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_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 /**
  64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  65  */
  66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  67     MpegEncContext * const s = &h->s;
  68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  70     int i;
  71
  72     if(!(h->top_samples_available&0x8000)){
  73         for(i=0; i<4; i++){
  74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  75             if(status<0){
  76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  77                 return -1;
  78             } else if(status){
  79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  80             }
  81         }
  82     }
  83
  84     if((h->left_samples_available&0x8888)!=0x8888){
  85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  86         for(i=0; i<4; i++){
  87             if(!(h->left_samples_available&mask[i])){
  88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  89                 if(status<0){
  90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  91                     return -1;
  92                 } else if(status){
  93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  94                 }
  95             }
  96         }
  97     }
  98
  99     return 0;
 100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 101
 102 /**
 103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 104  */
 105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 106     MpegEncContext * const s = &h->s;
 107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 109
 110     if(mode > 6U) {
 111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 112         return -1;
 113     }
 114
 115     if(!(h->top_samples_available&0x8000)){
 116         mode= top[ mode ];
 117         if(mode<0){
 118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 119             return -1;
 120         }
 121     }
 122
 123     if((h->left_samples_available&0x8080) != 0x8080){
 124         mode= left[ mode ];
 125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 127         }
 128         if(mode<0){
 129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 130             return -1;
 131         }
 132     }
 133
 134     return mode;
 135 }
 136
 137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 138     int i, si, di;
 139     uint8_t *dst;
 140     int bufidx;
 141
 142 //    src[0]&0x80;                //forbidden bit
 143     h->nal_ref_idc= src[0]>>5;
 144     h->nal_unit_type= src[0]&0x1F;
 145
 146     src++; length--;
 147
 148 #if HAVE_FAST_UNALIGNED
 149 # if HAVE_FAST_64BIT
 150 #   define RS 7
 151     for(i=0; i+1<length; i+=9){
 152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 153 # else
 154 #   define RS 3
 155     for(i=0; i+1<length; i+=5){
 156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 157 # endif
 158             continue;
 159         if(i>0 && !src[i]) i--;
 160         while(src[i]) i++;
 161 #else
 162 #   define RS 0
 163     for(i=0; i+1<length; i+=2){
 164         if(src[i]) continue;
 165         if(i>0 && src[i-1]==0) i--;
 166 #endif
 167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 168             if(src[i+2]!=3){
 169                 /* startcode, so we must be past the end */
 170                 length=i;
 171             }
 172             break;
 173         }
 174         i-= RS;
 175     }
 176
 177     if(i>=length-1){ //no escaped 0
 178         *dst_length= length;
 179         *consumed= length+1; //+1 for the header
 180         return src;
 181     }
 182
 183     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 184     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 185     dst= h->rbsp_buffer[bufidx];
 186
 187     if (dst == NULL){
 188         return NULL;
 189     }
 190
 191 //printf("decoding esc\n");
 192     memcpy(dst, src, i);
 193     si=di=i;
 194     while(si+2<length){
 195         //remove escapes (very rare 1:2^22)
 196         if(src[si+2]>3){
 197             dst[di++]= src[si++];
 198             dst[di++]= src[si++];
 199         }else if(src[si]==0 && src[si+1]==0){
 200             if(src[si+2]==3){ //escape
 201                 dst[di++]= 0;
 202                 dst[di++]= 0;
 203                 si+=3;
 204                 continue;
 205             }else //next start code
 206                 goto nsc;
 207         }
 208
 209         dst[di++]= src[si++];
 210     }
 211     while(si<length)
 212         dst[di++]= src[si++];
 213 nsc:
 214
 215     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 216
 217     *dst_length= di;
 218     *consumed= si + 1;//+1 for the header
 219 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 220     return dst;
 221 }
 222
 223 /**
 224  * Identify the exact end of the bitstream
 225  * @return the length of the trailing, or 0 if damaged
 226  */
 227 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 228     int v= *src;
 229     int r;
 230
 231     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 232
 233     for(r=1; r<9; r++){
 234         if(v&1) return r;
 235         v>>=1;
 236     }
 237     return 0;
 238 }
 239
 240 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 241                                  int y_offset, int list){
 242     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 243     int filter_height= (raw_my&3) ? 2 : 0;
 244     int full_my= (raw_my>>2) + y_offset;
 245     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 246
 247     return FFMAX(abs(top), bottom);
 248 }
 249
 250 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 251                                int y_offset, int list0, int list1, int *nrefs){
 252     MpegEncContext * const s = &h->s;
 253     int my;
 254
 255     y_offset += 16*(s->mb_y >> MB_FIELD);
 256
 257     if(list0){
 258         int ref_n = h->ref_cache[0][ scan8[n] ];
 259         Picture *ref= &h->ref_list[0][ref_n];
 260
 261         // Error resilience puts the current picture in the ref list.
 262         // Don't try to wait on these as it will cause a deadlock.
 263         // Fields can wait on each other, though.
 264         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 265            (ref->f.reference & 3) != s->picture_structure) {
 266             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 267             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 268             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 269         }
 270     }
 271
 272     if(list1){
 273         int ref_n = h->ref_cache[1][ scan8[n] ];
 274         Picture *ref= &h->ref_list[1][ref_n];
 275
 276         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 277            (ref->f.reference & 3) != s->picture_structure) {
 278             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 279             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 280             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 281         }
 282     }
 283 }
 284
 285 /**
 286  * Wait until all reference frames are available for MC operations.
 287  *
 288  * @param h the H264 context
 289  */
 290 static void await_references(H264Context *h){
 291     MpegEncContext * const s = &h->s;
 292     const int mb_xy= h->mb_xy;
 293     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 294     int refs[2][48];
 295     int nrefs[2] = {0};
 296     int ref, list;
 297
 298     memset(refs, -1, sizeof(refs));
 299
 300     if(IS_16X16(mb_type)){
 301         get_lowest_part_y(h, refs, 0, 16, 0,
 302                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 303     }else if(IS_16X8(mb_type)){
 304         get_lowest_part_y(h, refs, 0, 8, 0,
 305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 306         get_lowest_part_y(h, refs, 8, 8, 8,
 307                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 308     }else if(IS_8X16(mb_type)){
 309         get_lowest_part_y(h, refs, 0, 16, 0,
 310                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 311         get_lowest_part_y(h, refs, 4, 16, 0,
 312                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 313     }else{
 314         int i;
 315
 316         assert(IS_8X8(mb_type));
 317
 318         for(i=0; i<4; i++){
 319             const int sub_mb_type= h->sub_mb_type[i];
 320             const int n= 4*i;
 321             int y_offset= (i&2)<<2;
 322
 323             if(IS_SUB_8X8(sub_mb_type)){
 324                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 325                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 326             }else if(IS_SUB_8X4(sub_mb_type)){
 327                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 329                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 330                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 331             }else if(IS_SUB_4X8(sub_mb_type)){
 332                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 334                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 335                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 336             }else{
 337                 int j;
 338                 assert(IS_SUB_4X4(sub_mb_type));
 339                 for(j=0; j<4; j++){
 340                     int sub_y_offset= y_offset + 2*(j&2);
 341                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 342                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 343                 }
 344             }
 345         }
 346     }
 347
 348     for(list=h->list_count-1; list>=0; list--){
 349         for(ref=0; ref<48 && nrefs[list]; ref++){
 350             int row = refs[list][ref];
 351             if(row >= 0){
 352                 Picture *ref_pic = &h->ref_list[list][ref];
 353                 int ref_field = ref_pic->f.reference - 1;
 354                 int ref_field_picture = ref_pic->field_picture;
 355                 int pic_height = 16*s->mb_height >> ref_field_picture;
 356
 357                 row <<= MB_MBAFF;
 358                 nrefs[list]--;
 359
 360                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 361                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 363                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 365                 }else if(FIELD_PICTURE){
 366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 367                 }else{
 368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 369                 }
 370             }
 371         }
 372     }
 373 }
 374
 375 #if 0
 376 /**
 377  * DCT transforms the 16 dc values.
 378  * @param qp quantization parameter ??? FIXME
 379  */
 380 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 381 //    const int qmul= dequant_coeff[qp][0];
 382     int i;
 383     int temp[16]; //FIXME check if this is a good idea
 384     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 385     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 386
 387     for(i=0; i<4; i++){
 388         const int offset= y_offset[i];
 389         const int z0= block[offset+stride*0] + block[offset+stride*4];
 390         const int z1= block[offset+stride*0] - block[offset+stride*4];
 391         const int z2= block[offset+stride*1] - block[offset+stride*5];
 392         const int z3= block[offset+stride*1] + block[offset+stride*5];
 393
 394         temp[4*i+0]= z0+z3;
 395         temp[4*i+1]= z1+z2;
 396         temp[4*i+2]= z1-z2;
 397         temp[4*i+3]= z0-z3;
 398     }
 399
 400     for(i=0; i<4; i++){
 401         const int offset= x_offset[i];
 402         const int z0= temp[4*0+i] + temp[4*2+i];
 403         const int z1= temp[4*0+i] - temp[4*2+i];
 404         const int z2= temp[4*1+i] - temp[4*3+i];
 405         const int z3= temp[4*1+i] + temp[4*3+i];
 406
 407         block[stride*0 +offset]= (z0 + z3)>>1;
 408         block[stride*2 +offset]= (z1 + z2)>>1;
 409         block[stride*8 +offset]= (z1 - z2)>>1;
 410         block[stride*10+offset]= (z0 - z3)>>1;
 411     }
 412 }
 413 #endif
 414
 415 #undef xStride
 416 #undef stride
 417
 418 #if 0
 419 static void chroma_dc_dct_c(DCTELEM *block){
 420     const int stride= 16*2;
 421     const int xStride= 16;
 422     int a,b,c,d,e;
 423
 424     a= block[stride*0 + xStride*0];
 425     b= block[stride*0 + xStride*1];
 426     c= block[stride*1 + xStride*0];
 427     d= block[stride*1 + xStride*1];
 428
 429     e= a-b;
 430     a= a+b;
 431     b= c-d;
 432     c= c+d;
 433
 434     block[stride*0 + xStride*0]= (a+c);
 435     block[stride*0 + xStride*1]= (e+b);
 436     block[stride*1 + xStride*0]= (a-c);
 437     block[stride*1 + xStride*1]= (e-b);
 438 }
 439 #endif
 440
 441 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 442                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 443                            int src_x_offset, int src_y_offset,
 444                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 445                            int pixel_shift, int chroma444){
 446     MpegEncContext * const s = &h->s;
 447     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 448     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 449     const int luma_xy= (mx&3) + ((my&3)<<2);
 450     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 451     uint8_t * src_y = pic->f.data[0] + offset;
 452     uint8_t * src_cb, * src_cr;
 453     int extra_width= h->emu_edge_width;
 454     int extra_height= h->emu_edge_height;
 455     int emu=0;
 456     const int full_mx= mx>>2;
 457     const int full_my= my>>2;
 458     const int pic_width  = 16*s->mb_width;
 459     const int pic_height = 16*s->mb_height >> MB_FIELD;
 460
 461     if(mx&7) extra_width -= 3;
 462     if(my&7) extra_height -= 3;
 463
 464     if(   full_mx < 0-extra_width
 465        || full_my < 0-extra_height
 466        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 467        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 468         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 469             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 470         emu=1;
 471     }
 472
 473     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 474     if(!square){
 475         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 476     }
 477
 478     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 479
 480     if(chroma444){
 481         src_cb = pic->f.data[1] + offset;
 482         if(emu){
 483             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 484                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 485             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 486         }
 487         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 488         if(!square){
 489             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 490         }
 491
 492         src_cr = pic->f.data[2] + offset;
 493         if(emu){
 494             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 495                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 496             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 497         }
 498         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 499         if(!square){
 500             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 501         }
 502         return;
 503     }
 504
 505     if(MB_FIELD){
 506         // chroma offset when predicting from a field of opposite parity
 507         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 509     }
 510     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 511     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 512
 513     if(emu){
 514         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 515             src_cb= s->edge_emu_buffer;
 516     }
 517     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 518
 519     if(emu){
 520         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 521             src_cr= s->edge_emu_buffer;
 522     }
 523     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 524 }
 525
 526 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 527                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 528                            int x_offset, int y_offset,
 529                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 530                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 531                            int list0, int list1, int pixel_shift, int chroma444){
 532     MpegEncContext * const s = &h->s;
 533     qpel_mc_func *qpix_op=  qpix_put;
 534     h264_chroma_mc_func chroma_op= chroma_put;
 535
 536     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 537     if(chroma444){
 538         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 539         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 540     }else{
 541         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 542         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 543     }
 544     x_offset += 8*s->mb_x;
 545     y_offset += 8*(s->mb_y >> MB_FIELD);
 546
 547     if(list0){
 548         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 549         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 550                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 551                            qpix_op, chroma_op, pixel_shift, chroma444);
 552
 553         qpix_op=  qpix_avg;
 554         chroma_op= chroma_avg;
 555     }
 556
 557     if(list1){
 558         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 559         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 560                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 561                            qpix_op, chroma_op, pixel_shift, chroma444);
 562     }
 563 }
 564
 565 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 566                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 567                            int x_offset, int y_offset,
 568                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 569                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 570                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 571                            int list0, int list1, int pixel_shift, int chroma444){
 572     MpegEncContext * const s = &h->s;
 573
 574     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 575     if(chroma444){
 576         chroma_weight_avg = luma_weight_avg;
 577         chroma_weight_op = luma_weight_op;
 578         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 579         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 580     }else{
 581         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 582         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 583     }
 584     x_offset += 8*s->mb_x;
 585     y_offset += 8*(s->mb_y >> MB_FIELD);
 586
 587     if(list0 && list1){
 588         /* don't optimize for luma-only case, since B-frames usually
 589          * use implicit weights => chroma too. */
 590         uint8_t *tmp_cb = s->obmc_scratchpad;
 591         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 592         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 593         int refn0 = h->ref_cache[0][ scan8[n] ];
 594         int refn1 = h->ref_cache[1][ scan8[n] ];
 595
 596         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 597                     dest_y, dest_cb, dest_cr,
 598                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 599         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 600                     tmp_y, tmp_cb, tmp_cr,
 601                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 602
 603         if(h->use_weight == 2){
 604             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 605             int weight1 = 64 - weight0;
 606             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 607             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 608             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 609         }else{
 610             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 611                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 612                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 613             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 614                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 615                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 616             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 617                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 618                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 619         }
 620     }else{
 621         int list = list1 ? 1 : 0;
 622         int refn = h->ref_cache[list][ scan8[n] ];
 623         Picture *ref= &h->ref_list[list][refn];
 624         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 625                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 626                     qpix_put, chroma_put, pixel_shift, chroma444);
 627
 628         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 629                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 630         if(h->use_weight_chroma){
 631             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 632                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 633             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 634                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 635         }
 636     }
 637 }
 638
 639 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 640                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 641                            int x_offset, int y_offset,
 642                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 643                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 644                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 645                            int list0, int list1, int pixel_shift, int chroma444){
 646     if((h->use_weight==2 && list0 && list1
 647         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 648        || h->use_weight==1)
 649         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 650                          x_offset, y_offset, qpix_put, chroma_put,
 651                          weight_op[0], weight_op[3], weight_avg[0],
 652                          weight_avg[3], list0, list1, pixel_shift, chroma444);
 653     else
 654         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 655                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 656                     chroma_avg, list0, list1, pixel_shift, chroma444);
 657 }
 658
 659 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
 660     /* fetch pixels for estimated mv 4 macroblocks ahead
 661      * optimized for 64byte cache lines */
 662     MpegEncContext * const s = &h->s;
 663     const int refn = h->ref_cache[list][scan8[0]];
 664     if(refn >= 0){
 665         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 666         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 667         uint8_t **src = h->ref_list[list][refn].f.data;
 668         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 669         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 670         if(chroma444){
 671             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 672             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 673         }else{
 674             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
 675             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 676         }
 677     }
 678 }
 679
 680 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 681                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 682                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 683                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 684                       int pixel_shift, int chroma444){
 685     MpegEncContext * const s = &h->s;
 686     const int mb_xy= h->mb_xy;
 687     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 688
 689     assert(IS_INTER(mb_type));
 690
 691     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 692         await_references(h);
 693     prefetch_motion(h, 0, pixel_shift, chroma444);
 694
 695     if(IS_16X16(mb_type)){
 696         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 697                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 698                 weight_op, weight_avg,
 699                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 700                 pixel_shift, chroma444);
 701     }else if(IS_16X8(mb_type)){
 702         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 703                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 704                 &weight_op[1], &weight_avg[1],
 705                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 706                 pixel_shift, chroma444);
 707         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 708                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 709                 &weight_op[1], &weight_avg[1],
 710                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 711                 pixel_shift, chroma444);
 712     }else if(IS_8X16(mb_type)){
 713         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 714                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 715                 &weight_op[2], &weight_avg[2],
 716                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 717                 pixel_shift, chroma444);
 718         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 719                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 720                 &weight_op[2], &weight_avg[2],
 721                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 722                 pixel_shift, chroma444);
 723     }else{
 724         int i;
 725
 726         assert(IS_8X8(mb_type));
 727
 728         for(i=0; i<4; i++){
 729             const int sub_mb_type= h->sub_mb_type[i];
 730             const int n= 4*i;
 731             int x_offset= (i&1)<<2;
 732             int y_offset= (i&2)<<1;
 733
 734             if(IS_SUB_8X8(sub_mb_type)){
 735                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 736                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 737                     &weight_op[3], &weight_avg[3],
 738                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 739                     pixel_shift, chroma444);
 740             }else if(IS_SUB_8X4(sub_mb_type)){
 741                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 742                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 743                     &weight_op[4], &weight_avg[4],
 744                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 745                     pixel_shift, chroma444);
 746                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 747                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 748                     &weight_op[4], &weight_avg[4],
 749                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 750                     pixel_shift, chroma444);
 751             }else if(IS_SUB_4X8(sub_mb_type)){
 752                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 753                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 754                     &weight_op[5], &weight_avg[5],
 755                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 756                     pixel_shift, chroma444);
 757                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 758                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 759                     &weight_op[5], &weight_avg[5],
 760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 761                     pixel_shift, chroma444);
 762             }else{
 763                 int j;
 764                 assert(IS_SUB_4X4(sub_mb_type));
 765                 for(j=0; j<4; j++){
 766                     int sub_x_offset= x_offset + 2*(j&1);
 767                     int sub_y_offset= y_offset +   (j&2);
 768                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 769                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 770                         &weight_op[6], &weight_avg[6],
 771                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 772                         pixel_shift, chroma444);
 773                 }
 774             }
 775         }
 776     }
 777
 778     prefetch_motion(h, 1, pixel_shift, chroma444);
 779 }
 780
 781 static void free_tables(H264Context *h, int free_rbsp){
 782     int i;
 783     H264Context *hx;
 784
 785     av_freep(&h->intra4x4_pred_mode);
 786     av_freep(&h->chroma_pred_mode_table);
 787     av_freep(&h->cbp_table);
 788     av_freep(&h->mvd_table[0]);
 789     av_freep(&h->mvd_table[1]);
 790     av_freep(&h->direct_table);
 791     av_freep(&h->non_zero_count);
 792     av_freep(&h->slice_table_base);
 793     h->slice_table= NULL;
 794     av_freep(&h->list_counts);
 795
 796     av_freep(&h->mb2b_xy);
 797     av_freep(&h->mb2br_xy);
 798
 799     for(i = 0; i < MAX_THREADS; i++) {
 800         hx = h->thread_context[i];
 801         if(!hx) continue;
 802         av_freep(&hx->top_borders[1]);
 803         av_freep(&hx->top_borders[0]);
 804         av_freep(&hx->s.obmc_scratchpad);
 805         if (free_rbsp){
 806             av_freep(&hx->rbsp_buffer[1]);
 807             av_freep(&hx->rbsp_buffer[0]);
 808             hx->rbsp_buffer_size[0] = 0;
 809             hx->rbsp_buffer_size[1] = 0;
 810         }
 811         if (i) av_freep(&h->thread_context[i]);
 812     }
 813 }
 814
 815 static void init_dequant8_coeff_table(H264Context *h){
 816     int i,j,q,x;
 817     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 818
 819     for(i=0; i<6; i++ ){
 820         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 821         for(j=0; j<i; j++){
 822             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 823                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 824                 break;
 825             }
 826         }
 827         if(j<i)
 828             continue;
 829
 830         for(q=0; q<max_qp+1; q++){
 831             int shift = div6[q];
 832             int idx = rem6[q];
 833             for(x=0; x<64; x++)
 834                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 835                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 836                     h->pps.scaling_matrix8[i][x]) << shift;
 837         }
 838     }
 839 }
 840
 841 static void init_dequant4_coeff_table(H264Context *h){
 842     int i,j,q,x;
 843     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 844     for(i=0; i<6; i++ ){
 845         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 846         for(j=0; j<i; j++){
 847             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 848                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 849                 break;
 850             }
 851         }
 852         if(j<i)
 853             continue;
 854
 855         for(q=0; q<max_qp+1; q++){
 856             int shift = div6[q] + 2;
 857             int idx = rem6[q];
 858             for(x=0; x<16; x++)
 859                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 860                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 861                     h->pps.scaling_matrix4[i][x]) << shift;
 862         }
 863     }
 864 }
 865
 866 static void init_dequant_tables(H264Context *h){
 867     int i,x;
 868     init_dequant4_coeff_table(h);
 869     if(h->pps.transform_8x8_mode)
 870         init_dequant8_coeff_table(h);
 871     if(h->sps.transform_bypass){
 872         for(i=0; i<6; i++)
 873             for(x=0; x<16; x++)
 874                 h->dequant4_coeff[i][0][x] = 1<<6;
 875         if(h->pps.transform_8x8_mode)
 876             for(i=0; i<6; i++)
 877                 for(x=0; x<64; x++)
 878                     h->dequant8_coeff[i][0][x] = 1<<6;
 879     }
 880 }
 881
 882
 883 int ff_h264_alloc_tables(H264Context *h){
 884     MpegEncContext * const s = &h->s;
 885     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 886     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 887     int x,y;
 888
 889     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 890
 891     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 892     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 894
 895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 896     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 898     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 900
 901     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 902     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 903
 904     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 905     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 906     for(y=0; y<s->mb_height; y++){
 907         for(x=0; x<s->mb_width; x++){
 908             const int mb_xy= x + y*s->mb_stride;
 909             const int b_xy = 4*x + 4*y*h->b_stride;
 910
 911             h->mb2b_xy [mb_xy]= b_xy;
 912             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 913         }
 914     }
 915
 916     s->obmc_scratchpad = NULL;
 917
 918     if(!h->dequant4_coeff[0])
 919         init_dequant_tables(h);
 920
 921     return 0;
 922 fail:
 923     free_tables(h, 1);
 924     return -1;
 925 }
 926
 927 /**
 928  * Mimic alloc_tables(), but for every context thread.
 929  */
 930 static void clone_tables(H264Context *dst, H264Context *src, int i){
 931     MpegEncContext * const s = &src->s;
 932     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 933     dst->non_zero_count           = src->non_zero_count;
 934     dst->slice_table              = src->slice_table;
 935     dst->cbp_table                = src->cbp_table;
 936     dst->mb2b_xy                  = src->mb2b_xy;
 937     dst->mb2br_xy                 = src->mb2br_xy;
 938     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 939     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 940     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 941     dst->direct_table             = src->direct_table;
 942     dst->list_counts              = src->list_counts;
 943
 944     dst->s.obmc_scratchpad = NULL;
 945     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
 946 }
 947
 948 /**
 949  * Init context
 950  * Allocate buffers which are not shared amongst multiple threads.
 951  */
 952 static int context_init(H264Context *h){
 953     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 955
 956     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 957     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 958
 959     return 0;
 960 fail:
 961     return -1; // free_tables will clean up for us
 962 }
 963
 964 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 965
 966 static av_cold void common_init(H264Context *h){
 967     MpegEncContext * const s = &h->s;
 968
 969     s->width = s->avctx->width;
 970     s->height = s->avctx->height;
 971     s->codec_id= s->avctx->codec->id;
 972
 973     ff_h264dsp_init(&h->h264dsp, 8);
 974     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
 975
 976     h->dequant_coeff_pps= -1;
 977     s->unrestricted_mv=1;
 978     s->decode=1; //FIXME
 979
 980     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 981
 982     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 983     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 984 }
 985
 986 int ff_h264_decode_extradata(H264Context *h)
 987 {
 988     AVCodecContext *avctx = h->s.avctx;
 989
 990     if(avctx->extradata[0] == 1){
 991         int i, cnt, nalsize;
 992         unsigned char *p = avctx->extradata;
 993
 994         h->is_avc = 1;
 995
 996         if(avctx->extradata_size < 7) {
 997             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
 998             return -1;
 999         }
1000         /* sps and pps in the avcC always have length coded with 2 bytes,
1001            so put a fake nal_length_size = 2 while parsing them */
1002         h->nal_length_size = 2;
1003         // Decode sps from avcC
1004         cnt = *(p+5) & 0x1f; // Number of sps
1005         p += 6;
1006         for (i = 0; i < cnt; i++) {
1007             nalsize = AV_RB16(p) + 2;
1008             if(decode_nal_units(h, p, nalsize) < 0) {
1009                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1010                 return -1;
1011             }
1012             p += nalsize;
1013         }
1014         // Decode pps from avcC
1015         cnt = *(p++); // Number of pps
1016         for (i = 0; i < cnt; i++) {
1017             nalsize = AV_RB16(p) + 2;
1018             if (decode_nal_units(h, p, nalsize) < 0) {
1019                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1020                 return -1;
1021             }
1022             p += nalsize;
1023         }
1024         // Now store right nal length size, that will be use to parse all other nals
1025         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1026     } else {
1027         h->is_avc = 0;
1028         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1029             return -1;
1030     }
1031     return 0;
1032 }
1033
1034 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1035     H264Context *h= avctx->priv_data;
1036     MpegEncContext * const s = &h->s;
1037
1038     MPV_decode_defaults(s);
1039
1040     s->avctx = avctx;
1041     common_init(h);
1042
1043     s->out_format = FMT_H264;
1044     s->workaround_bugs= avctx->workaround_bugs;
1045
1046     // set defaults
1047 //    s->decode_mb= ff_h263_decode_mb;
1048     s->quarter_sample = 1;
1049     if(!avctx->has_b_frames)
1050     s->low_delay= 1;
1051
1052     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1053
1054     ff_h264_decode_init_vlc();
1055
1056     h->pixel_shift = 0;
1057     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1058
1059     h->thread_context[0] = h;
1060     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1061     h->prev_poc_msb= 1<<16;
1062     h->x264_build = -1;
1063     ff_h264_reset_sei(h);
1064     if(avctx->codec_id == CODEC_ID_H264){
1065         if(avctx->ticks_per_frame == 1){
1066             s->avctx->time_base.den *=2;
1067         }
1068         avctx->ticks_per_frame = 2;
1069     }
1070
1071     if(avctx->extradata_size > 0 && avctx->extradata &&
1072         ff_h264_decode_extradata(h))
1073         return -1;
1074
1075     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1076         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1077         s->low_delay = 0;
1078     }
1079
1080     return 0;
1081 }
1082
1083 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1084 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1085 {
1086     int i;
1087
1088     for (i=0; i<count; i++){
1089         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1090                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1091                 !from[i]));
1092         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1093     }
1094 }
1095
1096 static void copy_parameter_set(void **to, void **from, int count, int size)
1097 {
1098     int i;
1099
1100     for (i=0; i<count; i++){
1101         if (to[i] && !from[i]) av_freep(&to[i]);
1102         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1103
1104         if (from[i]) memcpy(to[i], from[i], size);
1105     }
1106 }
1107
1108 static int decode_init_thread_copy(AVCodecContext *avctx){
1109     H264Context *h= avctx->priv_data;
1110
1111     if (!avctx->is_copy) return 0;
1112     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1113     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1114
1115     return 0;
1116 }
1117
1118 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1119 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1120     H264Context *h= dst->priv_data, *h1= src->priv_data;
1121     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1122     int inited = s->context_initialized, err;
1123     int i;
1124
1125     if(dst == src || !s1->context_initialized) return 0;
1126
1127     err = ff_mpeg_update_thread_context(dst, src);
1128     if(err) return err;
1129
1130     //FIXME handle width/height changing
1131     if(!inited){
1132         for(i = 0; i < MAX_SPS_COUNT; i++)
1133             av_freep(h->sps_buffers + i);
1134
1135         for(i = 0; i < MAX_PPS_COUNT; i++)
1136             av_freep(h->pps_buffers + i);
1137
1138         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1139         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1140         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1141         ff_h264_alloc_tables(h);
1142         context_init(h);
1143
1144         for(i=0; i<2; i++){
1145             h->rbsp_buffer[i] = NULL;
1146             h->rbsp_buffer_size[i] = 0;
1147         }
1148
1149         h->thread_context[0] = h;
1150
1151         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1152         // so this has to be allocated here
1153         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1154
1155         s->dsp.clear_blocks(h->mb);
1156         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1157     }
1158
1159     //extradata/NAL handling
1160     h->is_avc          = h1->is_avc;
1161
1162     //SPS/PPS
1163     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1164     h->sps             = h1->sps;
1165     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1166     h->pps             = h1->pps;
1167
1168     //Dequantization matrices
1169     //FIXME these are big - can they be only copied when PPS changes?
1170     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1171
1172     for(i=0; i<6; i++)
1173         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1174
1175     for(i=0; i<6; i++)
1176         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1177
1178     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1179
1180     //POC timing
1181     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1182
1183     //reference lists
1184     copy_fields(h, h1, ref_count, list_count);
1185     copy_fields(h, h1, ref_list,  intra_gb);
1186     copy_fields(h, h1, short_ref, cabac_init_idc);
1187
1188     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1189     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1190     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1191
1192     h->last_slice_type = h1->last_slice_type;
1193
1194     if(!s->current_picture_ptr) return 0;
1195
1196     if(!s->dropable) {
1197         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1198         h->prev_poc_msb     = h->poc_msb;
1199         h->prev_poc_lsb     = h->poc_lsb;
1200     }
1201     h->prev_frame_num_offset= h->frame_num_offset;
1202     h->prev_frame_num       = h->frame_num;
1203     h->outputed_poc         = h->next_outputed_poc;
1204
1205     return 0;
1206 }
1207
1208 int ff_h264_frame_start(H264Context *h){
1209     MpegEncContext * const s = &h->s;
1210     int i;
1211     const int pixel_shift = h->pixel_shift;
1212     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1213
1214     if(MPV_frame_start(s, s->avctx) < 0)
1215         return -1;
1216     ff_er_frame_start(s);
1217     /*
1218      * MPV_frame_start uses pict_type to derive key_frame.
1219      * This is incorrect for H.264; IDR markings must be used.
1220      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1221      * See decode_nal_units().
1222      */
1223     s->current_picture_ptr->f.key_frame = 0;
1224     s->current_picture_ptr->mmco_reset= 0;
1225
1226     assert(s->linesize && s->uvlinesize);
1227
1228     for(i=0; i<16; i++){
1229         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1230         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1231     }
1232     for(i=0; i<16; i++){
1233         h->block_offset[16+i]=
1234         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1235         h->block_offset[48+16+i]=
1236         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1237     }
1238
1239     /* can't be in alloc_tables because linesize isn't known there.
1240      * FIXME: redo bipred weight to not require extra buffer? */
1241     for(i = 0; i < thread_count; i++)
1242         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1243             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1244
1245     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1246     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1247
1248 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1249
1250     // We mark the current picture as non-reference after allocating it, so
1251     // that if we break out due to an error it can be released automatically
1252     // in the next MPV_frame_start().
1253     // SVQ3 as well as most other codecs have only last/next/current and thus
1254     // get released even with set reference, besides SVQ3 and others do not
1255     // mark frames as reference later "naturally".
1256     if(s->codec_id != CODEC_ID_SVQ3)
1257         s->current_picture_ptr->f.reference = 0;
1258
1259     s->current_picture_ptr->field_poc[0]=
1260     s->current_picture_ptr->field_poc[1]= INT_MAX;
1261
1262     h->next_output_pic = NULL;
1263
1264     assert(s->current_picture_ptr->long_ref==0);
1265
1266     return 0;
1267 }
1268
1269 /**
1270   * Run setup operations that must be run after slice header decoding.
1271   * This includes finding the next displayed frame.
1272   *
1273   * @param h h264 master context
1274   * @param setup_finished enough NALs have been read that we can call
1275   * ff_thread_finish_setup()
1276   */
1277 static void decode_postinit(H264Context *h, int setup_finished){
1278     MpegEncContext * const s = &h->s;
1279     Picture *out = s->current_picture_ptr;
1280     Picture *cur = s->current_picture_ptr;
1281     int i, pics, out_of_order, out_idx;
1282
1283     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1284     s->current_picture_ptr->f.pict_type   = s->pict_type;
1285
1286     if (h->next_output_pic) return;
1287
1288     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1289         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1290         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1291         //to find this yet, so we assume the worst for now.
1292         //if (setup_finished)
1293         //    ff_thread_finish_setup(s->avctx);
1294         return;
1295     }
1296
1297     cur->f.interlaced_frame = 0;
1298     cur->f.repeat_pict      = 0;
1299
1300     /* Signal interlacing information externally. */
1301     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1302
1303     if(h->sps.pic_struct_present_flag){
1304         switch (h->sei_pic_struct)
1305         {
1306         case SEI_PIC_STRUCT_FRAME:
1307             break;
1308         case SEI_PIC_STRUCT_TOP_FIELD:
1309         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1310             cur->f.interlaced_frame = 1;
1311             break;
1312         case SEI_PIC_STRUCT_TOP_BOTTOM:
1313         case SEI_PIC_STRUCT_BOTTOM_TOP:
1314             if (FIELD_OR_MBAFF_PICTURE)
1315                 cur->f.interlaced_frame = 1;
1316             else
1317                 // try to flag soft telecine progressive
1318                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1319             break;
1320         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1321         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1322             // Signal the possibility of telecined film externally (pic_struct 5,6)
1323             // From these hints, let the applications decide if they apply deinterlacing.
1324             cur->f.repeat_pict = 1;
1325             break;
1326         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1327             // Force progressive here, as doubling interlaced frame is a bad idea.
1328             cur->f.repeat_pict = 2;
1329             break;
1330         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1331             cur->f.repeat_pict = 4;
1332             break;
1333         }
1334
1335         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1336             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1337     }else{
1338         /* Derive interlacing flag from used decoding process. */
1339         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1340     }
1341     h->prev_interlaced_frame = cur->f.interlaced_frame;
1342
1343     if (cur->field_poc[0] != cur->field_poc[1]){
1344         /* Derive top_field_first from field pocs. */
1345         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1346     }else{
1347         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1348             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1349             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1350               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1351                 cur->f.top_field_first = 1;
1352             else
1353                 cur->f.top_field_first = 0;
1354         }else{
1355             /* Most likely progressive */
1356             cur->f.top_field_first = 0;
1357         }
1358     }
1359
1360     //FIXME do something with unavailable reference frames
1361
1362     /* Sort B-frames into display order */
1363
1364     if(h->sps.bitstream_restriction_flag
1365        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1366         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1367         s->low_delay = 0;
1368     }
1369
1370     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1371        && !h->sps.bitstream_restriction_flag){
1372         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1373         s->low_delay= 0;
1374     }
1375
1376     pics = 0;
1377     while(h->delayed_pic[pics]) pics++;
1378
1379     assert(pics <= MAX_DELAYED_PIC_COUNT);
1380
1381     h->delayed_pic[pics++] = cur;
1382     if (cur->f.reference == 0)
1383         cur->f.reference = DELAYED_PIC_REF;
1384
1385     out = h->delayed_pic[0];
1386     out_idx = 0;
1387     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1388         if(h->delayed_pic[i]->poc < out->poc){
1389             out = h->delayed_pic[i];
1390             out_idx = i;
1391         }
1392     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1393         h->next_outputed_poc= INT_MIN;
1394     out_of_order = out->poc < h->next_outputed_poc;
1395
1396     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1397         { }
1398     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1399        || (s->low_delay &&
1400         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1401          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1402     {
1403         s->low_delay = 0;
1404         s->avctx->has_b_frames++;
1405     }
1406
1407     if(out_of_order || pics > s->avctx->has_b_frames){
1408         out->f.reference &= ~DELAYED_PIC_REF;
1409         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1410                          // or else the first thread can release the picture and reuse it unsafely
1411         for(i=out_idx; h->delayed_pic[i]; i++)
1412             h->delayed_pic[i] = h->delayed_pic[i+1];
1413     }
1414     if(!out_of_order && pics > s->avctx->has_b_frames){
1415         h->next_output_pic = out;
1416         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1417             h->next_outputed_poc = INT_MIN;
1418         } else
1419             h->next_outputed_poc = out->poc;
1420     }else{
1421         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1422     }
1423
1424     if (setup_finished)
1425         ff_thread_finish_setup(s->avctx);
1426 }
1427
1428 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int chroma444, int simple){
1429     MpegEncContext * const s = &h->s;
1430     uint8_t *top_border;
1431     int top_idx = 1;
1432     const int pixel_shift = h->pixel_shift;
1433
1434     src_y  -=   linesize;
1435     src_cb -= uvlinesize;
1436     src_cr -= uvlinesize;
1437
1438     if(!simple && FRAME_MBAFF){
1439         if(s->mb_y&1){
1440             if(!MB_MBAFF){
1441                 top_border = h->top_borders[0][s->mb_x];
1442                 AV_COPY128(top_border, src_y + 15*linesize);
1443                 if (pixel_shift)
1444                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1445                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1446                     if(chroma444){
1447                         if (pixel_shift){
1448                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1449                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1450                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1451                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1452                         } else {
1453                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1454                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1455                         }
1456                     } else {
1457                         if (pixel_shift) {
1458                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1459                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1460                         } else {
1461                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1462                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1463                         }
1464                     }
1465                 }
1466             }
1467         }else if(MB_MBAFF){
1468             top_idx = 0;
1469         }else
1470             return;
1471     }
1472
1473     top_border = h->top_borders[top_idx][s->mb_x];
1474     // There are two lines saved, the line above the the top macroblock of a pair,
1475     // and the line above the bottom macroblock
1476     AV_COPY128(top_border, src_y + 16*linesize);
1477     if (pixel_shift)
1478         AV_COPY128(top_border+16, src_y+16*linesize+16);
1479
1480     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1481         if(chroma444){
1482             if (pixel_shift){
1483                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1484                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1485                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1486                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1487             } else {
1488                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1489                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1490             }
1491         } else {
1492             if (pixel_shift) {
1493                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1494                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1495             } else {
1496                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1497                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1498             }
1499         }
1500     }
1501 }
1502
1503 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1504                                   uint8_t *src_cb, uint8_t *src_cr,
1505                                   int linesize, int uvlinesize,
1506                                   int xchg, int chroma444,
1507                                   int simple, int pixel_shift){
1508     MpegEncContext * const s = &h->s;
1509     int deblock_topleft;
1510     int deblock_top;
1511     int top_idx = 1;
1512     uint8_t *top_border_m1;
1513     uint8_t *top_border;
1514
1515     if(!simple && FRAME_MBAFF){
1516         if(s->mb_y&1){
1517             if(!MB_MBAFF)
1518                 return;
1519         }else{
1520             top_idx = MB_MBAFF ? 0 : 1;
1521         }
1522     }
1523
1524     if(h->deblocking_filter == 2) {
1525         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1526         deblock_top     = h->top_type;
1527     } else {
1528         deblock_topleft = (s->mb_x > 0);
1529         deblock_top     = (s->mb_y > !!MB_FIELD);
1530     }
1531
1532     src_y  -=   linesize + 1 + pixel_shift;
1533     src_cb -= uvlinesize + 1 + pixel_shift;
1534     src_cr -= uvlinesize + 1 + pixel_shift;
1535
1536     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1537     top_border    = h->top_borders[top_idx][s->mb_x];
1538
1539 #define XCHG(a,b,xchg)\
1540     if (pixel_shift) {\
1541         if (xchg) {\
1542             AV_SWAP64(b+0,a+0);\
1543             AV_SWAP64(b+8,a+8);\
1544         } else {\
1545             AV_COPY128(b,a); \
1546         }\
1547     } else \
1548 if (xchg) AV_SWAP64(b,a);\
1549 else      AV_COPY64(b,a);
1550
1551     if(deblock_top){
1552         if(deblock_topleft){
1553             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1554         }
1555         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1556         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1557         if(s->mb_x+1 < s->mb_width){
1558             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1559         }
1560     }
1561     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1562         if(chroma444){
1563             if(deblock_topleft){
1564                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1565                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1566             }
1567             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1568             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1569             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1570             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1571             if(s->mb_x+1 < s->mb_width){
1572                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1573                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1574             }
1575         } else {
1576             if(deblock_top){
1577                 if(deblock_topleft){
1578                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1579                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1580                 }
1581                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1582                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1583             }
1584         }
1585     }
1586 }
1587
1588 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1589     if (high_bit_depth) {
1590         return AV_RN32A(((int32_t*)mb) + index);
1591     } else
1592         return AV_RN16A(mb + index);
1593 }
1594
1595 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1596     if (high_bit_depth) {
1597         AV_WN32A(((int32_t*)mb) + index, value);
1598     } else
1599         AV_WN16A(mb + index, value);
1600 }
1601
1602 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1603                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1604 {
1605     MpegEncContext * const s = &h->s;
1606     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1607     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1608     int i;
1609     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1610     block_offset += 16*p;
1611     if(IS_INTRA4x4(mb_type)){
1612         if(simple || !s->encoding){
1613             if(IS_8x8DCT(mb_type)){
1614                 if(transform_bypass){
1615                     idct_dc_add =
1616                     idct_add    = s->dsp.add_pixels8;
1617                 }else{
1618                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1619                     idct_add    = h->h264dsp.h264_idct8_add;
1620                 }
1621                 for(i=0; i<16; i+=4){
1622                     uint8_t * const ptr= dest_y + block_offset[i];
1623                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1624                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1625                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1626                     }else{
1627                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1628                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1629                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1630                         if(nnz){
1631                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1632                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1633                             else
1634                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1635                         }
1636                     }
1637                 }
1638             }else{
1639                 if(transform_bypass){
1640                     idct_dc_add =
1641                     idct_add    = s->dsp.add_pixels4;
1642                 }else{
1643                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1644                     idct_add    = h->h264dsp.h264_idct_add;
1645                 }
1646                 for(i=0; i<16; i++){
1647                     uint8_t * const ptr= dest_y + block_offset[i];
1648                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1649
1650                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1651                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1652                     }else{
1653                         uint8_t *topright;
1654                         int nnz, tr;
1655                         uint64_t tr_high;
1656                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1657                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1658                             assert(s->mb_y || linesize <= block_offset[i]);
1659                             if(!topright_avail){
1660                                 if (pixel_shift) {
1661                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1662                                     topright= (uint8_t*) &tr_high;
1663                                 } else {
1664                                     tr= ptr[3 - linesize]*0x01010101;
1665                                     topright= (uint8_t*) &tr;
1666                                 }
1667                             }else
1668                                 topright= ptr + (4 << pixel_shift) - linesize;
1669                         }else
1670                             topright= NULL;
1671
1672                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1673                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1674                         if(nnz){
1675                             if(is_h264){
1676                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1677                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1678                                 else
1679                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1680                             }else
1681                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1682                         }
1683                     }
1684                 }
1685             }
1686         }
1687     }else{
1688         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1689         if(is_h264){
1690             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1691                 if(!transform_bypass)
1692                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1693                 else{
1694                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1695                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1696                     for(i = 0; i < 16; i++)
1697                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1698                 }
1699             }
1700         }else
1701             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1702     }
1703 }
1704
1705 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1706                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1707 {
1708     MpegEncContext * const s = &h->s;
1709     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1710     int i;
1711     block_offset += 16*p;
1712     if(!IS_INTRA4x4(mb_type)){
1713         if(is_h264){
1714             if(IS_INTRA16x16(mb_type)){
1715                 if(transform_bypass){
1716                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1717                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1718                     }else{
1719                         for(i=0; i<16; i++){
1720                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1721                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1722                         }
1723                     }
1724                 }else{
1725                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1726                 }
1727             }else if(h->cbp&15){
1728                 if(transform_bypass){
1729                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1730                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1731                     for(i=0; i<16; i+=di){
1732                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1733                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1734                         }
1735                     }
1736                 }else{
1737                     if(IS_8x8DCT(mb_type)){
1738                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1739                     }else{
1740                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1741                     }
1742                 }
1743             }
1744         }else{
1745             for(i=0; i<16; i++){
1746                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1747                     uint8_t * const ptr= dest_y + block_offset[i];
1748                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1749                 }
1750             }
1751         }
1752     }
1753 }
1754
1755 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1756     MpegEncContext * const s = &h->s;
1757     const int mb_x= s->mb_x;
1758     const int mb_y= s->mb_y;
1759     const int mb_xy= h->mb_xy;
1760     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1761     uint8_t  *dest_y, *dest_cb, *dest_cr;
1762     int linesize, uvlinesize /*dct_offset*/;
1763     int i, j;
1764     int *block_offset = &h->block_offset[0];
1765     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1766     /* is_h264 should always be true if SVQ3 is disabled. */
1767     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1768     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1769
1770     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1771     dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1772     dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1773
1774     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1775     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1776
1777     h->list_counts[mb_xy]= h->list_count;
1778
1779     if (!simple && MB_FIELD) {
1780         linesize   = h->mb_linesize   = s->linesize * 2;
1781         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1782         block_offset = &h->block_offset[48];
1783         if(mb_y&1){ //FIXME move out of this function?
1784             dest_y -= s->linesize*15;
1785             dest_cb-= s->uvlinesize*7;
1786             dest_cr-= s->uvlinesize*7;
1787         }
1788         if(FRAME_MBAFF) {
1789             int list;
1790             for(list=0; list<h->list_count; list++){
1791                 if(!USES_LIST(mb_type, list))
1792                     continue;
1793                 if(IS_16X16(mb_type)){
1794                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1795                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1796                 }else{
1797                     for(i=0; i<16; i+=4){
1798                         int ref = h->ref_cache[list][scan8[i]];
1799                         if(ref >= 0)
1800                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1801                     }
1802                 }
1803             }
1804         }
1805     } else {
1806         linesize   = h->mb_linesize   = s->linesize;
1807         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1808 //        dct_offset = s->linesize * 16;
1809     }
1810
1811     if (!simple && IS_INTRA_PCM(mb_type)) {
1812         if (pixel_shift) {
1813             const int bit_depth = h->sps.bit_depth_luma;
1814             int j;
1815             GetBitContext gb;
1816             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1817
1818             for (i = 0; i < 16; i++) {
1819                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1820                 for (j = 0; j < 16; j++)
1821                     tmp_y[j] = get_bits(&gb, bit_depth);
1822             }
1823             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1824                 for (i = 0; i < 8; i++) {
1825                     uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1826                     for (j = 0; j < 8; j++)
1827                         tmp_cb[j] = get_bits(&gb, bit_depth);
1828                 }
1829                 for (i = 0; i < 8; i++) {
1830                     uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1831                     for (j = 0; j < 8; j++)
1832                         tmp_cr[j] = get_bits(&gb, bit_depth);
1833                 }
1834             }
1835         } else {
1836             for (i=0; i<16; i++) {
1837                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1838             }
1839             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1840                 for (i=0; i<8; i++) {
1841                     memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1842                     memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1843                 }
1844             }
1845         }
1846     } else {
1847         if(IS_INTRA(mb_type)){
1848             if(h->deblocking_filter)
1849                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1850
1851             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1852                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1853                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1854             }
1855
1856             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1857
1858             if(h->deblocking_filter)
1859                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1860         }else if(is_h264){
1861             hl_motion(h, dest_y, dest_cb, dest_cr,
1862                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1863                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1864                       h->h264dsp.weight_h264_pixels_tab,
1865                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1866         }
1867
1868         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1869
1870         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1871             uint8_t *dest[2] = {dest_cb, dest_cr};
1872             if(transform_bypass){
1873                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1874                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1875                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1876                 }else{
1877                     idct_add = s->dsp.add_pixels4;
1878                     for(j=1; j<3; j++){
1879                         for(i=j*16; i<j*16+4; i++){
1880                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1881                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1882                         }
1883                     }
1884                 }
1885             }else{
1886                 if(is_h264){
1887                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1888                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1889                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1890                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1891                     h->h264dsp.h264_idct_add8(dest, block_offset,
1892                                               h->mb, uvlinesize,
1893                                               h->non_zero_count_cache);
1894                 }else{
1895                     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]);
1896                     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]);
1897                     for(j=1; j<3; j++){
1898                         for(i=j*16; i<j*16+4; i++){
1899                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1900                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1901                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1902                             }
1903                         }
1904                     }
1905                 }
1906             }
1907         }
1908     }
1909     if(h->cbp || IS_INTRA(mb_type))
1910     {
1911         s->dsp.clear_blocks(h->mb);
1912         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1913     }
1914 }
1915
1916 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1917     MpegEncContext * const s = &h->s;
1918     const int mb_x= s->mb_x;
1919     const int mb_y= s->mb_y;
1920     const int mb_xy= h->mb_xy;
1921     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1922     uint8_t  *dest[3];
1923     int linesize;
1924     int i, j, p;
1925     int *block_offset = &h->block_offset[0];
1926     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1927     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1928
1929     for (p = 0; p < plane_count; p++)
1930     {
1931         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1932         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1933     }
1934
1935     h->list_counts[mb_xy]= h->list_count;
1936
1937     if (!simple && MB_FIELD) {
1938         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1939         block_offset = &h->block_offset[48];
1940         if(mb_y&1) //FIXME move out of this function?
1941             for (p = 0; p < 3; p++)
1942                 dest[p] -= s->linesize*15;
1943         if(FRAME_MBAFF) {
1944             int list;
1945             for(list=0; list<h->list_count; list++){
1946                 if(!USES_LIST(mb_type, list))
1947                     continue;
1948                 if(IS_16X16(mb_type)){
1949                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1950                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1951                 }else{
1952                     for(i=0; i<16; i+=4){
1953                         int ref = h->ref_cache[list][scan8[i]];
1954                         if(ref >= 0)
1955                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1956                     }
1957                 }
1958             }
1959         }
1960     } else {
1961         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1962     }
1963
1964     if (!simple && IS_INTRA_PCM(mb_type)) {
1965         if (pixel_shift) {
1966             const int bit_depth = h->sps.bit_depth_luma;
1967             GetBitContext gb;
1968             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1969
1970             for (p = 0; p < plane_count; p++) {
1971                 for (i = 0; i < 16; i++) {
1972                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
1973                     for (j = 0; j < 16; j++)
1974                         tmp[j] = get_bits(&gb, bit_depth);
1975                 }
1976             }
1977         } else {
1978             for (p = 0; p < plane_count; p++) {
1979                 for (i = 0; i < 16; i++) {
1980                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
1981                 }
1982             }
1983         }
1984     } else {
1985         if(IS_INTRA(mb_type)){
1986             if(h->deblocking_filter)
1987                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
1988
1989             for (p = 0; p < plane_count; p++)
1990                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
1991
1992             if(h->deblocking_filter)
1993                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
1994         }else{
1995             hl_motion(h, dest[0], dest[1], dest[2],
1996                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1997                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1998                       h->h264dsp.weight_h264_pixels_tab,
1999                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2000         }
2001
2002         for (p = 0; p < plane_count; p++)
2003             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2004     }
2005     if(h->cbp || IS_INTRA(mb_type))
2006     {
2007         s->dsp.clear_blocks(h->mb);
2008         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2009     }
2010 }
2011
2012 /**
2013  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2014  */
2015 #define hl_decode_mb_simple(sh, bits) \
2016 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2017     hl_decode_mb_internal(h, 1, sh); \
2018 }
2019 hl_decode_mb_simple(0, 8);
2020 hl_decode_mb_simple(1, 16);
2021
2022 /**
2023  * Process a macroblock; this handles edge cases, such as interlacing.
2024  */
2025 static void av_noinline hl_decode_mb_complex(H264Context *h){
2026     hl_decode_mb_internal(h, 0, h->pixel_shift);
2027 }
2028
2029 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2030     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2031 }
2032
2033 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2034     hl_decode_mb_444_internal(h, 1, 0);
2035 }
2036
2037 void ff_h264_hl_decode_mb(H264Context *h){
2038     MpegEncContext * const s = &h->s;
2039     const int mb_xy= h->mb_xy;
2040     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2041     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2042
2043     if (CHROMA444) {
2044         if(is_complex || h->pixel_shift)
2045             hl_decode_mb_444_complex(h);
2046         else
2047             hl_decode_mb_444_simple(h);
2048     } else if (is_complex) {
2049         hl_decode_mb_complex(h);
2050     } else if (h->pixel_shift) {
2051         hl_decode_mb_simple_16(h);
2052     } else
2053         hl_decode_mb_simple_8(h);
2054 }
2055
2056 static int pred_weight_table(H264Context *h){
2057     MpegEncContext * const s = &h->s;
2058     int list, i;
2059     int luma_def, chroma_def;
2060
2061     h->use_weight= 0;
2062     h->use_weight_chroma= 0;
2063     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2064     if(h->sps.chroma_format_idc)
2065         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2066     luma_def = 1<<h->luma_log2_weight_denom;
2067     chroma_def = 1<<h->chroma_log2_weight_denom;
2068
2069     for(list=0; list<2; list++){
2070         h->luma_weight_flag[list]   = 0;
2071         h->chroma_weight_flag[list] = 0;
2072         for(i=0; i<h->ref_count[list]; i++){
2073             int luma_weight_flag, chroma_weight_flag;
2074
2075             luma_weight_flag= get_bits1(&s->gb);
2076             if(luma_weight_flag){
2077                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2078                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2079                 if(   h->luma_weight[i][list][0] != luma_def
2080                    || h->luma_weight[i][list][1] != 0) {
2081                     h->use_weight= 1;
2082                     h->luma_weight_flag[list]= 1;
2083                 }
2084             }else{
2085                 h->luma_weight[i][list][0]= luma_def;
2086                 h->luma_weight[i][list][1]= 0;
2087             }
2088
2089             if(h->sps.chroma_format_idc){
2090                 chroma_weight_flag= get_bits1(&s->gb);
2091                 if(chroma_weight_flag){
2092                     int j;
2093                     for(j=0; j<2; j++){
2094                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2095                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2096                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2097                            || h->chroma_weight[i][list][j][1] != 0) {
2098                             h->use_weight_chroma= 1;
2099                             h->chroma_weight_flag[list]= 1;
2100                         }
2101                     }
2102                 }else{
2103                     int j;
2104                     for(j=0; j<2; j++){
2105                         h->chroma_weight[i][list][j][0]= chroma_def;
2106                         h->chroma_weight[i][list][j][1]= 0;
2107                     }
2108                 }
2109             }
2110         }
2111         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2112     }
2113     h->use_weight= h->use_weight || h->use_weight_chroma;
2114     return 0;
2115 }
2116
2117 /**
2118  * Initialize implicit_weight table.
2119  * @param field  0/1 initialize the weight for interlaced MBAFF
2120  *                -1 initializes the rest
2121  */
2122 static void implicit_weight_table(H264Context *h, int field){
2123     MpegEncContext * const s = &h->s;
2124     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2125
2126     for (i = 0; i < 2; i++) {
2127         h->luma_weight_flag[i]   = 0;
2128         h->chroma_weight_flag[i] = 0;
2129     }
2130
2131     if(field < 0){
2132         cur_poc = s->current_picture_ptr->poc;
2133     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2134        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2135         h->use_weight= 0;
2136         h->use_weight_chroma= 0;
2137         return;
2138     }
2139         ref_start= 0;
2140         ref_count0= h->ref_count[0];
2141         ref_count1= h->ref_count[1];
2142     }else{
2143         cur_poc = s->current_picture_ptr->field_poc[field];
2144         ref_start= 16;
2145         ref_count0= 16+2*h->ref_count[0];
2146         ref_count1= 16+2*h->ref_count[1];
2147     }
2148
2149     h->use_weight= 2;
2150     h->use_weight_chroma= 2;
2151     h->luma_log2_weight_denom= 5;
2152     h->chroma_log2_weight_denom= 5;
2153
2154     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2155         int poc0 = h->ref_list[0][ref0].poc;
2156         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2157             int poc1 = h->ref_list[1][ref1].poc;
2158             int td = av_clip(poc1 - poc0, -128, 127);
2159             int w= 32;
2160             if(td){
2161                 int tb = av_clip(cur_poc - poc0, -128, 127);
2162                 int tx = (16384 + (FFABS(td) >> 1)) / td;
2163                 int dist_scale_factor = (tb*tx + 32) >> 8;
2164                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2165                     w = 64 - dist_scale_factor;
2166             }
2167             if(field<0){
2168                 h->implicit_weight[ref0][ref1][0]=
2169                 h->implicit_weight[ref0][ref1][1]= w;
2170             }else{
2171                 h->implicit_weight[ref0][ref1][field]=w;
2172             }
2173         }
2174     }
2175 }
2176
2177 /**
2178  * instantaneous decoder refresh.
2179  */
2180 static void idr(H264Context *h){
2181     ff_h264_remove_all_refs(h);
2182     h->prev_frame_num= 0;
2183     h->prev_frame_num_offset= 0;
2184     h->prev_poc_msb=
2185     h->prev_poc_lsb= 0;
2186 }
2187
2188 /* forget old pics after a seek */
2189 static void flush_dpb(AVCodecContext *avctx){
2190     H264Context *h= avctx->priv_data;
2191     int i;
2192     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2193         if(h->delayed_pic[i])
2194             h->delayed_pic[i]->f.reference = 0;
2195         h->delayed_pic[i]= NULL;
2196     }
2197     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2198     h->prev_interlaced_frame = 1;
2199     idr(h);
2200     if(h->s.current_picture_ptr)
2201         h->s.current_picture_ptr->f.reference = 0;
2202     h->s.first_field= 0;
2203     ff_h264_reset_sei(h);
2204     ff_mpeg_flush(avctx);
2205 }
2206
2207 static int init_poc(H264Context *h){
2208     MpegEncContext * const s = &h->s;
2209     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2210     int field_poc[2];
2211     Picture *cur = s->current_picture_ptr;
2212
2213     h->frame_num_offset= h->prev_frame_num_offset;
2214     if(h->frame_num < h->prev_frame_num)
2215         h->frame_num_offset += max_frame_num;
2216
2217     if(h->sps.poc_type==0){
2218         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2219
2220         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2221             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2222         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2223             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2224         else
2225             h->poc_msb = h->prev_poc_msb;
2226 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2227         field_poc[0] =
2228         field_poc[1] = h->poc_msb + h->poc_lsb;
2229         if(s->picture_structure == PICT_FRAME)
2230             field_poc[1] += h->delta_poc_bottom;
2231     }else if(h->sps.poc_type==1){
2232         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2233         int i;
2234
2235         if(h->sps.poc_cycle_length != 0)
2236             abs_frame_num = h->frame_num_offset + h->frame_num;
2237         else
2238             abs_frame_num = 0;
2239
2240         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2241             abs_frame_num--;
2242
2243         expected_delta_per_poc_cycle = 0;
2244         for(i=0; i < h->sps.poc_cycle_length; i++)
2245             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2246
2247         if(abs_frame_num > 0){
2248             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2249             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2250
2251             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2252             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2253                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2254         } else
2255             expectedpoc = 0;
2256
2257         if(h->nal_ref_idc == 0)
2258             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2259
2260         field_poc[0] = expectedpoc + h->delta_poc[0];
2261         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2262
2263         if(s->picture_structure == PICT_FRAME)
2264             field_poc[1] += h->delta_poc[1];
2265     }else{
2266         int poc= 2*(h->frame_num_offset + h->frame_num);
2267
2268         if(!h->nal_ref_idc)
2269             poc--;
2270
2271         field_poc[0]= poc;
2272         field_poc[1]= poc;
2273     }
2274
2275     if(s->picture_structure != PICT_BOTTOM_FIELD)
2276         s->current_picture_ptr->field_poc[0]= field_poc[0];
2277     if(s->picture_structure != PICT_TOP_FIELD)
2278         s->current_picture_ptr->field_poc[1]= field_poc[1];
2279     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2280
2281     return 0;
2282 }
2283
2284
2285 /**
2286  * initialize scan tables
2287  */
2288 static void init_scan_tables(H264Context *h){
2289     int i;
2290     for(i=0; i<16; i++){
2291 #define T(x) (x>>2) | ((x<<2) & 0xF)
2292         h->zigzag_scan[i] = T(zigzag_scan[i]);
2293         h-> field_scan[i] = T( field_scan[i]);
2294 #undef T
2295     }
2296     for(i=0; i<64; i++){
2297 #define T(x) (x>>3) | ((x&7)<<3)
2298         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2299         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2300         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2301         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2302 #undef T
2303     }
2304     if(h->sps.transform_bypass){ //FIXME same ugly
2305         h->zigzag_scan_q0          = zigzag_scan;
2306         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2307         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2308         h->field_scan_q0           = field_scan;
2309         h->field_scan8x8_q0        = field_scan8x8;
2310         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2311     }else{
2312         h->zigzag_scan_q0          = h->zigzag_scan;
2313         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2314         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2315         h->field_scan_q0           = h->field_scan;
2316         h->field_scan8x8_q0        = h->field_scan8x8;
2317         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2318     }
2319 }
2320
2321 static void field_end(H264Context *h, int in_setup){
2322     MpegEncContext * const s = &h->s;
2323     AVCodecContext * const avctx= s->avctx;
2324     s->mb_y= 0;
2325
2326     if (!in_setup && !s->dropable)
2327         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2328                                  s->picture_structure==PICT_BOTTOM_FIELD);
2329
2330     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2331         ff_vdpau_h264_set_reference_frames(s);
2332
2333     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2334         if(!s->dropable) {
2335             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2336             h->prev_poc_msb= h->poc_msb;
2337             h->prev_poc_lsb= h->poc_lsb;
2338         }
2339         h->prev_frame_num_offset= h->frame_num_offset;
2340         h->prev_frame_num= h->frame_num;
2341         h->outputed_poc = h->next_outputed_poc;
2342     }
2343
2344     if (avctx->hwaccel) {
2345         if (avctx->hwaccel->end_frame(avctx) < 0)
2346             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2347     }
2348
2349     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2350         ff_vdpau_h264_picture_complete(s);
2351
2352     /*
2353      * FIXME: Error handling code does not seem to support interlaced
2354      * when slices span multiple rows
2355      * The ff_er_add_slice calls don't work right for bottom
2356      * fields; they cause massive erroneous error concealing
2357      * Error marking covers both fields (top and bottom).
2358      * This causes a mismatched s->error_count
2359      * and a bad error table. Further, the error count goes to
2360      * INT_MAX when called for bottom field, because mb_y is
2361      * past end by one (callers fault) and resync_mb_y != 0
2362      * causes problems for the first MB line, too.
2363      */
2364     if (!FIELD_PICTURE)
2365         ff_er_frame_end(s);
2366
2367     MPV_frame_end(s);
2368
2369     h->current_slice=0;
2370 }
2371
2372 /**
2373  * Replicate H264 "master" context to thread contexts.
2374  */
2375 static void clone_slice(H264Context *dst, H264Context *src)
2376 {
2377     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2378     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2379     dst->s.current_picture      = src->s.current_picture;
2380     dst->s.linesize             = src->s.linesize;
2381     dst->s.uvlinesize           = src->s.uvlinesize;
2382     dst->s.first_field          = src->s.first_field;
2383
2384     dst->prev_poc_msb           = src->prev_poc_msb;
2385     dst->prev_poc_lsb           = src->prev_poc_lsb;
2386     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2387     dst->prev_frame_num         = src->prev_frame_num;
2388     dst->short_ref_count        = src->short_ref_count;
2389
2390     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2391     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2392     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2393     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2394
2395     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2396     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2397 }
2398
2399 /**
2400  * computes profile from profile_idc and constraint_set?_flags
2401  *
2402  * @param sps SPS
2403  *
2404  * @return profile as defined by FF_PROFILE_H264_*
2405  */
2406 int ff_h264_get_profile(SPS *sps)
2407 {
2408     int profile = sps->profile_idc;
2409
2410     switch(sps->profile_idc) {
2411     case FF_PROFILE_H264_BASELINE:
2412         // constraint_set1_flag set to 1
2413         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2414         break;
2415     case FF_PROFILE_H264_HIGH_10:
2416     case FF_PROFILE_H264_HIGH_422:
2417     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2418         // constraint_set3_flag set to 1
2419         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2420         break;
2421     }
2422
2423     return profile;
2424 }
2425
2426 /**
2427  * decodes a slice header.
2428  * This will also call MPV_common_init() and frame_start() as needed.
2429  *
2430  * @param h h264context
2431  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2432  *
2433  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2434  */
2435 static int decode_slice_header(H264Context *h, H264Context *h0){
2436     MpegEncContext * const s = &h->s;
2437     MpegEncContext * const s0 = &h0->s;
2438     unsigned int first_mb_in_slice;
2439     unsigned int pps_id;
2440     int num_ref_idx_active_override_flag;
2441     unsigned int slice_type, tmp, i, j;
2442     int default_ref_list_done = 0;
2443     int last_pic_structure;
2444
2445     s->dropable= h->nal_ref_idc == 0;
2446
2447     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2448     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2449         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2450         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2451     }else{
2452         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2453         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2454     }
2455
2456     first_mb_in_slice= get_ue_golomb(&s->gb);
2457
2458     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2459         if(h0->current_slice && FIELD_PICTURE){
2460             field_end(h, 1);
2461         }
2462
2463         h0->current_slice = 0;
2464         if (!s0->first_field)
2465             s->current_picture_ptr= NULL;
2466     }
2467
2468     slice_type= get_ue_golomb_31(&s->gb);
2469     if(slice_type > 9){
2470         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);
2471         return -1;
2472     }
2473     if(slice_type > 4){
2474         slice_type -= 5;
2475         h->slice_type_fixed=1;
2476     }else
2477         h->slice_type_fixed=0;
2478
2479     slice_type= golomb_to_pict_type[ slice_type ];
2480     if (slice_type == AV_PICTURE_TYPE_I
2481         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2482         default_ref_list_done = 1;
2483     }
2484     h->slice_type= slice_type;
2485     h->slice_type_nos= slice_type & 3;
2486
2487     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2488
2489     pps_id= get_ue_golomb(&s->gb);
2490     if(pps_id>=MAX_PPS_COUNT){
2491         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2492         return -1;
2493     }
2494     if(!h0->pps_buffers[pps_id]) {
2495         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2496         return -1;
2497     }
2498     h->pps= *h0->pps_buffers[pps_id];
2499
2500     if(!h0->sps_buffers[h->pps.sps_id]) {
2501         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2502         return -1;
2503     }
2504     h->sps = *h0->sps_buffers[h->pps.sps_id];
2505
2506     s->avctx->profile = ff_h264_get_profile(&h->sps);
2507     s->avctx->level   = h->sps.level_idc;
2508     s->avctx->refs    = h->sps.ref_frame_count;
2509
2510     if(h == h0 && h->dequant_coeff_pps != pps_id){
2511         h->dequant_coeff_pps = pps_id;
2512         init_dequant_tables(h);
2513     }
2514
2515     s->mb_width= h->sps.mb_width;
2516     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2517
2518     h->b_stride=  s->mb_width*4;
2519
2520     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2521     if(h->sps.frame_mbs_only_flag)
2522         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2523     else
2524         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2525
2526     if (s->context_initialized
2527         && (   s->width != s->avctx->width || s->height != s->avctx->height
2528             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2529         if(h != h0) {
2530             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2531             return -1;   // width / height changed during parallelized decoding
2532         }
2533         free_tables(h, 0);
2534         flush_dpb(s->avctx);
2535         MPV_common_end(s);
2536     }
2537     if (!s->context_initialized) {
2538         if (h != h0) {
2539             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2540             return -1;
2541         }
2542
2543         avcodec_set_dimensions(s->avctx, s->width, s->height);
2544         s->avctx->sample_aspect_ratio= h->sps.sar;
2545         av_assert0(s->avctx->sample_aspect_ratio.den);
2546
2547         h->s.avctx->coded_width = 16*s->mb_width;
2548         h->s.avctx->coded_height = 16*s->mb_height;
2549
2550         if(h->sps.video_signal_type_present_flag){
2551             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2552             if(h->sps.colour_description_present_flag){
2553                 s->avctx->color_primaries = h->sps.color_primaries;
2554                 s->avctx->color_trc       = h->sps.color_trc;
2555                 s->avctx->colorspace      = h->sps.colorspace;
2556             }
2557         }
2558
2559         if(h->sps.timing_info_present_flag){
2560             int64_t den= h->sps.time_scale;
2561             if(h->x264_build < 44U)
2562                 den *= 2;
2563             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2564                       h->sps.num_units_in_tick, den, 1<<30);
2565         }
2566
2567         switch (h->sps.bit_depth_luma) {
2568             case 9 :
2569                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2570                 break;
2571             case 10 :
2572                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2573                 break;
2574             default:
2575                 if (CHROMA444){
2576                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2577                 }else{
2578                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2579                                                              s->avctx->codec->pix_fmts ?
2580                                                              s->avctx->codec->pix_fmts :
2581                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2582                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2583                                                              ff_hwaccel_pixfmt_list_420);
2584                 }
2585         }
2586
2587         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2588
2589         if (MPV_common_init(s) < 0) {
2590             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2591             return -1;
2592         }
2593         s->first_field = 0;
2594         h->prev_interlaced_frame = 1;
2595
2596         init_scan_tables(h);
2597         ff_h264_alloc_tables(h);
2598
2599         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2600             if (context_init(h) < 0) {
2601                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2602                 return -1;
2603             }
2604         } else {
2605             for(i = 1; i < s->avctx->thread_count; i++) {
2606                 H264Context *c;
2607                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2608                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2609                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2610                 c->h264dsp = h->h264dsp;
2611                 c->sps = h->sps;
2612                 c->pps = h->pps;
2613                 c->pixel_shift = h->pixel_shift;
2614                 init_scan_tables(c);
2615                 clone_tables(c, h, i);
2616             }
2617
2618             for(i = 0; i < s->avctx->thread_count; i++)
2619                 if (context_init(h->thread_context[i]) < 0) {
2620                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2621                     return -1;
2622                 }
2623         }
2624     }
2625
2626     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2627
2628     h->mb_mbaff = 0;
2629     h->mb_aff_frame = 0;
2630     last_pic_structure = s0->picture_structure;
2631     if(h->sps.frame_mbs_only_flag){
2632         s->picture_structure= PICT_FRAME;
2633     }else{
2634         if(get_bits1(&s->gb)) { //field_pic_flag
2635             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2636         } else {
2637             s->picture_structure= PICT_FRAME;
2638             h->mb_aff_frame = h->sps.mb_aff;
2639         }
2640     }
2641     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2642
2643     if(h0->current_slice == 0){
2644         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2645         if(h->frame_num != h->prev_frame_num) {
2646             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2647
2648             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2649
2650             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2651                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2652                 if (unwrap_prev_frame_num < 0)
2653                     unwrap_prev_frame_num += max_frame_num;
2654
2655                 h->prev_frame_num = unwrap_prev_frame_num;
2656             }
2657         }
2658
2659         while(h->frame_num !=  h->prev_frame_num &&
2660               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2661             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2662             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2663             if (ff_h264_frame_start(h) < 0)
2664                 return -1;
2665             h->prev_frame_num++;
2666             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2667             s->current_picture_ptr->frame_num= h->prev_frame_num;
2668             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2669             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2670             ff_generate_sliding_window_mmcos(h);
2671             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2672             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2673              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2674              * about there being no actual duplicates.
2675              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2676              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2677              * be fixed. */
2678             if (h->short_ref_count) {
2679                 if (prev) {
2680                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2681                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2682                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2683                     h->short_ref[0]->poc = prev->poc+2;
2684                 }
2685                 h->short_ref[0]->frame_num = h->prev_frame_num;
2686             }
2687         }
2688
2689         /* See if we have a decoded first field looking for a pair... */
2690         if (s0->first_field) {
2691             assert(s0->current_picture_ptr);
2692             assert(s0->current_picture_ptr->f.data[0]);
2693             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2694
2695             /* figure out if we have a complementary field pair */
2696             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2697                 /*
2698                  * Previous field is unmatched. Don't display it, but let it
2699                  * remain for reference if marked as such.
2700                  */
2701                 s0->current_picture_ptr = NULL;
2702                 s0->first_field = FIELD_PICTURE;
2703
2704             } else {
2705                 if (h->nal_ref_idc &&
2706                         s0->current_picture_ptr->f.reference &&
2707                         s0->current_picture_ptr->frame_num != h->frame_num) {
2708                     /*
2709                      * This and previous field were reference, but had
2710                      * different frame_nums. Consider this field first in
2711                      * pair. Throw away previous field except for reference
2712                      * purposes.
2713                      */
2714                     s0->first_field = 1;
2715                     s0->current_picture_ptr = NULL;
2716
2717                 } else {
2718                     /* Second field in complementary pair */
2719                     s0->first_field = 0;
2720                 }
2721             }
2722
2723         } else {
2724             /* Frame or first field in a potentially complementary pair */
2725             assert(!s0->current_picture_ptr);
2726             s0->first_field = FIELD_PICTURE;
2727         }
2728
2729         if(!FIELD_PICTURE || s0->first_field) {
2730             if (ff_h264_frame_start(h) < 0) {
2731                 s0->first_field = 0;
2732                 return -1;
2733             }
2734         } else {
2735             ff_release_unused_pictures(s, 0);
2736         }
2737     }
2738     if(h != h0)
2739         clone_slice(h, h0);
2740
2741     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2742
2743     assert(s->mb_num == s->mb_width * s->mb_height);
2744     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2745        first_mb_in_slice                    >= s->mb_num){
2746         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2747         return -1;
2748     }
2749     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2750     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2751     if (s->picture_structure == PICT_BOTTOM_FIELD)
2752         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2753     assert(s->mb_y < s->mb_height);
2754
2755     if(s->picture_structure==PICT_FRAME){
2756         h->curr_pic_num=   h->frame_num;
2757         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2758     }else{
2759         h->curr_pic_num= 2*h->frame_num + 1;
2760         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2761     }
2762
2763     if(h->nal_unit_type == NAL_IDR_SLICE){
2764         get_ue_golomb(&s->gb); /* idr_pic_id */
2765     }
2766
2767     if(h->sps.poc_type==0){
2768         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2769
2770         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2771             h->delta_poc_bottom= get_se_golomb(&s->gb);
2772         }
2773     }
2774
2775     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2776         h->delta_poc[0]= get_se_golomb(&s->gb);
2777
2778         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2779             h->delta_poc[1]= get_se_golomb(&s->gb);
2780     }
2781
2782     init_poc(h);
2783
2784     if(h->pps.redundant_pic_cnt_present){
2785         h->redundant_pic_count= get_ue_golomb(&s->gb);
2786     }
2787
2788     //set defaults, might be overridden a few lines later
2789     h->ref_count[0]= h->pps.ref_count[0];
2790     h->ref_count[1]= h->pps.ref_count[1];
2791
2792     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2793         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2794             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2795         }
2796         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2797
2798         if(num_ref_idx_active_override_flag){
2799             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2800             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2801                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2802
2803             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2804                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2805                 h->ref_count[0]= h->ref_count[1]= 1;
2806                 return -1;
2807             }
2808         }
2809         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2810             h->list_count= 2;
2811         else
2812             h->list_count= 1;
2813     }else
2814         h->list_count= 0;
2815
2816     if(!default_ref_list_done){
2817         ff_h264_fill_default_ref_list(h);
2818     }
2819
2820     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2821         return -1;
2822
2823     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2824         s->last_picture_ptr= &h->ref_list[0][0];
2825         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2826     }
2827     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2828         s->next_picture_ptr= &h->ref_list[1][0];
2829         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2830     }
2831
2832     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2833        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2834         pred_weight_table(h);
2835     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2836         implicit_weight_table(h, -1);
2837     }else {
2838         h->use_weight = 0;
2839         for (i = 0; i < 2; i++) {
2840             h->luma_weight_flag[i]   = 0;
2841             h->chroma_weight_flag[i] = 0;
2842         }
2843     }
2844
2845     if(h->nal_ref_idc)
2846         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2847
2848     if(FRAME_MBAFF){
2849         ff_h264_fill_mbaff_ref_list(h);
2850
2851         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2852             implicit_weight_table(h, 0);
2853             implicit_weight_table(h, 1);
2854         }
2855     }
2856
2857     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2858         ff_h264_direct_dist_scale_factor(h);
2859     ff_h264_direct_ref_list_init(h);
2860
2861     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2862         tmp = get_ue_golomb_31(&s->gb);
2863         if(tmp > 2){
2864             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2865             return -1;
2866         }
2867         h->cabac_init_idc= tmp;
2868     }
2869
2870     h->last_qscale_diff = 0;
2871     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2872     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2873         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2874         return -1;
2875     }
2876     s->qscale= tmp;
2877     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2878     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2879     //FIXME qscale / qp ... stuff
2880     if(h->slice_type == AV_PICTURE_TYPE_SP){
2881         get_bits1(&s->gb); /* sp_for_switch_flag */
2882     }
2883     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2884         get_se_golomb(&s->gb); /* slice_qs_delta */
2885     }
2886
2887     h->deblocking_filter = 1;
2888     h->slice_alpha_c0_offset = 52;
2889     h->slice_beta_offset = 52;
2890     if( h->pps.deblocking_filter_parameters_present ) {
2891         tmp= get_ue_golomb_31(&s->gb);
2892         if(tmp > 2){
2893             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2894             return -1;
2895         }
2896         h->deblocking_filter= tmp;
2897         if(h->deblocking_filter < 2)
2898             h->deblocking_filter^= 1; // 1<->0
2899
2900         if( h->deblocking_filter ) {
2901             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2902             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2903             if(   h->slice_alpha_c0_offset > 104U
2904                || h->slice_beta_offset     > 104U){
2905                 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);
2906                 return -1;
2907             }
2908         }
2909     }
2910
2911     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2912        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2913        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2914        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2915         h->deblocking_filter= 0;
2916
2917     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2918         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2919             /* Cheat slightly for speed:
2920                Do not bother to deblock across slices. */
2921             h->deblocking_filter = 2;
2922         } else {
2923             h0->max_contexts = 1;
2924             if(!h0->single_decode_warning) {
2925                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2926                 h0->single_decode_warning = 1;
2927             }
2928             if (h != h0) {
2929                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2930                 return 1;
2931             }
2932         }
2933     }
2934     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
2935                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
2936                  + 6 * (h->sps.bit_depth_luma - 8);
2937
2938 #if 0 //FMO
2939     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2940         slice_group_change_cycle= get_bits(&s->gb, ?);
2941 #endif
2942
2943     h0->last_slice_type = slice_type;
2944     h->slice_num = ++h0->current_slice;
2945     if(h->slice_num >= MAX_SLICES){
2946         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2947     }
2948
2949     for(j=0; j<2; j++){
2950         int id_list[16];
2951         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2952         for(i=0; i<16; i++){
2953             id_list[i]= 60;
2954             if (h->ref_list[j][i].f.data[0]) {
2955                 int k;
2956                 uint8_t *base = h->ref_list[j][i].f.base[0];
2957                 for(k=0; k<h->short_ref_count; k++)
2958                     if (h->short_ref[k]->f.base[0] == base) {
2959                         id_list[i]= k;
2960                         break;
2961                     }
2962                 for(k=0; k<h->long_ref_count; k++)
2963                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
2964                         id_list[i]= h->short_ref_count + k;
2965                         break;
2966                     }
2967             }
2968         }
2969
2970         ref2frm[0]=
2971         ref2frm[1]= -1;
2972         for(i=0; i<16; i++)
2973             ref2frm[i+2]= 4*id_list[i]
2974                           + (h->ref_list[j][i].f.reference & 3);
2975         ref2frm[18+0]=
2976         ref2frm[18+1]= -1;
2977         for(i=16; i<48; i++)
2978             ref2frm[i+4]= 4*id_list[(i-16)>>1]
2979                           + (h->ref_list[j][i].f.reference & 3);
2980     }
2981
2982     //FIXME: fix draw_edges+PAFF+frame threads
2983     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
2984     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
2985
2986     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
2987         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",
2988                h->slice_num,
2989                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
2990                first_mb_in_slice,
2991                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
2992                pps_id, h->frame_num,
2993                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
2994                h->ref_count[0], h->ref_count[1],
2995                s->qscale,
2996                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
2997                h->use_weight,
2998                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
2999                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3000                );
3001     }
3002
3003     return 0;
3004 }
3005
3006 int ff_h264_get_slice_type(const H264Context *h)
3007 {
3008     switch (h->slice_type) {
3009     case AV_PICTURE_TYPE_P:  return 0;
3010     case AV_PICTURE_TYPE_B:  return 1;
3011     case AV_PICTURE_TYPE_I:  return 2;
3012     case AV_PICTURE_TYPE_SP: return 3;
3013     case AV_PICTURE_TYPE_SI: return 4;
3014     default:         return -1;
3015     }
3016 }
3017
3018 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3019                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3020 {
3021     int b_stride = h->b_stride;
3022     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3023     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3024     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3025         if(USES_LIST(top_type, list)){
3026             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3027             const int b8_xy= 4*top_xy + 2;
3028             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3029             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3030             ref_cache[0 - 1*8]=
3031             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3032             ref_cache[2 - 1*8]=
3033             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3034         }else{
3035             AV_ZERO128(mv_dst - 1*8);
3036             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3037         }
3038
3039         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3040             if(USES_LIST(left_type[LTOP], list)){
3041                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3042                 const int b8_xy= 4*left_xy[LTOP] + 1;
3043                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3044                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3045                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3046                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3047                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3048                 ref_cache[-1 +  0]=
3049                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3050                 ref_cache[-1 + 16]=
3051                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3052             }else{
3053                 AV_ZERO32(mv_dst - 1 + 0);
3054                 AV_ZERO32(mv_dst - 1 + 8);
3055                 AV_ZERO32(mv_dst - 1 +16);
3056                 AV_ZERO32(mv_dst - 1 +24);
3057                 ref_cache[-1 +  0]=
3058                 ref_cache[-1 +  8]=
3059                 ref_cache[-1 + 16]=
3060                 ref_cache[-1 + 24]= LIST_NOT_USED;
3061             }
3062         }
3063     }
3064
3065     if(!USES_LIST(mb_type, list)){
3066         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3067         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3068         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3069         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3070         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3071         return;
3072     }
3073
3074     {
3075         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3076         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3077         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3078         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3079         AV_WN32A(&ref_cache[0*8], ref01);
3080         AV_WN32A(&ref_cache[1*8], ref01);
3081         AV_WN32A(&ref_cache[2*8], ref23);
3082         AV_WN32A(&ref_cache[3*8], ref23);
3083     }
3084
3085     {
3086         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3087         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3088         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3089         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3090         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3091     }
3092 }
3093
3094 /**
3095  *
3096  * @return non zero if the loop filter can be skiped
3097  */
3098 static int fill_filter_caches(H264Context *h, int mb_type){
3099     MpegEncContext * const s = &h->s;
3100     const int mb_xy= h->mb_xy;
3101     int top_xy, left_xy[LEFT_MBS];
3102     int top_type, left_type[LEFT_MBS];
3103     uint8_t *nnz;
3104     uint8_t *nnz_cache;
3105
3106     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3107
3108     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3109      * stuff, I can't imagine that these complex rules are worth it. */
3110
3111     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3112     if(FRAME_MBAFF){
3113         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3114         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3115         if(s->mb_y&1){
3116             if (left_mb_field_flag != curr_mb_field_flag) {
3117                 left_xy[LTOP] -= s->mb_stride;
3118             }
3119         }else{
3120             if(curr_mb_field_flag){
3121                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3122             }
3123             if (left_mb_field_flag != curr_mb_field_flag) {
3124                 left_xy[LBOT] += s->mb_stride;
3125             }
3126         }
3127     }
3128
3129     h->top_mb_xy = top_xy;
3130     h->left_mb_xy[LTOP] = left_xy[LTOP];
3131     h->left_mb_xy[LBOT] = left_xy[LBOT];
3132     {
3133         //for sufficiently low qp, filtering wouldn't do anything
3134         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3135         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3136         int qp = s->current_picture.f.qscale_table[mb_xy];
3137         if(qp <= qp_thresh
3138            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3139            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3140             if(!FRAME_MBAFF)
3141                 return 1;
3142             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3143                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3144                 return 1;
3145         }
3146     }
3147
3148     top_type        = s->current_picture.f.mb_type[top_xy];
3149     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3150     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3151     if(h->deblocking_filter == 2){
3152         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3153         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3154     }else{
3155         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3156         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3157     }
3158     h->top_type       = top_type;
3159     h->left_type[LTOP]= left_type[LTOP];
3160     h->left_type[LBOT]= left_type[LBOT];
3161
3162     if(IS_INTRA(mb_type))
3163         return 0;
3164
3165     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3166     if(h->list_count == 2)
3167         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3168
3169     nnz = h->non_zero_count[mb_xy];
3170     nnz_cache = h->non_zero_count_cache;
3171     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3172     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3173     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3174     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3175     h->cbp= h->cbp_table[mb_xy];
3176
3177     if(top_type){
3178         nnz = h->non_zero_count[top_xy];
3179         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3180     }
3181
3182     if(left_type[LTOP]){
3183         nnz = h->non_zero_count[left_xy[LTOP]];
3184         nnz_cache[3+8*1]= nnz[3+0*4];
3185         nnz_cache[3+8*2]= nnz[3+1*4];
3186         nnz_cache[3+8*3]= nnz[3+2*4];
3187         nnz_cache[3+8*4]= nnz[3+3*4];
3188     }
3189
3190     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3191     if(!CABAC && h->pps.transform_8x8_mode){
3192         if(IS_8x8DCT(top_type)){
3193             nnz_cache[4+8*0]=
3194             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3195             nnz_cache[6+8*0]=
3196             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3197         }
3198         if(IS_8x8DCT(left_type[LTOP])){
3199             nnz_cache[3+8*1]=
3200             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3201         }
3202         if(IS_8x8DCT(left_type[LBOT])){
3203             nnz_cache[3+8*3]=
3204             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3205         }
3206
3207         if(IS_8x8DCT(mb_type)){
3208             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3209             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3210
3211             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3212             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3213
3214             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3215             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3216
3217             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3218             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3219         }
3220     }
3221
3222     return 0;
3223 }
3224
3225 static void loop_filter(H264Context *h, int start_x, int end_x){
3226     MpegEncContext * const s = &h->s;
3227     uint8_t  *dest_y, *dest_cb, *dest_cr;
3228     int linesize, uvlinesize, mb_x, mb_y;
3229     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3230     const int old_slice_type= h->slice_type;
3231     const int pixel_shift = h->pixel_shift;
3232
3233     if(h->deblocking_filter) {
3234         for(mb_x= start_x; mb_x<end_x; mb_x++){
3235             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3236                 int mb_xy, mb_type;
3237                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3238                 h->slice_num= h->slice_table[mb_xy];
3239                 mb_type = s->current_picture.f.mb_type[mb_xy];
3240                 h->list_count= h->list_counts[mb_xy];
3241
3242                 if(FRAME_MBAFF)
3243                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3244
3245                 s->mb_x= mb_x;
3246                 s->mb_y= mb_y;
3247                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3248                 dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3249                 dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3250                     //FIXME simplify above
3251
3252                 if (MB_FIELD) {
3253                     linesize   = h->mb_linesize   = s->linesize * 2;
3254                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3255                     if(mb_y&1){ //FIXME move out of this function?
3256                         dest_y -= s->linesize*15;
3257                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3258                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3259                     }
3260                 } else {
3261                     linesize   = h->mb_linesize   = s->linesize;
3262                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3263                 }
3264                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3265                 if(fill_filter_caches(h, mb_type))
3266                     continue;
3267                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3268                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3269
3270                 if (FRAME_MBAFF) {
3271                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3272                 } else {
3273                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3274                 }
3275             }
3276         }
3277     }
3278     h->slice_type= old_slice_type;
3279     s->mb_x= end_x;
3280     s->mb_y= end_mb_y - FRAME_MBAFF;
3281     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3282     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3283 }
3284
3285 static void predict_field_decoding_flag(H264Context *h){
3286     MpegEncContext * const s = &h->s;
3287     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3288     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3289                 ? s->current_picture.f.mb_type[mb_xy - 1]
3290                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3291                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3292                 : 0;
3293     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3294 }
3295
3296 /**
3297  * Draw edges and report progress for the last MB row.
3298  */
3299 static void decode_finish_row(H264Context *h){
3300     MpegEncContext * const s = &h->s;
3301     int top = 16*(s->mb_y >> FIELD_PICTURE);
3302     int height = 16 << FRAME_MBAFF;
3303     int deblock_border = (16 + 4) << FRAME_MBAFF;
3304     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3305
3306     if (h->deblocking_filter) {
3307         if((top + height) >= pic_height)
3308             height += deblock_border;
3309
3310         top -= deblock_border;
3311     }
3312
3313     if (top >= pic_height || (top + height) < h->emu_edge_height)
3314         return;
3315
3316     height = FFMIN(height, pic_height - top);
3317     if (top < h->emu_edge_height) {
3318         height = top+height;
3319         top = 0;
3320     }
3321
3322     ff_draw_horiz_band(s, top, height);
3323
3324     if (s->dropable) return;
3325
3326     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3327                              s->picture_structure==PICT_BOTTOM_FIELD);
3328 }
3329
3330 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3331     H264Context *h = *(void**)arg;
3332     MpegEncContext * const s = &h->s;
3333     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3334     int lf_x_start = s->mb_x;
3335
3336     s->mb_skip_run= -1;
3337
3338     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3339                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3340
3341     if( h->pps.cabac ) {
3342         /* realign */
3343         align_get_bits( &s->gb );
3344
3345         /* init cabac */
3346         ff_init_cabac_states( &h->cabac);
3347         ff_init_cabac_decoder( &h->cabac,
3348                                s->gb.buffer + get_bits_count(&s->gb)/8,
3349                                (get_bits_left(&s->gb) + 7)/8);
3350
3351         ff_h264_init_cabac_states(h);
3352
3353         for(;;){
3354 //START_TIMER
3355             int ret = ff_h264_decode_mb_cabac(h);
3356             int eos;
3357 //STOP_TIMER("decode_mb_cabac")
3358
3359             if(ret>=0) ff_h264_hl_decode_mb(h);
3360
3361             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3362                 s->mb_y++;
3363
3364                 ret = ff_h264_decode_mb_cabac(h);
3365
3366                 if(ret>=0) ff_h264_hl_decode_mb(h);
3367                 s->mb_y--;
3368             }
3369             eos = get_cabac_terminate( &h->cabac );
3370
3371             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3372                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3373                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3374                 return 0;
3375             }
3376             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3377                 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);
3378                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3379                 return -1;
3380             }
3381
3382             if( ++s->mb_x >= s->mb_width ) {
3383                 loop_filter(h, lf_x_start, s->mb_x);
3384                 s->mb_x = lf_x_start = 0;
3385                 decode_finish_row(h);
3386                 ++s->mb_y;
3387                 if(FIELD_OR_MBAFF_PICTURE) {
3388                     ++s->mb_y;
3389                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3390                         predict_field_decoding_flag(h);
3391                 }
3392             }
3393
3394             if( eos || s->mb_y >= s->mb_height ) {
3395                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3396                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3397                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3398                 return 0;
3399             }
3400         }
3401
3402     } else {
3403         for(;;){
3404             int ret = ff_h264_decode_mb_cavlc(h);
3405
3406             if(ret>=0) ff_h264_hl_decode_mb(h);
3407
3408             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3409                 s->mb_y++;
3410                 ret = ff_h264_decode_mb_cavlc(h);
3411
3412                 if(ret>=0) ff_h264_hl_decode_mb(h);
3413                 s->mb_y--;
3414             }
3415
3416             if(ret<0){
3417                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3418                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3419                 return -1;
3420             }
3421
3422             if(++s->mb_x >= s->mb_width){
3423                 loop_filter(h, lf_x_start, s->mb_x);
3424                 s->mb_x = lf_x_start = 0;
3425                 decode_finish_row(h);
3426                 ++s->mb_y;
3427                 if(FIELD_OR_MBAFF_PICTURE) {
3428                     ++s->mb_y;
3429                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3430                         predict_field_decoding_flag(h);
3431                 }
3432                 if(s->mb_y >= s->mb_height){
3433                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3434
3435                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3436                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3437
3438                         return 0;
3439                     }else{
3440                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3441
3442                         return -1;
3443                     }
3444                 }
3445             }
3446
3447             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3448                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3449                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3450                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3451                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3452
3453                     return 0;
3454                 }else{
3455                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3456
3457                     return -1;
3458                 }
3459             }
3460         }
3461     }
3462
3463 #if 0
3464     for(;s->mb_y < s->mb_height; s->mb_y++){
3465         for(;s->mb_x < s->mb_width; s->mb_x++){
3466             int ret= decode_mb(h);
3467
3468             ff_h264_hl_decode_mb(h);
3469
3470             if(ret<0){
3471                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3472                 ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3473
3474                 return -1;
3475             }
3476
3477             if(++s->mb_x >= s->mb_width){
3478                 s->mb_x=0;
3479                 if(++s->mb_y >= s->mb_height){
3480                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
3481                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3482
3483                         return 0;
3484                     }else{
3485                         ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3486
3487                         return -1;
3488                     }
3489                 }
3490             }
3491
3492             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
3493                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
3494                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x-1, s->mb_y, (AC_END|DC_END|MV_END)&part_mask);
3495
3496                     return 0;
3497                 }else{
3498                     ff_er_add_slice(s, s->resync_mb_x, s->resync_mb_y, s->mb_x, s->mb_y, (AC_ERROR|DC_ERROR|MV_ERROR)&part_mask);
3499
3500                     return -1;
3501                 }
3502             }
3503         }
3504         s->mb_x=0;
3505         ff_draw_horiz_band(s, 16*s->mb_y, 16);
3506     }
3507 #endif
3508 }
3509
3510 /**
3511  * Call decode_slice() for each context.
3512  *
3513  * @param h h264 master context
3514  * @param context_count number of contexts to execute
3515  */
3516 static void execute_decode_slices(H264Context *h, int context_count){
3517     MpegEncContext * const s = &h->s;
3518     AVCodecContext * const avctx= s->avctx;
3519     H264Context *hx;
3520     int i;
3521
3522     if (s->avctx->hwaccel)
3523         return;
3524     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3525         return;
3526     if(context_count == 1) {
3527         decode_slice(avctx, &h);
3528     } else {
3529         for(i = 1; i < context_count; i++) {
3530             hx = h->thread_context[i];
3531             hx->s.error_recognition = avctx->error_recognition;
3532             hx->s.error_count = 0;
3533         }
3534
3535         avctx->execute(avctx, (void *)decode_slice,
3536                        h->thread_context, NULL, context_count, sizeof(void*));
3537
3538         /* pull back stuff from slices to master context */
3539         hx = h->thread_context[context_count - 1];
3540         s->mb_x = hx->s.mb_x;
3541         s->mb_y = hx->s.mb_y;
3542         s->dropable = hx->s.dropable;
3543         s->picture_structure = hx->s.picture_structure;
3544         for(i = 1; i < context_count; i++)
3545             h->s.error_count += h->thread_context[i]->s.error_count;
3546     }
3547 }
3548
3549
3550 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3551     MpegEncContext * const s = &h->s;
3552     AVCodecContext * const avctx= s->avctx;
3553     H264Context *hx; ///< thread context
3554     int buf_index;
3555     int context_count;
3556     int next_avc;
3557     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3558     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3559     int nal_index;
3560
3561     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3562     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3563         h->current_slice = 0;
3564         if (!s->first_field)
3565             s->current_picture_ptr= NULL;
3566         ff_h264_reset_sei(h);
3567     }
3568
3569     for(;pass <= 1;pass++){
3570         buf_index = 0;
3571         context_count = 0;
3572         next_avc = h->is_avc ? 0 : buf_size;
3573         nal_index = 0;
3574     for(;;){
3575         int consumed;
3576         int dst_length;
3577         int bit_length;
3578         const uint8_t *ptr;
3579         int i, nalsize = 0;
3580         int err;
3581
3582         if(buf_index >= next_avc) {
3583             if(buf_index >= buf_size) break;
3584             nalsize = 0;
3585             for(i = 0; i < h->nal_length_size; i++)
3586                 nalsize = (nalsize << 8) | buf[buf_index++];
3587             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3588                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3589                 break;
3590             }
3591             next_avc= buf_index + nalsize;
3592         } else {
3593             // start code prefix search
3594             for(; buf_index + 3 < next_avc; buf_index++){
3595                 // This should always succeed in the first iteration.
3596                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3597                     break;
3598             }
3599
3600             if(buf_index+3 >= buf_size) break;
3601
3602             buf_index+=3;
3603             if(buf_index >= next_avc) continue;
3604         }
3605
3606         hx = h->thread_context[context_count];
3607
3608         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3609         if (ptr==NULL || dst_length < 0){
3610             return -1;
3611         }
3612         i= buf_index + consumed;
3613         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3614            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3615             s->workaround_bugs |= FF_BUG_TRUNCATED;
3616
3617         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3618         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3619             dst_length--;
3620         }
3621         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3622
3623         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3624             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);
3625         }
3626
3627         if (h->is_avc && (nalsize != consumed) && nalsize){
3628             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3629         }
3630
3631         buf_index += consumed;
3632         nal_index++;
3633
3634         if(pass == 0) {
3635             // packets can sometimes contain multiple PPS/SPS
3636             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3637             // if so, when frame threading we can't start the next thread until we've read all of them
3638             switch (hx->nal_unit_type) {
3639                 case NAL_SPS:
3640                 case NAL_PPS:
3641                 case NAL_IDR_SLICE:
3642                 case NAL_SLICE:
3643                     nals_needed = nal_index;
3644             }
3645             continue;
3646         }
3647
3648         //FIXME do not discard SEI id
3649         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3650             continue;
3651
3652       again:
3653         err = 0;
3654         switch(hx->nal_unit_type){
3655         case NAL_IDR_SLICE:
3656             if (h->nal_unit_type != NAL_IDR_SLICE) {
3657                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3658                 return -1;
3659             }
3660             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3661         case NAL_SLICE:
3662             init_get_bits(&hx->s.gb, ptr, bit_length);
3663             hx->intra_gb_ptr=
3664             hx->inter_gb_ptr= &hx->s.gb;
3665             hx->s.data_partitioning = 0;
3666
3667             if((err = decode_slice_header(hx, h)))
3668                break;
3669
3670             s->current_picture_ptr->f.key_frame |=
3671                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3672                     (h->sei_recovery_frame_cnt >= 0);
3673
3674             if (h->current_slice == 1) {
3675                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3676                     decode_postinit(h, nal_index >= nals_needed);
3677                 }
3678
3679                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3680                     return -1;
3681                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3682                     ff_vdpau_h264_picture_start(s);
3683             }
3684
3685             if(hx->redundant_pic_count==0
3686                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3687                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3688                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3689                && avctx->skip_frame < AVDISCARD_ALL){
3690                 if(avctx->hwaccel) {
3691                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3692                         return -1;
3693                 }else
3694                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3695                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3696                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3697                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3698                 }else
3699                     context_count++;
3700             }
3701             break;
3702         case NAL_DPA:
3703             init_get_bits(&hx->s.gb, ptr, bit_length);
3704             hx->intra_gb_ptr=
3705             hx->inter_gb_ptr= NULL;
3706
3707             if ((err = decode_slice_header(hx, h)) < 0)
3708                 break;
3709
3710             hx->s.data_partitioning = 1;
3711
3712             break;
3713         case NAL_DPB:
3714             init_get_bits(&hx->intra_gb, ptr, bit_length);
3715             hx->intra_gb_ptr= &hx->intra_gb;
3716             break;
3717         case NAL_DPC:
3718             init_get_bits(&hx->inter_gb, ptr, bit_length);
3719             hx->inter_gb_ptr= &hx->inter_gb;
3720
3721             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3722                && s->context_initialized
3723                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3724                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3725                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3726                && avctx->skip_frame < AVDISCARD_ALL)
3727                 context_count++;
3728             break;
3729         case NAL_SEI:
3730             init_get_bits(&s->gb, ptr, bit_length);
3731             ff_h264_decode_sei(h);
3732             break;
3733         case NAL_SPS:
3734             init_get_bits(&s->gb, ptr, bit_length);
3735             ff_h264_decode_seq_parameter_set(h);
3736
3737             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3738                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3739                 s->low_delay=1;
3740
3741             if(avctx->has_b_frames < 2)
3742                 avctx->has_b_frames= !s->low_delay;
3743
3744             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3745                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3746                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3747                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3748
3749                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3750                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3751                     dsputil_init(&s->dsp, s->avctx);
3752                 } else {
3753                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3754                     return -1;
3755                 }
3756             }
3757             break;
3758         case NAL_PPS:
3759             init_get_bits(&s->gb, ptr, bit_length);
3760
3761             ff_h264_decode_picture_parameter_set(h, bit_length);
3762
3763             break;
3764         case NAL_AUD:
3765         case NAL_END_SEQUENCE:
3766         case NAL_END_STREAM:
3767         case NAL_FILLER_DATA:
3768         case NAL_SPS_EXT:
3769         case NAL_AUXILIARY_SLICE:
3770             break;
3771         default:
3772             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3773         }
3774
3775         if(context_count == h->max_contexts) {
3776             execute_decode_slices(h, context_count);
3777             context_count = 0;
3778         }
3779
3780         if (err < 0)
3781             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3782         else if(err == 1) {
3783             /* Slice could not be decoded in parallel mode, copy down
3784              * NAL unit stuff to context 0 and restart. Note that
3785              * rbsp_buffer is not transferred, but since we no longer
3786              * run in parallel mode this should not be an issue. */
3787             h->nal_unit_type = hx->nal_unit_type;
3788             h->nal_ref_idc   = hx->nal_ref_idc;
3789             hx = h;
3790             goto again;
3791         }
3792     }
3793     }
3794     if(context_count)
3795         execute_decode_slices(h, context_count);
3796     return buf_index;
3797 }
3798
3799 /**
3800  * returns the number of bytes consumed for building the current frame
3801  */
3802 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3803         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3804         if(pos+10>buf_size) pos=buf_size; // oops ;)
3805
3806         return pos;
3807 }
3808
3809 static int decode_frame(AVCodecContext *avctx,
3810                              void *data, int *data_size,
3811                              AVPacket *avpkt)
3812 {
3813     const uint8_t *buf = avpkt->data;
3814     int buf_size = avpkt->size;
3815     H264Context *h = avctx->priv_data;
3816     MpegEncContext *s = &h->s;
3817     AVFrame *pict = data;
3818     int buf_index;
3819
3820     s->flags= avctx->flags;
3821     s->flags2= avctx->flags2;
3822
3823    /* end of stream, output what is still in the buffers */
3824  out:
3825     if (buf_size == 0) {
3826         Picture *out;
3827         int i, out_idx;
3828
3829         s->current_picture_ptr = NULL;
3830
3831 //FIXME factorize this with the output code below
3832         out = h->delayed_pic[0];
3833         out_idx = 0;
3834         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3835             if(h->delayed_pic[i]->poc < out->poc){
3836                 out = h->delayed_pic[i];
3837                 out_idx = i;
3838             }
3839
3840         for(i=out_idx; h->delayed_pic[i]; i++)
3841             h->delayed_pic[i] = h->delayed_pic[i+1];
3842
3843         if(out){
3844             *data_size = sizeof(AVFrame);
3845             *pict= *(AVFrame*)out;
3846         }
3847
3848         return 0;
3849     }
3850
3851     buf_index=decode_nal_units(h, buf, buf_size);
3852     if(buf_index < 0)
3853         return -1;
3854
3855     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3856         buf_size = 0;
3857         goto out;
3858     }
3859
3860     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3861         if (avctx->skip_frame >= AVDISCARD_NONREF)
3862             return 0;
3863         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3864         return -1;
3865     }
3866
3867     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3868
3869         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3870
3871         field_end(h, 0);
3872
3873         if (!h->next_output_pic) {
3874             /* Wait for second field. */
3875             *data_size = 0;
3876
3877         } else {
3878             *data_size = sizeof(AVFrame);
3879             *pict = *(AVFrame*)h->next_output_pic;
3880         }
3881     }
3882
3883     assert(pict->data[0] || !*data_size);
3884     ff_print_debug_info(s, pict);
3885 //printf("out %d\n", (int)pict->data[0]);
3886
3887     return get_consumed_bytes(s, buf_index, buf_size);
3888 }
3889 #if 0
3890 static inline void fill_mb_avail(H264Context *h){
3891     MpegEncContext * const s = &h->s;
3892     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3893
3894     if(s->mb_y){
3895         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3896         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3897         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3898     }else{
3899         h->mb_avail[0]=
3900         h->mb_avail[1]=
3901         h->mb_avail[2]= 0;
3902     }
3903     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3904     h->mb_avail[4]= 1; //FIXME move out
3905     h->mb_avail[5]= 0; //FIXME move out
3906 }
3907 #endif
3908
3909 #ifdef TEST
3910 #undef printf
3911 #undef random
3912 #define COUNT 8000
3913 #define SIZE (COUNT*40)
3914 int main(void){
3915     int i;
3916     uint8_t temp[SIZE];
3917     PutBitContext pb;
3918     GetBitContext gb;
3919 //    int int_temp[10000];
3920     DSPContext dsp;
3921     AVCodecContext avctx;
3922
3923     dsputil_init(&dsp, &avctx);
3924
3925     init_put_bits(&pb, temp, SIZE);
3926     printf("testing unsigned exp golomb\n");
3927     for(i=0; i<COUNT; i++){
3928         START_TIMER
3929         set_ue_golomb(&pb, i);
3930         STOP_TIMER("set_ue_golomb");
3931     }
3932     flush_put_bits(&pb);
3933
3934     init_get_bits(&gb, temp, 8*SIZE);
3935     for(i=0; i<COUNT; i++){
3936         int j, s;
3937
3938         s= show_bits(&gb, 24);
3939
3940         START_TIMER
3941         j= get_ue_golomb(&gb);
3942         if(j != i){
3943             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3944 //            return -1;
3945         }
3946         STOP_TIMER("get_ue_golomb");
3947     }
3948
3949
3950     init_put_bits(&pb, temp, SIZE);
3951     printf("testing signed exp golomb\n");
3952     for(i=0; i<COUNT; i++){
3953         START_TIMER
3954         set_se_golomb(&pb, i - COUNT/2);
3955         STOP_TIMER("set_se_golomb");
3956     }
3957     flush_put_bits(&pb);
3958
3959     init_get_bits(&gb, temp, 8*SIZE);
3960     for(i=0; i<COUNT; i++){
3961         int j, s;
3962
3963         s= show_bits(&gb, 24);
3964
3965         START_TIMER
3966         j= get_se_golomb(&gb);
3967         if(j != i - COUNT/2){
3968             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3969 //            return -1;
3970         }
3971         STOP_TIMER("get_se_golomb");
3972     }
3973
3974 #if 0
3975     printf("testing 4x4 (I)DCT\n");
3976
3977     DCTELEM block[16];
3978     uint8_t src[16], ref[16];
3979     uint64_t error= 0, max_error=0;
3980
3981     for(i=0; i<COUNT; i++){
3982         int j;
3983 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
3984         for(j=0; j<16; j++){
3985             ref[j]= random()%255;
3986             src[j]= random()%255;
3987         }
3988
3989         h264_diff_dct_c(block, src, ref, 4);
3990
3991         //normalize
3992         for(j=0; j<16; j++){
3993 //            printf("%d ", block[j]);
3994             block[j]= block[j]*4;
3995             if(j&1) block[j]= (block[j]*4 + 2)/5;
3996             if(j&4) block[j]= (block[j]*4 + 2)/5;
3997         }
3998 //        printf("\n");
3999
4000         h->h264dsp.h264_idct_add(ref, block, 4);
4001 /*        for(j=0; j<16; j++){
4002             printf("%d ", ref[j]);
4003         }
4004         printf("\n");*/
4005
4006         for(j=0; j<16; j++){
4007             int diff= FFABS(src[j] - ref[j]);
4008
4009             error+= diff*diff;
4010             max_error= FFMAX(max_error, diff);
4011         }
4012     }
4013     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
4014     printf("testing quantizer\n");
4015     for(qp=0; qp<52; qp++){
4016         for(i=0; i<16; i++)
4017             src1_block[i]= src2_block[i]= random()%255;
4018
4019     }
4020     printf("Testing NAL layer\n");
4021
4022     uint8_t bitstream[COUNT];
4023     uint8_t nal[COUNT*2];
4024     H264Context h;
4025     memset(&h, 0, sizeof(H264Context));
4026
4027     for(i=0; i<COUNT; i++){
4028         int zeros= i;
4029         int nal_length;
4030         int consumed;
4031         int out_length;
4032         uint8_t *out;
4033         int j;
4034
4035         for(j=0; j<COUNT; j++){
4036             bitstream[j]= (random() % 255) + 1;
4037         }
4038
4039         for(j=0; j<zeros; j++){
4040             int pos= random() % COUNT;
4041             while(bitstream[pos] == 0){
4042                 pos++;
4043                 pos %= COUNT;
4044             }
4045             bitstream[pos]=0;
4046         }
4047
4048         START_TIMER
4049
4050         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
4051         if(nal_length<0){
4052             printf("encoding failed\n");
4053             return -1;
4054         }
4055
4056         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
4057
4058         STOP_TIMER("NAL")
4059
4060         if(out_length != COUNT){
4061             printf("incorrect length %d %d\n", out_length, COUNT);
4062             return -1;
4063         }
4064
4065         if(consumed != nal_length){
4066             printf("incorrect consumed length %d %d\n", nal_length, consumed);
4067             return -1;
4068         }
4069
4070         if(memcmp(bitstream, out, COUNT)){
4071             printf("mismatch\n");
4072             return -1;
4073         }
4074     }
4075 #endif
4076
4077     printf("Testing RBSP\n");
4078
4079
4080     return 0;
4081 }
4082 #endif /* TEST */
4083
4084
4085 av_cold void ff_h264_free_context(H264Context *h)
4086 {
4087     int i;
4088
4089     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4090
4091     for(i = 0; i < MAX_SPS_COUNT; i++)
4092         av_freep(h->sps_buffers + i);
4093
4094     for(i = 0; i < MAX_PPS_COUNT; i++)
4095         av_freep(h->pps_buffers + i);
4096 }
4097
4098 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4099 {
4100     H264Context *h = avctx->priv_data;
4101     MpegEncContext *s = &h->s;
4102
4103     ff_h264_free_context(h);
4104
4105     MPV_common_end(s);
4106
4107 //    memset(h, 0, sizeof(H264Context));
4108
4109     return 0;
4110 }
4111
4112 static const AVProfile profiles[] = {
4113     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4114     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4115     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4116     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4117     { FF_PROFILE_H264_HIGH,                 "High"                  },
4118     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4119     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4120     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4121     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4122     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4123     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4124     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4125     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4126     { FF_PROFILE_UNKNOWN },
4127 };
4128
4129 AVCodec ff_h264_decoder = {
4130     "h264",
4131     AVMEDIA_TYPE_VIDEO,
4132     CODEC_ID_H264,
4133     sizeof(H264Context),
4134     ff_h264_decode_init,
4135     NULL,
4136     ff_h264_decode_end,
4137     decode_frame,
4138     /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4139         CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4140     .flush= flush_dpb,
4141     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4142     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4143     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4144     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4145 };
4146
4147 #if CONFIG_H264_VDPAU_DECODER
4148 AVCodec ff_h264_vdpau_decoder = {
4149     "h264_vdpau",
4150     AVMEDIA_TYPE_VIDEO,
4151     CODEC_ID_H264,
4152     sizeof(H264Context),
4153     ff_h264_decode_init,
4154     NULL,
4155     ff_h264_decode_end,
4156     decode_frame,
4157     CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4158     .flush= flush_dpb,
4159     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4160     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4161     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4162 };
4163 #endif