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 Libav.
   6  *
   7  * Libav 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  * Libav 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 Libav; 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_VDA_VLD,
  60     PIX_FMT_YUVJ420P,
  61     PIX_FMT_NONE
  62 };
  63
  64 /**
  65  * Check if the top & left blocks are available if needed and
  66  * change the dc mode so it only uses the available blocks.
  67  */
  68 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  69     MpegEncContext * const s = &h->s;
  70     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  71     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  72     int i;
  73
  74     if(!(h->top_samples_available&0x8000)){
  75         for(i=0; i<4; i++){
  76             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  77             if(status<0){
  78                 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);
  79                 return -1;
  80             } else if(status){
  81                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  82             }
  83         }
  84     }
  85
  86     if((h->left_samples_available&0x8888)!=0x8888){
  87         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  88         for(i=0; i<4; i++){
  89             if(!(h->left_samples_available&mask[i])){
  90                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  91                 if(status<0){
  92                     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);
  93                     return -1;
  94                 } else if(status){
  95                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  96                 }
  97             }
  98         }
  99     }
 100
 101     return 0;
 102 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 103
 104 /**
 105  * Check if the top & left blocks are available if needed and
 106  * change the dc mode so it only uses the available blocks.
 107  */
 108 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 109     MpegEncContext * const s = &h->s;
 110     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 111     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 112
 113     if(mode > 6U) {
 114         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);
 115         return -1;
 116     }
 117
 118     if(!(h->top_samples_available&0x8000)){
 119         mode= top[ mode ];
 120         if(mode<0){
 121             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);
 122             return -1;
 123         }
 124     }
 125
 126     if((h->left_samples_available&0x8080) != 0x8080){
 127         mode= left[ mode ];
 128         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 129             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 130         }
 131         if(mode<0){
 132             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);
 133             return -1;
 134         }
 135     }
 136
 137     return mode;
 138 }
 139
 140 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src,
 141                                   int *dst_length, int *consumed, int length,
 142                                   int nalsize_known)
 143 {
 144     int i, si, di;
 145     uint8_t *dst;
 146     int bufidx;
 147
 148 //    src[0]&0x80;                //forbidden bit
 149     h->nal_ref_idc= src[0]>>5;
 150     h->nal_unit_type= src[0]&0x1F;
 151
 152     src++; length--;
 153
 154     if (nalsize_known) {
 155         i = length;
 156     } else
 157 #if HAVE_FAST_UNALIGNED
 158 # if HAVE_FAST_64BIT
 159 #   define RS 7
 160     for(i=0; i+1<length; i+=9){
 161         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 162 # else
 163 #   define RS 3
 164     for(i=0; i+1<length; i+=5){
 165         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 166 # endif
 167             continue;
 168         if(i>0 && !src[i]) i--;
 169         while(src[i]) i++;
 170 #else
 171 #   define RS 0
 172     for(i=0; i+1<length; i+=2){
 173         if(src[i]) continue;
 174         if(i>0 && src[i-1]==0) i--;
 175 #endif
 176         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 177             if(src[i+2]!=3){
 178                 /* startcode, so we must be past the end */
 179                 length=i;
 180             }
 181             break;
 182         }
 183         i-= RS;
 184     }
 185
 186     if(i>=length-1){ //no escaped 0
 187         *dst_length= length;
 188         *consumed= length+1; //+1 for the header
 189         return src;
 190     }
 191
 192     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 193     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 194     dst= h->rbsp_buffer[bufidx];
 195
 196     if (dst == NULL){
 197         return NULL;
 198     }
 199
 200 //printf("decoding esc\n");
 201     memcpy(dst, src, i);
 202     si=di=i;
 203     while(si+2<length){
 204         //remove escapes (very rare 1:2^22)
 205         if(src[si+2]>3){
 206             dst[di++]= src[si++];
 207             dst[di++]= src[si++];
 208         }else if(src[si]==0 && src[si+1]==0){
 209             if(src[si+2]==3){ //escape
 210                 dst[di++]= 0;
 211                 dst[di++]= 0;
 212                 si+=3;
 213                 continue;
 214             }else //next start code
 215                 goto nsc;
 216         }
 217
 218         dst[di++]= src[si++];
 219     }
 220     while(si<length)
 221         dst[di++]= src[si++];
 222 nsc:
 223
 224     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 225
 226     *dst_length= di;
 227     *consumed= si + 1;//+1 for the header
 228 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 229     return dst;
 230 }
 231
 232 /**
 233  * Identify the exact end of the bitstream
 234  * @return the length of the trailing, or 0 if damaged
 235  */
 236 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 237     int v= *src;
 238     int r;
 239
 240     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 241
 242     for(r=1; r<9; r++){
 243         if(v&1) return r;
 244         v>>=1;
 245     }
 246     return 0;
 247 }
 248
 249 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 250                                  int y_offset, int list){
 251     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 252     int filter_height= (raw_my&3) ? 2 : 0;
 253     int full_my= (raw_my>>2) + y_offset;
 254     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 255
 256     return FFMAX(abs(top), bottom);
 257 }
 258
 259 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 260                                int y_offset, int list0, int list1, int *nrefs){
 261     MpegEncContext * const s = &h->s;
 262     int my;
 263
 264     y_offset += 16*(s->mb_y >> MB_FIELD);
 265
 266     if(list0){
 267         int ref_n = h->ref_cache[0][ scan8[n] ];
 268         Picture *ref= &h->ref_list[0][ref_n];
 269
 270         // Error resilience puts the current picture in the ref list.
 271         // Don't try to wait on these as it will cause a deadlock.
 272         // Fields can wait on each other, though.
 273         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 274            (ref->f.reference & 3) != s->picture_structure) {
 275             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 276             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 277             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 278         }
 279     }
 280
 281     if(list1){
 282         int ref_n = h->ref_cache[1][ scan8[n] ];
 283         Picture *ref= &h->ref_list[1][ref_n];
 284
 285         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 286            (ref->f.reference & 3) != s->picture_structure) {
 287             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 288             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 289             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 290         }
 291     }
 292 }
 293
 294 /**
 295  * Wait until all reference frames are available for MC operations.
 296  *
 297  * @param h the H264 context
 298  */
 299 static void await_references(H264Context *h){
 300     MpegEncContext * const s = &h->s;
 301     const int mb_xy= h->mb_xy;
 302     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 303     int refs[2][48];
 304     int nrefs[2] = {0};
 305     int ref, list;
 306
 307     memset(refs, -1, sizeof(refs));
 308
 309     if(IS_16X16(mb_type)){
 310         get_lowest_part_y(h, refs, 0, 16, 0,
 311                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 312     }else if(IS_16X8(mb_type)){
 313         get_lowest_part_y(h, refs, 0, 8, 0,
 314                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 315         get_lowest_part_y(h, refs, 8, 8, 8,
 316                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 317     }else if(IS_8X16(mb_type)){
 318         get_lowest_part_y(h, refs, 0, 16, 0,
 319                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 320         get_lowest_part_y(h, refs, 4, 16, 0,
 321                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 322     }else{
 323         int i;
 324
 325         assert(IS_8X8(mb_type));
 326
 327         for(i=0; i<4; i++){
 328             const int sub_mb_type= h->sub_mb_type[i];
 329             const int n= 4*i;
 330             int y_offset= (i&2)<<2;
 331
 332             if(IS_SUB_8X8(sub_mb_type)){
 333                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 334                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 335             }else if(IS_SUB_8X4(sub_mb_type)){
 336                 get_lowest_part_y(h, refs, n  , 4, 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+2, 4, y_offset+4,
 339                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 340             }else if(IS_SUB_4X8(sub_mb_type)){
 341                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 342                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 343                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 344                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 345             }else{
 346                 int j;
 347                 assert(IS_SUB_4X4(sub_mb_type));
 348                 for(j=0; j<4; j++){
 349                     int sub_y_offset= y_offset + 2*(j&2);
 350                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 351                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 352                 }
 353             }
 354         }
 355     }
 356
 357     for(list=h->list_count-1; list>=0; list--){
 358         for(ref=0; ref<48 && nrefs[list]; ref++){
 359             int row = refs[list][ref];
 360             if(row >= 0){
 361                 Picture *ref_pic = &h->ref_list[list][ref];
 362                 int ref_field = ref_pic->f.reference - 1;
 363                 int ref_field_picture = ref_pic->field_picture;
 364                 int pic_height = 16*s->mb_height >> ref_field_picture;
 365
 366                 row <<= MB_MBAFF;
 367                 nrefs[list]--;
 368
 369                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 370                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 371                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 372                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 373                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 374                 }else if(FIELD_PICTURE){
 375                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 376                 }else{
 377                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 378                 }
 379             }
 380         }
 381     }
 382 }
 383
 384 #if 0
 385 /**
 386  * DCT transforms the 16 dc values.
 387  * @param qp quantization parameter ??? FIXME
 388  */
 389 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 390 //    const int qmul= dequant_coeff[qp][0];
 391     int i;
 392     int temp[16]; //FIXME check if this is a good idea
 393     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 394     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 395
 396     for(i=0; i<4; i++){
 397         const int offset= y_offset[i];
 398         const int z0= block[offset+stride*0] + block[offset+stride*4];
 399         const int z1= block[offset+stride*0] - block[offset+stride*4];
 400         const int z2= block[offset+stride*1] - block[offset+stride*5];
 401         const int z3= block[offset+stride*1] + block[offset+stride*5];
 402
 403         temp[4*i+0]= z0+z3;
 404         temp[4*i+1]= z1+z2;
 405         temp[4*i+2]= z1-z2;
 406         temp[4*i+3]= z0-z3;
 407     }
 408
 409     for(i=0; i<4; i++){
 410         const int offset= x_offset[i];
 411         const int z0= temp[4*0+i] + temp[4*2+i];
 412         const int z1= temp[4*0+i] - temp[4*2+i];
 413         const int z2= temp[4*1+i] - temp[4*3+i];
 414         const int z3= temp[4*1+i] + temp[4*3+i];
 415
 416         block[stride*0 +offset]= (z0 + z3)>>1;
 417         block[stride*2 +offset]= (z1 + z2)>>1;
 418         block[stride*8 +offset]= (z1 - z2)>>1;
 419         block[stride*10+offset]= (z0 - z3)>>1;
 420     }
 421 }
 422 #endif
 423
 424 #undef xStride
 425 #undef stride
 426
 427 #if 0
 428 static void chroma_dc_dct_c(DCTELEM *block){
 429     const int stride= 16*2;
 430     const int xStride= 16;
 431     int a,b,c,d,e;
 432
 433     a= block[stride*0 + xStride*0];
 434     b= block[stride*0 + xStride*1];
 435     c= block[stride*1 + xStride*0];
 436     d= block[stride*1 + xStride*1];
 437
 438     e= a-b;
 439     a= a+b;
 440     b= c-d;
 441     c= c+d;
 442
 443     block[stride*0 + xStride*0]= (a+c);
 444     block[stride*0 + xStride*1]= (e+b);
 445     block[stride*1 + xStride*0]= (a-c);
 446     block[stride*1 + xStride*1]= (e-b);
 447 }
 448 #endif
 449
 450 static av_always_inline void
 451 mc_dir_part(H264Context *h, Picture *pic, int n, int square,
 452             int height, int delta, int list,
 453             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 454             int src_x_offset, int src_y_offset,
 455             qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 456             int pixel_shift, int chroma_idc)
 457 {
 458     MpegEncContext * const s = &h->s;
 459     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 460     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 461     const int luma_xy= (mx&3) + ((my&3)<<2);
 462     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 463     uint8_t * src_y = pic->f.data[0] + offset;
 464     uint8_t * src_cb, * src_cr;
 465     int extra_width= h->emu_edge_width;
 466     int extra_height= h->emu_edge_height;
 467     int emu=0;
 468     const int full_mx= mx>>2;
 469     const int full_my= my>>2;
 470     const int pic_width  = 16*s->mb_width;
 471     const int pic_height = 16*s->mb_height >> MB_FIELD;
 472     int ysh;
 473
 474     if(mx&7) extra_width -= 3;
 475     if(my&7) extra_height -= 3;
 476
 477     if(   full_mx < 0-extra_width
 478        || full_my < 0-extra_height
 479        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 480        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 481         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 482                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 483             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 484         emu=1;
 485     }
 486
 487     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 488     if(!square){
 489         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 490     }
 491
 492     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 493
 494     if(chroma_idc == 3 /* yuv444 */){
 495         src_cb = pic->f.data[1] + offset;
 496         if(emu){
 497             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 498                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 499             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 500         }
 501         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 502         if(!square){
 503             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 504         }
 505
 506         src_cr = pic->f.data[2] + offset;
 507         if(emu){
 508             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 509                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 510             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 511         }
 512         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 513         if(!square){
 514             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 515         }
 516         return;
 517     }
 518
 519     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 520     if(chroma_idc == 1 /* yuv420 */ && MB_FIELD){
 521         // chroma offset when predicting from a field of opposite parity
 522         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 523         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 524     }
 525
 526     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 527     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 528
 529     if(emu){
 530         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 531                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 532                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 533             src_cb= s->edge_emu_buffer;
 534     }
 535     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 536               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 537
 538     if(emu){
 539         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 540                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 541                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 542             src_cr= s->edge_emu_buffer;
 543     }
 544     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 545               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 546 }
 547
 548 static av_always_inline void
 549 mc_part_std(H264Context *h, int n, int square, int height, int delta,
 550             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 551             int x_offset, int y_offset,
 552             qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 553             qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 554             int list0, int list1, int pixel_shift, int chroma_idc)
 555 {
 556     MpegEncContext * const s = &h->s;
 557     qpel_mc_func *qpix_op=  qpix_put;
 558     h264_chroma_mc_func chroma_op= chroma_put;
 559
 560     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 561     if (chroma_idc == 3 /* yuv444 */) {
 562         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 563         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 564     } else if (chroma_idc == 2 /* yuv422 */) {
 565         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 566         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 567     } else /* yuv420 */ {
 568         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 569         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 570     }
 571     x_offset += 8*s->mb_x;
 572     y_offset += 8*(s->mb_y >> MB_FIELD);
 573
 574     if(list0){
 575         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 576         mc_dir_part(h, ref, n, square, height, delta, 0,
 577                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 578                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 579
 580         qpix_op=  qpix_avg;
 581         chroma_op= chroma_avg;
 582     }
 583
 584     if(list1){
 585         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 586         mc_dir_part(h, ref, n, square, height, delta, 1,
 587                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 588                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 589     }
 590 }
 591
 592 static av_always_inline void
 593 mc_part_weighted(H264Context *h, int n, int square, int height, int delta,
 594                  uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 595                  int x_offset, int y_offset,
 596                  qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 597                  h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 598                  h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 599                  int list0, int list1, int pixel_shift, int chroma_idc){
 600     MpegEncContext * const s = &h->s;
 601     int chroma_height;
 602
 603     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 604     if (chroma_idc == 3 /* yuv444 */) {
 605         chroma_height = height;
 606         chroma_weight_avg = luma_weight_avg;
 607         chroma_weight_op = luma_weight_op;
 608         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 609         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 610     } else if (chroma_idc == 2 /* yuv422 */) {
 611         chroma_height = height;
 612         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 613         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 614     } else /* yuv420 */ {
 615         chroma_height = height >> 1;
 616         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 617         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 618     }
 619     x_offset += 8*s->mb_x;
 620     y_offset += 8*(s->mb_y >> MB_FIELD);
 621
 622     if(list0 && list1){
 623         /* don't optimize for luma-only case, since B-frames usually
 624          * use implicit weights => chroma too. */
 625         uint8_t *tmp_cb = s->obmc_scratchpad;
 626         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 627         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 628         int refn0 = h->ref_cache[0][ scan8[n] ];
 629         int refn1 = h->ref_cache[1][ scan8[n] ];
 630
 631         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 632                     dest_y, dest_cb, dest_cr,
 633                     x_offset, y_offset, qpix_put, chroma_put,
 634                     pixel_shift, chroma_idc);
 635         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 636                     tmp_y, tmp_cb, tmp_cr,
 637                     x_offset, y_offset, qpix_put, chroma_put,
 638                     pixel_shift, chroma_idc);
 639
 640         if(h->use_weight == 2){
 641             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 642             int weight1 = 64 - weight0;
 643             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize,
 644                               height,        5, weight0, weight1, 0);
 645             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 646                               chroma_height, 5, weight0, weight1, 0);
 647             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 648                               chroma_height, 5, weight0, weight1, 0);
 649         }else{
 650             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 651                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 652                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 653             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 654                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 655                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 656             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 657                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 658                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 659         }
 660     }else{
 661         int list = list1 ? 1 : 0;
 662         int refn = h->ref_cache[list][ scan8[n] ];
 663         Picture *ref= &h->ref_list[list][refn];
 664         mc_dir_part(h, ref, n, square, height, delta, list,
 665                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 666                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 667
 668         luma_weight_op(dest_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 669                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 670         if(h->use_weight_chroma){
 671             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 672                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 673             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 674                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 675         }
 676     }
 677 }
 678
 679 static av_always_inline void
 680 mc_part(H264Context *h, int n, int square, int height, int delta,
 681         uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 682         int x_offset, int y_offset,
 683         qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 684         qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 685         h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 686         int list0, int list1, int pixel_shift, int chroma_idc)
 687 {
 688     if((h->use_weight==2 && list0 && list1
 689         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 690        || h->use_weight==1)
 691         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 692                          x_offset, y_offset, qpix_put, chroma_put,
 693                          weight_op[0], weight_op[1], weight_avg[0],
 694                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
 695     else
 696         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 697                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 698                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
 699 }
 700
 701 static av_always_inline void
 702 prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma_idc)
 703 {
 704     /* fetch pixels for estimated mv 4 macroblocks ahead
 705      * optimized for 64byte cache lines */
 706     MpegEncContext * const s = &h->s;
 707     const int refn = h->ref_cache[list][scan8[0]];
 708     if(refn >= 0){
 709         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 710         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 711         uint8_t **src = h->ref_list[list][refn].f.data;
 712         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 713         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 714         if (chroma_idc == 3 /* yuv444 */) {
 715             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 716             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 717         }else{
 718             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
 719             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 720         }
 721     }
 722 }
 723
 724 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 725                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 726                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 727                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 728                       int pixel_shift, int chroma_idc)
 729 {
 730     MpegEncContext * const s = &h->s;
 731     const int mb_xy= h->mb_xy;
 732     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 733
 734     assert(IS_INTER(mb_type));
 735
 736     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 737         await_references(h);
 738     prefetch_motion(h, 0, pixel_shift, chroma_idc);
 739
 740     if(IS_16X16(mb_type)){
 741         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
 742                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 743                 weight_op, weight_avg,
 744                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 745                 pixel_shift, chroma_idc);
 746     }else if(IS_16X8(mb_type)){
 747         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 748                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 749                 weight_op, weight_avg,
 750                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 751                 pixel_shift, chroma_idc);
 752         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 753                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 754                 weight_op, weight_avg,
 755                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 756                 pixel_shift, chroma_idc);
 757     }else if(IS_8X16(mb_type)){
 758         mc_part(h, 0, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 759                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 760                 &weight_op[1], &weight_avg[1],
 761                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 762                 pixel_shift, chroma_idc);
 763         mc_part(h, 4, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 764                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 765                 &weight_op[1], &weight_avg[1],
 766                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 767                 pixel_shift, chroma_idc);
 768     }else{
 769         int i;
 770
 771         assert(IS_8X8(mb_type));
 772
 773         for(i=0; i<4; i++){
 774             const int sub_mb_type= h->sub_mb_type[i];
 775             const int n= 4*i;
 776             int x_offset= (i&1)<<2;
 777             int y_offset= (i&2)<<1;
 778
 779             if(IS_SUB_8X8(sub_mb_type)){
 780                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 781                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 782                     &weight_op[1], &weight_avg[1],
 783                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 784                     pixel_shift, chroma_idc);
 785             }else if(IS_SUB_8X4(sub_mb_type)){
 786                 mc_part(h, n  , 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 787                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 788                     &weight_op[1], &weight_avg[1],
 789                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 790                     pixel_shift, chroma_idc);
 791                 mc_part(h, n+2, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 792                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 793                     &weight_op[1], &weight_avg[1],
 794                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 795                     pixel_shift, chroma_idc);
 796             }else if(IS_SUB_4X8(sub_mb_type)){
 797                 mc_part(h, n  , 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 798                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 799                     &weight_op[2], &weight_avg[2],
 800                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 801                     pixel_shift, chroma_idc);
 802                 mc_part(h, n+1, 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 803                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 804                     &weight_op[2], &weight_avg[2],
 805                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 806                     pixel_shift, chroma_idc);
 807             }else{
 808                 int j;
 809                 assert(IS_SUB_4X4(sub_mb_type));
 810                 for(j=0; j<4; j++){
 811                     int sub_x_offset= x_offset + 2*(j&1);
 812                     int sub_y_offset= y_offset +   (j&2);
 813                     mc_part(h, n+j, 1, 4, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 814                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 815                         &weight_op[2], &weight_avg[2],
 816                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 817                         pixel_shift, chroma_idc);
 818                 }
 819             }
 820         }
 821     }
 822
 823     prefetch_motion(h, 1, pixel_shift, chroma_idc);
 824 }
 825
 826 static av_always_inline void
 827 hl_motion_420(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 828               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 829               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 830               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 831               int pixel_shift)
 832 {
 833     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 834               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
 835 }
 836
 837 static av_always_inline void
 838 hl_motion_422(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 839               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 840               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 841               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 842               int pixel_shift)
 843 {
 844     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 845               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
 846 }
 847
 848 static void free_tables(H264Context *h, int free_rbsp){
 849     int i;
 850     H264Context *hx;
 851
 852     av_freep(&h->intra4x4_pred_mode);
 853     av_freep(&h->chroma_pred_mode_table);
 854     av_freep(&h->cbp_table);
 855     av_freep(&h->mvd_table[0]);
 856     av_freep(&h->mvd_table[1]);
 857     av_freep(&h->direct_table);
 858     av_freep(&h->non_zero_count);
 859     av_freep(&h->slice_table_base);
 860     h->slice_table= NULL;
 861     av_freep(&h->list_counts);
 862
 863     av_freep(&h->mb2b_xy);
 864     av_freep(&h->mb2br_xy);
 865
 866     for(i = 0; i < MAX_THREADS; i++) {
 867         hx = h->thread_context[i];
 868         if(!hx) continue;
 869         av_freep(&hx->top_borders[1]);
 870         av_freep(&hx->top_borders[0]);
 871         av_freep(&hx->s.obmc_scratchpad);
 872         if (free_rbsp){
 873             av_freep(&hx->rbsp_buffer[1]);
 874             av_freep(&hx->rbsp_buffer[0]);
 875             hx->rbsp_buffer_size[0] = 0;
 876             hx->rbsp_buffer_size[1] = 0;
 877         }
 878         if (i) av_freep(&h->thread_context[i]);
 879     }
 880 }
 881
 882 static void init_dequant8_coeff_table(H264Context *h){
 883     int i,j,q,x;
 884     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 885
 886     for(i=0; i<6; i++ ){
 887         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 888         for(j=0; j<i; j++){
 889             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 890                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 891                 break;
 892             }
 893         }
 894         if(j<i)
 895             continue;
 896
 897         for(q=0; q<max_qp+1; q++){
 898             int shift = div6[q];
 899             int idx = rem6[q];
 900             for(x=0; x<64; x++)
 901                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 902                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 903                     h->pps.scaling_matrix8[i][x]) << shift;
 904         }
 905     }
 906 }
 907
 908 static void init_dequant4_coeff_table(H264Context *h){
 909     int i,j,q,x;
 910     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 911     for(i=0; i<6; i++ ){
 912         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 913         for(j=0; j<i; j++){
 914             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 915                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 916                 break;
 917             }
 918         }
 919         if(j<i)
 920             continue;
 921
 922         for(q=0; q<max_qp+1; q++){
 923             int shift = div6[q] + 2;
 924             int idx = rem6[q];
 925             for(x=0; x<16; x++)
 926                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 927                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 928                     h->pps.scaling_matrix4[i][x]) << shift;
 929         }
 930     }
 931 }
 932
 933 static void init_dequant_tables(H264Context *h){
 934     int i,x;
 935     init_dequant4_coeff_table(h);
 936     if(h->pps.transform_8x8_mode)
 937         init_dequant8_coeff_table(h);
 938     if(h->sps.transform_bypass){
 939         for(i=0; i<6; i++)
 940             for(x=0; x<16; x++)
 941                 h->dequant4_coeff[i][0][x] = 1<<6;
 942         if(h->pps.transform_8x8_mode)
 943             for(i=0; i<6; i++)
 944                 for(x=0; x<64; x++)
 945                     h->dequant8_coeff[i][0][x] = 1<<6;
 946     }
 947 }
 948
 949
 950 int ff_h264_alloc_tables(H264Context *h){
 951     MpegEncContext * const s = &h->s;
 952     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 953     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 954     int x,y;
 955
 956     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 957
 958     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 959     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 960     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 961
 962     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 963     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 964     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 965     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 966     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 967
 968     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 969     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 970
 971     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 972     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 973     for(y=0; y<s->mb_height; y++){
 974         for(x=0; x<s->mb_width; x++){
 975             const int mb_xy= x + y*s->mb_stride;
 976             const int b_xy = 4*x + 4*y*h->b_stride;
 977
 978             h->mb2b_xy [mb_xy]= b_xy;
 979             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 980         }
 981     }
 982
 983     s->obmc_scratchpad = NULL;
 984
 985     if(!h->dequant4_coeff[0])
 986         init_dequant_tables(h);
 987
 988     return 0;
 989 fail:
 990     free_tables(h, 1);
 991     return -1;
 992 }
 993
 994 /**
 995  * Mimic alloc_tables(), but for every context thread.
 996  */
 997 static void clone_tables(H264Context *dst, H264Context *src, int i){
 998     MpegEncContext * const s = &src->s;
 999     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
1000     dst->non_zero_count           = src->non_zero_count;
1001     dst->slice_table              = src->slice_table;
1002     dst->cbp_table                = src->cbp_table;
1003     dst->mb2b_xy                  = src->mb2b_xy;
1004     dst->mb2br_xy                 = src->mb2br_xy;
1005     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
1006     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
1007     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
1008     dst->direct_table             = src->direct_table;
1009     dst->list_counts              = src->list_counts;
1010
1011     dst->s.obmc_scratchpad = NULL;
1012     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
1013 }
1014
1015 /**
1016  * Init context
1017  * Allocate buffers which are not shared amongst multiple threads.
1018  */
1019 static int context_init(H264Context *h){
1020     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1021     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1022
1023     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
1024     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
1025
1026     return 0;
1027 fail:
1028     return -1; // free_tables will clean up for us
1029 }
1030
1031 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1032
1033 static av_cold void common_init(H264Context *h){
1034     MpegEncContext * const s = &h->s;
1035
1036     s->width = s->avctx->width;
1037     s->height = s->avctx->height;
1038     s->codec_id= s->avctx->codec->id;
1039
1040     ff_h264dsp_init(&h->h264dsp, 8, 1);
1041     ff_h264_pred_init(&h->hpc, s->codec_id, 8, 1);
1042
1043     h->dequant_coeff_pps= -1;
1044     s->unrestricted_mv=1;
1045
1046     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1047
1048     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1049     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1050 }
1051
1052 int ff_h264_decode_extradata(H264Context *h)
1053 {
1054     AVCodecContext *avctx = h->s.avctx;
1055
1056     if(avctx->extradata[0] == 1){
1057         int i, cnt, nalsize;
1058         unsigned char *p = avctx->extradata;
1059
1060         h->is_avc = 1;
1061
1062         if(avctx->extradata_size < 7) {
1063             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1064             return -1;
1065         }
1066         /* sps and pps in the avcC always have length coded with 2 bytes,
1067            so put a fake nal_length_size = 2 while parsing them */
1068         h->nal_length_size = 2;
1069         // Decode sps from avcC
1070         cnt = *(p+5) & 0x1f; // Number of sps
1071         p += 6;
1072         for (i = 0; i < cnt; i++) {
1073             nalsize = AV_RB16(p) + 2;
1074             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1075                 return -1;
1076             if(decode_nal_units(h, p, nalsize) < 0) {
1077                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1078                 return -1;
1079             }
1080             p += nalsize;
1081         }
1082         // Decode pps from avcC
1083         cnt = *(p++); // Number of pps
1084         for (i = 0; i < cnt; i++) {
1085             nalsize = AV_RB16(p) + 2;
1086             if (p - avctx->extradata + nalsize > avctx->extradata_size)
1087                 return -1;
1088             if (decode_nal_units(h, p, nalsize) < 0) {
1089                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1090                 return -1;
1091             }
1092             p += nalsize;
1093         }
1094         // Now store right nal length size, that will be use to parse all other nals
1095         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1096     } else {
1097         h->is_avc = 0;
1098         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1099             return -1;
1100     }
1101     return 0;
1102 }
1103
1104 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1105     H264Context *h= avctx->priv_data;
1106     MpegEncContext * const s = &h->s;
1107     int i;
1108
1109     MPV_decode_defaults(s);
1110
1111     s->avctx = avctx;
1112     common_init(h);
1113
1114     s->out_format = FMT_H264;
1115     s->workaround_bugs= avctx->workaround_bugs;
1116
1117     // set defaults
1118 //    s->decode_mb= ff_h263_decode_mb;
1119     s->quarter_sample = 1;
1120     if(!avctx->has_b_frames)
1121     s->low_delay= 1;
1122
1123     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1124
1125     ff_h264_decode_init_vlc();
1126
1127     h->pixel_shift = 0;
1128     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1129
1130     h->thread_context[0] = h;
1131     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1132     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1133         h->last_pocs[i] = INT_MIN;
1134     h->prev_poc_msb= 1<<16;
1135     h->x264_build = -1;
1136     ff_h264_reset_sei(h);
1137     if(avctx->codec_id == CODEC_ID_H264){
1138         if(avctx->ticks_per_frame == 1){
1139             s->avctx->time_base.den *=2;
1140         }
1141         avctx->ticks_per_frame = 2;
1142     }
1143
1144     if(avctx->extradata_size > 0 && avctx->extradata &&
1145         ff_h264_decode_extradata(h))
1146         return -1;
1147
1148     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1149         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1150         s->low_delay = 0;
1151     }
1152
1153     return 0;
1154 }
1155
1156 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1157 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1158 {
1159     int i;
1160
1161     for (i=0; i<count; i++){
1162         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1163                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1164                 !from[i]));
1165         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1166     }
1167 }
1168
1169 static void copy_parameter_set(void **to, void **from, int count, int size)
1170 {
1171     int i;
1172
1173     for (i=0; i<count; i++){
1174         if (to[i] && !from[i]) av_freep(&to[i]);
1175         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1176
1177         if (from[i]) memcpy(to[i], from[i], size);
1178     }
1179 }
1180
1181 static int decode_init_thread_copy(AVCodecContext *avctx){
1182     H264Context *h= avctx->priv_data;
1183
1184     if (!avctx->internal->is_copy)
1185         return 0;
1186     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1187     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1188
1189     return 0;
1190 }
1191
1192 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1193 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1194     H264Context *h= dst->priv_data, *h1= src->priv_data;
1195     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1196     int inited = s->context_initialized, err;
1197     int i;
1198
1199     if(dst == src || !s1->context_initialized) return 0;
1200
1201     err = ff_mpeg_update_thread_context(dst, src);
1202     if(err) return err;
1203
1204     //FIXME handle width/height changing
1205     if(!inited){
1206         for(i = 0; i < MAX_SPS_COUNT; i++)
1207             av_freep(h->sps_buffers + i);
1208
1209         for(i = 0; i < MAX_PPS_COUNT; i++)
1210             av_freep(h->pps_buffers + i);
1211
1212         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1213         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1214         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1215         if (ff_h264_alloc_tables(h) < 0) {
1216             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1217             return AVERROR(ENOMEM);
1218         }
1219         context_init(h);
1220
1221         for(i=0; i<2; i++){
1222             h->rbsp_buffer[i] = NULL;
1223             h->rbsp_buffer_size[i] = 0;
1224         }
1225
1226         h->thread_context[0] = h;
1227
1228         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1229         // so this has to be allocated here
1230         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1231
1232         s->dsp.clear_blocks(h->mb);
1233         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1234     }
1235
1236     //extradata/NAL handling
1237     h->is_avc          = h1->is_avc;
1238
1239     //SPS/PPS
1240     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1241     h->sps             = h1->sps;
1242     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1243     h->pps             = h1->pps;
1244
1245     //Dequantization matrices
1246     //FIXME these are big - can they be only copied when PPS changes?
1247     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1248
1249     for(i=0; i<6; i++)
1250         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1251
1252     for(i=0; i<6; i++)
1253         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1254
1255     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1256
1257     //POC timing
1258     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1259
1260     //reference lists
1261     copy_fields(h, h1, ref_count, list_count);
1262     copy_fields(h, h1, ref_list,  intra_gb);
1263     copy_fields(h, h1, short_ref, cabac_init_idc);
1264
1265     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1266     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1267     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1268
1269     h->last_slice_type = h1->last_slice_type;
1270
1271     if(!s->current_picture_ptr) return 0;
1272
1273     if(!s->dropable) {
1274         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1275         h->prev_poc_msb     = h->poc_msb;
1276         h->prev_poc_lsb     = h->poc_lsb;
1277     }
1278     h->prev_frame_num_offset= h->frame_num_offset;
1279     h->prev_frame_num       = h->frame_num;
1280     h->outputed_poc         = h->next_outputed_poc;
1281
1282     return err;
1283 }
1284
1285 int ff_h264_frame_start(H264Context *h){
1286     MpegEncContext * const s = &h->s;
1287     int i;
1288     const int pixel_shift = h->pixel_shift;
1289     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1290
1291     if(MPV_frame_start(s, s->avctx) < 0)
1292         return -1;
1293     ff_er_frame_start(s);
1294     /*
1295      * MPV_frame_start uses pict_type to derive key_frame.
1296      * This is incorrect for H.264; IDR markings must be used.
1297      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1298      * See decode_nal_units().
1299      */
1300     s->current_picture_ptr->f.key_frame = 0;
1301     s->current_picture_ptr->mmco_reset= 0;
1302
1303     assert(s->linesize && s->uvlinesize);
1304
1305     for(i=0; i<16; i++){
1306         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1307         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1308     }
1309     for(i=0; i<16; i++){
1310         h->block_offset[16+i]=
1311         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1312         h->block_offset[48+16+i]=
1313         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1314     }
1315
1316     /* can't be in alloc_tables because linesize isn't known there.
1317      * FIXME: redo bipred weight to not require extra buffer? */
1318     for(i = 0; i < thread_count; i++)
1319         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1320             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1321
1322     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1323     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1324
1325 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1326
1327     // We mark the current picture as non-reference after allocating it, so
1328     // that if we break out due to an error it can be released automatically
1329     // in the next MPV_frame_start().
1330     // SVQ3 as well as most other codecs have only last/next/current and thus
1331     // get released even with set reference, besides SVQ3 and others do not
1332     // mark frames as reference later "naturally".
1333     if(s->codec_id != CODEC_ID_SVQ3)
1334         s->current_picture_ptr->f.reference = 0;
1335
1336     s->current_picture_ptr->field_poc[0]=
1337     s->current_picture_ptr->field_poc[1]= INT_MAX;
1338
1339     h->next_output_pic = NULL;
1340
1341     assert(s->current_picture_ptr->long_ref==0);
1342
1343     return 0;
1344 }
1345
1346 /**
1347   * Run setup operations that must be run after slice header decoding.
1348   * This includes finding the next displayed frame.
1349   *
1350   * @param h h264 master context
1351   * @param setup_finished enough NALs have been read that we can call
1352   * ff_thread_finish_setup()
1353   */
1354 static void decode_postinit(H264Context *h, int setup_finished){
1355     MpegEncContext * const s = &h->s;
1356     Picture *out = s->current_picture_ptr;
1357     Picture *cur = s->current_picture_ptr;
1358     int i, pics, out_of_order, out_idx;
1359     int invalid = 0, cnt = 0;
1360
1361     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1362     s->current_picture_ptr->f.pict_type   = s->pict_type;
1363
1364     if (h->next_output_pic) return;
1365
1366     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1367         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1368         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1369         //to find this yet, so we assume the worst for now.
1370         //if (setup_finished)
1371         //    ff_thread_finish_setup(s->avctx);
1372         return;
1373     }
1374
1375     cur->f.interlaced_frame = 0;
1376     cur->f.repeat_pict      = 0;
1377
1378     /* Signal interlacing information externally. */
1379     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1380
1381     if(h->sps.pic_struct_present_flag){
1382         switch (h->sei_pic_struct)
1383         {
1384         case SEI_PIC_STRUCT_FRAME:
1385             break;
1386         case SEI_PIC_STRUCT_TOP_FIELD:
1387         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1388             cur->f.interlaced_frame = 1;
1389             break;
1390         case SEI_PIC_STRUCT_TOP_BOTTOM:
1391         case SEI_PIC_STRUCT_BOTTOM_TOP:
1392             if (FIELD_OR_MBAFF_PICTURE)
1393                 cur->f.interlaced_frame = 1;
1394             else
1395                 // try to flag soft telecine progressive
1396                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1397             break;
1398         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1399         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1400             // Signal the possibility of telecined film externally (pic_struct 5,6)
1401             // From these hints, let the applications decide if they apply deinterlacing.
1402             cur->f.repeat_pict = 1;
1403             break;
1404         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1405             // Force progressive here, as doubling interlaced frame is a bad idea.
1406             cur->f.repeat_pict = 2;
1407             break;
1408         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1409             cur->f.repeat_pict = 4;
1410             break;
1411         }
1412
1413         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1414             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1415     }else{
1416         /* Derive interlacing flag from used decoding process. */
1417         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1418     }
1419     h->prev_interlaced_frame = cur->f.interlaced_frame;
1420
1421     if (cur->field_poc[0] != cur->field_poc[1]){
1422         /* Derive top_field_first from field pocs. */
1423         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1424     }else{
1425         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1426             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1427             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1428               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1429                 cur->f.top_field_first = 1;
1430             else
1431                 cur->f.top_field_first = 0;
1432         }else{
1433             /* Most likely progressive */
1434             cur->f.top_field_first = 0;
1435         }
1436     }
1437
1438     //FIXME do something with unavailable reference frames
1439
1440     /* Sort B-frames into display order */
1441
1442     if(h->sps.bitstream_restriction_flag
1443        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1444         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1445         s->low_delay = 0;
1446     }
1447
1448     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1449        && !h->sps.bitstream_restriction_flag){
1450         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1451         s->low_delay= 0;
1452     }
1453
1454     pics = 0;
1455     while(h->delayed_pic[pics]) pics++;
1456
1457     assert(pics <= MAX_DELAYED_PIC_COUNT);
1458
1459     h->delayed_pic[pics++] = cur;
1460     if (cur->f.reference == 0)
1461         cur->f.reference = DELAYED_PIC_REF;
1462
1463     /* Frame reordering. This code takes pictures from coding order and sorts
1464      * them by their incremental POC value into display order. It supports POC
1465      * gaps, MMCO reset codes and random resets.
1466      * A "display group" can start either with a IDR frame (f.key_frame = 1),
1467      * and/or can be closed down with a MMCO reset code. In sequences where
1468      * there is no delay, we can't detect that (since the frame was already
1469      * output to the user), so we also set h->mmco_reset to detect the MMCO
1470      * reset code.
1471      * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1472      * we increase the delay between input and output. All frames affected by
1473      * the lag (e.g. those that should have been output before another frame
1474      * that we already returned to the user) will be dropped. This is a bug
1475      * that we will fix later. */
1476     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1477         cnt     += out->poc < h->last_pocs[i];
1478         invalid += out->poc == INT_MIN;
1479     }
1480     if (!h->mmco_reset && !cur->f.key_frame && cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1481         h->mmco_reset = 2;
1482         if (pics > 1)
1483             h->delayed_pic[pics - 2]->mmco_reset = 2;
1484     }
1485     if (h->mmco_reset || cur->f.key_frame) {
1486         for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1487             h->last_pocs[i] = INT_MIN;
1488         cnt     = 0;
1489         invalid = MAX_DELAYED_PIC_COUNT;
1490     }
1491     out = h->delayed_pic[0];
1492     out_idx = 0;
1493     for (i = 1; i < MAX_DELAYED_PIC_COUNT && h->delayed_pic[i] &&
1494          !h->delayed_pic[i-1]->mmco_reset && !h->delayed_pic[i]->f.key_frame; i++)
1495     {
1496         if(h->delayed_pic[i]->poc < out->poc){
1497             out = h->delayed_pic[i];
1498             out_idx = i;
1499         }
1500     }
1501     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1502         h->next_outputed_poc = INT_MIN;
1503     out_of_order = !out->f.key_frame && !h->mmco_reset && (out->poc < h->next_outputed_poc);
1504
1505     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1506         { }
1507     else if (out_of_order && pics-1 == s->avctx->has_b_frames &&
1508              s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1509         if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1510             s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1511         }
1512         s->low_delay = 0;
1513     } else if (s->low_delay &&
1514                ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2) ||
1515                 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1516         s->low_delay = 0;
1517         s->avctx->has_b_frames++;
1518     }
1519
1520     if(pics > s->avctx->has_b_frames){
1521         out->f.reference &= ~DELAYED_PIC_REF;
1522         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1523                          // or else the first thread can release the picture and reuse it unsafely
1524         for(i=out_idx; h->delayed_pic[i]; i++)
1525             h->delayed_pic[i] = h->delayed_pic[i+1];
1526     }
1527     memmove(h->last_pocs, &h->last_pocs[1], sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1528     h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1529     if(!out_of_order && pics > s->avctx->has_b_frames){
1530         h->next_output_pic = out;
1531         if (out->mmco_reset) {
1532             if (out_idx > 0) {
1533                 h->next_outputed_poc = out->poc;
1534                 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1535             } else {
1536                 h->next_outputed_poc = INT_MIN;
1537             }
1538         } else {
1539             if (out_idx == 0 && pics > 1 && h->delayed_pic[0]->f.key_frame) {
1540                 h->next_outputed_poc = INT_MIN;
1541             } else {
1542                 h->next_outputed_poc = out->poc;
1543             }
1544         }
1545         h->mmco_reset = 0;
1546     }else{
1547         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1548     }
1549
1550     if (setup_finished)
1551         ff_thread_finish_setup(s->avctx);
1552 }
1553
1554 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1555                                               uint8_t *src_cb, uint8_t *src_cr,
1556                                               int linesize, int uvlinesize, int simple)
1557 {
1558     MpegEncContext * const s = &h->s;
1559     uint8_t *top_border;
1560     int top_idx = 1;
1561     const int pixel_shift = h->pixel_shift;
1562     int chroma444 = CHROMA444;
1563     int chroma422 = CHROMA422;
1564
1565     src_y  -=   linesize;
1566     src_cb -= uvlinesize;
1567     src_cr -= uvlinesize;
1568
1569     if(!simple && FRAME_MBAFF){
1570         if(s->mb_y&1){
1571             if(!MB_MBAFF){
1572                 top_border = h->top_borders[0][s->mb_x];
1573                 AV_COPY128(top_border, src_y + 15*linesize);
1574                 if (pixel_shift)
1575                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1576                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1577                     if(chroma444){
1578                         if (pixel_shift){
1579                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1580                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1581                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1582                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1583                         } else {
1584                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1585                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1586                         }
1587                     } else if(chroma422) {
1588                         if (pixel_shift) {
1589                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1590                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1591                         } else {
1592                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1593                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1594                         }
1595                     } else {
1596                         if (pixel_shift) {
1597                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1598                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1599                         } else {
1600                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1601                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1602                         }
1603                     }
1604                 }
1605             }
1606         }else if(MB_MBAFF){
1607             top_idx = 0;
1608         }else
1609             return;
1610     }
1611
1612     top_border = h->top_borders[top_idx][s->mb_x];
1613     // There are two lines saved, the line above the the top macroblock of a pair,
1614     // and the line above the bottom macroblock
1615     AV_COPY128(top_border, src_y + 16*linesize);
1616     if (pixel_shift)
1617         AV_COPY128(top_border+16, src_y+16*linesize+16);
1618
1619     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1620         if(chroma444){
1621             if (pixel_shift){
1622                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1623                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1624                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1625                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1626             } else {
1627                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1628                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1629             }
1630         } else if(chroma422) {
1631             if (pixel_shift) {
1632                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1633                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1634             } else {
1635                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1636                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1637             }
1638         } else {
1639             if (pixel_shift) {
1640                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1641                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1642             } else {
1643                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1644                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1645             }
1646         }
1647     }
1648 }
1649
1650 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1651                                   uint8_t *src_cb, uint8_t *src_cr,
1652                                   int linesize, int uvlinesize,
1653                                   int xchg, int chroma444,
1654                                   int simple, int pixel_shift){
1655     MpegEncContext * const s = &h->s;
1656     int deblock_topleft;
1657     int deblock_top;
1658     int top_idx = 1;
1659     uint8_t *top_border_m1;
1660     uint8_t *top_border;
1661
1662     if(!simple && FRAME_MBAFF){
1663         if(s->mb_y&1){
1664             if(!MB_MBAFF)
1665                 return;
1666         }else{
1667             top_idx = MB_MBAFF ? 0 : 1;
1668         }
1669     }
1670
1671     if(h->deblocking_filter == 2) {
1672         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1673         deblock_top     = h->top_type;
1674     } else {
1675         deblock_topleft = (s->mb_x > 0);
1676         deblock_top     = (s->mb_y > !!MB_FIELD);
1677     }
1678
1679     src_y  -=   linesize + 1 + pixel_shift;
1680     src_cb -= uvlinesize + 1 + pixel_shift;
1681     src_cr -= uvlinesize + 1 + pixel_shift;
1682
1683     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1684     top_border    = h->top_borders[top_idx][s->mb_x];
1685
1686 #define XCHG(a,b,xchg)\
1687     if (pixel_shift) {\
1688         if (xchg) {\
1689             AV_SWAP64(b+0,a+0);\
1690             AV_SWAP64(b+8,a+8);\
1691         } else {\
1692             AV_COPY128(b,a); \
1693         }\
1694     } else \
1695 if (xchg) AV_SWAP64(b,a);\
1696 else      AV_COPY64(b,a);
1697
1698     if(deblock_top){
1699         if(deblock_topleft){
1700             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1701         }
1702         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1703         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1704         if(s->mb_x+1 < s->mb_width){
1705             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1706         }
1707     }
1708     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1709         if(chroma444){
1710             if(deblock_topleft){
1711                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1712                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1713             }
1714             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1715             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1716             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1717             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1718             if(s->mb_x+1 < s->mb_width){
1719                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1720                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1721             }
1722         } else {
1723             if(deblock_top){
1724                 if(deblock_topleft){
1725                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1726                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1727                 }
1728                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1729                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1730             }
1731         }
1732     }
1733 }
1734
1735 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1736     if (high_bit_depth) {
1737         return AV_RN32A(((int32_t*)mb) + index);
1738     } else
1739         return AV_RN16A(mb + index);
1740 }
1741
1742 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1743     if (high_bit_depth) {
1744         AV_WN32A(((int32_t*)mb) + index, value);
1745     } else
1746         AV_WN16A(mb + index, value);
1747 }
1748
1749 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1750                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1751 {
1752     MpegEncContext * const s = &h->s;
1753     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1754     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1755     int i;
1756     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1757     block_offset += 16*p;
1758     if(IS_INTRA4x4(mb_type)){
1759         if(simple || !s->encoding){
1760             if(IS_8x8DCT(mb_type)){
1761                 if(transform_bypass){
1762                     idct_dc_add =
1763                     idct_add    = s->dsp.add_pixels8;
1764                 }else{
1765                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1766                     idct_add    = h->h264dsp.h264_idct8_add;
1767                 }
1768                 for(i=0; i<16; i+=4){
1769                     uint8_t * const ptr= dest_y + block_offset[i];
1770                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1771                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1772                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1773                     }else{
1774                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1775                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1776                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1777                         if(nnz){
1778                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1779                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1780                             else
1781                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1782                         }
1783                     }
1784                 }
1785             }else{
1786                 if(transform_bypass){
1787                     idct_dc_add =
1788                     idct_add    = s->dsp.add_pixels4;
1789                 }else{
1790                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1791                     idct_add    = h->h264dsp.h264_idct_add;
1792                 }
1793                 for(i=0; i<16; i++){
1794                     uint8_t * const ptr= dest_y + block_offset[i];
1795                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1796
1797                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1798                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1799                     }else{
1800                         uint8_t *topright;
1801                         int nnz, tr;
1802                         uint64_t tr_high;
1803                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1804                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1805                             assert(s->mb_y || linesize <= block_offset[i]);
1806                             if(!topright_avail){
1807                                 if (pixel_shift) {
1808                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1809                                     topright= (uint8_t*) &tr_high;
1810                                 } else {
1811                                     tr= ptr[3 - linesize]*0x01010101u;
1812                                     topright= (uint8_t*) &tr;
1813                                 }
1814                             }else
1815                                 topright= ptr + (4 << pixel_shift) - linesize;
1816                         }else
1817                             topright= NULL;
1818
1819                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1820                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1821                         if(nnz){
1822                             if(is_h264){
1823                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1824                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1825                                 else
1826                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1827                             }else
1828                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1829                         }
1830                     }
1831                 }
1832             }
1833         }
1834     }else{
1835         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1836         if(is_h264){
1837             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1838                 if(!transform_bypass)
1839                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1840                 else{
1841                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1842                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1843                     for(i = 0; i < 16; i++)
1844                         dctcoef_set(h->mb+(p*256 << pixel_shift), pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1845                 }
1846             }
1847         }else
1848             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1849     }
1850 }
1851
1852 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1853                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1854 {
1855     MpegEncContext * const s = &h->s;
1856     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1857     int i;
1858     block_offset += 16*p;
1859     if(!IS_INTRA4x4(mb_type)){
1860         if(is_h264){
1861             if(IS_INTRA16x16(mb_type)){
1862                 if(transform_bypass){
1863                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1864                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1865                     }else{
1866                         for(i=0; i<16; i++){
1867                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1868                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1869                         }
1870                     }
1871                 }else{
1872                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1873                 }
1874             }else if(h->cbp&15){
1875                 if(transform_bypass){
1876                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1877                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1878                     for(i=0; i<16; i+=di){
1879                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1880                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1881                         }
1882                     }
1883                 }else{
1884                     if(IS_8x8DCT(mb_type)){
1885                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1886                     }else{
1887                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1888                     }
1889                 }
1890             }
1891         }else{
1892             for(i=0; i<16; i++){
1893                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1894                     uint8_t * const ptr= dest_y + block_offset[i];
1895                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1896                 }
1897             }
1898         }
1899     }
1900 }
1901
1902 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift)
1903 {
1904     MpegEncContext * const s = &h->s;
1905     const int mb_x= s->mb_x;
1906     const int mb_y= s->mb_y;
1907     const int mb_xy= h->mb_xy;
1908     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1909     uint8_t  *dest_y, *dest_cb, *dest_cr;
1910     int linesize, uvlinesize /*dct_offset*/;
1911     int i, j;
1912     int *block_offset = &h->block_offset[0];
1913     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1914     /* is_h264 should always be true if SVQ3 is disabled. */
1915     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1916     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1917     const int block_h = 16 >> s->chroma_y_shift;
1918     const int chroma422 = CHROMA422;
1919
1920     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1921     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1922     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1923
1924     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1925     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1926
1927     h->list_counts[mb_xy]= h->list_count;
1928
1929     if (!simple && MB_FIELD) {
1930         linesize   = h->mb_linesize   = s->linesize * 2;
1931         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1932         block_offset = &h->block_offset[48];
1933         if(mb_y&1){ //FIXME move out of this function?
1934             dest_y -= s->linesize*15;
1935             dest_cb-= s->uvlinesize * (block_h - 1);
1936             dest_cr-= s->uvlinesize * (block_h - 1);
1937         }
1938         if(FRAME_MBAFF) {
1939             int list;
1940             for(list=0; list<h->list_count; list++){
1941                 if(!USES_LIST(mb_type, list))
1942                     continue;
1943                 if(IS_16X16(mb_type)){
1944                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1945                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1946                 }else{
1947                     for(i=0; i<16; i+=4){
1948                         int ref = h->ref_cache[list][scan8[i]];
1949                         if(ref >= 0)
1950                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1951                     }
1952                 }
1953             }
1954         }
1955     } else {
1956         linesize   = h->mb_linesize   = s->linesize;
1957         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1958 //        dct_offset = s->linesize * 16;
1959     }
1960
1961     if (!simple && IS_INTRA_PCM(mb_type)) {
1962         if (pixel_shift) {
1963             const int bit_depth = h->sps.bit_depth_luma;
1964             int j;
1965             GetBitContext gb;
1966             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1967
1968             for (i = 0; i < 16; i++) {
1969                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1970                 for (j = 0; j < 16; j++)
1971                     tmp_y[j] = get_bits(&gb, bit_depth);
1972             }
1973             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1974                 if (!h->sps.chroma_format_idc) {
1975                     for (i = 0; i < block_h; i++) {
1976                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1977                         for (j = 0; j < 8; j++) {
1978                             tmp_cb[j] = 1 << (bit_depth - 1);
1979                         }
1980                     }
1981                     for (i = 0; i < block_h; i++) {
1982                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1983                         for (j = 0; j < 8; j++) {
1984                             tmp_cr[j] = 1 << (bit_depth - 1);
1985                         }
1986                     }
1987                 } else {
1988                     for (i = 0; i < block_h; i++) {
1989                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1990                         for (j = 0; j < 8; j++)
1991                             tmp_cb[j] = get_bits(&gb, bit_depth);
1992                     }
1993                     for (i = 0; i < block_h; i++) {
1994                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1995                         for (j = 0; j < 8; j++)
1996                             tmp_cr[j] = get_bits(&gb, bit_depth);
1997                     }
1998                 }
1999             }
2000         } else {
2001             for (i=0; i<16; i++) {
2002                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2003             }
2004             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2005                 if (!h->sps.chroma_format_idc) {
2006                     for (i = 0; i < block_h; i++) {
2007                         memset(dest_cb + i*uvlinesize, 128, 8);
2008                         memset(dest_cr + i*uvlinesize, 128, 8);
2009                     }
2010                 } else {
2011                     for (i = 0; i < block_h; i++) {
2012                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
2013                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
2014                     }
2015                 }
2016             }
2017         }
2018     } else {
2019         if(IS_INTRA(mb_type)){
2020             if(h->deblocking_filter)
2021                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
2022
2023             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2024                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2025                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2026             }
2027
2028             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2029
2030             if(h->deblocking_filter)
2031                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
2032         }else if(is_h264){
2033             if (chroma422) {
2034                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2035                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2036                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2037                               h->h264dsp.weight_h264_pixels_tab,
2038                               h->h264dsp.biweight_h264_pixels_tab,
2039                               pixel_shift);
2040             } else {
2041                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2042                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2043                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2044                               h->h264dsp.weight_h264_pixels_tab,
2045                               h->h264dsp.biweight_h264_pixels_tab,
2046                               pixel_shift);
2047             }
2048         }
2049
2050         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2051
2052         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2053             uint8_t *dest[2] = {dest_cb, dest_cr};
2054             if(transform_bypass){
2055                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2056                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
2057                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
2058                 }else{
2059                     idct_add = s->dsp.add_pixels4;
2060                     for(j=1; j<3; j++){
2061                         for(i=j*16; i<j*16+4; i++){
2062                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2063                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
2064                         }
2065                         if (chroma422) {
2066                             for(i=j*16+4; i<j*16+8; i++){
2067                                 if(h->non_zero_count_cache[ scan8[i+4] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2068                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
2069                             }
2070                         }
2071                     }
2072                 }
2073             }else{
2074                 if(is_h264){
2075                     int qp[2];
2076                     if (chroma422) {
2077                         qp[0] = h->chroma_qp[0] + 3;
2078                         qp[1] = h->chroma_qp[1] + 3;
2079                     } else {
2080                         qp[0] = h->chroma_qp[0];
2081                         qp[1] = h->chroma_qp[1];
2082                     }
2083                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
2084                         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]);
2085                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
2086                         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]);
2087                     h->h264dsp.h264_idct_add8(dest, block_offset,
2088                                               h->mb, uvlinesize,
2089                                               h->non_zero_count_cache);
2090                 }else{
2091                     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]);
2092                     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]);
2093                     for(j=1; j<3; j++){
2094                         for(i=j*16; i<j*16+4; i++){
2095                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2096                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2097                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2098                             }
2099                         }
2100                     }
2101                 }
2102             }
2103         }
2104     }
2105     if(h->cbp || IS_INTRA(mb_type))
2106     {
2107         s->dsp.clear_blocks(h->mb);
2108         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2109     }
2110 }
2111
2112 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2113     MpegEncContext * const s = &h->s;
2114     const int mb_x= s->mb_x;
2115     const int mb_y= s->mb_y;
2116     const int mb_xy= h->mb_xy;
2117     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2118     uint8_t  *dest[3];
2119     int linesize;
2120     int i, j, p;
2121     int *block_offset = &h->block_offset[0];
2122     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2123     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2124
2125     for (p = 0; p < plane_count; p++)
2126     {
2127         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2128         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2129     }
2130
2131     h->list_counts[mb_xy]= h->list_count;
2132
2133     if (!simple && MB_FIELD) {
2134         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2135         block_offset = &h->block_offset[48];
2136         if(mb_y&1) //FIXME move out of this function?
2137             for (p = 0; p < 3; p++)
2138                 dest[p] -= s->linesize*15;
2139         if(FRAME_MBAFF) {
2140             int list;
2141             for(list=0; list<h->list_count; list++){
2142                 if(!USES_LIST(mb_type, list))
2143                     continue;
2144                 if(IS_16X16(mb_type)){
2145                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2146                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2147                 }else{
2148                     for(i=0; i<16; i+=4){
2149                         int ref = h->ref_cache[list][scan8[i]];
2150                         if(ref >= 0)
2151                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2152                     }
2153                 }
2154             }
2155         }
2156     } else {
2157         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2158     }
2159
2160     if (!simple && IS_INTRA_PCM(mb_type)) {
2161         if (pixel_shift) {
2162             const int bit_depth = h->sps.bit_depth_luma;
2163             GetBitContext gb;
2164             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2165
2166             for (p = 0; p < plane_count; p++) {
2167                 for (i = 0; i < 16; i++) {
2168                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2169                     for (j = 0; j < 16; j++)
2170                         tmp[j] = get_bits(&gb, bit_depth);
2171                 }
2172             }
2173         } else {
2174             for (p = 0; p < plane_count; p++) {
2175                 for (i = 0; i < 16; i++) {
2176                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2177                 }
2178             }
2179         }
2180     } else {
2181         if(IS_INTRA(mb_type)){
2182             if(h->deblocking_filter)
2183                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2184
2185             for (p = 0; p < plane_count; p++)
2186                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2187
2188             if(h->deblocking_filter)
2189                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2190         }else{
2191             hl_motion(h, dest[0], dest[1], dest[2],
2192                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2193                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2194                       h->h264dsp.weight_h264_pixels_tab,
2195                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2196         }
2197
2198         for (p = 0; p < plane_count; p++)
2199             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2200     }
2201     if(h->cbp || IS_INTRA(mb_type))
2202     {
2203         s->dsp.clear_blocks(h->mb);
2204         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2205     }
2206 }
2207
2208 /**
2209  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2210  */
2211 #define hl_decode_mb_simple(sh, bits) \
2212 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2213     hl_decode_mb_internal(h, 1, sh); \
2214 }
2215 hl_decode_mb_simple(0, 8)
2216 hl_decode_mb_simple(1, 16)
2217
2218 /**
2219  * Process a macroblock; this handles edge cases, such as interlacing.
2220  */
2221 static void av_noinline hl_decode_mb_complex(H264Context *h){
2222     hl_decode_mb_internal(h, 0, h->pixel_shift);
2223 }
2224
2225 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2226     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2227 }
2228
2229 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2230     hl_decode_mb_444_internal(h, 1, 0);
2231 }
2232
2233 void ff_h264_hl_decode_mb(H264Context *h){
2234     MpegEncContext * const s = &h->s;
2235     const int mb_xy= h->mb_xy;
2236     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2237     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2238
2239     if (CHROMA444) {
2240         if(is_complex || h->pixel_shift)
2241             hl_decode_mb_444_complex(h);
2242         else
2243             hl_decode_mb_444_simple(h);
2244     } else if (is_complex) {
2245         hl_decode_mb_complex(h);
2246     } else if (h->pixel_shift) {
2247         hl_decode_mb_simple_16(h);
2248     } else
2249         hl_decode_mb_simple_8(h);
2250 }
2251
2252 static int pred_weight_table(H264Context *h){
2253     MpegEncContext * const s = &h->s;
2254     int list, i;
2255     int luma_def, chroma_def;
2256
2257     h->use_weight= 0;
2258     h->use_weight_chroma= 0;
2259     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2260     if(h->sps.chroma_format_idc)
2261         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2262     luma_def = 1<<h->luma_log2_weight_denom;
2263     chroma_def = 1<<h->chroma_log2_weight_denom;
2264
2265     for(list=0; list<2; list++){
2266         h->luma_weight_flag[list]   = 0;
2267         h->chroma_weight_flag[list] = 0;
2268         for(i=0; i<h->ref_count[list]; i++){
2269             int luma_weight_flag, chroma_weight_flag;
2270
2271             luma_weight_flag= get_bits1(&s->gb);
2272             if(luma_weight_flag){
2273                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2274                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2275                 if(   h->luma_weight[i][list][0] != luma_def
2276                    || h->luma_weight[i][list][1] != 0) {
2277                     h->use_weight= 1;
2278                     h->luma_weight_flag[list]= 1;
2279                 }
2280             }else{
2281                 h->luma_weight[i][list][0]= luma_def;
2282                 h->luma_weight[i][list][1]= 0;
2283             }
2284
2285             if(h->sps.chroma_format_idc){
2286                 chroma_weight_flag= get_bits1(&s->gb);
2287                 if(chroma_weight_flag){
2288                     int j;
2289                     for(j=0; j<2; j++){
2290                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2291                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2292                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2293                            || h->chroma_weight[i][list][j][1] != 0) {
2294                             h->use_weight_chroma= 1;
2295                             h->chroma_weight_flag[list]= 1;
2296                         }
2297                     }
2298                 }else{
2299                     int j;
2300                     for(j=0; j<2; j++){
2301                         h->chroma_weight[i][list][j][0]= chroma_def;
2302                         h->chroma_weight[i][list][j][1]= 0;
2303                     }
2304                 }
2305             }
2306         }
2307         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2308     }
2309     h->use_weight= h->use_weight || h->use_weight_chroma;
2310     return 0;
2311 }
2312
2313 /**
2314  * Initialize implicit_weight table.
2315  * @param field  0/1 initialize the weight for interlaced MBAFF
2316  *                -1 initializes the rest
2317  */
2318 static void implicit_weight_table(H264Context *h, int field){
2319     MpegEncContext * const s = &h->s;
2320     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2321
2322     for (i = 0; i < 2; i++) {
2323         h->luma_weight_flag[i]   = 0;
2324         h->chroma_weight_flag[i] = 0;
2325     }
2326
2327     if(field < 0){
2328         if (s->picture_structure == PICT_FRAME) {
2329             cur_poc = s->current_picture_ptr->poc;
2330         } else {
2331             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2332         }
2333     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2334        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2335         h->use_weight= 0;
2336         h->use_weight_chroma= 0;
2337         return;
2338     }
2339         ref_start= 0;
2340         ref_count0= h->ref_count[0];
2341         ref_count1= h->ref_count[1];
2342     }else{
2343         cur_poc = s->current_picture_ptr->field_poc[field];
2344         ref_start= 16;
2345         ref_count0= 16+2*h->ref_count[0];
2346         ref_count1= 16+2*h->ref_count[1];
2347     }
2348
2349     h->use_weight= 2;
2350     h->use_weight_chroma= 2;
2351     h->luma_log2_weight_denom= 5;
2352     h->chroma_log2_weight_denom= 5;
2353
2354     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2355         int poc0 = h->ref_list[0][ref0].poc;
2356         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2357             int w = 32;
2358             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2359                 int poc1 = h->ref_list[1][ref1].poc;
2360                 int td = av_clip(poc1 - poc0, -128, 127);
2361                 if(td){
2362                     int tb = av_clip(cur_poc - poc0, -128, 127);
2363                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2364                     int dist_scale_factor = (tb*tx + 32) >> 8;
2365                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2366                         w = 64 - dist_scale_factor;
2367                 }
2368             }
2369             if(field<0){
2370                 h->implicit_weight[ref0][ref1][0]=
2371                 h->implicit_weight[ref0][ref1][1]= w;
2372             }else{
2373                 h->implicit_weight[ref0][ref1][field]=w;
2374             }
2375         }
2376     }
2377 }
2378
2379 /**
2380  * instantaneous decoder refresh.
2381  */
2382 static void idr(H264Context *h){
2383     ff_h264_remove_all_refs(h);
2384     h->prev_frame_num= 0;
2385     h->prev_frame_num_offset= 0;
2386     h->prev_poc_msb=
2387     h->prev_poc_lsb= 0;
2388 }
2389
2390 /* forget old pics after a seek */
2391 static void flush_dpb(AVCodecContext *avctx){
2392     H264Context *h= avctx->priv_data;
2393     int i;
2394     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2395         if(h->delayed_pic[i])
2396             h->delayed_pic[i]->f.reference = 0;
2397         h->delayed_pic[i]= NULL;
2398     }
2399     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2400         h->last_pocs[i] = INT_MIN;
2401     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2402     h->prev_interlaced_frame = 1;
2403     idr(h);
2404     if(h->s.current_picture_ptr)
2405         h->s.current_picture_ptr->f.reference = 0;
2406     h->s.first_field= 0;
2407     ff_h264_reset_sei(h);
2408     ff_mpeg_flush(avctx);
2409 }
2410
2411 static int init_poc(H264Context *h){
2412     MpegEncContext * const s = &h->s;
2413     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2414     int field_poc[2];
2415     Picture *cur = s->current_picture_ptr;
2416
2417     h->frame_num_offset= h->prev_frame_num_offset;
2418     if(h->frame_num < h->prev_frame_num)
2419         h->frame_num_offset += max_frame_num;
2420
2421     if(h->sps.poc_type==0){
2422         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2423
2424         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2425             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2426         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2427             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2428         else
2429             h->poc_msb = h->prev_poc_msb;
2430 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2431         field_poc[0] =
2432         field_poc[1] = h->poc_msb + h->poc_lsb;
2433         if(s->picture_structure == PICT_FRAME)
2434             field_poc[1] += h->delta_poc_bottom;
2435     }else if(h->sps.poc_type==1){
2436         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2437         int i;
2438
2439         if(h->sps.poc_cycle_length != 0)
2440             abs_frame_num = h->frame_num_offset + h->frame_num;
2441         else
2442             abs_frame_num = 0;
2443
2444         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2445             abs_frame_num--;
2446
2447         expected_delta_per_poc_cycle = 0;
2448         for(i=0; i < h->sps.poc_cycle_length; i++)
2449             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2450
2451         if(abs_frame_num > 0){
2452             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2453             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2454
2455             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2456             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2457                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2458         } else
2459             expectedpoc = 0;
2460
2461         if(h->nal_ref_idc == 0)
2462             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2463
2464         field_poc[0] = expectedpoc + h->delta_poc[0];
2465         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2466
2467         if(s->picture_structure == PICT_FRAME)
2468             field_poc[1] += h->delta_poc[1];
2469     }else{
2470         int poc= 2*(h->frame_num_offset + h->frame_num);
2471
2472         if(!h->nal_ref_idc)
2473             poc--;
2474
2475         field_poc[0]= poc;
2476         field_poc[1]= poc;
2477     }
2478
2479     if(s->picture_structure != PICT_BOTTOM_FIELD)
2480         s->current_picture_ptr->field_poc[0]= field_poc[0];
2481     if(s->picture_structure != PICT_TOP_FIELD)
2482         s->current_picture_ptr->field_poc[1]= field_poc[1];
2483     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2484
2485     return 0;
2486 }
2487
2488
2489 /**
2490  * initialize scan tables
2491  */
2492 static void init_scan_tables(H264Context *h){
2493     int i;
2494     for(i=0; i<16; i++){
2495 #define T(x) (x>>2) | ((x<<2) & 0xF)
2496         h->zigzag_scan[i] = T(zigzag_scan[i]);
2497         h-> field_scan[i] = T( field_scan[i]);
2498 #undef T
2499     }
2500     for(i=0; i<64; i++){
2501 #define T(x) (x>>3) | ((x&7)<<3)
2502         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2503         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2504         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2505         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2506 #undef T
2507     }
2508     if(h->sps.transform_bypass){ //FIXME same ugly
2509         h->zigzag_scan_q0          = zigzag_scan;
2510         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2511         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2512         h->field_scan_q0           = field_scan;
2513         h->field_scan8x8_q0        = field_scan8x8;
2514         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2515     }else{
2516         h->zigzag_scan_q0          = h->zigzag_scan;
2517         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2518         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2519         h->field_scan_q0           = h->field_scan;
2520         h->field_scan8x8_q0        = h->field_scan8x8;
2521         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2522     }
2523 }
2524
2525 static int field_end(H264Context *h, int in_setup){
2526     MpegEncContext * const s = &h->s;
2527     AVCodecContext * const avctx= s->avctx;
2528     int err = 0;
2529     s->mb_y= 0;
2530
2531     if (!in_setup && !s->dropable)
2532         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2533                                  s->picture_structure==PICT_BOTTOM_FIELD);
2534
2535     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2536         ff_vdpau_h264_set_reference_frames(s);
2537
2538     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2539         if(!s->dropable) {
2540             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2541             h->prev_poc_msb= h->poc_msb;
2542             h->prev_poc_lsb= h->poc_lsb;
2543         }
2544         h->prev_frame_num_offset= h->frame_num_offset;
2545         h->prev_frame_num= h->frame_num;
2546         h->outputed_poc = h->next_outputed_poc;
2547     }
2548
2549     if (avctx->hwaccel) {
2550         if (avctx->hwaccel->end_frame(avctx) < 0)
2551             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2552     }
2553
2554     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2555         ff_vdpau_h264_picture_complete(s);
2556
2557     /*
2558      * FIXME: Error handling code does not seem to support interlaced
2559      * when slices span multiple rows
2560      * The ff_er_add_slice calls don't work right for bottom
2561      * fields; they cause massive erroneous error concealing
2562      * Error marking covers both fields (top and bottom).
2563      * This causes a mismatched s->error_count
2564      * and a bad error table. Further, the error count goes to
2565      * INT_MAX when called for bottom field, because mb_y is
2566      * past end by one (callers fault) and resync_mb_y != 0
2567      * causes problems for the first MB line, too.
2568      */
2569     if (!FIELD_PICTURE)
2570         ff_er_frame_end(s);
2571
2572     MPV_frame_end(s);
2573
2574     h->current_slice=0;
2575
2576     return err;
2577 }
2578
2579 /**
2580  * Replicate H264 "master" context to thread contexts.
2581  */
2582 static void clone_slice(H264Context *dst, H264Context *src)
2583 {
2584     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2585     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2586     dst->s.current_picture      = src->s.current_picture;
2587     dst->s.linesize             = src->s.linesize;
2588     dst->s.uvlinesize           = src->s.uvlinesize;
2589     dst->s.first_field          = src->s.first_field;
2590
2591     dst->prev_poc_msb           = src->prev_poc_msb;
2592     dst->prev_poc_lsb           = src->prev_poc_lsb;
2593     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2594     dst->prev_frame_num         = src->prev_frame_num;
2595     dst->short_ref_count        = src->short_ref_count;
2596
2597     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2598     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2599     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2600     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2601
2602     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2603     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2604 }
2605
2606 /**
2607  * Compute profile from profile_idc and constraint_set?_flags.
2608  *
2609  * @param sps SPS
2610  *
2611  * @return profile as defined by FF_PROFILE_H264_*
2612  */
2613 int ff_h264_get_profile(SPS *sps)
2614 {
2615     int profile = sps->profile_idc;
2616
2617     switch(sps->profile_idc) {
2618     case FF_PROFILE_H264_BASELINE:
2619         // constraint_set1_flag set to 1
2620         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2621         break;
2622     case FF_PROFILE_H264_HIGH_10:
2623     case FF_PROFILE_H264_HIGH_422:
2624     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2625         // constraint_set3_flag set to 1
2626         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2627         break;
2628     }
2629
2630     return profile;
2631 }
2632
2633 /**
2634  * Decode a slice header.
2635  * This will also call MPV_common_init() and frame_start() as needed.
2636  *
2637  * @param h h264context
2638  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2639  *
2640  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2641  */
2642 static int decode_slice_header(H264Context *h, H264Context *h0){
2643     MpegEncContext * const s = &h->s;
2644     MpegEncContext * const s0 = &h0->s;
2645     unsigned int first_mb_in_slice;
2646     unsigned int pps_id;
2647     int num_ref_idx_active_override_flag;
2648     unsigned int slice_type, tmp, i, j;
2649     int default_ref_list_done = 0;
2650     int last_pic_structure;
2651
2652     s->dropable= h->nal_ref_idc == 0;
2653
2654     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2655     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2656         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2657         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2658     }else{
2659         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2660         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2661     }
2662
2663     first_mb_in_slice= get_ue_golomb(&s->gb);
2664
2665     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2666         if(h0->current_slice && FIELD_PICTURE){
2667             field_end(h, 1);
2668         }
2669
2670         h0->current_slice = 0;
2671         if (!s0->first_field)
2672             s->current_picture_ptr= NULL;
2673     }
2674
2675     slice_type= get_ue_golomb_31(&s->gb);
2676     if(slice_type > 9){
2677         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);
2678         return -1;
2679     }
2680     if(slice_type > 4){
2681         slice_type -= 5;
2682         h->slice_type_fixed=1;
2683     }else
2684         h->slice_type_fixed=0;
2685
2686     slice_type= golomb_to_pict_type[ slice_type ];
2687     if (slice_type == AV_PICTURE_TYPE_I
2688         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2689         default_ref_list_done = 1;
2690     }
2691     h->slice_type= slice_type;
2692     h->slice_type_nos= slice_type & 3;
2693
2694     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2695
2696     pps_id= get_ue_golomb(&s->gb);
2697     if(pps_id>=MAX_PPS_COUNT){
2698         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2699         return -1;
2700     }
2701     if(!h0->pps_buffers[pps_id]) {
2702         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2703         return -1;
2704     }
2705     h->pps= *h0->pps_buffers[pps_id];
2706
2707     if(!h0->sps_buffers[h->pps.sps_id]) {
2708         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2709         return -1;
2710     }
2711     h->sps = *h0->sps_buffers[h->pps.sps_id];
2712
2713     s->avctx->profile = ff_h264_get_profile(&h->sps);
2714     s->avctx->level   = h->sps.level_idc;
2715     s->avctx->refs    = h->sps.ref_frame_count;
2716
2717     if(h == h0 && h->dequant_coeff_pps != pps_id){
2718         h->dequant_coeff_pps = pps_id;
2719         init_dequant_tables(h);
2720     }
2721
2722     s->mb_width= h->sps.mb_width;
2723     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2724
2725     h->b_stride=  s->mb_width*4;
2726
2727     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2728
2729     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2730     if(h->sps.frame_mbs_only_flag)
2731         s->height= 16*s->mb_height - (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2732     else
2733         s->height= 16*s->mb_height - (2<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1);
2734
2735     if (s->context_initialized
2736         && (   s->width != s->avctx->width || s->height != s->avctx->height
2737             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2738         if(h != h0) {
2739             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2740             return -1;   // width / height changed during parallelized decoding
2741         }
2742         free_tables(h, 0);
2743         flush_dpb(s->avctx);
2744         MPV_common_end(s);
2745     }
2746     if (!s->context_initialized) {
2747         if (h != h0) {
2748             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2749             return -1;
2750         }
2751
2752         avcodec_set_dimensions(s->avctx, s->width, s->height);
2753         s->avctx->sample_aspect_ratio= h->sps.sar;
2754         av_assert0(s->avctx->sample_aspect_ratio.den);
2755
2756         if(h->sps.video_signal_type_present_flag){
2757             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2758             if(h->sps.colour_description_present_flag){
2759                 s->avctx->color_primaries = h->sps.color_primaries;
2760                 s->avctx->color_trc       = h->sps.color_trc;
2761                 s->avctx->colorspace      = h->sps.colorspace;
2762             }
2763         }
2764
2765         if(h->sps.timing_info_present_flag){
2766             int64_t den= h->sps.time_scale;
2767             if(h->x264_build < 44U)
2768                 den *= 2;
2769             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2770                       h->sps.num_units_in_tick, den, 1<<30);
2771         }
2772
2773         switch (h->sps.bit_depth_luma) {
2774             case 9 :
2775                 if (CHROMA444) {
2776                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2777                         s->avctx->pix_fmt = PIX_FMT_GBRP9;
2778                     } else
2779                         s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2780                 } else if (CHROMA422)
2781                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2782                 else
2783                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2784                 break;
2785             case 10 :
2786                 if (CHROMA444) {
2787                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2788                         s->avctx->pix_fmt = PIX_FMT_GBRP10;
2789                     } else
2790                         s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2791                 } else if (CHROMA422)
2792                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2793                 else
2794                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2795                 break;
2796             default:
2797                 if (CHROMA444){
2798                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2799                         s->avctx->pix_fmt = PIX_FMT_GBRP;
2800                     } else
2801                         s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2802                 } else if (CHROMA422) {
2803                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2804                 }else{
2805                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2806                                                              s->avctx->codec->pix_fmts ?
2807                                                              s->avctx->codec->pix_fmts :
2808                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2809                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2810                                                              ff_hwaccel_pixfmt_list_420);
2811                 }
2812         }
2813
2814         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2815
2816         if (MPV_common_init(s) < 0) {
2817             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2818             return -1;
2819         }
2820         s->first_field = 0;
2821         h->prev_interlaced_frame = 1;
2822
2823         init_scan_tables(h);
2824         if (ff_h264_alloc_tables(h) < 0) {
2825             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2826             return AVERROR(ENOMEM);
2827         }
2828
2829         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2830             if (context_init(h) < 0) {
2831                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2832                 return -1;
2833             }
2834         } else {
2835             for(i = 1; i < s->avctx->thread_count; i++) {
2836                 H264Context *c;
2837                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2838                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2839                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2840                 c->h264dsp = h->h264dsp;
2841                 c->sps = h->sps;
2842                 c->pps = h->pps;
2843                 c->pixel_shift = h->pixel_shift;
2844                 init_scan_tables(c);
2845                 clone_tables(c, h, i);
2846             }
2847
2848             for(i = 0; i < s->avctx->thread_count; i++)
2849                 if (context_init(h->thread_context[i]) < 0) {
2850                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2851                     return -1;
2852                 }
2853         }
2854     }
2855
2856     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2857
2858     h->mb_mbaff = 0;
2859     h->mb_aff_frame = 0;
2860     last_pic_structure = s0->picture_structure;
2861     if(h->sps.frame_mbs_only_flag){
2862         s->picture_structure= PICT_FRAME;
2863     }else{
2864         if(get_bits1(&s->gb)) { //field_pic_flag
2865             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2866         } else {
2867             s->picture_structure= PICT_FRAME;
2868             h->mb_aff_frame = h->sps.mb_aff;
2869         }
2870     }
2871     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2872
2873     if(h0->current_slice == 0){
2874         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2875         if(h->frame_num != h->prev_frame_num) {
2876             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2877
2878             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2879
2880             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2881                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2882                 if (unwrap_prev_frame_num < 0)
2883                     unwrap_prev_frame_num += max_frame_num;
2884
2885                 h->prev_frame_num = unwrap_prev_frame_num;
2886             }
2887         }
2888
2889         while(h->frame_num !=  h->prev_frame_num &&
2890               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2891             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2892             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2893             if (ff_h264_frame_start(h) < 0)
2894                 return -1;
2895             h->prev_frame_num++;
2896             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2897             s->current_picture_ptr->frame_num= h->prev_frame_num;
2898             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2899             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2900             ff_generate_sliding_window_mmcos(h);
2901             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2902                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2903                 return AVERROR_INVALIDDATA;
2904             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2905              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2906              * about there being no actual duplicates.
2907              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2908              * concealing a lost frame, this probably isn't noticeable by comparison, but it should
2909              * be fixed. */
2910             if (h->short_ref_count) {
2911                 if (prev) {
2912                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2913                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2914                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2915                     h->short_ref[0]->poc = prev->poc+2;
2916                 }
2917                 h->short_ref[0]->frame_num = h->prev_frame_num;
2918             }
2919         }
2920
2921         /* See if we have a decoded first field looking for a pair... */
2922         if (s0->first_field) {
2923             assert(s0->current_picture_ptr);
2924             assert(s0->current_picture_ptr->f.data[0]);
2925             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2926
2927             /* figure out if we have a complementary field pair */
2928             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2929                 /*
2930                  * Previous field is unmatched. Don't display it, but let it
2931                  * remain for reference if marked as such.
2932                  */
2933                 s0->current_picture_ptr = NULL;
2934                 s0->first_field = FIELD_PICTURE;
2935
2936             } else {
2937                 if (h->nal_ref_idc &&
2938                         s0->current_picture_ptr->f.reference &&
2939                         s0->current_picture_ptr->frame_num != h->frame_num) {
2940                     /*
2941                      * This and previous field were reference, but had
2942                      * different frame_nums. Consider this field first in
2943                      * pair. Throw away previous field except for reference
2944                      * purposes.
2945                      */
2946                     s0->first_field = 1;
2947                     s0->current_picture_ptr = NULL;
2948
2949                 } else {
2950                     /* Second field in complementary pair */
2951                     s0->first_field = 0;
2952                 }
2953             }
2954
2955         } else {
2956             /* Frame or first field in a potentially complementary pair */
2957             assert(!s0->current_picture_ptr);
2958             s0->first_field = FIELD_PICTURE;
2959         }
2960
2961         if(!FIELD_PICTURE || s0->first_field) {
2962             if (ff_h264_frame_start(h) < 0) {
2963                 s0->first_field = 0;
2964                 return -1;
2965             }
2966         } else {
2967             ff_release_unused_pictures(s, 0);
2968         }
2969     }
2970     if(h != h0)
2971         clone_slice(h, h0);
2972
2973     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2974
2975     assert(s->mb_num == s->mb_width * s->mb_height);
2976     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2977        first_mb_in_slice                    >= s->mb_num){
2978         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2979         return -1;
2980     }
2981     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2982     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2983     if (s->picture_structure == PICT_BOTTOM_FIELD)
2984         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2985     assert(s->mb_y < s->mb_height);
2986
2987     if(s->picture_structure==PICT_FRAME){
2988         h->curr_pic_num=   h->frame_num;
2989         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2990     }else{
2991         h->curr_pic_num= 2*h->frame_num + 1;
2992         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2993     }
2994
2995     if(h->nal_unit_type == NAL_IDR_SLICE){
2996         get_ue_golomb(&s->gb); /* idr_pic_id */
2997     }
2998
2999     if(h->sps.poc_type==0){
3000         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3001
3002         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3003             h->delta_poc_bottom= get_se_golomb(&s->gb);
3004         }
3005     }
3006
3007     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3008         h->delta_poc[0]= get_se_golomb(&s->gb);
3009
3010         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3011             h->delta_poc[1]= get_se_golomb(&s->gb);
3012     }
3013
3014     init_poc(h);
3015
3016     if(h->pps.redundant_pic_cnt_present){
3017         h->redundant_pic_count= get_ue_golomb(&s->gb);
3018     }
3019
3020     //set defaults, might be overridden a few lines later
3021     h->ref_count[0]= h->pps.ref_count[0];
3022     h->ref_count[1]= h->pps.ref_count[1];
3023
3024     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
3025         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
3026             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3027         }
3028         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3029
3030         if(num_ref_idx_active_override_flag){
3031             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3032             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
3033                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3034
3035             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
3036                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3037                 h->ref_count[0]= h->ref_count[1]= 1;
3038                 return -1;
3039             }
3040         }
3041         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
3042             h->list_count= 2;
3043         else
3044             h->list_count= 1;
3045     }else
3046         h->list_count= 0;
3047
3048     if(!default_ref_list_done){
3049         ff_h264_fill_default_ref_list(h);
3050     }
3051
3052     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3053         h->ref_count[1]= h->ref_count[0]= 0;
3054         return -1;
3055     }
3056
3057     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
3058         s->last_picture_ptr= &h->ref_list[0][0];
3059         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3060     }
3061     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
3062         s->next_picture_ptr= &h->ref_list[1][0];
3063         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3064     }
3065
3066     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
3067        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
3068         pred_weight_table(h);
3069     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3070         implicit_weight_table(h, -1);
3071     }else {
3072         h->use_weight = 0;
3073         for (i = 0; i < 2; i++) {
3074             h->luma_weight_flag[i]   = 0;
3075             h->chroma_weight_flag[i] = 0;
3076         }
3077     }
3078
3079     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3080        (s->avctx->err_recognition & AV_EF_EXPLODE))
3081         return AVERROR_INVALIDDATA;
3082
3083     if(FRAME_MBAFF){
3084         ff_h264_fill_mbaff_ref_list(h);
3085
3086         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3087             implicit_weight_table(h, 0);
3088             implicit_weight_table(h, 1);
3089         }
3090     }
3091
3092     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3093         ff_h264_direct_dist_scale_factor(h);
3094     ff_h264_direct_ref_list_init(h);
3095
3096     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3097         tmp = get_ue_golomb_31(&s->gb);
3098         if(tmp > 2){
3099             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3100             return -1;
3101         }
3102         h->cabac_init_idc= tmp;
3103     }
3104
3105     h->last_qscale_diff = 0;
3106     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3107     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3108         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3109         return -1;
3110     }
3111     s->qscale= tmp;
3112     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3113     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3114     //FIXME qscale / qp ... stuff
3115     if(h->slice_type == AV_PICTURE_TYPE_SP){
3116         get_bits1(&s->gb); /* sp_for_switch_flag */
3117     }
3118     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3119         get_se_golomb(&s->gb); /* slice_qs_delta */
3120     }
3121
3122     h->deblocking_filter = 1;
3123     h->slice_alpha_c0_offset = 52;
3124     h->slice_beta_offset = 52;
3125     if( h->pps.deblocking_filter_parameters_present ) {
3126         tmp= get_ue_golomb_31(&s->gb);
3127         if(tmp > 2){
3128             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3129             return -1;
3130         }
3131         h->deblocking_filter= tmp;
3132         if(h->deblocking_filter < 2)
3133             h->deblocking_filter^= 1; // 1<->0
3134
3135         if( h->deblocking_filter ) {
3136             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3137             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3138             if(   h->slice_alpha_c0_offset > 104U
3139                || h->slice_beta_offset     > 104U){
3140                 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);
3141                 return -1;
3142             }
3143         }
3144     }
3145
3146     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3147        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3148        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3149        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3150         h->deblocking_filter= 0;
3151
3152     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3153         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3154             /* Cheat slightly for speed:
3155                Do not bother to deblock across slices. */
3156             h->deblocking_filter = 2;
3157         } else {
3158             h0->max_contexts = 1;
3159             if(!h0->single_decode_warning) {
3160                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3161                 h0->single_decode_warning = 1;
3162             }
3163             if (h != h0) {
3164                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3165                 return 1;
3166             }
3167         }
3168     }
3169     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3170                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3171                  + 6 * (h->sps.bit_depth_luma - 8);
3172
3173 #if 0 //FMO
3174     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3175         slice_group_change_cycle= get_bits(&s->gb, ?);
3176 #endif
3177
3178     h0->last_slice_type = slice_type;
3179     h->slice_num = ++h0->current_slice;
3180     if(h->slice_num >= MAX_SLICES){
3181         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
3182     }
3183
3184     for(j=0; j<2; j++){
3185         int id_list[16];
3186         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3187         for(i=0; i<16; i++){
3188             id_list[i]= 60;
3189             if (h->ref_list[j][i].f.data[0]) {
3190                 int k;
3191                 uint8_t *base = h->ref_list[j][i].f.base[0];
3192                 for(k=0; k<h->short_ref_count; k++)
3193                     if (h->short_ref[k]->f.base[0] == base) {
3194                         id_list[i]= k;
3195                         break;
3196                     }
3197                 for(k=0; k<h->long_ref_count; k++)
3198                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3199                         id_list[i]= h->short_ref_count + k;
3200                         break;
3201                     }
3202             }
3203         }
3204
3205         ref2frm[0]=
3206         ref2frm[1]= -1;
3207         for(i=0; i<16; i++)
3208             ref2frm[i+2]= 4*id_list[i]
3209                           + (h->ref_list[j][i].f.reference & 3);
3210         ref2frm[18+0]=
3211         ref2frm[18+1]= -1;
3212         for(i=16; i<48; i++)
3213             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3214                           + (h->ref_list[j][i].f.reference & 3);
3215     }
3216
3217     //FIXME: fix draw_edges+PAFF+frame threads
3218     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3219     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3220
3221     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3222         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",
3223                h->slice_num,
3224                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3225                first_mb_in_slice,
3226                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3227                pps_id, h->frame_num,
3228                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3229                h->ref_count[0], h->ref_count[1],
3230                s->qscale,
3231                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3232                h->use_weight,
3233                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3234                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3235                );
3236     }
3237
3238     return 0;
3239 }
3240
3241 int ff_h264_get_slice_type(const H264Context *h)
3242 {
3243     switch (h->slice_type) {
3244     case AV_PICTURE_TYPE_P:  return 0;
3245     case AV_PICTURE_TYPE_B:  return 1;
3246     case AV_PICTURE_TYPE_I:  return 2;
3247     case AV_PICTURE_TYPE_SP: return 3;
3248     case AV_PICTURE_TYPE_SI: return 4;
3249     default:         return -1;
3250     }
3251 }
3252
3253 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3254                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3255 {
3256     int b_stride = h->b_stride;
3257     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3258     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3259     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3260         if(USES_LIST(top_type, list)){
3261             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3262             const int b8_xy= 4*top_xy + 2;
3263             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3264             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3265             ref_cache[0 - 1*8]=
3266             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3267             ref_cache[2 - 1*8]=
3268             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3269         }else{
3270             AV_ZERO128(mv_dst - 1*8);
3271             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3272         }
3273
3274         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3275             if(USES_LIST(left_type[LTOP], list)){
3276                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3277                 const int b8_xy= 4*left_xy[LTOP] + 1;
3278                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3279                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3280                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3281                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3282                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3283                 ref_cache[-1 +  0]=
3284                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3285                 ref_cache[-1 + 16]=
3286                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3287             }else{
3288                 AV_ZERO32(mv_dst - 1 + 0);
3289                 AV_ZERO32(mv_dst - 1 + 8);
3290                 AV_ZERO32(mv_dst - 1 +16);
3291                 AV_ZERO32(mv_dst - 1 +24);
3292                 ref_cache[-1 +  0]=
3293                 ref_cache[-1 +  8]=
3294                 ref_cache[-1 + 16]=
3295                 ref_cache[-1 + 24]= LIST_NOT_USED;
3296             }
3297         }
3298     }
3299
3300     if(!USES_LIST(mb_type, list)){
3301         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3302         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3303         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3304         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3305         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3306         return;
3307     }
3308
3309     {
3310         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3311         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3312         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3313         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3314         AV_WN32A(&ref_cache[0*8], ref01);
3315         AV_WN32A(&ref_cache[1*8], ref01);
3316         AV_WN32A(&ref_cache[2*8], ref23);
3317         AV_WN32A(&ref_cache[3*8], ref23);
3318     }
3319
3320     {
3321         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3322         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3323         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3324         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3325         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3326     }
3327 }
3328
3329 /**
3330  *
3331  * @return non zero if the loop filter can be skipped
3332  */
3333 static int fill_filter_caches(H264Context *h, int mb_type){
3334     MpegEncContext * const s = &h->s;
3335     const int mb_xy= h->mb_xy;
3336     int top_xy, left_xy[LEFT_MBS];
3337     int top_type, left_type[LEFT_MBS];
3338     uint8_t *nnz;
3339     uint8_t *nnz_cache;
3340
3341     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3342
3343     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3344      * stuff, I can't imagine that these complex rules are worth it. */
3345
3346     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3347     if(FRAME_MBAFF){
3348         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3349         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3350         if(s->mb_y&1){
3351             if (left_mb_field_flag != curr_mb_field_flag) {
3352                 left_xy[LTOP] -= s->mb_stride;
3353             }
3354         }else{
3355             if(curr_mb_field_flag){
3356                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3357             }
3358             if (left_mb_field_flag != curr_mb_field_flag) {
3359                 left_xy[LBOT] += s->mb_stride;
3360             }
3361         }
3362     }
3363
3364     h->top_mb_xy = top_xy;
3365     h->left_mb_xy[LTOP] = left_xy[LTOP];
3366     h->left_mb_xy[LBOT] = left_xy[LBOT];
3367     {
3368         //for sufficiently low qp, filtering wouldn't do anything
3369         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3370         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3371         int qp = s->current_picture.f.qscale_table[mb_xy];
3372         if(qp <= qp_thresh
3373            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3374            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3375             if(!FRAME_MBAFF)
3376                 return 1;
3377             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3378                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3379                 return 1;
3380         }
3381     }
3382
3383     top_type        = s->current_picture.f.mb_type[top_xy];
3384     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3385     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3386     if(h->deblocking_filter == 2){
3387         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3388         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3389     }else{
3390         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3391         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3392     }
3393     h->top_type       = top_type;
3394     h->left_type[LTOP]= left_type[LTOP];
3395     h->left_type[LBOT]= left_type[LBOT];
3396
3397     if(IS_INTRA(mb_type))
3398         return 0;
3399
3400     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3401     if(h->list_count == 2)
3402         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3403
3404     nnz = h->non_zero_count[mb_xy];
3405     nnz_cache = h->non_zero_count_cache;
3406     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3407     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3408     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3409     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3410     h->cbp= h->cbp_table[mb_xy];
3411
3412     if(top_type){
3413         nnz = h->non_zero_count[top_xy];
3414         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3415     }
3416
3417     if(left_type[LTOP]){
3418         nnz = h->non_zero_count[left_xy[LTOP]];
3419         nnz_cache[3+8*1]= nnz[3+0*4];
3420         nnz_cache[3+8*2]= nnz[3+1*4];
3421         nnz_cache[3+8*3]= nnz[3+2*4];
3422         nnz_cache[3+8*4]= nnz[3+3*4];
3423     }
3424
3425     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3426     if(!CABAC && h->pps.transform_8x8_mode){
3427         if(IS_8x8DCT(top_type)){
3428             nnz_cache[4+8*0]=
3429             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3430             nnz_cache[6+8*0]=
3431             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3432         }
3433         if(IS_8x8DCT(left_type[LTOP])){
3434             nnz_cache[3+8*1]=
3435             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3436         }
3437         if(IS_8x8DCT(left_type[LBOT])){
3438             nnz_cache[3+8*3]=
3439             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3440         }
3441
3442         if(IS_8x8DCT(mb_type)){
3443             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3444             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3445
3446             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3447             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3448
3449             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3450             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3451
3452             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3453             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3454         }
3455     }
3456
3457     return 0;
3458 }
3459
3460 static void loop_filter(H264Context *h, int start_x, int end_x){
3461     MpegEncContext * const s = &h->s;
3462     uint8_t  *dest_y, *dest_cb, *dest_cr;
3463     int linesize, uvlinesize, mb_x, mb_y;
3464     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3465     const int old_slice_type= h->slice_type;
3466     const int pixel_shift = h->pixel_shift;
3467     const int block_h = 16 >> s->chroma_y_shift;
3468
3469     if(h->deblocking_filter) {
3470         for(mb_x= start_x; mb_x<end_x; mb_x++){
3471             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3472                 int mb_xy, mb_type;
3473                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3474                 h->slice_num= h->slice_table[mb_xy];
3475                 mb_type = s->current_picture.f.mb_type[mb_xy];
3476                 h->list_count= h->list_counts[mb_xy];
3477
3478                 if(FRAME_MBAFF)
3479                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3480
3481                 s->mb_x= mb_x;
3482                 s->mb_y= mb_y;
3483                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3484                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3485                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3486                     //FIXME simplify above
3487
3488                 if (MB_FIELD) {
3489                     linesize   = h->mb_linesize   = s->linesize * 2;
3490                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3491                     if(mb_y&1){ //FIXME move out of this function?
3492                         dest_y -= s->linesize*15;
3493                         dest_cb-= s->uvlinesize * (block_h - 1);
3494                         dest_cr-= s->uvlinesize * (block_h - 1);
3495                     }
3496                 } else {
3497                     linesize   = h->mb_linesize   = s->linesize;
3498                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3499                 }
3500                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3501                 if(fill_filter_caches(h, mb_type))
3502                     continue;
3503                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3504                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3505
3506                 if (FRAME_MBAFF) {
3507                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3508                 } else {
3509                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3510                 }
3511             }
3512         }
3513     }
3514     h->slice_type= old_slice_type;
3515     s->mb_x= end_x;
3516     s->mb_y= end_mb_y - FRAME_MBAFF;
3517     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3518     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3519 }
3520
3521 static void predict_field_decoding_flag(H264Context *h){
3522     MpegEncContext * const s = &h->s;
3523     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3524     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3525                 ? s->current_picture.f.mb_type[mb_xy - 1]
3526                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3527                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3528                 : 0;
3529     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3530 }
3531
3532 /**
3533  * Draw edges and report progress for the last MB row.
3534  */
3535 static void decode_finish_row(H264Context *h){
3536     MpegEncContext * const s = &h->s;
3537     int top = 16*(s->mb_y >> FIELD_PICTURE);
3538     int height = 16 << FRAME_MBAFF;
3539     int deblock_border = (16 + 4) << FRAME_MBAFF;
3540     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3541
3542     if (h->deblocking_filter) {
3543         if((top + height) >= pic_height)
3544             height += deblock_border;
3545
3546         top -= deblock_border;
3547     }
3548
3549     if (top >= pic_height || (top + height) < h->emu_edge_height)
3550         return;
3551
3552     height = FFMIN(height, pic_height - top);
3553     if (top < h->emu_edge_height) {
3554         height = top+height;
3555         top = 0;
3556     }
3557
3558     ff_draw_horiz_band(s, top, height);
3559
3560     if (s->dropable) return;
3561
3562     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3563                              s->picture_structure==PICT_BOTTOM_FIELD);
3564 }
3565
3566 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3567     H264Context *h = *(void**)arg;
3568     MpegEncContext * const s = &h->s;
3569     const int part_mask= s->partitioned_frame ? (ER_AC_END|ER_AC_ERROR) : 0x7F;
3570     int lf_x_start = s->mb_x;
3571
3572     s->mb_skip_run= -1;
3573
3574     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3575                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3576
3577     if( h->pps.cabac ) {
3578         /* realign */
3579         align_get_bits( &s->gb );
3580
3581         /* init cabac */
3582         ff_init_cabac_states( &h->cabac);
3583         ff_init_cabac_decoder( &h->cabac,
3584                                s->gb.buffer + get_bits_count(&s->gb)/8,
3585                                (get_bits_left(&s->gb) + 7)/8);
3586
3587         ff_h264_init_cabac_states(h);
3588
3589         for(;;){
3590 //START_TIMER
3591             int ret = ff_h264_decode_mb_cabac(h);
3592             int eos;
3593 //STOP_TIMER("decode_mb_cabac")
3594
3595             if(ret>=0) ff_h264_hl_decode_mb(h);
3596
3597             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3598                 s->mb_y++;
3599
3600                 ret = ff_h264_decode_mb_cabac(h);
3601
3602                 if(ret>=0) ff_h264_hl_decode_mb(h);
3603                 s->mb_y--;
3604             }
3605             eos = get_cabac_terminate( &h->cabac );
3606
3607             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3608                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3609                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3610                 return 0;
3611             }
3612             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3613                 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);
3614                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3615                 return -1;
3616             }
3617
3618             if( ++s->mb_x >= s->mb_width ) {
3619                 loop_filter(h, lf_x_start, s->mb_x);
3620                 s->mb_x = lf_x_start = 0;
3621                 decode_finish_row(h);
3622                 ++s->mb_y;
3623                 if(FIELD_OR_MBAFF_PICTURE) {
3624                     ++s->mb_y;
3625                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3626                         predict_field_decoding_flag(h);
3627                 }
3628             }
3629
3630             if( eos || s->mb_y >= s->mb_height ) {
3631                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3632                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3633                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3634                 return 0;
3635             }
3636         }
3637
3638     } else {
3639         for(;;){
3640             int ret = ff_h264_decode_mb_cavlc(h);
3641
3642             if(ret>=0) ff_h264_hl_decode_mb(h);
3643
3644             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3645                 s->mb_y++;
3646                 ret = ff_h264_decode_mb_cavlc(h);
3647
3648                 if(ret>=0) ff_h264_hl_decode_mb(h);
3649                 s->mb_y--;
3650             }
3651
3652             if(ret<0){
3653                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3654                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3655                 return -1;
3656             }
3657
3658             if(++s->mb_x >= s->mb_width){
3659                 loop_filter(h, lf_x_start, s->mb_x);
3660                 s->mb_x = lf_x_start = 0;
3661                 decode_finish_row(h);
3662                 ++s->mb_y;
3663                 if(FIELD_OR_MBAFF_PICTURE) {
3664                     ++s->mb_y;
3665                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3666                         predict_field_decoding_flag(h);
3667                 }
3668                 if(s->mb_y >= s->mb_height){
3669                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3670
3671                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3672                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3673
3674                         return 0;
3675                     }else{
3676                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_END&part_mask);
3677
3678                         return -1;
3679                     }
3680                 }
3681             }
3682
3683             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3684                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3685                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3686                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, ER_MB_END&part_mask);
3687                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3688
3689                     return 0;
3690                 }else{
3691                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, ER_MB_ERROR&part_mask);
3692
3693                     return -1;
3694                 }
3695             }
3696         }
3697     }
3698 }
3699
3700 /**
3701  * Call decode_slice() for each context.
3702  *
3703  * @param h h264 master context
3704  * @param context_count number of contexts to execute
3705  */
3706 static int execute_decode_slices(H264Context *h, int context_count){
3707     MpegEncContext * const s = &h->s;
3708     AVCodecContext * const avctx= s->avctx;
3709     H264Context *hx;
3710     int i;
3711
3712     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3713         return 0;
3714     if(context_count == 1) {
3715         return decode_slice(avctx, &h);
3716     } else {
3717         for(i = 1; i < context_count; i++) {
3718             hx = h->thread_context[i];
3719             hx->s.err_recognition = avctx->err_recognition;
3720             hx->s.error_count = 0;
3721         }
3722
3723         avctx->execute(avctx, decode_slice,
3724                        h->thread_context, NULL, context_count, sizeof(void*));
3725
3726         /* pull back stuff from slices to master context */
3727         hx = h->thread_context[context_count - 1];
3728         s->mb_x = hx->s.mb_x;
3729         s->mb_y = hx->s.mb_y;
3730         s->dropable = hx->s.dropable;
3731         s->picture_structure = hx->s.picture_structure;
3732         for(i = 1; i < context_count; i++)
3733             h->s.error_count += h->thread_context[i]->s.error_count;
3734     }
3735
3736     return 0;
3737 }
3738
3739
3740 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3741     MpegEncContext * const s = &h->s;
3742     AVCodecContext * const avctx= s->avctx;
3743     H264Context *hx; ///< thread context
3744     int buf_index;
3745     int context_count;
3746     int next_avc;
3747     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3748     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3749     int nal_index;
3750
3751     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3752     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3753         h->current_slice = 0;
3754         if (!s->first_field)
3755             s->current_picture_ptr= NULL;
3756         ff_h264_reset_sei(h);
3757     }
3758
3759     for(;pass <= 1;pass++){
3760         buf_index = 0;
3761         context_count = 0;
3762         next_avc = h->is_avc ? 0 : buf_size;
3763         nal_index = 0;
3764     for(;;){
3765         int consumed;
3766         int dst_length;
3767         int bit_length;
3768         const uint8_t *ptr;
3769         int i, nalsize = 0;
3770         int err;
3771
3772         if(buf_index >= next_avc) {
3773             if(buf_index >= buf_size) break;
3774             nalsize = 0;
3775             for(i = 0; i < h->nal_length_size; i++)
3776                 nalsize = (nalsize << 8) | buf[buf_index++];
3777             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3778                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3779                 break;
3780             }
3781             next_avc= buf_index + nalsize;
3782         } else {
3783             // start code prefix search
3784             for(; buf_index + 3 < next_avc; buf_index++){
3785                 // This should always succeed in the first iteration.
3786                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3787                     break;
3788             }
3789
3790             if(buf_index+3 >= buf_size) break;
3791
3792             buf_index+=3;
3793             if(buf_index >= next_avc) continue;
3794         }
3795
3796         hx = h->thread_context[context_count];
3797
3798         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed,
3799                                 next_avc - buf_index, !!nalsize);
3800         if (ptr==NULL || dst_length < 0){
3801             return -1;
3802         }
3803         i= buf_index + consumed;
3804         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3805            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3806             s->workaround_bugs |= FF_BUG_TRUNCATED;
3807
3808         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3809         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3810             dst_length--;
3811         }
3812         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3813
3814         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3815             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
3816         }
3817
3818         if (h->is_avc && (nalsize != consumed) && nalsize){
3819             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3820         }
3821
3822         buf_index += consumed;
3823         nal_index++;
3824
3825         if(pass == 0) {
3826             // packets can sometimes contain multiple PPS/SPS
3827             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3828             // if so, when frame threading we can't start the next thread until we've read all of them
3829             switch (hx->nal_unit_type) {
3830                 case NAL_SPS:
3831                 case NAL_PPS:
3832                     nals_needed = nal_index;
3833                     break;
3834                 case NAL_IDR_SLICE:
3835                 case NAL_SLICE:
3836                     init_get_bits(&hx->s.gb, ptr, bit_length);
3837                     if (!get_ue_golomb(&hx->s.gb))
3838                         nals_needed = nal_index;
3839             }
3840             continue;
3841         }
3842
3843         //FIXME do not discard SEI id
3844         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3845             continue;
3846
3847       again:
3848         err = 0;
3849         switch(hx->nal_unit_type){
3850         case NAL_IDR_SLICE:
3851             if (h->nal_unit_type != NAL_IDR_SLICE) {
3852                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3853                 return -1;
3854             }
3855             idr(h); // FIXME ensure we don't lose some frames if there is reordering
3856         case NAL_SLICE:
3857             init_get_bits(&hx->s.gb, ptr, bit_length);
3858             hx->intra_gb_ptr=
3859             hx->inter_gb_ptr= &hx->s.gb;
3860             hx->s.data_partitioning = 0;
3861
3862             if((err = decode_slice_header(hx, h)))
3863                break;
3864
3865             s->current_picture_ptr->f.key_frame |=
3866                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3867                     (h->sei_recovery_frame_cnt >= 0);
3868
3869             if (h->current_slice == 1) {
3870                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3871                     decode_postinit(h, nal_index >= nals_needed);
3872                 }
3873
3874                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3875                     return -1;
3876                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3877                     ff_vdpau_h264_picture_start(s);
3878             }
3879
3880             if(hx->redundant_pic_count==0
3881                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3882                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3883                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3884                && avctx->skip_frame < AVDISCARD_ALL){
3885                 if(avctx->hwaccel) {
3886                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3887                         return -1;
3888                 }else
3889                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3890                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3891                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3892                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3893                 }else
3894                     context_count++;
3895             }
3896             break;
3897         case NAL_DPA:
3898             init_get_bits(&hx->s.gb, ptr, bit_length);
3899             hx->intra_gb_ptr=
3900             hx->inter_gb_ptr= NULL;
3901
3902             if ((err = decode_slice_header(hx, h)) < 0)
3903                 break;
3904
3905             hx->s.data_partitioning = 1;
3906
3907             break;
3908         case NAL_DPB:
3909             init_get_bits(&hx->intra_gb, ptr, bit_length);
3910             hx->intra_gb_ptr= &hx->intra_gb;
3911             break;
3912         case NAL_DPC:
3913             init_get_bits(&hx->inter_gb, ptr, bit_length);
3914             hx->inter_gb_ptr= &hx->inter_gb;
3915
3916             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3917                && s->context_initialized
3918                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3919                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3920                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3921                && avctx->skip_frame < AVDISCARD_ALL)
3922                 context_count++;
3923             break;
3924         case NAL_SEI:
3925             init_get_bits(&s->gb, ptr, bit_length);
3926             ff_h264_decode_sei(h);
3927             break;
3928         case NAL_SPS:
3929             init_get_bits(&s->gb, ptr, bit_length);
3930             ff_h264_decode_seq_parameter_set(h);
3931
3932             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3933                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3934                 s->low_delay=1;
3935
3936             if(avctx->has_b_frames < 2)
3937                 avctx->has_b_frames= !s->low_delay;
3938
3939             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
3940                 h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
3941                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3942                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3943                     h->cur_chroma_format_idc = h->sps.chroma_format_idc;
3944                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3945
3946                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3947                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
3948                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3949                     dsputil_init(&s->dsp, s->avctx);
3950                 } else {
3951                     av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3952                     return -1;
3953                 }
3954             }
3955             break;
3956         case NAL_PPS:
3957             init_get_bits(&s->gb, ptr, bit_length);
3958
3959             ff_h264_decode_picture_parameter_set(h, bit_length);
3960
3961             break;
3962         case NAL_AUD:
3963         case NAL_END_SEQUENCE:
3964         case NAL_END_STREAM:
3965         case NAL_FILLER_DATA:
3966         case NAL_SPS_EXT:
3967         case NAL_AUXILIARY_SLICE:
3968             break;
3969         default:
3970             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3971         }
3972
3973         if(context_count == h->max_contexts) {
3974             execute_decode_slices(h, context_count);
3975             context_count = 0;
3976         }
3977
3978         if (err < 0)
3979             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3980         else if(err == 1) {
3981             /* Slice could not be decoded in parallel mode, copy down
3982              * NAL unit stuff to context 0 and restart. Note that
3983              * rbsp_buffer is not transferred, but since we no longer
3984              * run in parallel mode this should not be an issue. */
3985             h->nal_unit_type = hx->nal_unit_type;
3986             h->nal_ref_idc   = hx->nal_ref_idc;
3987             hx = h;
3988             goto again;
3989         }
3990     }
3991     }
3992     if(context_count)
3993         execute_decode_slices(h, context_count);
3994     return buf_index;
3995 }
3996
3997 /**
3998  * Return the number of bytes consumed for building the current frame.
3999  */
4000 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
4001         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
4002         if(pos+10>buf_size) pos=buf_size; // oops ;)
4003
4004         return pos;
4005 }
4006
4007 static int decode_frame(AVCodecContext *avctx,
4008                              void *data, int *data_size,
4009                              AVPacket *avpkt)
4010 {
4011     const uint8_t *buf = avpkt->data;
4012     int buf_size = avpkt->size;
4013     H264Context *h = avctx->priv_data;
4014     MpegEncContext *s = &h->s;
4015     AVFrame *pict = data;
4016     int buf_index;
4017
4018     s->flags= avctx->flags;
4019     s->flags2= avctx->flags2;
4020
4021    /* end of stream, output what is still in the buffers */
4022  out:
4023     if (buf_size == 0) {
4024         Picture *out;
4025         int i, out_idx;
4026
4027         s->current_picture_ptr = NULL;
4028
4029 //FIXME factorize this with the output code below
4030         out = h->delayed_pic[0];
4031         out_idx = 0;
4032         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
4033             if(h->delayed_pic[i]->poc < out->poc){
4034                 out = h->delayed_pic[i];
4035                 out_idx = i;
4036             }
4037
4038         for(i=out_idx; h->delayed_pic[i]; i++)
4039             h->delayed_pic[i] = h->delayed_pic[i+1];
4040
4041         if(out){
4042             *data_size = sizeof(AVFrame);
4043             *pict= *(AVFrame*)out;
4044         }
4045
4046         return 0;
4047     }
4048
4049     buf_index=decode_nal_units(h, buf, buf_size);
4050     if(buf_index < 0)
4051         return -1;
4052
4053     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4054         buf_size = 0;
4055         goto out;
4056     }
4057
4058     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
4059         if (avctx->skip_frame >= AVDISCARD_NONREF)
4060             return 0;
4061         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4062         return -1;
4063     }
4064
4065     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
4066
4067         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
4068
4069         field_end(h, 0);
4070
4071         if (!h->next_output_pic) {
4072             /* Wait for second field. */
4073             *data_size = 0;
4074
4075         } else {
4076             *data_size = sizeof(AVFrame);
4077             *pict = *(AVFrame*)h->next_output_pic;
4078         }
4079     }
4080
4081     assert(pict->data[0] || !*data_size);
4082     ff_print_debug_info(s, pict);
4083 //printf("out %d\n", (int)pict->data[0]);
4084
4085     return get_consumed_bytes(s, buf_index, buf_size);
4086 }
4087 #if 0
4088 static inline void fill_mb_avail(H264Context *h){
4089     MpegEncContext * const s = &h->s;
4090     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
4091
4092     if(s->mb_y){
4093         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
4094         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
4095         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
4096     }else{
4097         h->mb_avail[0]=
4098         h->mb_avail[1]=
4099         h->mb_avail[2]= 0;
4100     }
4101     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
4102     h->mb_avail[4]= 1; //FIXME move out
4103     h->mb_avail[5]= 0; //FIXME move out
4104 }
4105 #endif
4106
4107 #ifdef TEST
4108 #undef printf
4109 #undef random
4110 #define COUNT 8000
4111 #define SIZE (COUNT*40)
4112 int main(void){
4113     int i;
4114     uint8_t temp[SIZE];
4115     PutBitContext pb;
4116     GetBitContext gb;
4117 //    int int_temp[10000];
4118     DSPContext dsp;
4119     AVCodecContext avctx;
4120
4121     dsputil_init(&dsp, &avctx);
4122
4123     init_put_bits(&pb, temp, SIZE);
4124     printf("testing unsigned exp golomb\n");
4125     for(i=0; i<COUNT; i++){
4126         START_TIMER
4127         set_ue_golomb(&pb, i);
4128         STOP_TIMER("set_ue_golomb");
4129     }
4130     flush_put_bits(&pb);
4131
4132     init_get_bits(&gb, temp, 8*SIZE);
4133     for(i=0; i<COUNT; i++){
4134         int j, s;
4135
4136         s= show_bits(&gb, 24);
4137
4138         START_TIMER
4139         j= get_ue_golomb(&gb);
4140         if(j != i){
4141             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4142 //            return -1;
4143         }
4144         STOP_TIMER("get_ue_golomb");
4145     }
4146
4147
4148     init_put_bits(&pb, temp, SIZE);
4149     printf("testing signed exp golomb\n");
4150     for(i=0; i<COUNT; i++){
4151         START_TIMER
4152         set_se_golomb(&pb, i - COUNT/2);
4153         STOP_TIMER("set_se_golomb");
4154     }
4155     flush_put_bits(&pb);
4156
4157     init_get_bits(&gb, temp, 8*SIZE);
4158     for(i=0; i<COUNT; i++){
4159         int j, s;
4160
4161         s= show_bits(&gb, 24);
4162
4163         START_TIMER
4164         j= get_se_golomb(&gb);
4165         if(j != i - COUNT/2){
4166             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4167 //            return -1;
4168         }
4169         STOP_TIMER("get_se_golomb");
4170     }
4171
4172     printf("Testing RBSP\n");
4173
4174
4175     return 0;
4176 }
4177 #endif /* TEST */
4178
4179
4180 av_cold void ff_h264_free_context(H264Context *h)
4181 {
4182     int i;
4183
4184     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4185
4186     for(i = 0; i < MAX_SPS_COUNT; i++)
4187         av_freep(h->sps_buffers + i);
4188
4189     for(i = 0; i < MAX_PPS_COUNT; i++)
4190         av_freep(h->pps_buffers + i);
4191 }
4192
4193 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4194 {
4195     H264Context *h = avctx->priv_data;
4196     MpegEncContext *s = &h->s;
4197
4198     ff_h264_free_context(h);
4199
4200     MPV_common_end(s);
4201
4202 //    memset(h, 0, sizeof(H264Context));
4203
4204     return 0;
4205 }
4206
4207 static const AVProfile profiles[] = {
4208     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4209     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4210     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4211     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4212     { FF_PROFILE_H264_HIGH,                 "High"                  },
4213     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4214     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4215     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4216     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4217     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4218     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4219     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4220     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4221     { FF_PROFILE_UNKNOWN },
4222 };
4223
4224 AVCodec ff_h264_decoder = {
4225     .name           = "h264",
4226     .type           = AVMEDIA_TYPE_VIDEO,
4227     .id             = CODEC_ID_H264,
4228     .priv_data_size = sizeof(H264Context),
4229     .init           = ff_h264_decode_init,
4230     .close          = ff_h264_decode_end,
4231     .decode         = decode_frame,
4232     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4233                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4234     .flush= flush_dpb,
4235     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4236     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4237     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4238     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4239 };
4240
4241 #if CONFIG_H264_VDPAU_DECODER
4242 AVCodec ff_h264_vdpau_decoder = {
4243     .name           = "h264_vdpau",
4244     .type           = AVMEDIA_TYPE_VIDEO,
4245     .id             = CODEC_ID_H264,
4246     .priv_data_size = sizeof(H264Context),
4247     .init           = ff_h264_decode_init,
4248     .close          = ff_h264_decode_end,
4249     .decode         = decode_frame,
4250     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4251     .flush= flush_dpb,
4252     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4253     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4254     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4255 };
4256 #endif