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