->encoding){
1638             if(IS_8x8DCT(mb_type)){
1639                 if(transform_bypass){
1640                     idct_dc_add =
1641                     idct_add    = s->dsp.add_pixels8;
1642                 }else{
1643                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1644                     idct_add    = h->h264dsp.h264_idct8_add;
1645                 }
1646                 for(i=0; i<16; i+=4){
1647                     uint8_t * const ptr= dest_y + block_offset[i];
1648                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1649                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1650                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1651                     }else{
1652                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1653                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1654                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1655                         if(nnz){
1656                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1657                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1658                             else
1659                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1660                         }
1661                     }
1662                 }
1663             }else{
1664                 if(transform_bypass){
1665                     idct_dc_add =
1666                     idct_add    = s->dsp.add_pixels4;
1667                 }else{
1668                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1669                     idct_add    = h->h264dsp.h264_idct_add;
1670                 }
1671                 for(i=0; i<16; i++){
1672                     uint8_t * const ptr= dest_y + block_offset[i];
1673                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1674
1675                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1676                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1677                     }else{
1678                         uint8_t *topright;
1679                         int nnz, tr;
1680                         uint64_t tr_high;
1681                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1682                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1683                             assert(s->mb_y || linesize <= block_offset[i]);
1684                             if(!topright_avail){
1685                                 if (pixel_shift) {
1686                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1687                                     topright= (uint8_t*) &tr_high;
1688                                 } else {
1689                                     tr= ptr[3 - linesize]*0x01010101;
1690                                     topright= (uint8_t*) &tr;
1691                                 }
1692                             }else
1693                                 topright= ptr + (4 << pixel_shift) - linesize;
1694                         }else
1695                             topright= NULL;
1696
1697                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1698                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1699                         if(nnz){
1700                             if(is_h264){
1701                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1702                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1703                                 else
1704                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1705                             }else
1706                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1707                         }
1708                     }
1709                 }
1710             }
1711         }
1712     }else{
1713         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1714         if(is_h264){
1715             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1716                 if(!transform_bypass)
1717                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1718                 else{
1719                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1720                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1721                     for(i = 0; i < 16; i++)
1722                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1723                 }
1724             }
1725         }else
1726             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1727     }
1728 }
1729
1730 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1731                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1732 {
1733     MpegEncContext * const s = &h->s;
1734     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1735     int i;
1736     block_offset += 16*p;
1737     if(!IS_INTRA4x4(mb_type)){
1738         if(is_h264){
1739             if(IS_INTRA16x16(mb_type)){
1740                 if(transform_bypass){
1741                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1742                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1743                     }else{
1744                         for(i=0; i<16; i++){
1745                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1746                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1747                         }
1748                     }
1749                 }else{
1750                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1751                 }
1752             }else if(h->cbp&15){
1753                 if(transform_bypass){
1754                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1755                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1756                     for(i=0; i<16; i+=di){
1757                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1758                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1759                         }
1760                     }
1761                 }else{
1762                     if(IS_8x8DCT(mb_type)){
1763                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1764                     }else{
1765                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1766                     }
1767                 }
1768             }
1769         }else{
1770             for(i=0; i<16; i++){
1771                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1772                     uint8_t * const ptr= dest_y + block_offset[i];
1773                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1774                 }
1775             }
1776         }
1777     }
1778 }
1779
1780 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1781     MpegEncContext * const s = &h->s;
1782     const int mb_x= s->mb_x;
1783     const int mb_y= s->mb_y;
1784     const int mb_xy= h->mb_xy;
1785     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1786     uint8_t  *dest_y, *dest_cb, *dest_cr;
1787     int linesize, uvlinesize /*dct_offset*/;
1788     int i, j;
1789     int *block_offset = &h->block_offset[0];
1790     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1791     /* is_h264 should always be true if SVQ3 is disabled. */
1792     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1793     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1794     const int block_h = 16>>s->chroma_y_shift;
1795
1796     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1797     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1798     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1799
1800     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1801     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1802
1803     h->list_counts[mb_xy]= h->list_count;
1804
1805     if (!simple && MB_FIELD) {
1806         linesize   = h->mb_linesize   = s->linesize * 2;
1807         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1808         block_offset = &h->block_offset[48];
1809         if(mb_y&1){ //FIXME move out of this function?
1810             dest_y -= s->linesize*15;
1811             dest_cb-= s->uvlinesize*(block_h-1);
1812             dest_cr-= s->uvlinesize*(block_h-1);
1813         }
1814         if(FRAME_MBAFF) {
1815             int list;
1816             for(list=0; list<h->list_count; list++){
1817                 if(!USES_LIST(mb_type, list))
1818                     continue;
1819                 if(IS_16X16(mb_type)){
1820                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1821                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1822                 }else{
1823                     for(i=0; i<16; i+=4){
1824                         int ref = h->ref_cache[list][scan8[i]];
1825                         if(ref >= 0)
1826                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1827                     }
1828                 }
1829             }
1830         }
1831     } else {
1832         linesize   = h->mb_linesize   = s->linesize;
1833         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1834 //        dct_offset = s->linesize * 16;
1835     }
1836
1837     if (!simple && IS_INTRA_PCM(mb_type)) {
1838         const int bit_depth = h->sps.bit_depth_luma;
1839         if (pixel_shift) {
1840             int j;
1841             GetBitContext gb;
1842             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1843
1844             for (i = 0; i < 16; i++) {
1845                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1846                 for (j = 0; j < 16; j++)
1847                     tmp_y[j] = get_bits(&gb, bit_depth);
1848             }
1849             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1850                 if (!h->sps.chroma_format_idc) {
1851                     for (i = 0; i < 8; i++) {
1852                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1853                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1854                         for (j = 0; j < 8; j++) {
1855                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
1856                         }
1857                     }
1858                 } else {
1859                     for (i = 0; i < block_h; i++) {
1860                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1861                         for (j = 0; j < 8; j++)
1862                             tmp_cb[j] = get_bits(&gb, bit_depth);
1863                     }
1864                     for (i = 0; i < block_h; i++) {
1865                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1866                         for (j = 0; j < 8; j++)
1867                             tmp_cr[j] = get_bits(&gb, bit_depth);
1868                     }
1869                 }
1870             }
1871         } else {
1872             for (i=0; i<16; i++) {
1873                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1874             }
1875             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1876                 if (!h->sps.chroma_format_idc) {
1877                     for (i=0; i<8; i++) {
1878                         memset(dest_cb+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1879                         memset(dest_cr+ i*uvlinesize, 1 << (bit_depth - 1), 8);
1880                     }
1881                 } else {
1882                     for (i=0; i<block_h; i++) {
1883                         memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1884                         memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1885                     }
1886                 }
1887             }
1888         }
1889     } else {
1890         if(IS_INTRA(mb_type)){
1891             if(h->deblocking_filter)
1892                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1893
1894             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1895                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1896                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1897             }
1898
1899             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1900
1901             if(h->deblocking_filter)
1902                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1903         }else if(is_h264){
1904             hl_motion(h, dest_y, dest_cb, dest_cr,
1905                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1906                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1907                       h->h264dsp.weight_h264_pixels_tab,
1908                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1909         }
1910
1911         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1912
1913         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1914             uint8_t *dest[2] = {dest_cb, dest_cr};
1915             if(transform_bypass){
1916                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1917                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1918                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1919                 }else{
1920                     idct_add = s->dsp.add_pixels4;
1921                     for(j=1; j<3; j++){
1922                         for(i=j*16; i<j*16+4; i++){
1923                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1924                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1925                         }
1926                     }
1927                 }
1928             }else{
1929                 if(is_h264){
1930                     int qp[2];
1931                     if (CHROMA422) {
1932                         qp[0] = h->chroma_qp[0]+3;
1933                         qp[1] = h->chroma_qp[1]+3;
1934                     } else {
1935                         qp[0] = h->chroma_qp[0];
1936                         qp[1] = h->chroma_qp[1];
1937                     }
1938                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1939                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][qp[0]][0]);
1940                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1941                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][qp[1]][0]);
1942                     h->h264dsp.h264_idct_add8(dest, block_offset,
1943                                               h->mb, uvlinesize,
1944                                               h->non_zero_count_cache);
1945                 }
1946 #if CONFIG_SVQ3_DECODER
1947                 else{
1948                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*1, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1949                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*2, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1950                     for(j=1; j<3; j++){
1951                         for(i=j*16; i<j*16+4; i++){
1952                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1953                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1954                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1955                             }
1956                         }
1957                     }
1958                 }
1959 #endif
1960             }
1961         }
1962     }
1963     if(h->cbp || IS_INTRA(mb_type))
1964     {
1965         s->dsp.clear_blocks(h->mb);
1966         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1967     }
1968 }
1969
1970 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1971     MpegEncContext * const s = &h->s;
1972     const int mb_x= s->mb_x;
1973     const int mb_y= s->mb_y;
1974     const int mb_xy= h->mb_xy;
1975     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1976     uint8_t  *dest[3];
1977     int linesize;
1978     int i, j, p;
1979     int *block_offset = &h->block_offset[0];
1980     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1981     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1982
1983     for (p = 0; p < plane_count; p++)
1984     {
1985         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1986         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1987     }
1988
1989     h->list_counts[mb_xy]= h->list_count;
1990
1991     if (!simple && MB_FIELD) {
1992         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1993         block_offset = &h->block_offset[48];
1994         if(mb_y&1) //FIXME move out of this function?
1995             for (p = 0; p < 3; p++)
1996                 dest[p] -= s->linesize*15;
1997         if(FRAME_MBAFF) {
1998             int list;
1999             for(list=0; list<h->list_count; list++){
2000                 if(!USES_LIST(mb_type, list))
2001                     continue;
2002                 if(IS_16X16(mb_type)){
2003                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2004                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2005                 }else{
2006                     for(i=0; i<16; i+=4){
2007                         int ref = h->ref_cache[list][scan8[i]];
2008                         if(ref >= 0)
2009                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2010                     }
2011                 }
2012             }
2013         }
2014     } else {
2015         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2016     }
2017
2018     if (!simple && IS_INTRA_PCM(mb_type)) {
2019         if (pixel_shift) {
2020             const int bit_depth = h->sps.bit_depth_luma;
2021             GetBitContext gb;
2022             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2023
2024             for (p = 0; p < plane_count; p++) {
2025                 for (i = 0; i < 16; i++) {
2026                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2027                     for (j = 0; j < 16; j++)
2028                         tmp[j] = get_bits(&gb, bit_depth);
2029                 }
2030             }
2031         } else {
2032             for (p = 0; p < plane_count; p++) {
2033                 for (i = 0; i < 16; i++) {
2034                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2035                 }
2036             }
2037         }
2038     } else {
2039         if(IS_INTRA(mb_type)){
2040             if(h->deblocking_filter)
2041                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2042
2043             for (p = 0; p < plane_count; p++)
2044                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2045
2046             if(h->deblocking_filter)
2047                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2048         }else{
2049             hl_motion(h, dest[0], dest[1], dest[2],
2050                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2051                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2052                       h->h264dsp.weight_h264_pixels_tab,
2053                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2054         }
2055
2056         for (p = 0; p < plane_count; p++)
2057             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2058     }
2059     if(h->cbp || IS_INTRA(mb_type))
2060     {
2061         s->dsp.clear_blocks(h->mb);
2062         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2063     }
2064 }
2065
2066 /**
2067  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2068  */
2069 #define hl_decode_mb_simple(sh, bits) \
2070 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2071     hl_decode_mb_internal(h, 1, sh); \
2072 }
2073 hl_decode_mb_simple(0, 8);
2074 hl_decode_mb_simple(1, 16);
2075
2076 /**
2077  * Process a macroblock; this handles edge cases, such as interlacing.
2078  */
2079 static void av_noinline hl_decode_mb_complex(H264Context *h){
2080     hl_decode_mb_internal(h, 0, h->pixel_shift);
2081 }
2082
2083 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2084     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2085 }
2086
2087 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2088     hl_decode_mb_444_internal(h, 1, 0);
2089 }
2090
2091 void ff_h264_hl_decode_mb(H264Context *h){
2092     MpegEncContext * const s = &h->s;
2093     const int mb_xy= h->mb_xy;
2094     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2095     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2096
2097     if (CHROMA444) {
2098         if(is_complex || h->pixel_shift)
2099             hl_decode_mb_444_complex(h);
2100         else
2101             hl_decode_mb_444_simple(h);
2102     } else if (is_complex) {
2103         hl_decode_mb_complex(h);
2104     } else if (h->pixel_shift) {
2105         hl_decode_mb_simple_16(h);
2106     } else
2107         hl_decode_mb_simple_8(h);
2108 }
2109
2110 static int pred_weight_table(H264Context *h){
2111     MpegEncContext * const s = &h->s;
2112     int list, i;
2113     int luma_def, chroma_def;
2114
2115     h->use_weight= 0;
2116     h->use_weight_chroma= 0;
2117     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2118     if(h->sps.chroma_format_idc)
2119         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2120     luma_def = 1<<h->luma_log2_weight_denom;
2121     chroma_def = 1<<h->chroma_log2_weight_denom;
2122
2123     for(list=0; list<2; list++){
2124         h->luma_weight_flag[list]   = 0;
2125         h->chroma_weight_flag[list] = 0;
2126         for(i=0; i<h->ref_count[list]; i++){
2127             int luma_weight_flag, chroma_weight_flag;
2128
2129             luma_weight_flag= get_bits1(&s->gb);
2130             if(luma_weight_flag){
2131                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2132                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2133                 if(   h->luma_weight[i][list][0] != luma_def
2134                    || h->luma_weight[i][list][1] != 0) {
2135                     h->use_weight= 1;
2136                     h->luma_weight_flag[list]= 1;
2137                 }
2138             }else{
2139                 h->luma_weight[i][list][0]= luma_def;
2140                 h->luma_weight[i][list][1]= 0;
2141             }
2142
2143             if(h->sps.chroma_format_idc){
2144                 chroma_weight_flag= get_bits1(&s->gb);
2145                 if(chroma_weight_flag){
2146                     int j;
2147                     for(j=0; j<2; j++){
2148                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2149                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2150                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2151                            || h->chroma_weight[i][list][j][1] != 0) {
2152                             h->use_weight_chroma= 1;
2153                             h->chroma_weight_flag[list]= 1;
2154                         }
2155                     }
2156                 }else{
2157                     int j;
2158                     for(j=0; j<2; j++){
2159                         h->chroma_weight[i][list][j][0]= chroma_def;
2160                         h->chroma_weight[i][list][j][1]= 0;
2161                     }
2162                 }
2163             }
2164         }
2165         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2166     }
2167     h->use_weight= h->use_weight || h->use_weight_chroma;
2168     return 0;
2169 }
2170
2171 /**
2172  * Initialize implicit_weight table.
2173  * @param field  0/1 initialize the weight for interlaced MBAFF
2174  *                -1 initializes the rest
2175  */
2176 static void implicit_weight_table(H264Context *h, int field){
2177     MpegEncContext * const s = &h->s;
2178     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2179
2180     for (i = 0; i < 2; i++) {
2181         h->luma_weight_flag[i]   = 0;
2182         h->chroma_weight_flag[i] = 0;
2183     }
2184
2185     if(field < 0){
2186         cur_poc = s->current_picture_ptr->poc;
2187     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2188        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2189         h->use_weight= 0;
2190         h->use_weight_chroma= 0;
2191         return;
2192     }
2193         ref_start= 0;
2194         ref_count0= h->ref_count[0];
2195         ref_count1= h->ref_count[1];
2196     }else{
2197         cur_poc = s->current_picture_ptr->field_poc[field];
2198         ref_start= 16;
2199         ref_count0= 16+2*h->ref_count[0];
2200         ref_count1= 16+2*h->ref_count[1];
2201     }
2202
2203     h->use_weight= 2;
2204     h->use_weight_chroma= 2;
2205     h->luma_log2_weight_denom= 5;
2206     h->chroma_log2_weight_denom= 5;
2207
2208     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2209         int poc0 = h->ref_list[0][ref0].poc;
2210         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2211             int w = 32;
2212             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2213                 int poc1 = h->ref_list[1][ref1].poc;
2214                 int td = av_clip(poc1 - poc0, -128, 127);
2215                 if(td){
2216                     int tb = av_clip(cur_poc - poc0, -128, 127);
2217                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2218                     int dist_scale_factor = (tb*tx + 32) >> 8;
2219                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2220                         w = 64 - dist_scale_factor;
2221                 }
2222             }
2223             if(field<0){
2224                 h->implicit_weight[ref0][ref1][0]=
2225                 h->implicit_weight[ref0][ref1][1]= w;
2226             }else{
2227                 h->implicit_weight[ref0][ref1][field]=w;
2228             }
2229         }
2230     }
2231 }
2232
2233 /**
2234  * instantaneous decoder refresh.
2235  */
2236 static void idr(H264Context *h){
2237     ff_h264_remove_all_refs(h);
2238     h->prev_frame_num= 0;
2239     h->prev_frame_num_offset= 0;
2240     h->prev_poc_msb=
2241     h->prev_poc_lsb= 0;
2242 }
2243
2244 /* forget old pics after a seek */
2245 static void flush_dpb(AVCodecContext *avctx){
2246     H264Context *h= avctx->priv_data;
2247     int i;
2248     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2249         if(h->delayed_pic[i])
2250             h->delayed_pic[i]->f.reference = 0;
2251         h->delayed_pic[i]= NULL;
2252     }
2253     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2254     h->prev_interlaced_frame = 1;
2255     idr(h);
2256     if(h->s.current_picture_ptr)
2257         h->s.current_picture_ptr->f.reference = 0;
2258     h->s.first_field= 0;
2259     ff_h264_reset_sei(h);
2260     ff_mpeg_flush(avctx);
2261 }
2262
2263 static int init_poc(H264Context *h){
2264     MpegEncContext * const s = &h->s;
2265     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2266     int field_poc[2];
2267     Picture *cur = s->current_picture_ptr;
2268
2269     h->frame_num_offset= h->prev_frame_num_offset;
2270     if(h->frame_num < h->prev_frame_num)
2271         h->frame_num_offset += max_frame_num;
2272
2273     if(h->sps.poc_type==0){
2274         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2275
2276         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2277             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2278         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2279             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2280         else
2281             h->poc_msb = h->prev_poc_msb;
2282 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2283         field_poc[0] =
2284         field_poc[1] = h->poc_msb + h->poc_lsb;
2285         if(s->picture_structure == PICT_FRAME)
2286             field_poc[1] += h->delta_poc_bottom;
2287     }else if(h->sps.poc_type==1){
2288         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2289         int i;
2290
2291         if(h->sps.poc_cycle_length != 0)
2292             abs_frame_num = h->frame_num_offset + h->frame_num;
2293         else
2294             abs_frame_num = 0;
2295
2296         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2297             abs_frame_num--;
2298
2299         expected_delta_per_poc_cycle = 0;
2300         for(i=0; i < h->sps.poc_cycle_length; i++)
2301             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2302
2303         if(abs_frame_num > 0){
2304             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2305             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2306
2307             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2308             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2309                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2310         } else
2311             expectedpoc = 0;
2312
2313         if(h->nal_ref_idc == 0)
2314             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2315
2316         field_poc[0] = expectedpoc + h->delta_poc[0];
2317         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2318
2319         if(s->picture_structure == PICT_FRAME)
2320             field_poc[1] += h->delta_poc[1];
2321     }else{
2322         int poc= 2*(h->frame_num_offset + h->frame_num);
2323
2324         if(!h->nal_ref_idc)
2325             poc--;
2326
2327         field_poc[0]= poc;
2328         field_poc[1]= poc;
2329     }
2330
2331     if(s->picture_structure != PICT_BOTTOM_FIELD)
2332         s->current_picture_ptr->field_poc[0]= field_poc[0];
2333     if(s->picture_structure != PICT_TOP_FIELD)
2334         s->current_picture_ptr->field_poc[1]= field_poc[1];
2335     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2336
2337     return 0;
2338 }
2339
2340
2341 /**
2342  * initialize scan tables
2343  */
2344 static void init_scan_tables(H264Context *h){
2345     int i;
2346     for(i=0; i<16; i++){
2347 #define T(x) (x>>2) | ((x<<2) & 0xF)
2348         h->zigzag_scan[i] = T(zigzag_scan[i]);
2349         h-> field_scan[i] = T( field_scan[i]);
2350 #undef T
2351     }
2352     for(i=0; i<64; i++){
2353 #define T(x) (x>>3) | ((x&7)<<3)
2354         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2355         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2356         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2357         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2358 #undef T
2359     }
2360     if(h->sps.transform_bypass){ //FIXME same ugly
2361         h->zigzag_scan_q0          = zigzag_scan;
2362         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2363         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2364         h->field_scan_q0           = field_scan;
2365         h->field_scan8x8_q0        = field_scan8x8;
2366         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2367     }else{
2368         h->zigzag_scan_q0          = h->zigzag_scan;
2369         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2370         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2371         h->field_scan_q0           = h->field_scan;
2372         h->field_scan8x8_q0        = h->field_scan8x8;
2373         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2374     }
2375 }
2376
2377 static int field_end(H264Context *h, int in_setup){
2378     MpegEncContext * const s = &h->s;
2379     AVCodecContext * const avctx= s->avctx;
2380     int err = 0;
2381     s->mb_y= 0;
2382
2383     if (!in_setup && !s->dropable)
2384         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2385                                  s->picture_structure==PICT_BOTTOM_FIELD);
2386
2387     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2388         ff_vdpau_h264_set_reference_frames(s);
2389
2390     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2391         if(!s->dropable) {
2392             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2393             h->prev_poc_msb= h->poc_msb;
2394             h->prev_poc_lsb= h->poc_lsb;
2395         }
2396         h->prev_frame_num_offset= h->frame_num_offset;
2397         h->prev_frame_num= h->frame_num;
2398         h->outputed_poc = h->next_outputed_poc;
2399     }
2400
2401     if (avctx->hwaccel) {
2402         if (avctx->hwaccel->end_frame(avctx) < 0)
2403             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2404     }
2405
2406     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2407         ff_vdpau_h264_picture_complete(s);
2408
2409     /*
2410      * FIXME: Error handling code does not seem to support interlaced
2411      * when slices span multiple rows
2412      * The ff_er_add_slice calls don't work right for bottom
2413      * fields; they cause massive erroneous error concealing
2414      * Error marking covers both fields (top and bottom).
2415      * This causes a mismatched s->error_count
2416      * and a bad error table. Further, the error count goes to
2417      * INT_MAX when called for bottom field, because mb_y is
2418      * past end by one (callers fault) and resync_mb_y != 0
2419      * causes problems for the first MB line, too.
2420      */
2421     if (!FIELD_PICTURE)
2422         ff_er_frame_end(s);
2423
2424     MPV_frame_end(s);
2425
2426     h->current_slice=0;
2427
2428     return err;
2429 }
2430
2431 /**
2432  * Replicate H264 "master" context to thread contexts.
2433  */
2434 static void clone_slice(H264Context *dst, H264Context *src)
2435 {
2436     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2437     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2438     dst->s.current_picture      = src->s.current_picture;
2439     dst->s.linesize             = src->s.linesize;
2440     dst->s.uvlinesize           = src->s.uvlinesize;
2441     dst->s.first_field          = src->s.first_field;
2442
2443     dst->prev_poc_msb           = src->prev_poc_msb;
2444     dst->prev_poc_lsb           = src->prev_poc_lsb;
2445     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2446     dst->prev_frame_num         = src->prev_frame_num;
2447     dst->short_ref_count        = src->short_ref_count;
2448
2449     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2450     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2451     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2452     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2453
2454     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2455     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2456 }
2457
2458 /**
2459  * computes profile from profile_idc and constraint_set?_flags
2460  *
2461  * @param sps SPS
2462  *
2463  * @return profile as defined by FF_PROFILE_H264_*
2464  */
2465 int ff_h264_get_profile(SPS *sps)
2466 {
2467     int profile = sps->profile_idc;
2468
2469     switch(sps->profile_idc) {
2470     case FF_PROFILE_H264_BASELINE:
2471         // constraint_set1_flag set to 1
2472         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2473         break;
2474     case FF_PROFILE_H264_HIGH_10:
2475     case FF_PROFILE_H264_HIGH_422:
2476     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2477         // constraint_set3_flag set to 1
2478         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2479         break;
2480     }
2481
2482     return profile;
2483 }
2484
2485 /**
2486  * decodes a slice header.
2487  * This will also call MPV_common_init() and frame_start() as needed.
2488  *
2489  * @param h h264context
2490  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2491  *
2492  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2493  */
2494 static int decode_slice_header(H264Context *h, H264Context *h0){
2495     MpegEncContext * const s = &h->s;
2496     MpegEncContext * const s0 = &h0->s;
2497     unsigned int first_mb_in_slice;
2498     unsigned int pps_id;
2499     int num_ref_idx_active_override_flag;
2500     unsigned int slice_type, tmp, i, j;
2501     int default_ref_list_done = 0;
2502     int last_pic_structure;
2503
2504     s->dropable= h->nal_ref_idc == 0;
2505
2506     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2507     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2508         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2509         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2510     }else{
2511         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2512         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2513     }
2514
2515     first_mb_in_slice= get_ue_golomb(&s->gb);
2516
2517     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2518         if(h0->current_slice && FIELD_PICTURE){
2519             field_end(h, 1);
2520         }
2521
2522         h0->current_slice = 0;
2523         if (!s0->first_field)
2524             s->current_picture_ptr= NULL;
2525     }
2526
2527     slice_type= get_ue_golomb_31(&s->gb);
2528     if(slice_type > 9){
2529         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);
2530         return -1;
2531     }
2532     if(slice_type > 4){
2533         slice_type -= 5;
2534         h->slice_type_fixed=1;
2535     }else
2536         h->slice_type_fixed=0;
2537
2538     slice_type= golomb_to_pict_type[ slice_type ];
2539     if (slice_type == AV_PICTURE_TYPE_I
2540         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2541         default_ref_list_done = 1;
2542     }
2543     h->slice_type= slice_type;
2544     h->slice_type_nos= slice_type & 3;
2545
2546     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2547
2548     pps_id= get_ue_golomb(&s->gb);
2549     if(pps_id>=MAX_PPS_COUNT){
2550         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2551         return -1;
2552     }
2553     if(!h0->pps_buffers[pps_id]) {
2554         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2555         return -1;
2556     }
2557     h->pps= *h0->pps_buffers[pps_id];
2558
2559     if(!h0->sps_buffers[h->pps.sps_id]) {
2560         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2561         return -1;
2562     }
2563     h->sps = *h0->sps_buffers[h->pps.sps_id];
2564
2565     s->avctx->profile = ff_h264_get_profile(&h->sps);
2566     s->avctx->level   = h->sps.level_idc;
2567     s->avctx->refs    = h->sps.ref_frame_count;
2568
2569     if(h == h0 && h->dequant_coeff_pps != pps_id){
2570         h->dequant_coeff_pps = pps_id;
2571         init_dequant_tables(h);
2572     }
2573
2574     s->mb_width= h->sps.mb_width;
2575     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2576
2577     h->b_stride=  s->mb_width*4;
2578
2579     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2580
2581     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2582     if(h->sps.frame_mbs_only_flag)
2583         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2584     else
2585         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2586
2587     if (s->context_initialized
2588         && (   s->width != s->avctx->width || s->height != s->avctx->height
2589             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2590         if(h != h0) {
2591             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2592             return -1;   // width / height changed during parallelized decoding
2593         }
2594         free_tables(h, 0);
2595         flush_dpb(s->avctx);
2596         MPV_common_end(s);
2597     }
2598     if (!s->context_initialized) {
2599         if (h != h0) {
2600             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2601             return -1;
2602         }
2603
2604         avcodec_set_dimensions(s->avctx, s->width, s->height);
2605         s->avctx->sample_aspect_ratio= h->sps.sar;
2606         av_assert0(s->avctx->sample_aspect_ratio.den);
2607
2608         if(h->sps.video_signal_type_present_flag){
2609             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2610             if(h->sps.colour_description_present_flag){
2611                 s->avctx->color_primaries = h->sps.color_primaries;
2612                 s->avctx->color_trc       = h->sps.color_trc;
2613                 s->avctx->colorspace      = h->sps.colorspace;
2614             }
2615         }
2616
2617         if(h->sps.timing_info_present_flag){
2618             int64_t den= h->sps.time_scale;
2619             if(h->x264_build < 44U)
2620                 den *= 2;
2621             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2622                       h->sps.num_units_in_tick, den, 1<<30);
2623         }
2624
2625         switch (h->sps.bit_depth_luma) {
2626             case 9 :
2627                 if (CHROMA444)
2628                     s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2629                 else
2630                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2631                 break;
2632             case 10 :
2633                 if (CHROMA444)
2634                     s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2635                 else if (CHROMA422)
2636                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2637                 else
2638                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2639                 break;
2640             default:
2641                 if (CHROMA444){
2642                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2643                 }else if (CHROMA422) {
2644                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2645                 }else{
2646                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2647                                                              s->avctx->codec->pix_fmts ?
2648                                                              s->avctx->codec->pix_fmts :
2649                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2650                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2651                                                              ff_hwaccel_pixfmt_list_420);
2652                 }
2653         }
2654
2655         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2656
2657         if (MPV_common_init(s) < 0) {
2658             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2659             return -1;
2660         }
2661         s->first_field = 0;
2662         h->prev_interlaced_frame = 1;
2663
2664         init_scan_tables(h);
2665         if (ff_h264_alloc_tables(h) < 0) {
2666             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2667             return AVERROR(ENOMEM);
2668         }
2669
2670         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2671             if (context_init(h) < 0) {
2672                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2673                 return -1;
2674             }
2675         } else {
2676             for(i = 1; i < s->avctx->thread_count; i++) {
2677                 H264Context *c;
2678                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2679                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2680                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2681                 c->h264dsp = h->h264dsp;
2682                 c->sps = h->sps;
2683                 c->pps = h->pps;
2684                 c->pixel_shift = h->pixel_shift;
2685                 init_scan_tables(c);
2686                 clone_tables(c, h, i);
2687             }
2688
2689             for(i = 0; i < s->avctx->thread_count; i++)
2690                 if (context_init(h->thread_context[i]) < 0) {
2691                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2692                     return -1;
2693                 }
2694         }
2695     }
2696
2697     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2698
2699     h->mb_mbaff = 0;
2700     h->mb_aff_frame = 0;
2701     last_pic_structure = s0->picture_structure;
2702     if(h->sps.frame_mbs_only_flag){
2703         s->picture_structure= PICT_FRAME;
2704     }else{
2705         if(get_bits1(&s->gb)) { //field_pic_flag
2706             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2707         } else {
2708             s->picture_structure= PICT_FRAME;
2709             h->mb_aff_frame = h->sps.mb_aff;
2710         }
2711     }
2712     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2713
2714     if(h0->current_slice == 0){
2715         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2716         if(h->frame_num != h->prev_frame_num) {
2717             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2718
2719             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2720
2721             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2722                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2723                 if (unwrap_prev_frame_num < 0)
2724                     unwrap_prev_frame_num += max_frame_num;
2725
2726                 h->prev_frame_num = unwrap_prev_frame_num;
2727             }
2728         }
2729
2730         while(h->frame_num !=  h->prev_frame_num &&
2731               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2732             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2733             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2734             if (ff_h264_frame_start(h) < 0)
2735                 return -1;
2736             h->prev_frame_num++;
2737             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2738             s->current_picture_ptr->frame_num= h->prev_frame_num;
2739             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2740             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2741             ff_generate_sliding_window_mmcos(h);
2742             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2743                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2744                 return AVERROR_INVALIDDATA;
2745             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2746              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2747              * about there being no actual duplicates.
2748              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2749              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2750              * be fixed. */
2751             if (h->short_ref_count) {
2752                 if (prev) {
2753                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2754                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2755                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2756                     h->short_ref[0]->poc = prev->poc+2;
2757                 }
2758                 h->short_ref[0]->frame_num = h->prev_frame_num;
2759             }
2760         }
2761
2762         /* See if we have a decoded first field looking for a pair... */
2763         if (s0->first_field) {
2764             assert(s0->current_picture_ptr);
2765             assert(s0->current_picture_ptr->f.data[0]);
2766             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2767
2768             /* figure out if we have a complementary field pair */
2769             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2770                 /*
2771                  * Previous field is unmatched. Don't display it, but let it
2772                  * remain for reference if marked as such.
2773                  */
2774                 s0->current_picture_ptr = NULL;
2775                 s0->first_field = FIELD_PICTURE;
2776
2777             } else {
2778                 if (h->nal_ref_idc &&
2779                         s0->current_picture_ptr->f.reference &&
2780                         s0->current_picture_ptr->frame_num != h->frame_num) {
2781                     /*
2782                      * This and previous field were reference, but had
2783                      * different frame_nums. Consider this field first in
2784                      * pair. Throw away previous field except for reference
2785                      * purposes.
2786                      */
2787                     s0->first_field = 1;
2788                     s0->current_picture_ptr = NULL;
2789
2790                 } else {
2791                     /* Second field in complementary pair */
2792                     s0->first_field = 0;
2793                 }
2794             }
2795
2796         } else {
2797             /* Frame or first field in a potentially complementary pair */
2798             assert(!s0->current_picture_ptr);
2799             s0->first_field = FIELD_PICTURE;
2800         }
2801
2802         if(!FIELD_PICTURE || s0->first_field) {
2803             if (ff_h264_frame_start(h) < 0) {
2804                 s0->first_field = 0;
2805                 return -1;
2806             }
2807         } else {
2808             ff_release_unused_pictures(s, 0);
2809         }
2810     }
2811     if(h != h0)
2812         clone_slice(h, h0);
2813
2814     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2815
2816     assert(s->mb_num == s->mb_width * s->mb_height);
2817     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2818        first_mb_in_slice                    >= s->mb_num){
2819         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2820         return -1;
2821     }
2822     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2823     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2824     if (s->picture_structure == PICT_BOTTOM_FIELD)
2825         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2826     assert(s->mb_y < s->mb_height);
2827
2828     if(s->picture_structure==PICT_FRAME){
2829         h->curr_pic_num=   h->frame_num;
2830         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2831     }else{
2832         h->curr_pic_num= 2*h->frame_num + 1;
2833         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2834     }
2835
2836     if(h->nal_unit_type == NAL_IDR_SLICE){
2837         get_ue_golomb(&s->gb); /* idr_pic_id */
2838     }
2839
2840     if(h->sps.poc_type==0){
2841         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2842
2843         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2844             h->delta_poc_bottom= get_se_golomb(&s->gb);
2845         }
2846     }
2847
2848     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2849         h->delta_poc[0]= get_se_golomb(&s->gb);
2850
2851         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2852             h->delta_poc[1]= get_se_golomb(&s->gb);
2853     }
2854
2855     init_poc(h);
2856
2857     if(h->pps.redundant_pic_cnt_present){
2858         h->redundant_pic_count= get_ue_golomb(&s->gb);
2859     }
2860
2861     //set defaults, might be overridden a few lines later
2862     h->ref_count[0]= h->pps.ref_count[0];
2863     h->ref_count[1]= h->pps.ref_count[1];
2864
2865     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2866         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2867             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2868         }
2869         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2870
2871         if(num_ref_idx_active_override_flag){
2872             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2873             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2874                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2875
2876             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2877                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2878                 h->ref_count[0]= h->ref_count[1]= 1;
2879                 return -1;
2880             }
2881         }
2882         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2883             h->list_count= 2;
2884         else
2885             h->list_count= 1;
2886     }else
2887         h->list_count= 0;
2888
2889     if(!default_ref_list_done){
2890         ff_h264_fill_default_ref_list(h);
2891     }
2892
2893     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2894         return -1;
2895
2896     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2897         s->last_picture_ptr= &h->ref_list[0][0];
2898         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2899     }
2900     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2901         s->next_picture_ptr= &h->ref_list[1][0];
2902         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2903     }
2904
2905     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2906        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2907         pred_weight_table(h);
2908     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2909         implicit_weight_table(h, -1);
2910     }else {
2911         h->use_weight = 0;
2912         for (i = 0; i < 2; i++) {
2913             h->luma_weight_flag[i]   = 0;
2914             h->chroma_weight_flag[i] = 0;
2915         }
2916     }
2917
2918     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2919        s->avctx->error_recognition >= FF_ER_EXPLODE)
2920         return AVERROR_INVALIDDATA;
2921
2922     if(FRAME_MBAFF){
2923         ff_h264_fill_mbaff_ref_list(h);
2924
2925         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2926             implicit_weight_table(h, 0);
2927             implicit_weight_table(h, 1);
2928         }
2929     }
2930
2931     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2932         ff_h264_direct_dist_scale_factor(h);
2933     ff_h264_direct_ref_list_init(h);
2934
2935     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2936         tmp = get_ue_golomb_31(&s->gb);
2937         if(tmp > 2){
2938             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2939             return -1;
2940         }
2941         h->cabac_init_idc= tmp;
2942     }
2943
2944     h->last_qscale_diff = 0;
2945     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2946     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2947         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2948         return -1;
2949     }
2950     s->qscale= tmp;
2951     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2952     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2953     //FIXME qscale / qp ... stuff
2954     if(h->slice_type == AV_PICTURE_TYPE_SP){
2955         get_bits1(&s->gb); /* sp_for_switch_flag */
2956     }
2957     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2958         get_se_golomb(&s->gb); /* slice_qs_delta */
2959     }
2960
2961     h->deblocking_filter = 1;
2962     h->slice_alpha_c0_offset = 52;
2963     h->slice_beta_offset = 52;
2964     if( h->pps.deblocking_filter_parameters_present ) {
2965         tmp= get_ue_golomb_31(&s->gb);
2966         if(tmp > 2){
2967             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2968             return -1;
2969         }
2970         h->deblocking_filter= tmp;
2971         if(h->deblocking_filter < 2)
2972             h->deblocking_filter^= 1; // 1<->0
2973
2974         if( h->deblocking_filter ) {
2975             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2976             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2977             if(   h->slice_alpha_c0_offset > 104U
2978                || h->slice_beta_offset     > 104U){
2979                 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);
2980                 return -1;
2981             }
2982         }
2983     }
2984
2985     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2986        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2987        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2988        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2989         h->deblocking_filter= 0;
2990
2991     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2992         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2993             /* Cheat slightly for speed:
2994                Do not bother to deblock across slices. */
2995             h->deblocking_filter = 2;
2996         } else {
2997             h0->max_contexts = 1;
2998             if(!h0->single_decode_warning) {
2999                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3000                 h0->single_decode_warning = 1;
3001             }
3002             if (h != h0) {
3003                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3004                 return 1;
3005             }
3006         }
3007     }
3008     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3009                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3010                  + 6 * (h->sps.bit_depth_luma - 8);
3011
3012 #if 0 //FMO
3013     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3014         slice_group_change_cycle= get_bits(&s->gb, ?);
3015 #endif
3016
3017     h0->last_slice_type = slice_type;
3018     h->slice_num = ++h0->current_slice;
3019
3020     if(h->slice_num)
3021         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3022     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3023         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3024         && h->slice_num >= MAX_SLICES) {
3025         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3026         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3027     }
3028
3029     for(j=0; j<2; j++){
3030         int id_list[16];
3031         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3032         for(i=0; i<16; i++){
3033             id_list[i]= 60;
3034             if (h->ref_list[j][i].f.data[0]) {
3035                 int k;
3036                 uint8_t *base = h->ref_list[j][i].f.base[0];
3037                 for(k=0; k<h->short_ref_count; k++)
3038                     if (h->short_ref[k]->f.base[0] == base) {
3039                         id_list[i]= k;
3040                         break;
3041                     }
3042                 for(k=0; k<h->long_ref_count; k++)
3043                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3044                         id_list[i]= h->short_ref_count + k;
3045                         break;
3046                     }
3047             }
3048         }
3049
3050         ref2frm[0]=
3051         ref2frm[1]= -1;
3052         for(i=0; i<16; i++)
3053             ref2frm[i+2]= 4*id_list[i]
3054                           + (h->ref_list[j][i].f.reference & 3);
3055         ref2frm[18+0]=
3056         ref2frm[18+1]= -1;
3057         for(i=16; i<48; i++)
3058             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3059                           + (h->ref_list[j][i].f.reference & 3);
3060     }
3061
3062     //FIXME: fix draw_edges+PAFF+frame threads
3063     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3064     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3065
3066     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3067         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",
3068                h->slice_num,
3069                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3070                first_mb_in_slice,
3071                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3072                pps_id, h->frame_num,
3073                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3074                h->ref_count[0], h->ref_count[1],
3075                s->qscale,
3076                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3077                h->use_weight,
3078                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3079                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3080                );
3081     }
3082
3083     return 0;
3084 }
3085
3086 int ff_h264_get_slice_type(const H264Context *h)
3087 {
3088     switch (h->slice_type) {
3089     case AV_PICTURE_TYPE_P:  return 0;
3090     case AV_PICTURE_TYPE_B:  return 1;
3091     case AV_PICTURE_TYPE_I:  return 2;
3092     case AV_PICTURE_TYPE_SP: return 3;
3093     case AV_PICTURE_TYPE_SI: return 4;
3094     default:         return -1;
3095     }
3096 }
3097
3098 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3099                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3100 {
3101     int b_stride = h->b_stride;
3102     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3103     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3104     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3105         if(USES_LIST(top_type, list)){
3106             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3107             const int b8_xy= 4*top_xy + 2;
3108             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3109             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3110             ref_cache[0 - 1*8]=
3111             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3112             ref_cache[2 - 1*8]=
3113             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3114         }else{
3115             AV_ZERO128(mv_dst - 1*8);
3116             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3117         }
3118
3119         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3120             if(USES_LIST(left_type[LTOP], list)){
3121                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3122                 const int b8_xy= 4*left_xy[LTOP] + 1;
3123                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3124                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3125                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3126                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3127                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3128                 ref_cache[-1 +  0]=
3129                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3130                 ref_cache[-1 + 16]=
3131                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3132             }else{
3133                 AV_ZERO32(mv_dst - 1 + 0);
3134                 AV_ZERO32(mv_dst - 1 + 8);
3135                 AV_ZERO32(mv_dst - 1 +16);
3136                 AV_ZERO32(mv_dst - 1 +24);
3137                 ref_cache[-1 +  0]=
3138                 ref_cache[-1 +  8]=
3139                 ref_cache[-1 + 16]=
3140                 ref_cache[-1 + 24]= LIST_NOT_USED;
3141             }
3142         }
3143     }
3144
3145     if(!USES_LIST(mb_type, list)){
3146         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3147         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3148         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3149         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3150         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3151         return;
3152     }
3153
3154     {
3155         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3156         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3157         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3158         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3159         AV_WN32A(&ref_cache[0*8], ref01);
3160         AV_WN32A(&ref_cache[1*8], ref01);
3161         AV_WN32A(&ref_cache[2*8], ref23);
3162         AV_WN32A(&ref_cache[3*8], ref23);
3163     }
3164
3165     {
3166         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3167         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3168         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3169         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3170         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3171     }
3172 }
3173
3174 /**
3175  *
3176  * @return non zero if the loop filter can be skiped
3177  */
3178 static int fill_filter_caches(H264Context *h, int mb_type){
3179     MpegEncContext * const s = &h->s;
3180     const int mb_xy= h->mb_xy;
3181     int top_xy, left_xy[LEFT_MBS];
3182     int top_type, left_type[LEFT_MBS];
3183     uint8_t *nnz;
3184     uint8_t *nnz_cache;
3185
3186     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3187
3188     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3189      * stuff, I can't imagine that these complex rules are worth it. */
3190
3191     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3192     if(FRAME_MBAFF){
3193         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3194         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3195         if(s->mb_y&1){
3196             if (left_mb_field_flag != curr_mb_field_flag) {
3197                 left_xy[LTOP] -= s->mb_stride;
3198             }
3199         }else{
3200             if(curr_mb_field_flag){
3201                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3202             }
3203             if (left_mb_field_flag != curr_mb_field_flag) {
3204                 left_xy[LBOT] += s->mb_stride;
3205             }
3206         }
3207     }
3208
3209     h->top_mb_xy = top_xy;
3210     h->left_mb_xy[LTOP] = left_xy[LTOP];
3211     h->left_mb_xy[LBOT] = left_xy[LBOT];
3212     {
3213         //for sufficiently low qp, filtering wouldn't do anything
3214         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3215         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3216         int qp = s->current_picture.f.qscale_table[mb_xy];
3217         if(qp <= qp_thresh
3218            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3219            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3220             if(!FRAME_MBAFF)
3221                 return 1;
3222             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3223                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3224                 return 1;
3225         }
3226     }
3227
3228     top_type        = s->current_picture.f.mb_type[top_xy];
3229     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3230     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3231     if(h->deblocking_filter == 2){
3232         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3233         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3234     }else{
3235         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3236         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3237     }
3238     h->top_type       = top_type;
3239     h->left_type[LTOP]= left_type[LTOP];
3240     h->left_type[LBOT]= left_type[LBOT];
3241
3242     if(IS_INTRA(mb_type))
3243         return 0;
3244
3245     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3246     if(h->list_count == 2)
3247         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3248
3249     nnz = h->non_zero_count[mb_xy];
3250     nnz_cache = h->non_zero_count_cache;
3251     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3252     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3253     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3254     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3255     h->cbp= h->cbp_table[mb_xy];
3256
3257     if(top_type){
3258         nnz = h->non_zero_count[top_xy];
3259         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3260     }
3261
3262     if(left_type[LTOP]){
3263         nnz = h->non_zero_count[left_xy[LTOP]];
3264         nnz_cache[3+8*1]= nnz[3+0*4];
3265         nnz_cache[3+8*2]= nnz[3+1*4];
3266         nnz_cache[3+8*3]= nnz[3+2*4];
3267         nnz_cache[3+8*4]= nnz[3+3*4];
3268     }
3269
3270     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3271     if(!CABAC && h->pps.transform_8x8_mode){
3272         if(IS_8x8DCT(top_type)){
3273             nnz_cache[4+8*0]=
3274             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3275             nnz_cache[6+8*0]=
3276             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3277         }
3278         if(IS_8x8DCT(left_type[LTOP])){
3279             nnz_cache[3+8*1]=
3280             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3281         }
3282         if(IS_8x8DCT(left_type[LBOT])){
3283             nnz_cache[3+8*3]=
3284             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3285         }
3286
3287         if(IS_8x8DCT(mb_type)){
3288             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3289             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3290
3291             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3292             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3293
3294             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3295             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3296
3297             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3298             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3299         }
3300     }
3301
3302     return 0;
3303 }
3304
3305 static void loop_filter(H264Context *h, int start_x, int end_x){
3306     MpegEncContext * const s = &h->s;
3307     uint8_t  *dest_y, *dest_cb, *dest_cr;
3308     int linesize, uvlinesize, mb_x, mb_y;
3309     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3310     const int old_slice_type= h->slice_type;
3311     const int pixel_shift = h->pixel_shift;
3312     const int block_h = 16>>s->chroma_y_shift;
3313
3314     if(h->deblocking_filter) {
3315         for(mb_x= start_x; mb_x<end_x; mb_x++){
3316             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3317                 int mb_xy, mb_type;
3318                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3319                 h->slice_num= h->slice_table[mb_xy];
3320                 mb_type = s->current_picture.f.mb_type[mb_xy];
3321                 h->list_count= h->list_counts[mb_xy];
3322
3323                 if(FRAME_MBAFF)
3324                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3325
3326                 s->mb_x= mb_x;
3327                 s->mb_y= mb_y;
3328                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3329                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3330                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*(8<<CHROMA444) + mb_y * s->uvlinesize * block_h;
3331                     //FIXME simplify above
3332
3333                 if (MB_FIELD) {
3334                     linesize   = h->mb_linesize   = s->linesize * 2;
3335                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3336                     if(mb_y&1){ //FIXME move out of this function?
3337                         dest_y -= s->linesize*15;
3338                         dest_cb-= s->uvlinesize*(block_h-1);
3339                         dest_cr-= s->uvlinesize*(block_h-1);
3340                     }
3341                 } else {
3342                     linesize   = h->mb_linesize   = s->linesize;
3343                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3344                 }
3345                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3346                 if(fill_filter_caches(h, mb_type))
3347                     continue;
3348                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3349                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3350
3351                 if (FRAME_MBAFF) {
3352                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3353                 } else {
3354                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3355                 }
3356             }
3357         }
3358     }
3359     h->slice_type= old_slice_type;
3360     s->mb_x= end_x;
3361     s->mb_y= end_mb_y - FRAME_MBAFF;
3362     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3363     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3364 }
3365
3366 static void predict_field_decoding_flag(H264Context *h){
3367     MpegEncContext * const s = &h->s;
3368     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3369     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3370                 ? s->current_picture.f.mb_type[mb_xy - 1]
3371                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3372                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3373                 : 0;
3374     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3375 }
3376
3377 /**
3378  * Draw edges and report progress for the last MB row.
3379  */
3380 static void decode_finish_row(H264Context *h){
3381     MpegEncContext * const s = &h->s;
3382     int top = 16*(s->mb_y >> FIELD_PICTURE);
3383     int height = 16 << FRAME_MBAFF;
3384     int deblock_border = (16 + 4) << FRAME_MBAFF;
3385     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3386
3387     if (h->deblocking_filter) {
3388         if((top + height) >= pic_height)
3389             height += deblock_border;
3390
3391         top -= deblock_border;
3392     }
3393
3394     if (top >= pic_height || (top + height) < h->emu_edge_height)
3395         return;
3396
3397     height = FFMIN(height, pic_height - top);
3398     if (top < h->emu_edge_height) {
3399         height = top+height;
3400         top = 0;
3401     }
3402
3403     ff_draw_horiz_band(s, top, height);
3404
3405     if (s->dropable) return;
3406
3407     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3408                              s->picture_structure==PICT_BOTTOM_FIELD);
3409 }
3410
3411 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3412     H264Context *h = *(void**)arg;
3413     MpegEncContext * const s = &h->s;
3414     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3415     int lf_x_start = s->mb_x;
3416
3417     s->mb_skip_run= -1;
3418
3419     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3420                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3421
3422     if( h->pps.cabac ) {
3423         /* realign */
3424         align_get_bits( &s->gb );
3425
3426         /* init cabac */
3427         ff_init_cabac_states( &h->cabac);
3428         ff_init_cabac_decoder( &h->cabac,
3429                                s->gb.buffer + get_bits_count(&s->gb)/8,
3430                                (get_bits_left(&s->gb) + 7)/8);
3431
3432         ff_h264_init_cabac_states(h);
3433
3434         for(;;){
3435 //START_TIMER
3436             int ret = ff_h264_decode_mb_cabac(h);
3437             int eos;
3438 //STOP_TIMER("decode_mb_cabac")
3439
3440             if(ret>=0) ff_h264_hl_decode_mb(h);
3441
3442             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3443                 s->mb_y++;
3444
3445                 ret = ff_h264_decode_mb_cabac(h);
3446
3447                 if(ret>=0) ff_h264_hl_decode_mb(h);
3448                 s->mb_y--;
3449             }
3450             eos = get_cabac_terminate( &h->cabac );
3451
3452             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3453                 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);
3454                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3455                 return 0;
3456             }
3457             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3458                 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);
3459                 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);
3460                 return -1;
3461             }
3462
3463             if( ++s->mb_x >= s->mb_width ) {
3464                 loop_filter(h, lf_x_start, s->mb_x);
3465                 s->mb_x = lf_x_start = 0;
3466                 decode_finish_row(h);
3467                 ++s->mb_y;
3468                 if(FIELD_OR_MBAFF_PICTURE) {
3469                     ++s->mb_y;
3470                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3471                         predict_field_decoding_flag(h);
3472                 }
3473             }
3474
3475             if( eos || s->mb_y >= s->mb_height ) {
3476                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3477                 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);
3478                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3479                 return 0;
3480             }
3481         }
3482
3483     } else {
3484         for(;;){
3485             int ret = ff_h264_decode_mb_cavlc(h);
3486
3487             if(ret>=0) ff_h264_hl_decode_mb(h);
3488
3489             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3490                 s->mb_y++;
3491                 ret = ff_h264_decode_mb_cavlc(h);
3492
3493                 if(ret>=0) ff_h264_hl_decode_mb(h);
3494                 s->mb_y--;
3495             }
3496
3497             if(ret<0){
3498                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3499                 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);
3500                 return -1;
3501             }
3502
3503             if(++s->mb_x >= s->mb_width){
3504                 loop_filter(h, lf_x_start, s->mb_x);
3505                 s->mb_x = lf_x_start = 0;
3506                 decode_finish_row(h);
3507                 ++s->mb_y;
3508                 if(FIELD_OR_MBAFF_PICTURE) {
3509                     ++s->mb_y;
3510                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3511                         predict_field_decoding_flag(h);
3512                 }
3513                 if(s->mb_y >= s->mb_height){
3514                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3515
3516                     if(   get_bits_count(&s->gb) == s->gb.size_in_bits
3517                        || get_bits_count(&s->gb) <  s->gb.size_in_bits && s->avctx->error_recognition < FF_ER_AGGRESSIVE) {
3518                         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);
3519
3520                         return 0;
3521                     }else{
3522                         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);
3523
3524                         return -1;
3525                     }
3526                 }
3527             }
3528
3529             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3530                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3531                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3532                     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);
3533                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3534
3535                     return 0;
3536                 }else{
3537                     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);
3538
3539                     return -1;
3540                 }
3541             }
3542         }
3543     }
3544 }
3545
3546 /**
3547  * Call decode_slice() for each context.
3548  *
3549  * @param h h264 master context
3550  * @param context_count number of contexts to execute
3551  */
3552 static int execute_decode_slices(H264Context *h, int context_count){
3553     MpegEncContext * const s = &h->s;
3554     AVCodecContext * const avctx= s->avctx;
3555     H264Context *hx;
3556     int i;
3557
3558     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3559         return 0;
3560     if(context_count == 1) {
3561         return decode_slice(avctx, &h);
3562     } else {
3563         for(i = 1; i < context_count; i++) {
3564             hx = h->thread_context[i];
3565             hx->s.error_recognition = avctx->error_recognition;
3566             hx->s.error_count = 0;
3567             hx->x264_build= h->x264_build;
3568         }
3569
3570         avctx->execute(avctx, (void *)decode_slice,
3571                        h->thread_context, NULL, context_count, sizeof(void*));
3572
3573         /* pull back stuff from slices to master context */
3574         hx = h->thread_context[context_count - 1];
3575         s->mb_x = hx->s.mb_x;
3576         s->mb_y = hx->s.mb_y;
3577         s->dropable = hx->s.dropable;
3578         s->picture_structure = hx->s.picture_structure;
3579         for(i = 1; i < context_count; i++)
3580             h->s.error_count += h->thread_context[i]->s.error_count;
3581     }
3582
3583     return 0;
3584 }
3585
3586
3587 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3588     MpegEncContext * const s = &h->s;
3589     AVCodecContext * const avctx= s->avctx;
3590     H264Context *hx; ///< thread context
3591     int buf_index;
3592     int context_count;
3593     int next_avc;
3594     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3595     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3596     int nal_index;
3597
3598     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3599     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3600         h->current_slice = 0;
3601         if (!s->first_field)
3602             s->current_picture_ptr= NULL;
3603         ff_h264_reset_sei(h);
3604     }
3605
3606     for(;pass <= 1;pass++){
3607         buf_index = 0;
3608         context_count = 0;
3609         next_avc = h->is_avc ? 0 : buf_size;
3610         nal_index = 0;
3611     for(;;){
3612         int consumed;
3613         int dst_length;
3614         int bit_length;
3615         const uint8_t *ptr;
3616         int i, nalsize = 0;
3617         int err;
3618
3619         if(buf_index >= next_avc) {
3620             if(buf_index >= buf_size) break;
3621             nalsize = 0;
3622             for(i = 0; i < h->nal_length_size; i++)
3623                 nalsize = (nalsize << 8) | buf[buf_index++];
3624             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3625                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3626                 break;
3627             }
3628             next_avc= buf_index + nalsize;
3629         } else {
3630             // start code prefix search
3631             for(; buf_index + 3 < next_avc; buf_index++){
3632                 // This should always succeed in the first iteration.
3633                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3634                     break;
3635             }
3636
3637             if(buf_index+3 >= buf_size) break;
3638
3639             buf_index+=3;
3640             if(buf_index >= next_avc) continue;
3641         }
3642
3643         hx = h->thread_context[context_count];
3644
3645         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3646         if (ptr==NULL || dst_length < 0){
3647             return -1;
3648         }
3649         i= buf_index + consumed;
3650         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3651            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3652             s->workaround_bugs |= FF_BUG_TRUNCATED;
3653
3654         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3655         while(dst_length > 0 && ptr[dst_length - 1] == 0)
3656             dst_length--;
3657         }
3658         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3659
3660         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3661             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length);
3662         }
3663
3664         if (h->is_avc && (nalsize != consumed) && nalsize){
3665             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3666         }
3667
3668         buf_index += consumed;
3669         nal_index++;
3670
3671         if(pass == 0) {
3672             // packets can sometimes contain multiple PPS/SPS
3673             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3674             // if so, when frame threading we can't start the next thread until we've read all of them
3675             switch (hx->nal_unit_type) {
3676                 case NAL_SPS:
3677                 case NAL_PPS:
3678                 case NAL_IDR_SLICE:
3679                 case NAL_SLICE:
3680                     nals_needed = nal_index;
3681             }
3682             continue;
3683         }
3684
3685         //FIXME do not discard SEI id
3686         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3687             continue;
3688
3689       again:
3690         err = 0;
3691         switch(hx->nal_unit_type){
3692         case NAL_IDR_SLICE:
3693             if (h->nal_unit_type != NAL_IDR_SLICE) {
3694                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3695                 return -1;
3696             }
3697             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3698         case NAL_SLICE:
3699             init_get_bits(&hx->s.gb, ptr, bit_length);
3700             hx->intra_gb_ptr=
3701             hx->inter_gb_ptr= &hx->s.gb;
3702             hx->s.data_partitioning = 0;
3703
3704             if((err = decode_slice_header(hx, h)))
3705                break;
3706
3707             s->current_picture_ptr->f.key_frame |=
3708                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3709                     (h->sei_recovery_frame_cnt >= 0);
3710
3711             if (h->current_slice == 1) {
3712                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3713                     decode_postinit(h, nal_index >= nals_needed);
3714                 }
3715
3716                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3717                     return -1;
3718                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3719                     ff_vdpau_h264_picture_start(s);
3720             }
3721
3722             if(hx->redundant_pic_count==0
3723                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3724                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3725                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3726                && avctx->skip_frame < AVDISCARD_ALL){
3727                 if(avctx->hwaccel) {
3728                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3729                         return -1;
3730                 }else
3731                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3732                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3733                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3734                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3735                 }else
3736                     context_count++;
3737             }
3738             break;
3739         case NAL_DPA:
3740             init_get_bits(&hx->s.gb, ptr, bit_length);
3741             hx->intra_gb_ptr=
3742             hx->inter_gb_ptr= NULL;
3743
3744             if ((err = decode_slice_header(hx, h)) < 0)
3745                 break;
3746
3747             hx->s.data_partitioning = 1;
3748
3749             break;
3750         case NAL_DPB:
3751             init_get_bits(&hx->intra_gb, ptr, bit_length);
3752             hx->intra_gb_ptr= &hx->intra_gb;
3753             break;
3754         case NAL_DPC:
3755             init_get_bits(&hx->inter_gb, ptr, bit_length);
3756             hx->inter_gb_ptr= &hx->inter_gb;
3757
3758             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3759                && s->context_initialized
3760                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3761                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3762                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3763                && avctx->skip_frame < AVDISCARD_ALL)
3764                 context_count++;
3765             break;
3766         case NAL_SEI:
3767             init_get_bits(&s->gb, ptr, bit_length);
3768             ff_h264_decode_sei(h);
3769             break;
3770         case NAL_SPS:
3771             init_get_bits(&s->gb, ptr, bit_length);
3772             ff_h264_decode_seq_parameter_set(h);
3773
3774             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3775                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3776                 s->low_delay=1;
3777
3778             if(avctx->has_b_frames < 2)
3779                 avctx->has_b_frames= !s->low_delay;
3780
3781             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3782                 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3783                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3784                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3785                     h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3786                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3787
3788                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3789                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3790                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3791                     dsputil_init(&s->dsp, s->avctx);
3792                 } else {
3793                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3794                     return -1;
3795                 }
3796             }
3797             break;
3798         case NAL_PPS:
3799             init_get_bits(&s->gb, ptr, bit_length);
3800
3801             ff_h264_decode_picture_parameter_set(h, bit_length);
3802
3803             break;
3804         case NAL_AUD:
3805         case NAL_END_SEQUENCE:
3806         case NAL_END_STREAM:
3807         case NAL_FILLER_DATA:
3808         case NAL_SPS_EXT:
3809         case NAL_AUXILIARY_SLICE:
3810             break;
3811         default:
3812             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3813         }
3814
3815         if(context_count == h->max_contexts) {
3816             execute_decode_slices(h, context_count);
3817             context_count = 0;
3818         }
3819
3820         if (err < 0)
3821             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3822         else if(err == 1) {
3823             /* Slice could not be decoded in parallel mode, copy down
3824              * NAL unit stuff to context 0 and restart. Note that
3825              * rbsp_buffer is not transferred, but since we no longer
3826              * run in parallel mode this should not be an issue. */
3827             h->nal_unit_type = hx->nal_unit_type;
3828             h->nal_ref_idc   = hx->nal_ref_idc;
3829             hx = h;
3830             goto again;
3831         }
3832     }
3833     }
3834     if(context_count)
3835         execute_decode_slices(h, context_count);
3836     return buf_index;
3837 }
3838
3839 /**
3840  * returns the number of bytes consumed for building the current frame
3841  */
3842 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3843         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3844         if(pos+10>buf_size) pos=buf_size; // oops ;)
3845
3846         return pos;
3847 }
3848
3849 static int decode_frame(AVCodecContext *avctx,
3850                              void *data, int *data_size,
3851                              AVPacket *avpkt)
3852 {
3853     const uint8_t *buf = avpkt->data;
3854     int buf_size = avpkt->size;
3855     H264Context *h = avctx->priv_data;
3856     MpegEncContext *s = &h->s;
3857     AVFrame *pict = data;
3858     int buf_index;
3859
3860     s->flags= avctx->flags;
3861     s->flags2= avctx->flags2;
3862
3863    /* end of stream, output what is still in the buffers */
3864  out:
3865     if (buf_size == 0) {
3866         Picture *out;
3867         int i, out_idx;
3868
3869         s->current_picture_ptr = NULL;
3870
3871 //FIXME factorize this with the output code below
3872         out = h->delayed_pic[0];
3873         out_idx = 0;
3874         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3875             if(h->delayed_pic[i]->poc < out->poc){
3876                 out = h->delayed_pic[i];
3877                 out_idx = i;
3878             }
3879
3880         for(i=out_idx; h->delayed_pic[i]; i++)
3881             h->delayed_pic[i] = h->delayed_pic[i+1];
3882
3883         if(out){
3884             *data_size = sizeof(AVFrame);
3885             *pict= *(AVFrame*)out;
3886         }
3887
3888         return 0;
3889     }
3890
3891     buf_index=decode_nal_units(h, buf, buf_size);
3892     if(buf_index < 0)
3893         return -1;
3894
3895     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3896         buf_size = 0;
3897         goto out;
3898     }
3899
3900     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3901         if (avctx->skip_frame >= AVDISCARD_NONREF)
3902             return 0;
3903         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3904         return -1;
3905     }
3906
3907     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3908
3909         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3910
3911         field_end(h, 0);
3912
3913         if (!h->next_output_pic) {
3914             /* Wait for second field. */
3915             *data_size = 0;
3916
3917         } else {
3918             *data_size = sizeof(AVFrame);
3919             *pict = *(AVFrame*)h->next_output_pic;
3920         }
3921     }
3922
3923     assert(pict->data[0] || !*data_size);
3924     ff_print_debug_info(s, pict);
3925 //printf("out %d\n", (int)pict->data[0]);
3926
3927     return get_consumed_bytes(s, buf_index, buf_size);
3928 }
3929 #if 0
3930 static inline void fill_mb_avail(H264Context *h){
3931     MpegEncContext * const s = &h->s;
3932     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3933
3934     if(s->mb_y){
3935         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3936         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3937         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3938     }else{
3939         h->mb_avail[0]=
3940         h->mb_avail[1]=
3941         h->mb_avail[2]= 0;
3942     }
3943     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3944     h->mb_avail[4]= 1; //FIXME move out
3945     h->mb_avail[5]= 0; //FIXME move out
3946 }
3947 #endif
3948
3949 #ifdef TEST
3950 #undef printf
3951 #undef random
3952 #define COUNT 8000
3953 #define SIZE (COUNT*40)
3954 extern AVCodec ff_h264_decoder;
3955 int main(void){
3956     int i;
3957     uint8_t temp[SIZE];
3958     PutBitContext pb;
3959     GetBitContext gb;
3960 //    int int_temp[10000];
3961     DSPContext dsp;
3962     AVCodecContext avctx;
3963
3964     avcodec_get_context_defaults3(&avctx, &ff_h264_decoder);
3965
3966     dsputil_init(&dsp, &avctx);
3967
3968     init_put_bits(&pb, temp, SIZE);
3969     printf("testing unsigned exp golomb\n");
3970     for(i=0; i<COUNT; i++){
3971         START_TIMER
3972         set_ue_golomb(&pb, i);
3973         STOP_TIMER("set_ue_golomb");
3974     }
3975     flush_put_bits(&pb);
3976
3977     init_get_bits(&gb, temp, 8*SIZE);
3978     for(i=0; i<COUNT; i++){
3979         int j, s;
3980
3981         s= show_bits(&gb, 24);
3982
3983         START_TIMER
3984         j= get_ue_golomb(&gb);
3985         if(j != i){
3986             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3987 //            return -1;
3988         }
3989         STOP_TIMER("get_ue_golomb");
3990     }
3991
3992
3993     init_put_bits(&pb, temp, SIZE);
3994     printf("testing signed exp golomb\n");
3995     for(i=0; i<COUNT; i++){
3996         START_TIMER
3997         set_se_golomb(&pb, i - COUNT/2);
3998         STOP_TIMER("set_se_golomb");
3999     }
4000     flush_put_bits(&pb);
4001
4002     init_get_bits(&gb, temp, 8*SIZE);
4003     for(i=0; i<COUNT; i++){
4004         int j, s;
4005
4006         s= show_bits(&gb, 24);
4007
4008         START_TIMER
4009         j= get_se_golomb(&gb);
4010         if(j != i - COUNT/2){
4011             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4012 //            return -1;
4013         }
4014         STOP_TIMER("get_se_golomb");
4015     }
4016
4017     printf("Testing RBSP\n");
4018
4019
4020     return 0;
4021 }
4022 #endif /* TEST */
4023
4024
4025 av_cold void ff_h264_free_context(H264Context *h)
4026 {
4027     int i;
4028
4029     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4030
4031     for(i = 0; i < MAX_SPS_COUNT; i++)
4032         av_freep(h->sps_buffers + i);
4033
4034     for(i = 0; i < MAX_PPS_COUNT; i++)
4035         av_freep(h->pps_buffers + i);
4036 }
4037
4038 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4039 {
4040     H264Context *h = avctx->priv_data;
4041     MpegEncContext *s = &h->s;
4042
4043     ff_h264_free_context(h);
4044
4045     MPV_common_end(s);
4046
4047 //    memset(h, 0, sizeof(H264Context));
4048
4049     return 0;
4050 }
4051
4052 static const AVProfile profiles[] = {
4053     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4054     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4055     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4056     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4057     { FF_PROFILE_H264_HIGH,                 "High"                  },
4058     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4059     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4060     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4061     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4062     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4063     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4064     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4065     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4066     { FF_PROFILE_UNKNOWN },
4067 };
4068
4069 AVCodec ff_h264_decoder = {
4070     .name           = "h264",
4071     .type           = AVMEDIA_TYPE_VIDEO,
4072     .id             = CODEC_ID_H264,
4073     .priv_data_size = sizeof(H264Context),
4074     .init           = ff_h264_decode_init,
4075     .close          = ff_h264_decode_end,
4076     .decode         = decode_frame,
4077     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4078                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4079     .flush= flush_dpb,
4080     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4081     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4082     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4083     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4084 };
4085
4086 #if CONFIG_H264_VDPAU_DECODER
4087 AVCodec ff_h264_vdpau_decoder = {
4088     .name           = "h264_vdpau",
4089     .type           = AVMEDIA_TYPE_VIDEO,
4090     .id             = CODEC_ID_H264,
4091     .priv_data_size = sizeof(H264Context),
4092     .init           = ff_h264_decode_init,
4093     .close          = ff_h264_decode_end,
4094     .decode         = decode_frame,
4095     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4096     .flush= flush_dpb,
4097     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4098     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4099     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4100 };
4101 #endif