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

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavutil/imgutils.h"
  29 #include "internal.h"
  30 #include "dsputil.h"
  31 #include "avcodec.h"
  32 #include "mpegvideo.h"
  33 #include "h264.h"
  34 #include "h264data.h"
  35 #include "h264_mvpred.h"
  36 #include "golomb.h"
  37 #include "mathops.h"
  38 #include "rectangle.h"
  39 #include "thread.h"
  40 #include "vdpau_internal.h"
  41 #include "libavutil/avassert.h"
  42
  43 #include "cabac.h"
  44
  45 //#undef NDEBUG
  46 #include <assert.h>
  47
  48 static const uint8_t rem6[QP_MAX_NUM+1]={
  49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[QP_MAX_NUM+1]={
  53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
  54 };
  55
  56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  57     PIX_FMT_DXVA2_VLD,
  58     PIX_FMT_VAAPI_VLD,
  59     PIX_FMT_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 /**
  64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  65  */
  66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  67     MpegEncContext * const s = &h->s;
  68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  70     int i;
  71
  72     if(!(h->top_samples_available&0x8000)){
  73         for(i=0; i<4; i++){
  74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  75             if(status<0){
  76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  77                 return -1;
  78             } else if(status){
  79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  80             }
  81         }
  82     }
  83
  84     if((h->left_samples_available&0x8888)!=0x8888){
  85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  86         for(i=0; i<4; i++){
  87             if(!(h->left_samples_available&mask[i])){
  88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  89                 if(status<0){
  90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  91                     return -1;
  92                 } else if(status){
  93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  94                 }
  95             }
  96         }
  97     }
  98
  99     return 0;
 100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 101
 102 /**
 103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 104  */
 105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 106     MpegEncContext * const s = &h->s;
 107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 109
 110     if(mode > 6U) {
 111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 112         return -1;
 113     }
 114
 115     if(!(h->top_samples_available&0x8000)){
 116         mode= top[ mode ];
 117         if(mode<0){
 118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 119             return -1;
 120         }
 121     }
 122
 123     if((h->left_samples_available&0x8080) != 0x8080){
 124         mode= left[ mode ];
 125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 127         }
 128         if(mode<0){
 129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 130             return -1;
 131         }
 132     }
 133
 134     return mode;
 135 }
 136
 137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 138     int i, si, di;
 139     uint8_t *dst;
 140     int bufidx;
 141
 142 //    src[0]&0x80;                //forbidden bit
 143     h->nal_ref_idc= src[0]>>5;
 144     h->nal_unit_type= src[0]&0x1F;
 145
 146     src++; length--;
 147
 148 #if HAVE_FAST_UNALIGNED
 149 # if HAVE_FAST_64BIT
 150 #   define RS 7
 151     for(i=0; i+1<length; i+=9){
 152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 153 # else
 154 #   define RS 3
 155     for(i=0; i+1<length; i+=5){
 156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 157 # endif
 158             continue;
 159         if(i>0 && !src[i]) i--;
 160         while(src[i]) i++;
 161 #else
 162 #   define RS 0
 163     for(i=0; i+1<length; i+=2){
 164         if(src[i]) continue;
 165         if(i>0 && src[i-1]==0) i--;
 166 #endif
 167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 168             if(src[i+2]!=3){
 169                 /* startcode, so we must be past the end */
 170                 length=i;
 171             }
 172             break;
 173         }
 174         i-= RS;
 175     }
 176
 177     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->thread_opaque != s->current_picture.thread_opaque ||
 265            (ref->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->thread_opaque != s->current_picture.thread_opaque ||
 277            (ref->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.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->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->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->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->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->reference - 1));
 508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 509     }
 510     src_cb= pic->data[1] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
 511     src_cr= pic->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].data;
 668         int off= ((mx+64)<<h->pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize;
 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)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 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.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 #define hl_motion_fn(sh, bits) \
 782 static av_always_inline void hl_motion_ ## bits(H264Context *h, \
 783                                        uint8_t *dest_y, \
 784                                        uint8_t *dest_cb, uint8_t *dest_cr, \
 785                                        qpel_mc_func (*qpix_put)[16], \
 786                                        h264_chroma_mc_func (*chroma_put), \
 787                                        qpel_mc_func (*qpix_avg)[16], \
 788                                        h264_chroma_mc_func (*chroma_avg), \
 789                                        h264_weight_func *weight_op, \
 790                                        h264_biweight_func *weight_avg, \
 791                                        int chroma444) \
 792 { \
 793     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put, \
 794               qpix_avg, chroma_avg, weight_op, weight_avg, sh, chroma444); \
 795 }
 796 hl_motion_fn(0, 8);
 797 hl_motion_fn(1, 16);
 798
 799 static void free_tables(H264Context *h, int free_rbsp){
 800     int i;
 801     H264Context *hx;
 802
 803     av_freep(&h->intra4x4_pred_mode);
 804     av_freep(&h->chroma_pred_mode_table);
 805     av_freep(&h->cbp_table);
 806     av_freep(&h->mvd_table[0]);
 807     av_freep(&h->mvd_table[1]);
 808     av_freep(&h->direct_table);
 809     av_freep(&h->non_zero_count);
 810     av_freep(&h->slice_table_base);
 811     h->slice_table= NULL;
 812     av_freep(&h->list_counts);
 813
 814     av_freep(&h->mb2b_xy);
 815     av_freep(&h->mb2br_xy);
 816
 817     for(i = 0; i < MAX_THREADS; i++) {
 818         hx = h->thread_context[i];
 819         if(!hx) continue;
 820         av_freep(&hx->top_borders[1]);
 821         av_freep(&hx->top_borders[0]);
 822         av_freep(&hx->s.obmc_scratchpad);
 823         if (free_rbsp){
 824             av_freep(&hx->rbsp_buffer[1]);
 825             av_freep(&hx->rbsp_buffer[0]);
 826             hx->rbsp_buffer_size[0] = 0;
 827             hx->rbsp_buffer_size[1] = 0;
 828         }
 829         if (i) av_freep(&h->thread_context[i]);
 830     }
 831 }
 832
 833 static void init_dequant8_coeff_table(H264Context *h){
 834     int i,j,q,x;
 835     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 836
 837     for(i=0; i<6; i++ ){
 838         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 839         for(j=0; j<i; j++){
 840             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 841                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 842                 break;
 843             }
 844         }
 845         if(j<i)
 846             continue;
 847
 848         for(q=0; q<max_qp+1; q++){
 849             int shift = div6[q];
 850             int idx = rem6[q];
 851             for(x=0; x<64; x++)
 852                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 853                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 854                     h->pps.scaling_matrix8[i][x]) << shift;
 855         }
 856     }
 857 }
 858
 859 static void init_dequant4_coeff_table(H264Context *h){
 860     int i,j,q,x;
 861     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 862     for(i=0; i<6; i++ ){
 863         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 864         for(j=0; j<i; j++){
 865             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 866                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 867                 break;
 868             }
 869         }
 870         if(j<i)
 871             continue;
 872
 873         for(q=0; q<max_qp+1; q++){
 874             int shift = div6[q] + 2;
 875             int idx = rem6[q];
 876             for(x=0; x<16; x++)
 877                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 878                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 879                     h->pps.scaling_matrix4[i][x]) << shift;
 880         }
 881     }
 882 }
 883
 884 static void init_dequant_tables(H264Context *h){
 885     int i,x;
 886     init_dequant4_coeff_table(h);
 887     if(h->pps.transform_8x8_mode)
 888         init_dequant8_coeff_table(h);
 889     if(h->sps.transform_bypass){
 890         for(i=0; i<6; i++)
 891             for(x=0; x<16; x++)
 892                 h->dequant4_coeff[i][0][x] = 1<<6;
 893         if(h->pps.transform_8x8_mode)
 894             for(i=0; i<6; i++)
 895                 for(x=0; x<64; x++)
 896                     h->dequant8_coeff[i][0][x] = 1<<6;
 897     }
 898 }
 899
 900
 901 int ff_h264_alloc_tables(H264Context *h){
 902     MpegEncContext * const s = &h->s;
 903     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 904     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 905     int x,y;
 906
 907     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 908
 909     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 910     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 911     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 912
 913     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 914     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 915     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 916     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 917     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 918
 919     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 920     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 921
 922     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 923     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 924     for(y=0; y<s->mb_height; y++){
 925         for(x=0; x<s->mb_width; x++){
 926             const int mb_xy= x + y*s->mb_stride;
 927             const int b_xy = 4*x + 4*y*h->b_stride;
 928
 929             h->mb2b_xy [mb_xy]= b_xy;
 930             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 931         }
 932     }
 933
 934     s->obmc_scratchpad = NULL;
 935
 936     if(!h->dequant4_coeff[0])
 937         init_dequant_tables(h);
 938
 939     return 0;
 940 fail:
 941     free_tables(h, 1);
 942     return -1;
 943 }
 944
 945 /**
 946  * Mimic alloc_tables(), but for every context thread.
 947  */
 948 static void clone_tables(H264Context *dst, H264Context *src, int i){
 949     MpegEncContext * const s = &src->s;
 950     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 951     dst->non_zero_count           = src->non_zero_count;
 952     dst->slice_table              = src->slice_table;
 953     dst->cbp_table                = src->cbp_table;
 954     dst->mb2b_xy                  = src->mb2b_xy;
 955     dst->mb2br_xy                 = src->mb2br_xy;
 956     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 957     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 958     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 959     dst->direct_table             = src->direct_table;
 960     dst->list_counts              = src->list_counts;
 961
 962     dst->s.obmc_scratchpad = NULL;
 963     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
 964 }
 965
 966 /**
 967  * Init context
 968  * Allocate buffers which are not shared amongst multiple threads.
 969  */
 970 static int context_init(H264Context *h){
 971     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 972     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 973
 974     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 975     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 976
 977     return 0;
 978 fail:
 979     return -1; // free_tables will clean up for us
 980 }
 981
 982 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 983
 984 static av_cold void common_init(H264Context *h){
 985     MpegEncContext * const s = &h->s;
 986
 987     s->width = s->avctx->width;
 988     s->height = s->avctx->height;
 989     s->codec_id= s->avctx->codec->id;
 990
 991     ff_h264dsp_init(&h->h264dsp, 8);
 992     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
 993
 994     h->dequant_coeff_pps= -1;
 995     s->unrestricted_mv=1;
 996     s->decode=1; //FIXME
 997
 998     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 999
1000     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1001     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1002 }
1003
1004 int ff_h264_decode_extradata(H264Context *h)
1005 {
1006     AVCodecContext *avctx = h->s.avctx;
1007
1008     if(avctx->extradata[0] == 1){
1009         int i, cnt, nalsize;
1010         unsigned char *p = avctx->extradata;
1011
1012         h->is_avc = 1;
1013
1014         if(avctx->extradata_size < 7) {
1015             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1016             return -1;
1017         }
1018         /* sps and pps in the avcC always have length coded with 2 bytes,
1019            so put a fake nal_length_size = 2 while parsing them */
1020         h->nal_length_size = 2;
1021         // Decode sps from avcC
1022         cnt = *(p+5) & 0x1f; // Number of sps
1023         p += 6;
1024         for (i = 0; i < cnt; i++) {
1025             nalsize = AV_RB16(p) + 2;
1026             if(decode_nal_units(h, p, nalsize) < 0) {
1027                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1028                 return -1;
1029             }
1030             p += nalsize;
1031         }
1032         // Decode pps from avcC
1033         cnt = *(p++); // Number of pps
1034         for (i = 0; i < cnt; i++) {
1035             nalsize = AV_RB16(p) + 2;
1036             if (decode_nal_units(h, p, nalsize) < 0) {
1037                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1038                 return -1;
1039             }
1040             p += nalsize;
1041         }
1042         // Now store right nal length size, that will be use to parse all other nals
1043         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1044     } else {
1045         h->is_avc = 0;
1046         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1047             return -1;
1048     }
1049     return 0;
1050 }
1051
1052 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1053     H264Context *h= avctx->priv_data;
1054     MpegEncContext * const s = &h->s;
1055
1056     MPV_decode_defaults(s);
1057
1058     s->avctx = avctx;
1059     common_init(h);
1060
1061     s->out_format = FMT_H264;
1062     s->workaround_bugs= avctx->workaround_bugs;
1063
1064     // set defaults
1065 //    s->decode_mb= ff_h263_decode_mb;
1066     s->quarter_sample = 1;
1067     if(!avctx->has_b_frames)
1068     s->low_delay= 1;
1069
1070     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1071
1072     ff_h264_decode_init_vlc();
1073
1074     h->pixel_shift = 0;
1075     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1076
1077     h->thread_context[0] = h;
1078     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1079     h->prev_poc_msb= 1<<16;
1080     h->x264_build = -1;
1081     ff_h264_reset_sei(h);
1082     if(avctx->codec_id == CODEC_ID_H264){
1083         if(avctx->ticks_per_frame == 1){
1084             s->avctx->time_base.den *=2;
1085         }
1086         avctx->ticks_per_frame = 2;
1087     }
1088
1089     if(avctx->extradata_size > 0 && avctx->extradata &&
1090         ff_h264_decode_extradata(h))
1091         return -1;
1092
1093     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1094         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1095         s->low_delay = 0;
1096     }
1097
1098     return 0;
1099 }
1100
1101 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1102 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1103 {
1104     int i;
1105
1106     for (i=0; i<count; i++){
1107         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1108                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1109                 !from[i]));
1110         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1111     }
1112 }
1113
1114 static void copy_parameter_set(void **to, void **from, int count, int size)
1115 {
1116     int i;
1117
1118     for (i=0; i<count; i++){
1119         if (to[i] && !from[i]) av_freep(&to[i]);
1120         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1121
1122         if (from[i]) memcpy(to[i], from[i], size);
1123     }
1124 }
1125
1126 static int decode_init_thread_copy(AVCodecContext *avctx){
1127     H264Context *h= avctx->priv_data;
1128
1129     if (!avctx->is_copy) return 0;
1130     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1131     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1132
1133     return 0;
1134 }
1135
1136 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1137 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1138     H264Context *h= dst->priv_data, *h1= src->priv_data;
1139     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1140     int inited = s->context_initialized, err;
1141     int i;
1142
1143     if(dst == src || !s1->context_initialized) return 0;
1144
1145     err = ff_mpeg_update_thread_context(dst, src);
1146     if(err) return err;
1147
1148     //FIXME handle width/height changing
1149     if(!inited){
1150         for(i = 0; i < MAX_SPS_COUNT; i++)
1151             av_freep(h->sps_buffers + i);
1152
1153         for(i = 0; i < MAX_PPS_COUNT; i++)
1154             av_freep(h->pps_buffers + i);
1155
1156         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1157         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1158         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1159         ff_h264_alloc_tables(h);
1160         context_init(h);
1161
1162         for(i=0; i<2; i++){
1163             h->rbsp_buffer[i] = NULL;
1164             h->rbsp_buffer_size[i] = 0;
1165         }
1166
1167         h->thread_context[0] = h;
1168
1169         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1170         // so this has to be allocated here
1171         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1172
1173         s->dsp.clear_blocks(h->mb);
1174         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1175     }
1176
1177     //extradata/NAL handling
1178     h->is_avc          = h1->is_avc;
1179
1180     //SPS/PPS
1181     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1182     h->sps             = h1->sps;
1183     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1184     h->pps             = h1->pps;
1185
1186     //Dequantization matrices
1187     //FIXME these are big - can they be only copied when PPS changes?
1188     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1189
1190     for(i=0; i<6; i++)
1191         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1192
1193     for(i=0; i<6; i++)
1194         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1195
1196     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1197
1198     //POC timing
1199     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1200
1201     //reference lists
1202     copy_fields(h, h1, ref_count, list_count);
1203     copy_fields(h, h1, ref_list,  intra_gb);
1204     copy_fields(h, h1, short_ref, cabac_init_idc);
1205
1206     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1207     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1208     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1209
1210     h->last_slice_type = h1->last_slice_type;
1211
1212     if(!s->current_picture_ptr) return 0;
1213
1214     if(!s->dropable) {
1215         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1216         h->prev_poc_msb     = h->poc_msb;
1217         h->prev_poc_lsb     = h->poc_lsb;
1218     }
1219     h->prev_frame_num_offset= h->frame_num_offset;
1220     h->prev_frame_num       = h->frame_num;
1221     h->outputed_poc         = h->next_outputed_poc;
1222
1223     return 0;
1224 }
1225
1226 int ff_h264_frame_start(H264Context *h){
1227     MpegEncContext * const s = &h->s;
1228     int i;
1229     const int pixel_shift = h->pixel_shift;
1230     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1231
1232     if(MPV_frame_start(s, s->avctx) < 0)
1233         return -1;
1234     ff_er_frame_start(s);
1235     /*
1236      * MPV_frame_start uses pict_type to derive key_frame.
1237      * This is incorrect for H.264; IDR markings must be used.
1238      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1239      * See decode_nal_units().
1240      */
1241     s->current_picture_ptr->key_frame= 0;
1242     s->current_picture_ptr->mmco_reset= 0;
1243
1244     assert(s->linesize && s->uvlinesize);
1245
1246     for(i=0; i<16; i++){
1247         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1248         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1249     }
1250     for(i=0; i<16; i++){
1251         h->block_offset[16+i]=
1252         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1253         h->block_offset[48+16+i]=
1254         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1255     }
1256
1257     /* can't be in alloc_tables because linesize isn't known there.
1258      * FIXME: redo bipred weight to not require extra buffer? */
1259     for(i = 0; i < thread_count; i++)
1260         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1261             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1262
1263     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1264     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1265
1266 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
1267
1268     // We mark the current picture as non-reference after allocating it, so
1269     // that if we break out due to an error it can be released automatically
1270     // in the next MPV_frame_start().
1271     // SVQ3 as well as most other codecs have only last/next/current and thus
1272     // get released even with set reference, besides SVQ3 and others do not
1273     // mark frames as reference later "naturally".
1274     if(s->codec_id != CODEC_ID_SVQ3)
1275         s->current_picture_ptr->reference= 0;
1276
1277     s->current_picture_ptr->field_poc[0]=
1278     s->current_picture_ptr->field_poc[1]= INT_MAX;
1279
1280     h->next_output_pic = NULL;
1281
1282     assert(s->current_picture_ptr->long_ref==0);
1283
1284     return 0;
1285 }
1286
1287 /**
1288   * Run setup operations that must be run after slice header decoding.
1289   * This includes finding the next displayed frame.
1290   *
1291   * @param h h264 master context
1292   * @param setup_finished enough NALs have been read that we can call
1293   * ff_thread_finish_setup()
1294   */
1295 static void decode_postinit(H264Context *h, int setup_finished){
1296     MpegEncContext * const s = &h->s;
1297     Picture *out = s->current_picture_ptr;
1298     Picture *cur = s->current_picture_ptr;
1299     int i, pics, out_of_order, out_idx;
1300
1301     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1302     s->current_picture_ptr->pict_type= s->pict_type;
1303
1304     if (h->next_output_pic) return;
1305
1306     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1307         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1308         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1309         //to find this yet, so we assume the worst for now.
1310         //if (setup_finished)
1311         //    ff_thread_finish_setup(s->avctx);
1312         return;
1313     }
1314
1315     cur->interlaced_frame = 0;
1316     cur->repeat_pict = 0;
1317
1318     /* Signal interlacing information externally. */
1319     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1320
1321     if(h->sps.pic_struct_present_flag){
1322         switch (h->sei_pic_struct)
1323         {
1324         case SEI_PIC_STRUCT_FRAME:
1325             break;
1326         case SEI_PIC_STRUCT_TOP_FIELD:
1327         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1328             cur->interlaced_frame = 1;
1329             break;
1330         case SEI_PIC_STRUCT_TOP_BOTTOM:
1331         case SEI_PIC_STRUCT_BOTTOM_TOP:
1332             if (FIELD_OR_MBAFF_PICTURE)
1333                 cur->interlaced_frame = 1;
1334             else
1335                 // try to flag soft telecine progressive
1336                 cur->interlaced_frame = h->prev_interlaced_frame;
1337             break;
1338         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1339         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1340             // Signal the possibility of telecined film externally (pic_struct 5,6)
1341             // From these hints, let the applications decide if they apply deinterlacing.
1342             cur->repeat_pict = 1;
1343             break;
1344         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1345             // Force progressive here, as doubling interlaced frame is a bad idea.
1346             cur->repeat_pict = 2;
1347             break;
1348         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1349             cur->repeat_pict = 4;
1350             break;
1351         }
1352
1353         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1354             cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
1355     }else{
1356         /* Derive interlacing flag from used decoding process. */
1357         cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1358     }
1359     h->prev_interlaced_frame = cur->interlaced_frame;
1360
1361     if (cur->field_poc[0] != cur->field_poc[1]){
1362         /* Derive top_field_first from field pocs. */
1363         cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
1364     }else{
1365         if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
1366             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1367             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1368               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1369                 cur->top_field_first = 1;
1370             else
1371                 cur->top_field_first = 0;
1372         }else{
1373             /* Most likely progressive */
1374             cur->top_field_first = 0;
1375         }
1376     }
1377
1378     //FIXME do something with unavailable reference frames
1379
1380     /* Sort B-frames into display order */
1381
1382     if(h->sps.bitstream_restriction_flag
1383        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1384         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1385         s->low_delay = 0;
1386     }
1387
1388     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1389        && !h->sps.bitstream_restriction_flag){
1390         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1391         s->low_delay= 0;
1392     }
1393
1394     pics = 0;
1395     while(h->delayed_pic[pics]) pics++;
1396
1397     assert(pics <= MAX_DELAYED_PIC_COUNT);
1398
1399     h->delayed_pic[pics++] = cur;
1400     if(cur->reference == 0)
1401         cur->reference = DELAYED_PIC_REF;
1402
1403     out = h->delayed_pic[0];
1404     out_idx = 0;
1405     for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1406         if(h->delayed_pic[i]->poc < out->poc){
1407             out = h->delayed_pic[i];
1408             out_idx = i;
1409         }
1410     if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
1411         h->next_outputed_poc= INT_MIN;
1412     out_of_order = out->poc < h->next_outputed_poc;
1413
1414     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1415         { }
1416     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1417        || (s->low_delay &&
1418         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1419          || cur->pict_type == AV_PICTURE_TYPE_B)))
1420     {
1421         s->low_delay = 0;
1422         s->avctx->has_b_frames++;
1423     }
1424
1425     if(out_of_order || pics > s->avctx->has_b_frames){
1426         out->reference &= ~DELAYED_PIC_REF;
1427         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1428                          // or else the first thread can release the picture and reuse it unsafely
1429         for(i=out_idx; h->delayed_pic[i]; i++)
1430             h->delayed_pic[i] = h->delayed_pic[i+1];
1431     }
1432     if(!out_of_order && pics > s->avctx->has_b_frames){
1433         h->next_output_pic = out;
1434         if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
1435             h->next_outputed_poc = INT_MIN;
1436         } else
1437             h->next_outputed_poc = out->poc;
1438     }else{
1439         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1440     }
1441
1442     if (setup_finished)
1443         ff_thread_finish_setup(s->avctx);
1444 }
1445
1446 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int chroma444, int simple){
1447     MpegEncContext * const s = &h->s;
1448     uint8_t *top_border;
1449     int top_idx = 1;
1450     const int pixel_shift = h->pixel_shift;
1451
1452     src_y  -=   linesize;
1453     src_cb -= uvlinesize;
1454     src_cr -= uvlinesize;
1455
1456     if(!simple && FRAME_MBAFF){
1457         if(s->mb_y&1){
1458             if(!MB_MBAFF){
1459                 top_border = h->top_borders[0][s->mb_x];
1460                 AV_COPY128(top_border, src_y + 15*linesize);
1461                 if (pixel_shift)
1462                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1463                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1464                     if(chroma444){
1465                         if (pixel_shift){
1466                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1467                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1468                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1469                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1470                         } else {
1471                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1472                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1473                         }
1474                     } else {
1475                         if (pixel_shift) {
1476                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1477                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1478                         } else {
1479                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1480                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1481                         }
1482                     }
1483                 }
1484             }
1485         }else if(MB_MBAFF){
1486             top_idx = 0;
1487         }else
1488             return;
1489     }
1490
1491     top_border = h->top_borders[top_idx][s->mb_x];
1492     // There are two lines saved, the line above the the top macroblock of a pair,
1493     // and the line above the bottom macroblock
1494     AV_COPY128(top_border, src_y + 16*linesize);
1495     if (pixel_shift)
1496         AV_COPY128(top_border+16, src_y+16*linesize+16);
1497
1498     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1499         if(chroma444){
1500             if (pixel_shift){
1501                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1502                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1503                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1504                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1505             } else {
1506                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1507                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1508             }
1509         } else {
1510             if (pixel_shift) {
1511                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1512                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1513             } else {
1514                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1515                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1516             }
1517         }
1518     }
1519 }
1520
1521 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1522                                   uint8_t *src_cb, uint8_t *src_cr,
1523                                   int linesize, int uvlinesize,
1524                                   int xchg, int chroma444,
1525                                   int simple, int pixel_shift){
1526     MpegEncContext * const s = &h->s;
1527     int deblock_topleft;
1528     int deblock_top;
1529     int top_idx = 1;
1530     uint8_t *top_border_m1;
1531     uint8_t *top_border;
1532
1533     if(!simple && FRAME_MBAFF){
1534         if(s->mb_y&1){
1535             if(!MB_MBAFF)
1536                 return;
1537         }else{
1538             top_idx = MB_MBAFF ? 0 : 1;
1539         }
1540     }
1541
1542     if(h->deblocking_filter == 2) {
1543         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1544         deblock_top     = h->top_type;
1545     } else {
1546         deblock_topleft = (s->mb_x > 0);
1547         deblock_top     = (s->mb_y > !!MB_FIELD);
1548     }
1549
1550     src_y  -=   linesize + 1 + pixel_shift;
1551     src_cb -= uvlinesize + 1 + pixel_shift;
1552     src_cr -= uvlinesize + 1 + pixel_shift;
1553
1554     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1555     top_border    = h->top_borders[top_idx][s->mb_x];
1556
1557 #define XCHG(a,b,xchg)\
1558     if (pixel_shift) {\
1559         if (xchg) {\
1560             AV_SWAP64(b+0,a+0);\
1561             AV_SWAP64(b+8,a+8);\
1562         } else {\
1563             AV_COPY128(b,a); \
1564         }\
1565     } else \
1566 if (xchg) AV_SWAP64(b,a);\
1567 else      AV_COPY64(b,a);
1568
1569     if(deblock_top){
1570         if(deblock_topleft){
1571             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1572         }
1573         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1574         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1575         if(s->mb_x+1 < s->mb_width){
1576             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1577         }
1578     }
1579     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1580         if(chroma444){
1581             if(deblock_topleft){
1582                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1583                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1584             }
1585             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1586             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1587             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1588             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1589             if(s->mb_x+1 < s->mb_width){
1590                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1591                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1592             }
1593         } else {
1594             if(deblock_top){
1595                 if(deblock_topleft){
1596                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1597                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1598                 }
1599                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1600                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1601             }
1602         }
1603     }
1604 }
1605
1606 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1607     if (high_bit_depth) {
1608         return AV_RN32A(((int32_t*)mb) + index);
1609     } else
1610         return AV_RN16A(mb + index);
1611 }
1612
1613 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1614     if (high_bit_depth) {
1615         AV_WN32A(((int32_t*)mb) + index, value);
1616     } else
1617         AV_WN16A(mb + index, value);
1618 }
1619
1620 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1621                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1622 {
1623     MpegEncContext * const s = &h->s;
1624     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1625     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1626     int i;
1627     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1628     block_offset += 16*p;
1629     if(IS_INTRA4x4(mb_type)){
1630         if(simple || !s->encoding){
1631             if(IS_8x8DCT(mb_type)){
1632                 if(transform_bypass){
1633                     idct_dc_add =
1634                     idct_add    = s->dsp.add_pixels8;
1635                 }else{
1636                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1637                     idct_add    = h->h264dsp.h264_idct8_add;
1638                 }
1639                 for(i=0; i<16; i+=4){
1640                     uint8_t * const ptr= dest_y + block_offset[i];
1641                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1642                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1643                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1644                     }else{
1645                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1646                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1647                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1648                         if(nnz){
1649                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1650                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1651                             else
1652                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1653                         }
1654                     }
1655                 }
1656             }else{
1657                 if(transform_bypass){
1658                     idct_dc_add =
1659                     idct_add    = s->dsp.add_pixels4;
1660                 }else{
1661                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1662                     idct_add    = h->h264dsp.h264_idct_add;
1663                 }
1664                 for(i=0; i<16; i++){
1665                     uint8_t * const ptr= dest_y + block_offset[i];
1666                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1667
1668                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1669                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1670                     }else{
1671                         uint8_t *topright;
1672                         int nnz, tr;
1673                         uint64_t tr_high;
1674                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1675                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1676                             assert(s->mb_y || linesize <= block_offset[i]);
1677                             if(!topright_avail){
1678                                 if (pixel_shift) {
1679                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1680                                     topright= (uint8_t*) &tr_high;
1681                                 } else {
1682                                     tr= ptr[3 - linesize]*0x01010101;
1683                                     topright= (uint8_t*) &tr;
1684                                 }
1685                             }else
1686                                 topright= ptr + (4 << pixel_shift) - linesize;
1687                         }else
1688                             topright= NULL;
1689
1690                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1691                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1692                         if(nnz){
1693                             if(is_h264){
1694                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1695                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1696                                 else
1697                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1698                             }else
1699                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1700                         }
1701                     }
1702                 }
1703             }
1704         }
1705     }else{
1706         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1707         if(is_h264){
1708             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1709                 if(!transform_bypass)
1710                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1711                 else{
1712                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1713                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1714                     for(i = 0; i < 16; i++)
1715                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1716                 }
1717             }
1718         }else
1719             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1720     }
1721 }
1722
1723 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1724                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1725 {
1726     MpegEncContext * const s = &h->s;
1727     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1728     int i;
1729     block_offset += 16*p;
1730     if(!IS_INTRA4x4(mb_type)){
1731         if(is_h264){
1732             if(IS_INTRA16x16(mb_type)){
1733                 if(transform_bypass){
1734                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1735                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1736                     }else{
1737                         for(i=0; i<16; i++){
1738                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1739                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1740                         }
1741                     }
1742                 }else{
1743                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1744                 }
1745             }else if(h->cbp&15){
1746                 if(transform_bypass){
1747                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1748                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1749                     for(i=0; i<16; i+=di){
1750                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1751                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1752                         }
1753                     }
1754                 }else{
1755                     if(IS_8x8DCT(mb_type)){
1756                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1757                     }else{
1758                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1759                     }
1760                 }
1761             }
1762         }else{
1763             for(i=0; i<16; i++){
1764                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1765                     uint8_t * const ptr= dest_y + block_offset[i];
1766                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1767                 }
1768             }
1769         }
1770     }
1771 }
1772
1773 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1774     MpegEncContext * const s = &h->s;
1775     const int mb_x= s->mb_x;
1776     const int mb_y= s->mb_y;
1777     const int mb_xy= h->mb_xy;
1778     const int mb_type= s->current_picture.mb_type[mb_xy];
1779     uint8_t  *dest_y, *dest_cb, *dest_cr;
1780     int linesize, uvlinesize /*dct_offset*/;
1781     int i, j;
1782     int *block_offset = &h->block_offset[0];
1783     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1784     /* is_h264 should always be true if SVQ3 is disabled. */
1785     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1786     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1787
1788     dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1789     dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1790     dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1791
1792     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1793     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1794
1795     h->list_counts[mb_xy]= h->list_count;
1796
1797     if (!simple && MB_FIELD) {
1798         linesize   = h->mb_linesize   = s->linesize * 2;
1799         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1800         block_offset = &h->block_offset[48];
1801         if(mb_y&1){ //FIXME move out of this function?
1802             dest_y -= s->linesize*15;
1803             dest_cb-= s->uvlinesize*7;
1804             dest_cr-= s->uvlinesize*7;
1805         }
1806         if(FRAME_MBAFF) {
1807             int list;
1808             for(list=0; list<h->list_count; list++){
1809                 if(!USES_LIST(mb_type, list))
1810                     continue;
1811                 if(IS_16X16(mb_type)){
1812                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1813                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1814                 }else{
1815                     for(i=0; i<16; i+=4){
1816                         int ref = h->ref_cache[list][scan8[i]];
1817                         if(ref >= 0)
1818                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1819                     }
1820                 }
1821             }
1822         }
1823     } else {
1824         linesize   = h->mb_linesize   = s->linesize;
1825         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1826 //        dct_offset = s->linesize * 16;
1827     }
1828
1829     if (!simple && IS_INTRA_PCM(mb_type)) {
1830         if (pixel_shift) {
1831             const int bit_depth = h->sps.bit_depth_luma;
1832             int j;
1833             GetBitContext gb;
1834             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1835
1836             for (i = 0; i < 16; i++) {
1837                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1838                 for (j = 0; j < 16; j++)
1839                     tmp_y[j] = get_bits(&gb, bit_depth);
1840             }
1841             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1842                 for (i = 0; i < 8; i++) {
1843                     uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1844                     for (j = 0; j < 8; j++)
1845                         tmp_cb[j] = get_bits(&gb, bit_depth);
1846                 }
1847                 for (i = 0; i < 8; i++) {
1848                     uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1849                     for (j = 0; j < 8; j++)
1850                         tmp_cr[j] = get_bits(&gb, bit_depth);
1851                 }
1852             }
1853         } else {
1854             for (i=0; i<16; i++) {
1855                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1856             }
1857             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1858                 for (i=0; i<8; i++) {
1859                     memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1860                     memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1861                 }
1862             }
1863         }
1864     } else {
1865         if(IS_INTRA(mb_type)){
1866             if(h->deblocking_filter)
1867                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1868
1869             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1870                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1871                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1872             }
1873
1874             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1875
1876             if(h->deblocking_filter)
1877                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1878         }else if(is_h264){
1879             if (pixel_shift) {
1880                 hl_motion_16(h, dest_y, dest_cb, dest_cr,
1881                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1882                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1883                              h->h264dsp.weight_h264_pixels_tab,
1884                              h->h264dsp.biweight_h264_pixels_tab, 0);
1885             } else
1886                 hl_motion_8(h, dest_y, dest_cb, dest_cr,
1887                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1888                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1889                             h->h264dsp.weight_h264_pixels_tab,
1890                             h->h264dsp.biweight_h264_pixels_tab, 0);
1891         }
1892
1893         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1894
1895         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1896             uint8_t *dest[2] = {dest_cb, dest_cr};
1897             if(transform_bypass){
1898                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1899                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1900                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1901                 }else{
1902                     idct_add = s->dsp.add_pixels4;
1903                     for(j=1; j<3; j++){
1904                         for(i=j*16; i<j*16+4; i++){
1905                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1906                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1907                         }
1908                     }
1909                 }
1910             }else{
1911                 if(is_h264){
1912                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1913                         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]);
1914                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1915                         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]);
1916                     h->h264dsp.h264_idct_add8(dest, block_offset,
1917                                               h->mb, uvlinesize,
1918                                               h->non_zero_count_cache);
1919                 }
1920 #if CONFIG_SVQ3_DECODER
1921                 else{
1922                     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]);
1923                     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]);
1924                     for(j=1; j<3; j++){
1925                         for(i=j*16; i<j*16+4; i++){
1926                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1927                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1928                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1929                             }
1930                         }
1931                     }
1932                 }
1933 #endif
1934             }
1935         }
1936     }
1937     if(h->cbp || IS_INTRA(mb_type))
1938     {
1939         s->dsp.clear_blocks(h->mb);
1940         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1941     }
1942 }
1943
1944 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1945     MpegEncContext * const s = &h->s;
1946     const int mb_x= s->mb_x;
1947     const int mb_y= s->mb_y;
1948     const int mb_xy= h->mb_xy;
1949     const int mb_type= s->current_picture.mb_type[mb_xy];
1950     uint8_t  *dest[3];
1951     int linesize;
1952     int i, j, p;
1953     int *block_offset = &h->block_offset[0];
1954     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1955     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1956
1957     for (p = 0; p < plane_count; p++)
1958     {
1959         dest[p] = s->current_picture.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1960         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1961     }
1962
1963     h->list_counts[mb_xy]= h->list_count;
1964
1965     if (!simple && MB_FIELD) {
1966         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1967         block_offset = &h->block_offset[48];
1968         if(mb_y&1) //FIXME move out of this function?
1969             for (p = 0; p < 3; p++)
1970                 dest[p] -= s->linesize*15;
1971         if(FRAME_MBAFF) {
1972             int list;
1973             for(list=0; list<h->list_count; list++){
1974                 if(!USES_LIST(mb_type, list))
1975                     continue;
1976                 if(IS_16X16(mb_type)){
1977                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1978                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1979                 }else{
1980                     for(i=0; i<16; i+=4){
1981                         int ref = h->ref_cache[list][scan8[i]];
1982                         if(ref >= 0)
1983                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1984                     }
1985                 }
1986             }
1987         }
1988     } else {
1989         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1990     }
1991
1992     if (!simple && IS_INTRA_PCM(mb_type)) {
1993         if (pixel_shift) {
1994             const int bit_depth = h->sps.bit_depth_luma;
1995             GetBitContext gb;
1996             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1997
1998             for (p = 0; p < plane_count; p++) {
1999                 for (i = 0; i < 16; i++) {
2000                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2001                     for (j = 0; j < 16; j++)
2002                         tmp[j] = get_bits(&gb, bit_depth);
2003                 }
2004             }
2005         } else {
2006             for (p = 0; p < plane_count; p++) {
2007                 for (i = 0; i < 16; i++) {
2008                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2009                 }
2010             }
2011         }
2012     } else {
2013         if(IS_INTRA(mb_type)){
2014             if(h->deblocking_filter)
2015                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2016
2017             for (p = 0; p < plane_count; p++)
2018                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2019
2020             if(h->deblocking_filter)
2021                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2022         }else{
2023             if (pixel_shift) {
2024                 hl_motion_16(h, dest[0], dest[1], dest[2],
2025                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2026                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2027                              h->h264dsp.weight_h264_pixels_tab,
2028                              h->h264dsp.biweight_h264_pixels_tab, 1);
2029             } else
2030                 hl_motion_8(h, dest[0], dest[1], dest[2],
2031                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2032                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2033                             h->h264dsp.weight_h264_pixels_tab,
2034                             h->h264dsp.biweight_h264_pixels_tab, 1);
2035         }
2036
2037         for (p = 0; p < plane_count; p++)
2038             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2039     }
2040     if(h->cbp || IS_INTRA(mb_type))
2041     {
2042         s->dsp.clear_blocks(h->mb);
2043         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2044     }
2045 }
2046
2047 /**
2048  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2049  */
2050 #define hl_decode_mb_simple(sh, bits) \
2051 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2052     hl_decode_mb_internal(h, 1, sh); \
2053 }
2054 hl_decode_mb_simple(0, 8);
2055 hl_decode_mb_simple(1, 16);
2056
2057 /**
2058  * Process a macroblock; this handles edge cases, such as interlacing.
2059  */
2060 static void av_noinline hl_decode_mb_complex(H264Context *h){
2061     hl_decode_mb_internal(h, 0, h->pixel_shift);
2062 }
2063
2064 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2065     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2066 }
2067
2068 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2069     hl_decode_mb_444_internal(h, 1, 0);
2070 }
2071
2072 void ff_h264_hl_decode_mb(H264Context *h){
2073     MpegEncContext * const s = &h->s;
2074     const int mb_xy= h->mb_xy;
2075     const int mb_type= s->current_picture.mb_type[mb_xy];
2076     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2077
2078     if (CHROMA444) {
2079         if(is_complex || h->pixel_shift)
2080             hl_decode_mb_444_complex(h);
2081         else
2082             hl_decode_mb_444_simple(h);
2083     } else if (is_complex) {
2084         hl_decode_mb_complex(h);
2085     } else if (h->pixel_shift) {
2086         hl_decode_mb_simple_16(h);
2087     } else
2088         hl_decode_mb_simple_8(h);
2089 }
2090
2091 static int pred_weight_table(H264Context *h){
2092     MpegEncContext * const s = &h->s;
2093     int list, i;
2094     int luma_def, chroma_def;
2095
2096     h->use_weight= 0;
2097     h->use_weight_chroma= 0;
2098     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2099     if(h->sps.chroma_format_idc)
2100         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2101     luma_def = 1<<h->luma_log2_weight_denom;
2102     chroma_def = 1<<h->chroma_log2_weight_denom;
2103
2104     for(list=0; list<2; list++){
2105         h->luma_weight_flag[list]   = 0;
2106         h->chroma_weight_flag[list] = 0;
2107         for(i=0; i<h->ref_count[list]; i++){
2108             int luma_weight_flag, chroma_weight_flag;
2109
2110             luma_weight_flag= get_bits1(&s->gb);
2111             if(luma_weight_flag){
2112                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2113                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2114                 if(   h->luma_weight[i][list][0] != luma_def
2115                    || h->luma_weight[i][list][1] != 0) {
2116                     h->use_weight= 1;
2117                     h->luma_weight_flag[list]= 1;
2118                 }
2119             }else{
2120                 h->luma_weight[i][list][0]= luma_def;
2121                 h->luma_weight[i][list][1]= 0;
2122             }
2123
2124             if(h->sps.chroma_format_idc){
2125                 chroma_weight_flag= get_bits1(&s->gb);
2126                 if(chroma_weight_flag){
2127                     int j;
2128                     for(j=0; j<2; j++){
2129                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2130                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2131                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2132                            || h->chroma_weight[i][list][j][1] != 0) {
2133                             h->use_weight_chroma= 1;
2134                             h->chroma_weight_flag[list]= 1;
2135                         }
2136                     }
2137                 }else{
2138                     int j;
2139                     for(j=0; j<2; j++){
2140                         h->chroma_weight[i][list][j][0]= chroma_def;
2141                         h->chroma_weight[i][list][j][1]= 0;
2142                     }
2143                 }
2144             }
2145         }
2146         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2147     }
2148     h->use_weight= h->use_weight || h->use_weight_chroma;
2149     return 0;
2150 }
2151
2152 /**
2153  * Initialize implicit_weight table.
2154  * @param field  0/1 initialize the weight for interlaced MBAFF
2155  *                -1 initializes the rest
2156  */
2157 static void implicit_weight_table(H264Context *h, int field){
2158     MpegEncContext * const s = &h->s;
2159     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2160
2161     for (i = 0; i < 2; i++) {
2162         h->luma_weight_flag[i]   = 0;
2163         h->chroma_weight_flag[i] = 0;
2164     }
2165
2166     if(field < 0){
2167         cur_poc = s->current_picture_ptr->poc;
2168     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2169        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2170         h->use_weight= 0;
2171         h->use_weight_chroma= 0;
2172         return;
2173     }
2174         ref_start= 0;
2175         ref_count0= h->ref_count[0];
2176         ref_count1= h->ref_count[1];
2177     }else{
2178         cur_poc = s->current_picture_ptr->field_poc[field];
2179         ref_start= 16;
2180         ref_count0= 16+2*h->ref_count[0];
2181         ref_count1= 16+2*h->ref_count[1];
2182     }
2183
2184     h->use_weight= 2;
2185     h->use_weight_chroma= 2;
2186     h->luma_log2_weight_denom= 5;
2187     h->chroma_log2_weight_denom= 5;
2188
2189     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2190         int poc0 = h->ref_list[0][ref0].poc;
2191         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2192             int poc1 = h->ref_list[1][ref1].poc;
2193             int td = av_clip(poc1 - poc0, -128, 127);
2194             int w= 32;
2195             if(td){
2196                 int tb = av_clip(cur_poc - poc0, -128, 127);
2197                 int tx = (16384 + (FFABS(td) >> 1)) / td;
2198                 int dist_scale_factor = (tb*tx + 32) >> 8;
2199                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2200                     w = 64 - dist_scale_factor;
2201             }
2202             if(field<0){
2203                 h->implicit_weight[ref0][ref1][0]=
2204                 h->implicit_weight[ref0][ref1][1]= w;
2205             }else{
2206                 h->implicit_weight[ref0][ref1][field]=w;
2207             }
2208         }
2209     }
2210 }
2211
2212 /**
2213  * instantaneous decoder refresh.
2214  */
2215 static void idr(H264Context *h){
2216     ff_h264_remove_all_refs(h);
2217     h->prev_frame_num= 0;
2218     h->prev_frame_num_offset= 0;
2219     h->prev_poc_msb=
2220     h->prev_poc_lsb= 0;
2221 }
2222
2223 /* forget old pics after a seek */
2224 static void flush_dpb(AVCodecContext *avctx){
2225     H264Context *h= avctx->priv_data;
2226     int i;
2227     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2228         if(h->delayed_pic[i])
2229             h->delayed_pic[i]->reference= 0;
2230         h->delayed_pic[i]= NULL;
2231     }
2232     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2233     h->prev_interlaced_frame = 1;
2234     idr(h);
2235     if(h->s.current_picture_ptr)
2236         h->s.current_picture_ptr->reference= 0;
2237     h->s.first_field= 0;
2238     ff_h264_reset_sei(h);
2239     ff_mpeg_flush(avctx);
2240 }
2241
2242 static int init_poc(H264Context *h){
2243     MpegEncContext * const s = &h->s;
2244     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2245     int field_poc[2];
2246     Picture *cur = s->current_picture_ptr;
2247
2248     h->frame_num_offset= h->prev_frame_num_offset;
2249     if(h->frame_num < h->prev_frame_num)
2250         h->frame_num_offset += max_frame_num;
2251
2252     if(h->sps.poc_type==0){
2253         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2254
2255         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2256             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2257         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2258             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2259         else
2260             h->poc_msb = h->prev_poc_msb;
2261 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2262         field_poc[0] =
2263         field_poc[1] = h->poc_msb + h->poc_lsb;
2264         if(s->picture_structure == PICT_FRAME)
2265             field_poc[1] += h->delta_poc_bottom;
2266     }else if(h->sps.poc_type==1){
2267         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2268         int i;
2269
2270         if(h->sps.poc_cycle_length != 0)
2271             abs_frame_num = h->frame_num_offset + h->frame_num;
2272         else
2273             abs_frame_num = 0;
2274
2275         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2276             abs_frame_num--;
2277
2278         expected_delta_per_poc_cycle = 0;
2279         for(i=0; i < h->sps.poc_cycle_length; i++)
2280             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2281
2282         if(abs_frame_num > 0){
2283             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2284             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2285
2286             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2287             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2288                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2289         } else
2290             expectedpoc = 0;
2291
2292         if(h->nal_ref_idc == 0)
2293             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2294
2295         field_poc[0] = expectedpoc + h->delta_poc[0];
2296         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2297
2298         if(s->picture_structure == PICT_FRAME)
2299             field_poc[1] += h->delta_poc[1];
2300     }else{
2301         int poc= 2*(h->frame_num_offset + h->frame_num);
2302
2303         if(!h->nal_ref_idc)
2304             poc--;
2305
2306         field_poc[0]= poc;
2307         field_poc[1]= poc;
2308     }
2309
2310     if(s->picture_structure != PICT_BOTTOM_FIELD)
2311         s->current_picture_ptr->field_poc[0]= field_poc[0];
2312     if(s->picture_structure != PICT_TOP_FIELD)
2313         s->current_picture_ptr->field_poc[1]= field_poc[1];
2314     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2315
2316     return 0;
2317 }
2318
2319
2320 /**
2321  * initialize scan tables
2322  */
2323 static void init_scan_tables(H264Context *h){
2324     int i;
2325     for(i=0; i<16; i++){
2326 #define T(x) (x>>2) | ((x<<2) & 0xF)
2327         h->zigzag_scan[i] = T(zigzag_scan[i]);
2328         h-> field_scan[i] = T( field_scan[i]);
2329 #undef T
2330     }
2331     for(i=0; i<64; i++){
2332 #define T(x) (x>>3) | ((x&7)<<3)
2333         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2334         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2335         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2336         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2337 #undef T
2338     }
2339     if(h->sps.transform_bypass){ //FIXME same ugly
2340         h->zigzag_scan_q0          = zigzag_scan;
2341         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2342         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2343         h->field_scan_q0           = field_scan;
2344         h->field_scan8x8_q0        = field_scan8x8;
2345         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2346     }else{
2347         h->zigzag_scan_q0          = h->zigzag_scan;
2348         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2349         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2350         h->field_scan_q0           = h->field_scan;
2351         h->field_scan8x8_q0        = h->field_scan8x8;
2352         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2353     }
2354 }
2355
2356 static void field_end(H264Context *h, int in_setup){
2357     MpegEncContext * const s = &h->s;
2358     AVCodecContext * const avctx= s->avctx;
2359     s->mb_y= 0;
2360
2361     if (!in_setup && !s->dropable)
2362         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2363                                  s->picture_structure==PICT_BOTTOM_FIELD);
2364
2365     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2366         ff_vdpau_h264_set_reference_frames(s);
2367
2368     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2369         if(!s->dropable) {
2370             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2371             h->prev_poc_msb= h->poc_msb;
2372             h->prev_poc_lsb= h->poc_lsb;
2373         }
2374         h->prev_frame_num_offset= h->frame_num_offset;
2375         h->prev_frame_num= h->frame_num;
2376         h->outputed_poc = h->next_outputed_poc;
2377     }
2378
2379     if (avctx->hwaccel) {
2380         if (avctx->hwaccel->end_frame(avctx) < 0)
2381             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2382     }
2383
2384     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2385         ff_vdpau_h264_picture_complete(s);
2386
2387     /*
2388      * FIXME: Error handling code does not seem to support interlaced
2389      * when slices span multiple rows
2390      * The ff_er_add_slice calls don't work right for bottom
2391      * fields; they cause massive erroneous error concealing
2392      * Error marking covers both fields (top and bottom).
2393      * This causes a mismatched s->error_count
2394      * and a bad error table. Further, the error count goes to
2395      * INT_MAX when called for bottom field, because mb_y is
2396      * past end by one (callers fault) and resync_mb_y != 0
2397      * causes problems for the first MB line, too.
2398      */
2399     if (!FIELD_PICTURE)
2400         ff_er_frame_end(s);
2401
2402     MPV_frame_end(s);
2403
2404     h->current_slice=0;
2405 }
2406
2407 /**
2408  * Replicate H264 "master" context to thread contexts.
2409  */
2410 static void clone_slice(H264Context *dst, H264Context *src)
2411 {
2412     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2413     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2414     dst->s.current_picture      = src->s.current_picture;
2415     dst->s.linesize             = src->s.linesize;
2416     dst->s.uvlinesize           = src->s.uvlinesize;
2417     dst->s.first_field          = src->s.first_field;
2418
2419     dst->prev_poc_msb           = src->prev_poc_msb;
2420     dst->prev_poc_lsb           = src->prev_poc_lsb;
2421     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2422     dst->prev_frame_num         = src->prev_frame_num;
2423     dst->short_ref_count        = src->short_ref_count;
2424
2425     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2426     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2427     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2428     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2429
2430     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2431     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2432 }
2433
2434 /**
2435  * computes profile from profile_idc and constraint_set?_flags
2436  *
2437  * @param sps SPS
2438  *
2439  * @return profile as defined by FF_PROFILE_H264_*
2440  */
2441 int ff_h264_get_profile(SPS *sps)
2442 {
2443     int profile = sps->profile_idc;
2444
2445     switch(sps->profile_idc) {
2446     case FF_PROFILE_H264_BASELINE:
2447         // constraint_set1_flag set to 1
2448         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2449         break;
2450     case FF_PROFILE_H264_HIGH_10:
2451     case FF_PROFILE_H264_HIGH_422:
2452     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2453         // constraint_set3_flag set to 1
2454         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2455         break;
2456     }
2457
2458     return profile;
2459 }
2460
2461 /**
2462  * decodes a slice header.
2463  * This will also call MPV_common_init() and frame_start() as needed.
2464  *
2465  * @param h h264context
2466  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2467  *
2468  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2469  */
2470 static int decode_slice_header(H264Context *h, H264Context *h0){
2471     MpegEncContext * const s = &h->s;
2472     MpegEncContext * const s0 = &h0->s;
2473     unsigned int first_mb_in_slice;
2474     unsigned int pps_id;
2475     int num_ref_idx_active_override_flag;
2476     unsigned int slice_type, tmp, i, j;
2477     int default_ref_list_done = 0;
2478     int last_pic_structure;
2479
2480     s->dropable= h->nal_ref_idc == 0;
2481
2482     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2483     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2484         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2485         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2486     }else{
2487         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2488         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2489     }
2490
2491     first_mb_in_slice= get_ue_golomb(&s->gb);
2492
2493     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2494         if(h0->current_slice && FIELD_PICTURE){
2495             field_end(h, 1);
2496         }
2497
2498         h0->current_slice = 0;
2499         if (!s0->first_field)
2500             s->current_picture_ptr= NULL;
2501     }
2502
2503     slice_type= get_ue_golomb_31(&s->gb);
2504     if(slice_type > 9){
2505         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);
2506         return -1;
2507     }
2508     if(slice_type > 4){
2509         slice_type -= 5;
2510         h->slice_type_fixed=1;
2511     }else
2512         h->slice_type_fixed=0;
2513
2514     slice_type= golomb_to_pict_type[ slice_type ];
2515     if (slice_type == AV_PICTURE_TYPE_I
2516         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2517         default_ref_list_done = 1;
2518     }
2519     h->slice_type= slice_type;
2520     h->slice_type_nos= slice_type & 3;
2521
2522     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2523
2524     pps_id= get_ue_golomb(&s->gb);
2525     if(pps_id>=MAX_PPS_COUNT){
2526         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2527         return -1;
2528     }
2529     if(!h0->pps_buffers[pps_id]) {
2530         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2531         return -1;
2532     }
2533     h->pps= *h0->pps_buffers[pps_id];
2534
2535     if(!h0->sps_buffers[h->pps.sps_id]) {
2536         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2537         return -1;
2538     }
2539     h->sps = *h0->sps_buffers[h->pps.sps_id];
2540
2541     s->avctx->profile = ff_h264_get_profile(&h->sps);
2542     s->avctx->level   = h->sps.level_idc;
2543     s->avctx->refs    = h->sps.ref_frame_count;
2544
2545     if(h == h0 && h->dequant_coeff_pps != pps_id){
2546         h->dequant_coeff_pps = pps_id;
2547         init_dequant_tables(h);
2548     }
2549
2550     s->mb_width= h->sps.mb_width;
2551     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2552
2553     h->b_stride=  s->mb_width*4;
2554
2555     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2556     if(h->sps.frame_mbs_only_flag)
2557         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2558     else
2559         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2560
2561     if (s->context_initialized
2562         && (   s->width != s->avctx->width || s->height != s->avctx->height
2563             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2564         if(h != h0) {
2565             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2566             return -1;   // width / height changed during parallelized decoding
2567         }
2568         free_tables(h, 0);
2569         flush_dpb(s->avctx);
2570         MPV_common_end(s);
2571     }
2572     if (!s->context_initialized) {
2573         if (h != h0) {
2574             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2575             return -1;
2576         }
2577
2578         avcodec_set_dimensions(s->avctx, s->width, s->height);
2579         s->avctx->sample_aspect_ratio= h->sps.sar;
2580         av_assert0(s->avctx->sample_aspect_ratio.den);
2581
2582         h->s.avctx->coded_width = 16*s->mb_width;
2583         h->s.avctx->coded_height = 16*s->mb_height;
2584
2585         if(h->sps.video_signal_type_present_flag){
2586             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2587             if(h->sps.colour_description_present_flag){
2588                 s->avctx->color_primaries = h->sps.color_primaries;
2589                 s->avctx->color_trc       = h->sps.color_trc;
2590                 s->avctx->colorspace      = h->sps.colorspace;
2591             }
2592         }
2593
2594         if(h->sps.timing_info_present_flag){
2595             int64_t den= h->sps.time_scale;
2596             if(h->x264_build < 44U)
2597                 den *= 2;
2598             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2599                       h->sps.num_units_in_tick, den, 1<<30);
2600         }
2601
2602         switch (h->sps.bit_depth_luma) {
2603             case 9 :
2604                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2605                 break;
2606             case 10 :
2607                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2608                 break;
2609             default:
2610                 if (CHROMA444){
2611                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2612                 }else{
2613                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2614                                                              s->avctx->codec->pix_fmts ?
2615                                                              s->avctx->codec->pix_fmts :
2616                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2617                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2618                                                              ff_hwaccel_pixfmt_list_420);
2619                 }
2620         }
2621
2622         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2623
2624         if (MPV_common_init(s) < 0) {
2625             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2626             return -1;
2627         }
2628         s->first_field = 0;
2629         h->prev_interlaced_frame = 1;
2630
2631         init_scan_tables(h);
2632         ff_h264_alloc_tables(h);
2633
2634         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2635             if (context_init(h) < 0) {
2636                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2637                 return -1;
2638             }
2639         } else {
2640             for(i = 1; i < s->avctx->thread_count; i++) {
2641                 H264Context *c;
2642                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2643                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2644                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2645                 c->h264dsp = h->h264dsp;
2646                 c->sps = h->sps;
2647                 c->pps = h->pps;
2648                 c->pixel_shift = h->pixel_shift;
2649                 init_scan_tables(c);
2650                 clone_tables(c, h, i);
2651             }
2652
2653             for(i = 0; i < s->avctx->thread_count; i++)
2654                 if (context_init(h->thread_context[i]) < 0) {
2655                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2656                     return -1;
2657                 }
2658         }
2659     }
2660
2661     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2662
2663     h->mb_mbaff = 0;
2664     h->mb_aff_frame = 0;
2665     last_pic_structure = s0->picture_structure;
2666     if(h->sps.frame_mbs_only_flag){
2667         s->picture_structure= PICT_FRAME;
2668     }else{
2669         if(get_bits1(&s->gb)) { //field_pic_flag
2670             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2671         } else {
2672             s->picture_structure= PICT_FRAME;
2673             h->mb_aff_frame = h->sps.mb_aff;
2674         }
2675     }
2676     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2677
2678     if(h0->current_slice == 0){
2679         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2680         if(h->frame_num != h->prev_frame_num) {
2681             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2682
2683             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2684
2685             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2686                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2687                 if (unwrap_prev_frame_num < 0)
2688                     unwrap_prev_frame_num += max_frame_num;
2689
2690                 h->prev_frame_num = unwrap_prev_frame_num;
2691             }
2692         }
2693
2694         while(h->frame_num !=  h->prev_frame_num &&
2695               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2696             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2697             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2698             if (ff_h264_frame_start(h) < 0)
2699                 return -1;
2700             h->prev_frame_num++;
2701             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2702             s->current_picture_ptr->frame_num= h->prev_frame_num;
2703             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2704             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2705             ff_generate_sliding_window_mmcos(h);
2706             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2707             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2708              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2709              * about there being no actual duplicates.
2710              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2711              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2712              * be fixed. */
2713             if (h->short_ref_count) {
2714                 if (prev) {
2715                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
2716                                   (const uint8_t**)prev->data, prev->linesize,
2717                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2718                     h->short_ref[0]->poc = prev->poc+2;
2719                 }
2720                 h->short_ref[0]->frame_num = h->prev_frame_num;
2721             }
2722         }
2723
2724         /* See if we have a decoded first field looking for a pair... */
2725         if (s0->first_field) {
2726             assert(s0->current_picture_ptr);
2727             assert(s0->current_picture_ptr->data[0]);
2728             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2729
2730             /* figure out if we have a complementary field pair */
2731             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2732                 /*
2733                  * Previous field is unmatched. Don't display it, but let it
2734                  * remain for reference if marked as such.
2735                  */
2736                 s0->current_picture_ptr = NULL;
2737                 s0->first_field = FIELD_PICTURE;
2738
2739             } else {
2740                 if (h->nal_ref_idc &&
2741                         s0->current_picture_ptr->reference &&
2742                         s0->current_picture_ptr->frame_num != h->frame_num) {
2743                     /*
2744                      * This and previous field were reference, but had
2745                      * different frame_nums. Consider this field first in
2746                      * pair. Throw away previous field except for reference
2747                      * purposes.
2748                      */
2749                     s0->first_field = 1;
2750                     s0->current_picture_ptr = NULL;
2751
2752                 } else {
2753                     /* Second field in complementary pair */
2754                     s0->first_field = 0;
2755                 }
2756             }
2757
2758         } else {
2759             /* Frame or first field in a potentially complementary pair */
2760             assert(!s0->current_picture_ptr);
2761             s0->first_field = FIELD_PICTURE;
2762         }
2763
2764         if(!FIELD_PICTURE || s0->first_field) {
2765             if (ff_h264_frame_start(h) < 0) {
2766                 s0->first_field = 0;
2767                 return -1;
2768             }
2769         } else {
2770             ff_release_unused_pictures(s, 0);
2771         }
2772     }
2773     if(h != h0)
2774         clone_slice(h, h0);
2775
2776     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2777
2778     assert(s->mb_num == s->mb_width * s->mb_height);
2779     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2780        first_mb_in_slice                    >= s->mb_num){
2781         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2782         return -1;
2783     }
2784     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2785     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2786     if (s->picture_structure == PICT_BOTTOM_FIELD)
2787         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2788     assert(s->mb_y < s->mb_height);
2789
2790     if(s->picture_structure==PICT_FRAME){
2791         h->curr_pic_num=   h->frame_num;
2792         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2793     }else{
2794         h->curr_pic_num= 2*h->frame_num + 1;
2795         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2796     }
2797
2798     if(h->nal_unit_type == NAL_IDR_SLICE){
2799         get_ue_golomb(&s->gb); /* idr_pic_id */
2800     }
2801
2802     if(h->sps.poc_type==0){
2803         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2804
2805         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2806             h->delta_poc_bottom= get_se_golomb(&s->gb);
2807         }
2808     }
2809
2810     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2811         h->delta_poc[0]= get_se_golomb(&s->gb);
2812
2813         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2814             h->delta_poc[1]= get_se_golomb(&s->gb);
2815     }
2816
2817     init_poc(h);
2818
2819     if(h->pps.redundant_pic_cnt_present){
2820         h->redundant_pic_count= get_ue_golomb(&s->gb);
2821     }
2822
2823     //set defaults, might be overridden a few lines later
2824     h->ref_count[0]= h->pps.ref_count[0];
2825     h->ref_count[1]= h->pps.ref_count[1];
2826
2827     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2828         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2829             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2830         }
2831         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2832
2833         if(num_ref_idx_active_override_flag){
2834             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2835             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2836                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2837
2838             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2839                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2840                 h->ref_count[0]= h->ref_count[1]= 1;
2841                 return -1;
2842             }
2843         }
2844         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2845             h->list_count= 2;
2846         else
2847             h->list_count= 1;
2848     }else
2849         h->list_count= 0;
2850
2851     if(!default_ref_list_done){
2852         ff_h264_fill_default_ref_list(h);
2853     }
2854
2855     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2856         return -1;
2857
2858     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2859         s->last_picture_ptr= &h->ref_list[0][0];
2860         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2861     }
2862     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2863         s->next_picture_ptr= &h->ref_list[1][0];
2864         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2865     }
2866
2867     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2868        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2869         pred_weight_table(h);
2870     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2871         implicit_weight_table(h, -1);
2872     }else {
2873         h->use_weight = 0;
2874         for (i = 0; i < 2; i++) {
2875             h->luma_weight_flag[i]   = 0;
2876             h->chroma_weight_flag[i] = 0;
2877         }
2878     }
2879
2880     if(h->nal_ref_idc)
2881         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2882
2883     if(FRAME_MBAFF){
2884         ff_h264_fill_mbaff_ref_list(h);
2885
2886         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2887             implicit_weight_table(h, 0);
2888             implicit_weight_table(h, 1);
2889         }
2890     }
2891
2892     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2893         ff_h264_direct_dist_scale_factor(h);
2894     ff_h264_direct_ref_list_init(h);
2895
2896     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2897         tmp = get_ue_golomb_31(&s->gb);
2898         if(tmp > 2){
2899             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2900             return -1;
2901         }
2902         h->cabac_init_idc= tmp;
2903     }
2904
2905     h->last_qscale_diff = 0;
2906     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2907     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2908         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2909         return -1;
2910     }
2911     s->qscale= tmp;
2912     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2913     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2914     //FIXME qscale / qp ... stuff
2915     if(h->slice_type == AV_PICTURE_TYPE_SP){
2916         get_bits1(&s->gb); /* sp_for_switch_flag */
2917     }
2918     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2919         get_se_golomb(&s->gb); /* slice_qs_delta */
2920     }
2921
2922     h->deblocking_filter = 1;
2923     h->slice_alpha_c0_offset = 52;
2924     h->slice_beta_offset = 52;
2925     if( h->pps.deblocking_filter_parameters_present ) {
2926         tmp= get_ue_golomb_31(&s->gb);
2927         if(tmp > 2){
2928             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2929             return -1;
2930         }
2931         h->deblocking_filter= tmp;
2932         if(h->deblocking_filter < 2)
2933             h->deblocking_filter^= 1; // 1<->0
2934
2935         if( h->deblocking_filter ) {
2936             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2937             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2938             if(   h->slice_alpha_c0_offset > 104U
2939                || h->slice_beta_offset     > 104U){
2940                 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);
2941                 return -1;
2942             }
2943         }
2944     }
2945
2946     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2947        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2948        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2949        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2950         h->deblocking_filter= 0;
2951
2952     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2953         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2954             /* Cheat slightly for speed:
2955                Do not bother to deblock across slices. */
2956             h->deblocking_filter = 2;
2957         } else {
2958             h0->max_contexts = 1;
2959             if(!h0->single_decode_warning) {
2960                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2961                 h0->single_decode_warning = 1;
2962             }
2963             if (h != h0) {
2964                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2965                 return 1;
2966             }
2967         }
2968     }
2969     h->qp_thresh= 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1]);
2970
2971 #if 0 //FMO
2972     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2973         slice_group_change_cycle= get_bits(&s->gb, ?);
2974 #endif
2975
2976     h0->last_slice_type = slice_type;
2977     h->slice_num = ++h0->current_slice;
2978     if(h->slice_num >= MAX_SLICES){
2979         av_log(s->avctx, AV_LOG_ERROR, "Too many slices (%d >= %d), increase MAX_SLICES and recompile\n", h->slice_num, MAX_SLICES);
2980     }
2981
2982     for(j=0; j<2; j++){
2983         int id_list[16];
2984         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2985         for(i=0; i<16; i++){
2986             id_list[i]= 60;
2987             if(h->ref_list[j][i].data[0]){
2988                 int k;
2989                 uint8_t *base= h->ref_list[j][i].base[0];
2990                 for(k=0; k<h->short_ref_count; k++)
2991                     if(h->short_ref[k]->base[0] == base){
2992                         id_list[i]= k;
2993                         break;
2994                     }
2995                 for(k=0; k<h->long_ref_count; k++)
2996                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2997                         id_list[i]= h->short_ref_count + k;
2998                         break;
2999                     }
3000             }
3001         }
3002
3003         ref2frm[0]=
3004         ref2frm[1]= -1;
3005         for(i=0; i<16; i++)
3006             ref2frm[i+2]= 4*id_list[i]
3007                           +(h->ref_list[j][i].reference&3);
3008         ref2frm[18+0]=
3009         ref2frm[18+1]= -1;
3010         for(i=16; i<48; i++)
3011             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3012                           +(h->ref_list[j][i].reference&3);
3013     }
3014
3015     //FIXME: fix draw_edges+PAFF+frame threads
3016     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3017     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3018
3019     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3020         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",
3021                h->slice_num,
3022                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3023                first_mb_in_slice,
3024                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3025                pps_id, h->frame_num,
3026                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3027                h->ref_count[0], h->ref_count[1],
3028                s->qscale,
3029                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3030                h->use_weight,
3031                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3032                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3033                );
3034     }
3035
3036     return 0;
3037 }
3038
3039 int ff_h264_get_slice_type(const H264Context *h)
3040 {
3041     switch (h->slice_type) {
3042     case AV_PICTURE_TYPE_P:  return 0;
3043     case AV_PICTURE_TYPE_B:  return 1;
3044     case AV_PICTURE_TYPE_I:  return 2;
3045     case AV_PICTURE_TYPE_SP: return 3;
3046     case AV_PICTURE_TYPE_SI: return 4;
3047     default:         return -1;
3048     }
3049 }
3050
3051 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3052                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3053 {
3054     int b_stride = h->b_stride;
3055     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3056     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3057     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3058         if(USES_LIST(top_type, list)){
3059             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3060             const int b8_xy= 4*top_xy + 2;
3061             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3062             AV_COPY128(mv_dst - 1*8, s->current_picture.motion_val[list][b_xy + 0]);
3063             ref_cache[0 - 1*8]=
3064             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
3065             ref_cache[2 - 1*8]=
3066             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
3067         }else{
3068             AV_ZERO128(mv_dst - 1*8);
3069             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3070         }
3071
3072         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3073             if(USES_LIST(left_type[LTOP], list)){
3074                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3075                 const int b8_xy= 4*left_xy[LTOP] + 1;
3076                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3077                 AV_COPY32(mv_dst - 1 + 0, s->current_picture.motion_val[list][b_xy + b_stride*0]);
3078                 AV_COPY32(mv_dst - 1 + 8, s->current_picture.motion_val[list][b_xy + b_stride*1]);
3079                 AV_COPY32(mv_dst - 1 +16, s->current_picture.motion_val[list][b_xy + b_stride*2]);
3080                 AV_COPY32(mv_dst - 1 +24, s->current_picture.motion_val[list][b_xy + b_stride*3]);
3081                 ref_cache[-1 +  0]=
3082                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
3083                 ref_cache[-1 + 16]=
3084                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
3085             }else{
3086                 AV_ZERO32(mv_dst - 1 + 0);
3087                 AV_ZERO32(mv_dst - 1 + 8);
3088                 AV_ZERO32(mv_dst - 1 +16);
3089                 AV_ZERO32(mv_dst - 1 +24);
3090                 ref_cache[-1 +  0]=
3091                 ref_cache[-1 +  8]=
3092                 ref_cache[-1 + 16]=
3093                 ref_cache[-1 + 24]= LIST_NOT_USED;
3094             }
3095         }
3096     }
3097
3098     if(!USES_LIST(mb_type, list)){
3099         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3100         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3101         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3102         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3103         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3104         return;
3105     }
3106
3107     {
3108         int8_t *ref = &s->current_picture.ref_index[list][4*mb_xy];
3109         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3110         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3111         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3112         AV_WN32A(&ref_cache[0*8], ref01);
3113         AV_WN32A(&ref_cache[1*8], ref01);
3114         AV_WN32A(&ref_cache[2*8], ref23);
3115         AV_WN32A(&ref_cache[3*8], ref23);
3116     }
3117
3118     {
3119         int16_t (*mv_src)[2] = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3120         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3121         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3122         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3123         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3124     }
3125 }
3126
3127 /**
3128  *
3129  * @return non zero if the loop filter can be skiped
3130  */
3131 static int fill_filter_caches(H264Context *h, int mb_type){
3132     MpegEncContext * const s = &h->s;
3133     const int mb_xy= h->mb_xy;
3134     int top_xy, left_xy[LEFT_MBS];
3135     int top_type, left_type[LEFT_MBS];
3136     uint8_t *nnz;
3137     uint8_t *nnz_cache;
3138
3139     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3140
3141     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3142      * stuff, I can't imagine that these complex rules are worth it. */
3143
3144     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3145     if(FRAME_MBAFF){
3146         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
3147         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3148         if(s->mb_y&1){
3149             if (left_mb_field_flag != curr_mb_field_flag) {
3150                 left_xy[LTOP] -= s->mb_stride;
3151             }
3152         }else{
3153             if(curr_mb_field_flag){
3154                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
3155             }
3156             if (left_mb_field_flag != curr_mb_field_flag) {
3157                 left_xy[LBOT] += s->mb_stride;
3158             }
3159         }
3160     }
3161
3162     h->top_mb_xy = top_xy;
3163     h->left_mb_xy[LTOP] = left_xy[LTOP];
3164     h->left_mb_xy[LBOT] = left_xy[LBOT];
3165     {
3166         //for sufficiently low qp, filtering wouldn't do anything
3167         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3168         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3169         int qp = s->current_picture.qscale_table[mb_xy];
3170         if(qp <= qp_thresh
3171            && (left_xy[LTOP]<0 || ((qp + s->current_picture.qscale_table[left_xy[LTOP]] + 1)>>1) <= qp_thresh)
3172            && (top_xy       <0 || ((qp + s->current_picture.qscale_table[top_xy       ] + 1)>>1) <= qp_thresh)){
3173             if(!FRAME_MBAFF)
3174                 return 1;
3175             if(   (left_xy[LTOP]< 0            || ((qp + s->current_picture.qscale_table[left_xy[LBOT]             ] + 1)>>1) <= qp_thresh)
3176                && (top_xy       < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy       -s->mb_stride] + 1)>>1) <= qp_thresh))
3177                 return 1;
3178         }
3179     }
3180
3181     top_type        = s->current_picture.mb_type[top_xy];
3182     left_type[LTOP] = s->current_picture.mb_type[left_xy[LTOP]];
3183     left_type[LBOT] = s->current_picture.mb_type[left_xy[LBOT]];
3184     if(h->deblocking_filter == 2){
3185         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3186         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3187     }else{
3188         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3189         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3190     }
3191     h->top_type       = top_type;
3192     h->left_type[LTOP]= left_type[LTOP];
3193     h->left_type[LBOT]= left_type[LBOT];
3194
3195     if(IS_INTRA(mb_type))
3196         return 0;
3197
3198     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3199     if(h->list_count == 2)
3200         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3201
3202     nnz = h->non_zero_count[mb_xy];
3203     nnz_cache = h->non_zero_count_cache;
3204     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3205     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3206     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3207     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3208     h->cbp= h->cbp_table[mb_xy];
3209
3210     if(top_type){
3211         nnz = h->non_zero_count[top_xy];
3212         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3213     }
3214
3215     if(left_type[LTOP]){
3216         nnz = h->non_zero_count[left_xy[LTOP]];
3217         nnz_cache[3+8*1]= nnz[3+0*4];
3218         nnz_cache[3+8*2]= nnz[3+1*4];
3219         nnz_cache[3+8*3]= nnz[3+2*4];
3220         nnz_cache[3+8*4]= nnz[3+3*4];
3221     }
3222
3223     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3224     if(!CABAC && h->pps.transform_8x8_mode){
3225         if(IS_8x8DCT(top_type)){
3226             nnz_cache[4+8*0]=
3227             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3228             nnz_cache[6+8*0]=
3229             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3230         }
3231         if(IS_8x8DCT(left_type[LTOP])){
3232             nnz_cache[3+8*1]=
3233             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3234         }
3235         if(IS_8x8DCT(left_type[LBOT])){
3236             nnz_cache[3+8*3]=
3237             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3238         }
3239
3240         if(IS_8x8DCT(mb_type)){
3241             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3242             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3243
3244             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3245             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3246
3247             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3248             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3249
3250             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3251             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3252         }
3253     }
3254
3255     return 0;
3256 }
3257
3258 static void loop_filter(H264Context *h, int start_x, int end_x){
3259     MpegEncContext * const s = &h->s;
3260     uint8_t  *dest_y, *dest_cb, *dest_cr;
3261     int linesize, uvlinesize, mb_x, mb_y;
3262     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3263     const int old_slice_type= h->slice_type;
3264     const int pixel_shift = h->pixel_shift;
3265
3266     if(h->deblocking_filter) {
3267         for(mb_x= start_x; mb_x<end_x; mb_x++){
3268             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3269                 int mb_xy, mb_type;
3270                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3271                 h->slice_num= h->slice_table[mb_xy];
3272                 mb_type= s->current_picture.mb_type[mb_xy];
3273                 h->list_count= h->list_counts[mb_xy];
3274
3275                 if(FRAME_MBAFF)
3276                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3277
3278                 s->mb_x= mb_x;
3279                 s->mb_y= mb_y;
3280                 dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3281                 dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3282                 dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3283                     //FIXME simplify above
3284
3285                 if (MB_FIELD) {
3286                     linesize   = h->mb_linesize   = s->linesize * 2;
3287                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3288                     if(mb_y&1){ //FIXME move out of this function?
3289                         dest_y -= s->linesize*15;
3290                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3291                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3292                     }
3293                 } else {
3294                     linesize   = h->mb_linesize   = s->linesize;
3295                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3296                 }
3297                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3298                 if(fill_filter_caches(h, mb_type))
3299                     continue;
3300                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
3301                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
3302
3303                 if (FRAME_MBAFF) {
3304                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3305                 } else {
3306                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3307                 }
3308             }
3309         }
3310     }
3311     h->slice_type= old_slice_type;
3312     s->mb_x= end_x;
3313     s->mb_y= end_mb_y - FRAME_MBAFF;
3314     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3315     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3316 }
3317
3318 static void predict_field_decoding_flag(H264Context *h){
3319     MpegEncContext * const s = &h->s;
3320     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3321     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3322                 ? s->current_picture.mb_type[mb_xy-1]
3323                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3324                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
3325                 : 0;
3326     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3327 }
3328
3329 /**
3330  * Draw edges and report progress for the last MB row.
3331  */
3332 static void decode_finish_row(H264Context *h){
3333     MpegEncContext * const s = &h->s;
3334     int top = 16*(s->mb_y >> FIELD_PICTURE);
3335     int height = 16 << FRAME_MBAFF;
3336     int deblock_border = (16 + 4) << FRAME_MBAFF;
3337     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3338
3339     if (h->deblocking_filter) {
3340         if((top + height) >= pic_height)
3341             height += deblock_border;
3342
3343         top -= deblock_border;
3344     }
3345
3346     if (top >= pic_height || (top + height) < h->emu_edge_height)
3347         return;
3348
3349     height = FFMIN(height, pic_height - top);
3350     if (top < h->emu_edge_height) {
3351         height = top+height;
3352         top = 0;
3353     }
3354
3355     ff_draw_horiz_band(s, top, height);
3356
3357     if (s->dropable) return;
3358
3359     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3360                              s->picture_structure==PICT_BOTTOM_FIELD);
3361 }
3362
3363 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3364     H264Context *h = *(void**)arg;
3365     MpegEncContext * const s = &h->s;
3366     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3367     int lf_x_start = s->mb_x;
3368
3369     s->mb_skip_run= -1;
3370
3371     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3372                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3373
3374     if( h->pps.cabac ) {
3375         /* realign */
3376         align_get_bits( &s->gb );
3377
3378         /* init cabac */
3379         ff_init_cabac_states( &h->cabac);
3380         ff_init_cabac_decoder( &h->cabac,
3381                                s->gb.buffer + get_bits_count(&s->gb)/8,
3382                                (get_bits_left(&s->gb) + 7)/8);
3383
3384         ff_h264_init_cabac_states(h);
3385
3386         for(;;){
3387 //START_TIMER
3388             int ret = ff_h264_decode_mb_cabac(h);
3389             int eos;
3390 //STOP_TIMER("decode_mb_cabac")
3391
3392             if(ret>=0) ff_h264_hl_decode_mb(h);
3393
3394             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3395                 s->mb_y++;
3396
3397                 ret = ff_h264_decode_mb_cabac(h);
3398
3399                 if(ret>=0) ff_h264_hl_decode_mb(h);
3400                 s->mb_y--;
3401             }
3402             eos = get_cabac_terminate( &h->cabac );
3403
3404             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3405                 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);
3406                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3407                 return 0;
3408             }
3409             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3410                 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);
3411                 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);
3412                 return -1;
3413             }
3414
3415             if( ++s->mb_x >= s->mb_width ) {
3416                 loop_filter(h, lf_x_start, s->mb_x);
3417                 s->mb_x = lf_x_start = 0;
3418                 decode_finish_row(h);
3419                 ++s->mb_y;
3420                 if(FIELD_OR_MBAFF_PICTURE) {
3421                     ++s->mb_y;
3422                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3423                         predict_field_decoding_flag(h);
3424                 }
3425             }
3426
3427             if( eos || s->mb_y >= s->mb_height ) {
3428                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3429                 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);
3430                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3431                 return 0;
3432             }
3433         }
3434
3435     } else {
3436         for(;;){
3437             int ret = ff_h264_decode_mb_cavlc(h);
3438
3439             if(ret>=0) ff_h264_hl_decode_mb(h);
3440
3441             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3442                 s->mb_y++;
3443                 ret = ff_h264_decode_mb_cavlc(h);
3444
3445                 if(ret>=0) ff_h264_hl_decode_mb(h);
3446                 s->mb_y--;
3447             }
3448
3449             if(ret<0){
3450                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3451                 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);
3452                 return -1;
3453             }
3454
3455             if(++s->mb_x >= s->mb_width){
3456                 loop_filter(h, lf_x_start, s->mb_x);
3457                 s->mb_x = lf_x_start = 0;
3458                 decode_finish_row(h);
3459                 ++s->mb_y;
3460                 if(FIELD_OR_MBAFF_PICTURE) {
3461                     ++s->mb_y;
3462                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3463                         predict_field_decoding_flag(h);
3464                 }
3465                 if(s->mb_y >= s->mb_height){
3466                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3467
3468                     if(   get_bits_count(&s->gb) == s->gb.size_in_bits
3469                        || get_bits_count(&s->gb) <  s->gb.size_in_bits && s->avctx->error_recognition < FF_ER_AGGRESSIVE) {
3470                         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);
3471
3472                         return 0;
3473                     }else{
3474                         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);
3475
3476                         return -1;
3477                     }
3478                 }
3479             }
3480
3481             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3482                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3483                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3484                     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);
3485                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3486
3487                     return 0;
3488                 }else{
3489                     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);
3490
3491                     return -1;
3492                 }
3493             }
3494         }
3495     }
3496
3497 #if 0
3498     for(;s->mb_y < s->mb_height; s->mb_y++){
3499         for(;s->mb_x < s->mb_width; s->mb_x++){
3500             int ret= decode_mb(h);
3501
3502             ff_h264_hl_decode_mb(h);
3503
3504             if(ret<0){
3505                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3506                 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);
3507
3508                 return -1;
3509             }
3510
3511             if(++s->mb_x >= s->mb_width){
3512                 s->mb_x=0;
3513                 if(++s->mb_y >= s->mb_height){
3514                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
3515                         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);
3516
3517                         return 0;
3518                     }else{
3519                         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);
3520
3521                         return -1;
3522                     }
3523                 }
3524             }
3525
3526             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
3527                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
3528                     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);
3529
3530                     return 0;
3531                 }else{
3532                     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);
3533
3534                     return -1;
3535                 }
3536             }
3537         }
3538         s->mb_x=0;
3539         ff_draw_horiz_band(s, 16*s->mb_y, 16);
3540     }
3541 #endif
3542 }
3543
3544 /**
3545  * Call decode_slice() for each context.
3546  *
3547  * @param h h264 master context
3548  * @param context_count number of contexts to execute
3549  */
3550 static void execute_decode_slices(H264Context *h, int context_count){
3551     MpegEncContext * const s = &h->s;
3552     AVCodecContext * const avctx= s->avctx;
3553     H264Context *hx;
3554     int i;
3555
3556     if (s->avctx->hwaccel)
3557         return;
3558     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3559         return;
3560     if(context_count == 1) {
3561         decode_slice(avctx, &h);
3562     } else {
3563         for(i = 1; i < context_count; i++) {
3564             hx = h->thread_context[i];
3565             hx->s.error_recognition = avctx->error_recognition;
3566             hx->s.error_count = 0;
3567             hx->x264_build= h->x264_build;
3568         }
3569
3570         avctx->execute(avctx, (void *)decode_slice,
3571                        h->thread_context, NULL, context_count, sizeof(void*));
3572
3573         /* pull back stuff from slices to master context */
3574         hx = h->thread_context[context_count - 1];
3575         s->mb_x = hx->s.mb_x;
3576         s->mb_y = hx->s.mb_y;
3577         s->dropable = hx->s.dropable;
3578         s->picture_structure = hx->s.picture_structure;
3579         for(i = 1; i < context_count; i++)
3580             h->s.error_count += h->thread_context[i]->s.error_count;
3581     }
3582 }
3583
3584
3585 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3586     MpegEncContext * const s = &h->s;
3587     AVCodecContext * const avctx= s->avctx;
3588     H264Context *hx; ///< thread context
3589     int buf_index;
3590     int context_count;
3591     int next_avc;
3592     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3593     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3594     int nal_index;
3595
3596     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3597     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3598         h->current_slice = 0;
3599         if (!s->first_field)
3600             s->current_picture_ptr= NULL;
3601         ff_h264_reset_sei(h);
3602     }
3603
3604     for(;pass <= 1;pass++){
3605         buf_index = 0;
3606         context_count = 0;
3607         next_avc = h->is_avc ? 0 : buf_size;
3608         nal_index = 0;
3609     for(;;){
3610         int consumed;
3611         int dst_length;
3612         int bit_length;
3613         const uint8_t *ptr;
3614         int i, nalsize = 0;
3615         int err;
3616
3617         if(buf_index >= next_avc) {
3618             if(buf_index >= buf_size) break;
3619             nalsize = 0;
3620             for(i = 0; i < h->nal_length_size; i++)
3621                 nalsize = (nalsize << 8) | buf[buf_index++];
3622             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3623                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3624                 break;
3625             }
3626             next_avc= buf_index + nalsize;
3627         } else {
3628             // start code prefix search
3629             for(; buf_index + 3 < next_avc; buf_index++){
3630                 // This should always succeed in the first iteration.
3631                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3632                     break;
3633             }
3634
3635             if(buf_index+3 >= buf_size) break;
3636
3637             buf_index+=3;
3638             if(buf_index >= next_avc) continue;
3639         }
3640
3641         hx = h->thread_context[context_count];
3642
3643         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3644         if (ptr==NULL || dst_length < 0){
3645             return -1;
3646         }
3647         i= buf_index + consumed;
3648         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3649            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3650             s->workaround_bugs |= FF_BUG_TRUNCATED;
3651
3652         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3653         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3654             dst_length--;
3655         }
3656         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3657
3658         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3659             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length);
3660         }
3661
3662         if (h->is_avc && (nalsize != consumed) && nalsize){
3663             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3664         }
3665
3666         buf_index += consumed;
3667         nal_index++;
3668
3669         if(pass == 0) {
3670             // packets can sometimes contain multiple PPS/SPS
3671             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3672             // if so, when frame threading we can't start the next thread until we've read all of them
3673             switch (hx->nal_unit_type) {
3674                 case NAL_SPS:
3675                 case NAL_PPS:
3676                 case NAL_IDR_SLICE:
3677                 case NAL_SLICE:
3678                     nals_needed = nal_index;
3679             }
3680             continue;
3681         }
3682
3683         //FIXME do not discard SEI id
3684         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3685             continue;
3686
3687       again:
3688         err = 0;
3689         switch(hx->nal_unit_type){
3690         case NAL_IDR_SLICE:
3691             if (h->nal_unit_type != NAL_IDR_SLICE) {
3692                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3693                 return -1;
3694             }
3695             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3696         case NAL_SLICE:
3697             init_get_bits(&hx->s.gb, ptr, bit_length);
3698             hx->intra_gb_ptr=
3699             hx->inter_gb_ptr= &hx->s.gb;
3700             hx->s.data_partitioning = 0;
3701
3702             if((err = decode_slice_header(hx, h)))
3703                break;
3704
3705             s->current_picture_ptr->key_frame |=
3706                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3707                     (h->sei_recovery_frame_cnt >= 0);
3708
3709             if (h->current_slice == 1) {
3710                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3711                     decode_postinit(h, nal_index >= nals_needed);
3712                 }
3713
3714                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3715                     return -1;
3716                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3717                     ff_vdpau_h264_picture_start(s);
3718             }
3719
3720             if(hx->redundant_pic_count==0
3721                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3722                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3723                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3724                && avctx->skip_frame < AVDISCARD_ALL){
3725                 if(avctx->hwaccel) {
3726                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3727                         return -1;
3728                 }else
3729                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3730                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3731                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3732                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3733                 }else
3734                     context_count++;
3735             }
3736             break;
3737         case NAL_DPA:
3738             init_get_bits(&hx->s.gb, ptr, bit_length);
3739             hx->intra_gb_ptr=
3740             hx->inter_gb_ptr= NULL;
3741
3742             if ((err = decode_slice_header(hx, h)) < 0)
3743                 break;
3744
3745             hx->s.data_partitioning = 1;
3746
3747             break;
3748         case NAL_DPB:
3749             init_get_bits(&hx->intra_gb, ptr, bit_length);
3750             hx->intra_gb_ptr= &hx->intra_gb;
3751             break;
3752         case NAL_DPC:
3753             init_get_bits(&hx->inter_gb, ptr, bit_length);
3754             hx->inter_gb_ptr= &hx->inter_gb;
3755
3756             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3757                && s->context_initialized
3758                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3759                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3760                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3761                && avctx->skip_frame < AVDISCARD_ALL)
3762                 context_count++;
3763             break;
3764         case NAL_SEI:
3765             init_get_bits(&s->gb, ptr, bit_length);
3766             ff_h264_decode_sei(h);
3767             break;
3768         case NAL_SPS:
3769             init_get_bits(&s->gb, ptr, bit_length);
3770             ff_h264_decode_seq_parameter_set(h);
3771
3772             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3773                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3774                 s->low_delay=1;
3775
3776             if(avctx->has_b_frames < 2)
3777                 avctx->has_b_frames= !s->low_delay;
3778
3779             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3780                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3781                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3782                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3783
3784                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3785                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3786                     dsputil_init(&s->dsp, s->avctx);
3787                 } else {
3788                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3789                     return -1;
3790                 }
3791             }
3792             break;
3793         case NAL_PPS:
3794             init_get_bits(&s->gb, ptr, bit_length);
3795
3796             ff_h264_decode_picture_parameter_set(h, bit_length);
3797
3798             break;
3799         case NAL_AUD:
3800         case NAL_END_SEQUENCE:
3801         case NAL_END_STREAM:
3802         case NAL_FILLER_DATA:
3803         case NAL_SPS_EXT:
3804         case NAL_AUXILIARY_SLICE:
3805             break;
3806         default:
3807             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3808         }
3809
3810         if(context_count == h->max_contexts) {
3811             execute_decode_slices(h, context_count);
3812             context_count = 0;
3813         }
3814
3815         if (err < 0)
3816             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3817         else if(err == 1) {
3818             /* Slice could not be decoded in parallel mode, copy down
3819              * NAL unit stuff to context 0 and restart. Note that
3820              * rbsp_buffer is not transferred, but since we no longer
3821              * run in parallel mode this should not be an issue. */
3822             h->nal_unit_type = hx->nal_unit_type;
3823             h->nal_ref_idc   = hx->nal_ref_idc;
3824             hx = h;
3825             goto again;
3826         }
3827     }
3828     }
3829     if(context_count)
3830         execute_decode_slices(h, context_count);
3831     return buf_index;
3832 }
3833
3834 /**
3835  * returns the number of bytes consumed for building the current frame
3836  */
3837 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3838         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3839         if(pos+10>buf_size) pos=buf_size; // oops ;)
3840
3841         return pos;
3842 }
3843
3844 static int decode_frame(AVCodecContext *avctx,
3845                              void *data, int *data_size,
3846                              AVPacket *avpkt)
3847 {
3848     const uint8_t *buf = avpkt->data;
3849     int buf_size = avpkt->size;
3850     H264Context *h = avctx->priv_data;
3851     MpegEncContext *s = &h->s;
3852     AVFrame *pict = data;
3853     int buf_index;
3854
3855     s->flags= avctx->flags;
3856     s->flags2= avctx->flags2;
3857
3858    /* end of stream, output what is still in the buffers */
3859  out:
3860     if (buf_size == 0) {
3861         Picture *out;
3862         int i, out_idx;
3863
3864         s->current_picture_ptr = NULL;
3865
3866 //FIXME factorize this with the output code below
3867         out = h->delayed_pic[0];
3868         out_idx = 0;
3869         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3870             if(h->delayed_pic[i]->poc < out->poc){
3871                 out = h->delayed_pic[i];
3872                 out_idx = i;
3873             }
3874
3875         for(i=out_idx; h->delayed_pic[i]; i++)
3876             h->delayed_pic[i] = h->delayed_pic[i+1];
3877
3878         if(out){
3879             *data_size = sizeof(AVFrame);
3880             *pict= *(AVFrame*)out;
3881         }
3882
3883         return 0;
3884     }
3885
3886     buf_index=decode_nal_units(h, buf, buf_size);
3887     if(buf_index < 0)
3888         return -1;
3889
3890     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3891         buf_size = 0;
3892         goto out;
3893     }
3894
3895     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3896         if (avctx->skip_frame >= AVDISCARD_NONREF)
3897             return 0;
3898         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3899         return -1;
3900     }
3901
3902     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3903
3904         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3905
3906         field_end(h, 0);
3907
3908         if (!h->next_output_pic) {
3909             /* Wait for second field. */
3910             *data_size = 0;
3911
3912         } else {
3913             *data_size = sizeof(AVFrame);
3914             *pict = *(AVFrame*)h->next_output_pic;
3915         }
3916     }
3917
3918     assert(pict->data[0] || !*data_size);
3919     ff_print_debug_info(s, pict);
3920 //printf("out %d\n", (int)pict->data[0]);
3921
3922     return get_consumed_bytes(s, buf_index, buf_size);
3923 }
3924 #if 0
3925 static inline void fill_mb_avail(H264Context *h){
3926     MpegEncContext * const s = &h->s;
3927     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3928
3929     if(s->mb_y){
3930         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3931         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3932         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3933     }else{
3934         h->mb_avail[0]=
3935         h->mb_avail[1]=
3936         h->mb_avail[2]= 0;
3937     }
3938     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3939     h->mb_avail[4]= 1; //FIXME move out
3940     h->mb_avail[5]= 0; //FIXME move out
3941 }
3942 #endif
3943
3944 #ifdef TEST
3945 #undef printf
3946 #undef random
3947 #define COUNT 8000
3948 #define SIZE (COUNT*40)
3949 int main(void){
3950     int i;
3951     uint8_t temp[SIZE];
3952     PutBitContext pb;
3953     GetBitContext gb;
3954 //    int int_temp[10000];
3955     DSPContext dsp;
3956     AVCodecContext avctx;
3957
3958     dsputil_init(&dsp, &avctx);
3959
3960     init_put_bits(&pb, temp, SIZE);
3961     printf("testing unsigned exp golomb\n");
3962     for(i=0; i<COUNT; i++){
3963         START_TIMER
3964         set_ue_golomb(&pb, i);
3965         STOP_TIMER("set_ue_golomb");
3966     }
3967     flush_put_bits(&pb);
3968
3969     init_get_bits(&gb, temp, 8*SIZE);
3970     for(i=0; i<COUNT; i++){
3971         int j, s;
3972
3973         s= show_bits(&gb, 24);
3974
3975         START_TIMER
3976         j= get_ue_golomb(&gb);
3977         if(j != i){
3978             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3979 //            return -1;
3980         }
3981         STOP_TIMER("get_ue_golomb");
3982     }
3983
3984
3985     init_put_bits(&pb, temp, SIZE);
3986     printf("testing signed exp golomb\n");
3987     for(i=0; i<COUNT; i++){
3988         START_TIMER
3989         set_se_golomb(&pb, i - COUNT/2);
3990         STOP_TIMER("set_se_golomb");
3991     }
3992     flush_put_bits(&pb);
3993
3994     init_get_bits(&gb, temp, 8*SIZE);
3995     for(i=0; i<COUNT; i++){
3996         int j, s;
3997
3998         s= show_bits(&gb, 24);
3999
4000         START_TIMER
4001         j= get_se_golomb(&gb);
4002         if(j != i - COUNT/2){
4003             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4004 //            return -1;
4005         }
4006         STOP_TIMER("get_se_golomb");
4007     }
4008
4009 #if 0
4010     printf("testing 4x4 (I)DCT\n");
4011
4012     DCTELEM block[16];
4013     uint8_t src[16], ref[16];
4014     uint64_t error= 0, max_error=0;
4015
4016     for(i=0; i<COUNT; i++){
4017         int j;
4018 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
4019         for(j=0; j<16; j++){
4020             ref[j]= random()%255;
4021             src[j]= random()%255;
4022         }
4023
4024         h264_diff_dct_c(block, src, ref, 4);
4025
4026         //normalize
4027         for(j=0; j<16; j++){
4028 //            printf("%d ", block[j]);
4029             block[j]= block[j]*4;
4030             if(j&1) block[j]= (block[j]*4 + 2)/5;
4031             if(j&4) block[j]= (block[j]*4 + 2)/5;
4032         }
4033 //        printf("\n");
4034
4035         h->h264dsp.h264_idct_add(ref, block, 4);
4036 /*        for(j=0; j<16; j++){
4037             printf("%d ", ref[j]);
4038         }
4039         printf("\n");*/
4040
4041         for(j=0; j<16; j++){
4042             int diff= FFABS(src[j] - ref[j]);
4043
4044             error+= diff*diff;
4045             max_error= FFMAX(max_error, diff);
4046         }
4047     }
4048     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
4049     printf("testing quantizer\n");
4050     for(qp=0; qp<52; qp++){
4051         for(i=0; i<16; i++)
4052             src1_block[i]= src2_block[i]= random()%255;
4053
4054     }
4055     printf("Testing NAL layer\n");
4056
4057     uint8_t bitstream[COUNT];
4058     uint8_t nal[COUNT*2];
4059     H264Context h;
4060     memset(&h, 0, sizeof(H264Context));
4061
4062     for(i=0; i<COUNT; i++){
4063         int zeros= i;
4064         int nal_length;
4065         int consumed;
4066         int out_length;
4067         uint8_t *out;
4068         int j;
4069
4070         for(j=0; j<COUNT; j++){
4071             bitstream[j]= (random() % 255) + 1;
4072         }
4073
4074         for(j=0; j<zeros; j++){
4075             int pos= random() % COUNT;
4076             while(bitstream[pos] == 0){
4077                 pos++;
4078                 pos %= COUNT;
4079             }
4080             bitstream[pos]=0;
4081         }
4082
4083         START_TIMER
4084
4085         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
4086         if(nal_length<0){
4087             printf("encoding failed\n");
4088             return -1;
4089         }
4090
4091         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
4092
4093         STOP_TIMER("NAL")
4094
4095         if(out_length != COUNT){
4096             printf("incorrect length %d %d\n", out_length, COUNT);
4097             return -1;
4098         }
4099
4100         if(consumed != nal_length){
4101             printf("incorrect consumed length %d %d\n", nal_length, consumed);
4102             return -1;
4103         }
4104
4105         if(memcmp(bitstream, out, COUNT)){
4106             printf("mismatch\n");
4107             return -1;
4108         }
4109     }
4110 #endif
4111
4112     printf("Testing RBSP\n");
4113
4114
4115     return 0;
4116 }
4117 #endif /* TEST */
4118
4119
4120 av_cold void ff_h264_free_context(H264Context *h)
4121 {
4122     int i;
4123
4124     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4125
4126     for(i = 0; i < MAX_SPS_COUNT; i++)
4127         av_freep(h->sps_buffers + i);
4128
4129     for(i = 0; i < MAX_PPS_COUNT; i++)
4130         av_freep(h->pps_buffers + i);
4131 }
4132
4133 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4134 {
4135     H264Context *h = avctx->priv_data;
4136     MpegEncContext *s = &h->s;
4137
4138     ff_h264_free_context(h);
4139
4140     MPV_common_end(s);
4141
4142 //    memset(h, 0, sizeof(H264Context));
4143
4144     return 0;
4145 }
4146
4147 static const AVProfile profiles[] = {
4148     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4149     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4150     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4151     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4152     { FF_PROFILE_H264_HIGH,                 "High"                  },
4153     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4154     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4155     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4156     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4157     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4158     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4159     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4160     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4161     { FF_PROFILE_UNKNOWN },
4162 };
4163
4164 AVCodec ff_h264_decoder = {
4165     "h264",
4166     AVMEDIA_TYPE_VIDEO,
4167     CODEC_ID_H264,
4168     sizeof(H264Context),
4169     ff_h264_decode_init,
4170     NULL,
4171     ff_h264_decode_end,
4172     decode_frame,
4173     /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4174         CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4175     .flush= flush_dpb,
4176     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4177     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4178     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4179     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4180 };
4181
4182 #if CONFIG_H264_VDPAU_DECODER
4183 AVCodec ff_h264_vdpau_decoder = {
4184     "h264_vdpau",
4185     AVMEDIA_TYPE_VIDEO,
4186     CODEC_ID_H264,
4187     sizeof(H264Context),
4188     ff_h264_decode_init,
4189     NULL,
4190     ff_h264_decode_end,
4191     decode_frame,
4192     CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4193     .flush= flush_dpb,
4194     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4195     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4196     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4197 };
4198 #endif