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

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * Libav is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with Libav; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavutil/imgutils.h"
  29 #include "internal.h"
  30 #include "dsputil.h"
  31 #include "avcodec.h"
  32 #include "mpegvideo.h"
  33 #include "h264.h"
  34 #include "h264data.h"
  35 #include "h264_mvpred.h"
  36 #include "golomb.h"
  37 #include "mathops.h"
  38 #include "rectangle.h"
  39 #include "thread.h"
  40 #include "vdpau_internal.h"
  41 #include "libavutil/avassert.h"
  42
  43 #include "cabac.h"
  44
  45 //#undef NDEBUG
  46 #include <assert.h>
  47
  48 static const uint8_t rem6[QP_MAX_NUM+1]={
  49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[QP_MAX_NUM+1]={
  53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
  54 };
  55
  56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  57     PIX_FMT_DXVA2_VLD,
  58     PIX_FMT_VAAPI_VLD,
  59     PIX_FMT_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 void ff_h264_write_back_intra_pred_mode(H264Context *h){
  64     int8_t *mode= h->intra4x4_pred_mode + h->mb2br_xy[h->mb_xy];
  65
  66     AV_COPY32(mode, h->intra4x4_pred_mode_cache + 4 + 8*4);
  67     mode[4]= h->intra4x4_pred_mode_cache[7+8*3];
  68     mode[5]= h->intra4x4_pred_mode_cache[7+8*2];
  69     mode[6]= h->intra4x4_pred_mode_cache[7+8*1];
  70 }
  71
  72 /**
  73  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  74  */
  75 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  76     MpegEncContext * const s = &h->s;
  77     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  78     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  79     int i;
  80
  81     if(!(h->top_samples_available&0x8000)){
  82         for(i=0; i<4; i++){
  83             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  84             if(status<0){
  85                 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);
  86                 return -1;
  87             } else if(status){
  88                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  89             }
  90         }
  91     }
  92
  93     if((h->left_samples_available&0x8888)!=0x8888){
  94         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  95         for(i=0; i<4; i++){
  96             if(!(h->left_samples_available&mask[i])){
  97                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  98                 if(status<0){
  99                     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);
 100                     return -1;
 101                 } else if(status){
 102                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
 103                 }
 104             }
 105         }
 106     }
 107
 108     return 0;
 109 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 110
 111 /**
 112  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 113  */
 114 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 115     MpegEncContext * const s = &h->s;
 116     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 117     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 118
 119     if(mode > 6U) {
 120         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);
 121         return -1;
 122     }
 123
 124     if(!(h->top_samples_available&0x8000)){
 125         mode= top[ mode ];
 126         if(mode<0){
 127             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);
 128             return -1;
 129         }
 130     }
 131
 132     if((h->left_samples_available&0x8080) != 0x8080){
 133         mode= left[ mode ];
 134         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 135             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 136         }
 137         if(mode<0){
 138             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);
 139             return -1;
 140         }
 141     }
 142
 143     return mode;
 144 }
 145
 146 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 147     int i, si, di;
 148     uint8_t *dst;
 149     int bufidx;
 150
 151 //    src[0]&0x80;                //forbidden bit
 152     h->nal_ref_idc= src[0]>>5;
 153     h->nal_unit_type= src[0]&0x1F;
 154
 155     src++; length--;
 156
 157 #if HAVE_FAST_UNALIGNED
 158 # if HAVE_FAST_64BIT
 159 #   define RS 7
 160     for(i=0; i+1<length; i+=9){
 161         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 162 # else
 163 #   define RS 3
 164     for(i=0; i+1<length; i+=5){
 165         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 166 # endif
 167             continue;
 168         if(i>0 && !src[i]) i--;
 169         while(src[i]) i++;
 170 #else
 171 #   define RS 0
 172     for(i=0; i+1<length; i+=2){
 173         if(src[i]) continue;
 174         if(i>0 && src[i-1]==0) i--;
 175 #endif
 176         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 177             if(src[i+2]!=3){
 178                 /* startcode, so we must be past the end */
 179                 length=i;
 180             }
 181             break;
 182         }
 183         i-= RS;
 184     }
 185
 186     if(i>=length-1){ //no escaped 0
 187         *dst_length= length;
 188         *consumed= length+1; //+1 for the header
 189         return src;
 190     }
 191
 192     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 193     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 194     dst= h->rbsp_buffer[bufidx];
 195
 196     if (dst == NULL){
 197         return NULL;
 198     }
 199
 200 //printf("decoding esc\n");
 201     memcpy(dst, src, i);
 202     si=di=i;
 203     while(si+2<length){
 204         //remove escapes (very rare 1:2^22)
 205         if(src[si+2]>3){
 206             dst[di++]= src[si++];
 207             dst[di++]= src[si++];
 208         }else if(src[si]==0 && src[si+1]==0){
 209             if(src[si+2]==3){ //escape
 210                 dst[di++]= 0;
 211                 dst[di++]= 0;
 212                 si+=3;
 213                 continue;
 214             }else //next start code
 215                 goto nsc;
 216         }
 217
 218         dst[di++]= src[si++];
 219     }
 220     while(si<length)
 221         dst[di++]= src[si++];
 222 nsc:
 223
 224     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 225
 226     *dst_length= di;
 227     *consumed= si + 1;//+1 for the header
 228 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 229     return dst;
 230 }
 231
 232 /**
 233  * Identify the exact end of the bitstream
 234  * @return the length of the trailing, or 0 if damaged
 235  */
 236 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 237     int v= *src;
 238     int r;
 239
 240     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 241
 242     for(r=1; r<9; r++){
 243         if(v&1) return r;
 244         v>>=1;
 245     }
 246     return 0;
 247 }
 248
 249 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 250                                  int y_offset, int list){
 251     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 252     int filter_height= (raw_my&3) ? 2 : 0;
 253     int full_my= (raw_my>>2) + y_offset;
 254     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 255
 256     return FFMAX(abs(top), bottom);
 257 }
 258
 259 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 260                                int y_offset, int list0, int list1, int *nrefs){
 261     MpegEncContext * const s = &h->s;
 262     int my;
 263
 264     y_offset += 16*(s->mb_y >> MB_FIELD);
 265
 266     if(list0){
 267         int ref_n = h->ref_cache[0][ scan8[n] ];
 268         Picture *ref= &h->ref_list[0][ref_n];
 269
 270         // Error resilience puts the current picture in the ref list.
 271         // Don't try to wait on these as it will cause a deadlock.
 272         // Fields can wait on each other, though.
 273         if(ref->thread_opaque != s->current_picture.thread_opaque ||
 274            (ref->reference&3) != s->picture_structure) {
 275             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 276             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 277             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 278         }
 279     }
 280
 281     if(list1){
 282         int ref_n = h->ref_cache[1][ scan8[n] ];
 283         Picture *ref= &h->ref_list[1][ref_n];
 284
 285         if(ref->thread_opaque != s->current_picture.thread_opaque ||
 286            (ref->reference&3) != s->picture_structure) {
 287             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 288             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 289             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 290         }
 291     }
 292 }
 293
 294 /**
 295  * Wait until all reference frames are available for MC operations.
 296  *
 297  * @param h the H264 context
 298  */
 299 static void await_references(H264Context *h){
 300     MpegEncContext * const s = &h->s;
 301     const int mb_xy= h->mb_xy;
 302     const int mb_type= s->current_picture.mb_type[mb_xy];
 303     int refs[2][48];
 304     int nrefs[2] = {0};
 305     int ref, list;
 306
 307     memset(refs, -1, sizeof(refs));
 308
 309     if(IS_16X16(mb_type)){
 310         get_lowest_part_y(h, refs, 0, 16, 0,
 311                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 312     }else if(IS_16X8(mb_type)){
 313         get_lowest_part_y(h, refs, 0, 8, 0,
 314                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 315         get_lowest_part_y(h, refs, 8, 8, 8,
 316                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 317     }else if(IS_8X16(mb_type)){
 318         get_lowest_part_y(h, refs, 0, 16, 0,
 319                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 320         get_lowest_part_y(h, refs, 4, 16, 0,
 321                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 322     }else{
 323         int i;
 324
 325         assert(IS_8X8(mb_type));
 326
 327         for(i=0; i<4; i++){
 328             const int sub_mb_type= h->sub_mb_type[i];
 329             const int n= 4*i;
 330             int y_offset= (i&2)<<2;
 331
 332             if(IS_SUB_8X8(sub_mb_type)){
 333                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 334                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 335             }else if(IS_SUB_8X4(sub_mb_type)){
 336                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 337                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 338                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 339                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 340             }else if(IS_SUB_4X8(sub_mb_type)){
 341                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 342                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 343                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 344                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 345             }else{
 346                 int j;
 347                 assert(IS_SUB_4X4(sub_mb_type));
 348                 for(j=0; j<4; j++){
 349                     int sub_y_offset= y_offset + 2*(j&2);
 350                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 351                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 352                 }
 353             }
 354         }
 355     }
 356
 357     for(list=h->list_count-1; list>=0; list--){
 358         for(ref=0; ref<48 && nrefs[list]; ref++){
 359             int row = refs[list][ref];
 360             if(row >= 0){
 361                 Picture *ref_pic = &h->ref_list[list][ref];
 362                 int ref_field = ref_pic->reference - 1;
 363                 int ref_field_picture = ref_pic->field_picture;
 364                 int pic_height = 16*s->mb_height >> ref_field_picture;
 365
 366                 row <<= MB_MBAFF;
 367                 nrefs[list]--;
 368
 369                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 370                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 371                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 372                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 373                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 374                 }else if(FIELD_PICTURE){
 375                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 376                 }else{
 377                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 378                 }
 379             }
 380         }
 381     }
 382 }
 383
 384 #if 0
 385 /**
 386  * DCT transforms the 16 dc values.
 387  * @param qp quantization parameter ??? FIXME
 388  */
 389 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 390 //    const int qmul= dequant_coeff[qp][0];
 391     int i;
 392     int temp[16]; //FIXME check if this is a good idea
 393     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 394     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 395
 396     for(i=0; i<4; i++){
 397         const int offset= y_offset[i];
 398         const int z0= block[offset+stride*0] + block[offset+stride*4];
 399         const int z1= block[offset+stride*0] - block[offset+stride*4];
 400         const int z2= block[offset+stride*1] - block[offset+stride*5];
 401         const int z3= block[offset+stride*1] + block[offset+stride*5];
 402
 403         temp[4*i+0]= z0+z3;
 404         temp[4*i+1]= z1+z2;
 405         temp[4*i+2]= z1-z2;
 406         temp[4*i+3]= z0-z3;
 407     }
 408
 409     for(i=0; i<4; i++){
 410         const int offset= x_offset[i];
 411         const int z0= temp[4*0+i] + temp[4*2+i];
 412         const int z1= temp[4*0+i] - temp[4*2+i];
 413         const int z2= temp[4*1+i] - temp[4*3+i];
 414         const int z3= temp[4*1+i] + temp[4*3+i];
 415
 416         block[stride*0 +offset]= (z0 + z3)>>1;
 417         block[stride*2 +offset]= (z1 + z2)>>1;
 418         block[stride*8 +offset]= (z1 - z2)>>1;
 419         block[stride*10+offset]= (z0 - z3)>>1;
 420     }
 421 }
 422 #endif
 423
 424 #undef xStride
 425 #undef stride
 426
 427 #if 0
 428 static void chroma_dc_dct_c(DCTELEM *block){
 429     const int stride= 16*2;
 430     const int xStride= 16;
 431     int a,b,c,d,e;
 432
 433     a= block[stride*0 + xStride*0];
 434     b= block[stride*0 + xStride*1];
 435     c= block[stride*1 + xStride*0];
 436     d= block[stride*1 + xStride*1];
 437
 438     e= a-b;
 439     a= a+b;
 440     b= c-d;
 441     c= c+d;
 442
 443     block[stride*0 + xStride*0]= (a+c);
 444     block[stride*0 + xStride*1]= (e+b);
 445     block[stride*1 + xStride*0]= (a-c);
 446     block[stride*1 + xStride*1]= (e-b);
 447 }
 448 #endif
 449
 450 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 451                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 452                            int src_x_offset, int src_y_offset,
 453                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 454                            int pixel_shift){
 455     MpegEncContext * const s = &h->s;
 456     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 457     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 458     const int luma_xy= (mx&3) + ((my&3)<<2);
 459     uint8_t * src_y = pic->data[0] + ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 460     uint8_t * src_cb, * src_cr;
 461     int extra_width= h->emu_edge_width;
 462     int extra_height= h->emu_edge_height;
 463     int emu=0;
 464     const int full_mx= mx>>2;
 465     const int full_my= my>>2;
 466     const int pic_width  = 16*s->mb_width;
 467     const int pic_height = 16*s->mb_height >> MB_FIELD;
 468
 469     if(mx&7) extra_width -= 3;
 470     if(my&7) extra_height -= 3;
 471
 472     if(   full_mx < 0-extra_width
 473        || full_my < 0-extra_height
 474        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 475        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 476         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);
 477             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 478         emu=1;
 479     }
 480
 481     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 482     if(!square){
 483         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 484     }
 485
 486     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 487
 488     if(MB_FIELD){
 489         // chroma offset when predicting from a field of opposite parity
 490         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
 491         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 492     }
 493     src_cb= pic->data[1] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
 494     src_cr= pic->data[2] + ((mx>>3) << pixel_shift) + (my>>3)*h->mb_uvlinesize;
 495
 496     if(emu){
 497         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);
 498             src_cb= s->edge_emu_buffer;
 499     }
 500     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 501
 502     if(emu){
 503         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);
 504             src_cr= s->edge_emu_buffer;
 505     }
 506     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 507 }
 508
 509 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 510                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 511                            int x_offset, int y_offset,
 512                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 513                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 514                            int list0, int list1, int pixel_shift){
 515     MpegEncContext * const s = &h->s;
 516     qpel_mc_func *qpix_op=  qpix_put;
 517     h264_chroma_mc_func chroma_op= chroma_put;
 518
 519     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->  mb_linesize;
 520     dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 521     dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 522     x_offset += 8*s->mb_x;
 523     y_offset += 8*(s->mb_y >> MB_FIELD);
 524
 525     if(list0){
 526         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 527         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 528                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 529                            qpix_op, chroma_op, pixel_shift);
 530
 531         qpix_op=  qpix_avg;
 532         chroma_op= chroma_avg;
 533     }
 534
 535     if(list1){
 536         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 537         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 538                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 539                            qpix_op, chroma_op, pixel_shift);
 540     }
 541 }
 542
 543 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 544                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 545                            int x_offset, int y_offset,
 546                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 547                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 548                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 549                            int list0, int list1, int pixel_shift){
 550     MpegEncContext * const s = &h->s;
 551
 552     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->  mb_linesize;
 553     dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 554     dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 555     x_offset += 8*s->mb_x;
 556     y_offset += 8*(s->mb_y >> MB_FIELD);
 557
 558     if(list0 && list1){
 559         /* don't optimize for luma-only case, since B-frames usually
 560          * use implicit weights => chroma too. */
 561         uint8_t *tmp_cb = s->obmc_scratchpad;
 562         uint8_t *tmp_cr = s->obmc_scratchpad + (8 << pixel_shift);
 563         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
 564         int refn0 = h->ref_cache[0][ scan8[n] ];
 565         int refn1 = h->ref_cache[1][ scan8[n] ];
 566
 567         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 568                     dest_y, dest_cb, dest_cr,
 569                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift);
 570         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 571                     tmp_y, tmp_cb, tmp_cr,
 572                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift);
 573
 574         if(h->use_weight == 2){
 575             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 576             int weight1 = 64 - weight0;
 577             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 578             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 579             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 580         }else{
 581             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 582                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 583                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 584             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 585                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 586                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 587             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 588                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 589                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 590         }
 591     }else{
 592         int list = list1 ? 1 : 0;
 593         int refn = h->ref_cache[list][ scan8[n] ];
 594         Picture *ref= &h->ref_list[list][refn];
 595         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 596                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 597                     qpix_put, chroma_put, pixel_shift);
 598
 599         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 600                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 601         if(h->use_weight_chroma){
 602             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 603                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 604             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 605                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 606         }
 607     }
 608 }
 609
 610 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 611                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 612                            int x_offset, int y_offset,
 613                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 614                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 615                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 616                            int list0, int list1, int pixel_shift){
 617     if((h->use_weight==2 && list0 && list1
 618         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 619        || h->use_weight==1)
 620         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 621                          x_offset, y_offset, qpix_put, chroma_put,
 622                          weight_op[0], weight_op[3], weight_avg[0],
 623                          weight_avg[3], list0, list1, pixel_shift);
 624     else
 625         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 626                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 627                     chroma_avg, list0, list1, pixel_shift);
 628 }
 629
 630 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift){
 631     /* fetch pixels for estimated mv 4 macroblocks ahead
 632      * optimized for 64byte cache lines */
 633     MpegEncContext * const s = &h->s;
 634     const int refn = h->ref_cache[list][scan8[0]];
 635     if(refn >= 0){
 636         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 637         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 638         uint8_t **src= h->ref_list[list][refn].data;
 639         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 640         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 641         off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
 642         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 643     }
 644 }
 645
 646 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 647                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 648                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 649                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 650                       int pixel_shift){
 651     MpegEncContext * const s = &h->s;
 652     const int mb_xy= h->mb_xy;
 653     const int mb_type= s->current_picture.mb_type[mb_xy];
 654
 655     assert(IS_INTER(mb_type));
 656
 657     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 658         await_references(h);
 659     prefetch_motion(h, 0, pixel_shift);
 660
 661     if(IS_16X16(mb_type)){
 662         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 663                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 664                 weight_op, weight_avg,
 665                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 666                 pixel_shift);
 667     }else if(IS_16X8(mb_type)){
 668         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 669                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 670                 &weight_op[1], &weight_avg[1],
 671                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 672                 pixel_shift);
 673         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 674                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 675                 &weight_op[1], &weight_avg[1],
 676                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 677                 pixel_shift);
 678     }else if(IS_8X16(mb_type)){
 679         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 680                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 681                 &weight_op[2], &weight_avg[2],
 682                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 683                 pixel_shift);
 684         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 685                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 686                 &weight_op[2], &weight_avg[2],
 687                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 688                 pixel_shift);
 689     }else{
 690         int i;
 691
 692         assert(IS_8X8(mb_type));
 693
 694         for(i=0; i<4; i++){
 695             const int sub_mb_type= h->sub_mb_type[i];
 696             const int n= 4*i;
 697             int x_offset= (i&1)<<2;
 698             int y_offset= (i&2)<<1;
 699
 700             if(IS_SUB_8X8(sub_mb_type)){
 701                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 702                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 703                     &weight_op[3], &weight_avg[3],
 704                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 705                     pixel_shift);
 706             }else if(IS_SUB_8X4(sub_mb_type)){
 707                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 708                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 709                     &weight_op[4], &weight_avg[4],
 710                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 711                     pixel_shift);
 712                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 713                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 714                     &weight_op[4], &weight_avg[4],
 715                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 716                     pixel_shift);
 717             }else if(IS_SUB_4X8(sub_mb_type)){
 718                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 719                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 720                     &weight_op[5], &weight_avg[5],
 721                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 722                     pixel_shift);
 723                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 724                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 725                     &weight_op[5], &weight_avg[5],
 726                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 727                     pixel_shift);
 728             }else{
 729                 int j;
 730                 assert(IS_SUB_4X4(sub_mb_type));
 731                 for(j=0; j<4; j++){
 732                     int sub_x_offset= x_offset + 2*(j&1);
 733                     int sub_y_offset= y_offset +   (j&2);
 734                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 735                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 736                         &weight_op[6], &weight_avg[6],
 737                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 738                         pixel_shift);
 739                 }
 740             }
 741         }
 742     }
 743
 744     prefetch_motion(h, 1, pixel_shift);
 745 }
 746
 747 #define hl_motion_fn(sh, bits) \
 748 static av_always_inline void hl_motion_ ## bits(H264Context *h, \
 749                                        uint8_t *dest_y, \
 750                                        uint8_t *dest_cb, uint8_t *dest_cr, \
 751                                        qpel_mc_func (*qpix_put)[16], \
 752                                        h264_chroma_mc_func (*chroma_put), \
 753                                        qpel_mc_func (*qpix_avg)[16], \
 754                                        h264_chroma_mc_func (*chroma_avg), \
 755                                        h264_weight_func *weight_op, \
 756                                        h264_biweight_func *weight_avg) \
 757 { \
 758     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put, \
 759               qpix_avg, chroma_avg, weight_op, weight_avg, sh); \
 760 }
 761 hl_motion_fn(0, 8);
 762 hl_motion_fn(1, 16);
 763
 764 static void free_tables(H264Context *h, int free_rbsp){
 765     int i;
 766     H264Context *hx;
 767
 768     av_freep(&h->intra4x4_pred_mode);
 769     av_freep(&h->chroma_pred_mode_table);
 770     av_freep(&h->cbp_table);
 771     av_freep(&h->mvd_table[0]);
 772     av_freep(&h->mvd_table[1]);
 773     av_freep(&h->direct_table);
 774     av_freep(&h->non_zero_count);
 775     av_freep(&h->slice_table_base);
 776     h->slice_table= NULL;
 777     av_freep(&h->list_counts);
 778
 779     av_freep(&h->mb2b_xy);
 780     av_freep(&h->mb2br_xy);
 781
 782     for(i = 0; i < MAX_THREADS; i++) {
 783         hx = h->thread_context[i];
 784         if(!hx) continue;
 785         av_freep(&hx->top_borders[1]);
 786         av_freep(&hx->top_borders[0]);
 787         av_freep(&hx->s.obmc_scratchpad);
 788         if (free_rbsp){
 789             av_freep(&hx->rbsp_buffer[1]);
 790             av_freep(&hx->rbsp_buffer[0]);
 791             hx->rbsp_buffer_size[0] = 0;
 792             hx->rbsp_buffer_size[1] = 0;
 793         }
 794         if (i) av_freep(&h->thread_context[i]);
 795     }
 796 }
 797
 798 static void init_dequant8_coeff_table(H264Context *h){
 799     int i,q,x;
 800     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 801     h->dequant8_coeff[0] = h->dequant8_buffer[0];
 802     h->dequant8_coeff[1] = h->dequant8_buffer[1];
 803
 804     for(i=0; i<2; i++ ){
 805         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
 806             h->dequant8_coeff[1] = h->dequant8_buffer[0];
 807             break;
 808         }
 809
 810         for(q=0; q<max_qp+1; q++){
 811             int shift = div6[q];
 812             int idx = rem6[q];
 813             for(x=0; x<64; x++)
 814                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 815                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 816                     h->pps.scaling_matrix8[i][x]) << shift;
 817         }
 818     }
 819 }
 820
 821 static void init_dequant4_coeff_table(H264Context *h){
 822     int i,j,q,x;
 823     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 824     for(i=0; i<6; i++ ){
 825         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 826         for(j=0; j<i; j++){
 827             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 828                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 829                 break;
 830             }
 831         }
 832         if(j<i)
 833             continue;
 834
 835         for(q=0; q<max_qp+1; q++){
 836             int shift = div6[q] + 2;
 837             int idx = rem6[q];
 838             for(x=0; x<16; x++)
 839                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 840                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 841                     h->pps.scaling_matrix4[i][x]) << shift;
 842         }
 843     }
 844 }
 845
 846 static void init_dequant_tables(H264Context *h){
 847     int i,x;
 848     init_dequant4_coeff_table(h);
 849     if(h->pps.transform_8x8_mode)
 850         init_dequant8_coeff_table(h);
 851     if(h->sps.transform_bypass){
 852         for(i=0; i<6; i++)
 853             for(x=0; x<16; x++)
 854                 h->dequant4_coeff[i][0][x] = 1<<6;
 855         if(h->pps.transform_8x8_mode)
 856             for(i=0; i<2; i++)
 857                 for(x=0; x<64; x++)
 858                     h->dequant8_coeff[i][0][x] = 1<<6;
 859     }
 860 }
 861
 862
 863 int ff_h264_alloc_tables(H264Context *h){
 864     MpegEncContext * const s = &h->s;
 865     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 866     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 867     int x,y;
 868
 869     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 870
 871     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
 872     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 873     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 874
 875     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 876     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 877     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 878     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 879     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 880
 881     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 882     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 883
 884     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 885     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 886     for(y=0; y<s->mb_height; y++){
 887         for(x=0; x<s->mb_width; x++){
 888             const int mb_xy= x + y*s->mb_stride;
 889             const int b_xy = 4*x + 4*y*h->b_stride;
 890
 891             h->mb2b_xy [mb_xy]= b_xy;
 892             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 893         }
 894     }
 895
 896     s->obmc_scratchpad = NULL;
 897
 898     if(!h->dequant4_coeff[0])
 899         init_dequant_tables(h);
 900
 901     return 0;
 902 fail:
 903     free_tables(h, 1);
 904     return -1;
 905 }
 906
 907 /**
 908  * Mimic alloc_tables(), but for every context thread.
 909  */
 910 static void clone_tables(H264Context *dst, H264Context *src, int i){
 911     MpegEncContext * const s = &src->s;
 912     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 913     dst->non_zero_count           = src->non_zero_count;
 914     dst->slice_table              = src->slice_table;
 915     dst->cbp_table                = src->cbp_table;
 916     dst->mb2b_xy                  = src->mb2b_xy;
 917     dst->mb2br_xy                 = src->mb2br_xy;
 918     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 919     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 920     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 921     dst->direct_table             = src->direct_table;
 922     dst->list_counts              = src->list_counts;
 923
 924     dst->s.obmc_scratchpad = NULL;
 925     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
 926 }
 927
 928 /**
 929  * Init context
 930  * Allocate buffers which are not shared amongst multiple threads.
 931  */
 932 static int context_init(H264Context *h){
 933     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t)*2, fail)
 934     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t)*2, fail)
 935
 936     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 937     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 938
 939     return 0;
 940 fail:
 941     return -1; // free_tables will clean up for us
 942 }
 943
 944 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 945
 946 static av_cold void common_init(H264Context *h){
 947     MpegEncContext * const s = &h->s;
 948
 949     s->width = s->avctx->width;
 950     s->height = s->avctx->height;
 951     s->codec_id= s->avctx->codec->id;
 952
 953     ff_h264dsp_init(&h->h264dsp, 8);
 954     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
 955
 956     h->dequant_coeff_pps= -1;
 957     s->unrestricted_mv=1;
 958     s->decode=1; //FIXME
 959
 960     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 961
 962     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 963     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 964 }
 965
 966 int ff_h264_decode_extradata(H264Context *h)
 967 {
 968     AVCodecContext *avctx = h->s.avctx;
 969
 970     if(*(char *)avctx->extradata == 1){
 971         int i, cnt, nalsize;
 972         unsigned char *p = avctx->extradata;
 973
 974         h->is_avc = 1;
 975
 976         if(avctx->extradata_size < 7) {
 977             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
 978             return -1;
 979         }
 980         /* sps and pps in the avcC always have length coded with 2 bytes,
 981            so put a fake nal_length_size = 2 while parsing them */
 982         h->nal_length_size = 2;
 983         // Decode sps from avcC
 984         cnt = *(p+5) & 0x1f; // Number of sps
 985         p += 6;
 986         for (i = 0; i < cnt; i++) {
 987             nalsize = AV_RB16(p) + 2;
 988             if(decode_nal_units(h, p, nalsize) < 0) {
 989                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
 990                 return -1;
 991             }
 992             p += nalsize;
 993         }
 994         // Decode pps from avcC
 995         cnt = *(p++); // Number of pps
 996         for (i = 0; i < cnt; i++) {
 997             nalsize = AV_RB16(p) + 2;
 998             if(decode_nal_units(h, p, nalsize)  != nalsize) {
 999                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1000                 return -1;
1001             }
1002             p += nalsize;
1003         }
1004         // Now store right nal length size, that will be use to parse all other nals
1005         h->nal_length_size = ((*(((char*)(avctx->extradata))+4))&0x03)+1;
1006     } else {
1007         h->is_avc = 0;
1008         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1009             return -1;
1010     }
1011     return 0;
1012 }
1013
1014 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1015     H264Context *h= avctx->priv_data;
1016     MpegEncContext * const s = &h->s;
1017
1018     MPV_decode_defaults(s);
1019
1020     s->avctx = avctx;
1021     common_init(h);
1022
1023     s->out_format = FMT_H264;
1024     s->workaround_bugs= avctx->workaround_bugs;
1025
1026     // set defaults
1027 //    s->decode_mb= ff_h263_decode_mb;
1028     s->quarter_sample = 1;
1029     if(!avctx->has_b_frames)
1030     s->low_delay= 1;
1031
1032     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1033
1034     ff_h264_decode_init_vlc();
1035
1036     h->pixel_shift = 0;
1037     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1038
1039     h->thread_context[0] = h;
1040     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1041     h->prev_poc_msb= 1<<16;
1042     h->x264_build = -1;
1043     ff_h264_reset_sei(h);
1044     if(avctx->codec_id == CODEC_ID_H264){
1045         if(avctx->ticks_per_frame == 1){
1046             s->avctx->time_base.den *=2;
1047         }
1048         avctx->ticks_per_frame = 2;
1049     }
1050
1051     if(avctx->extradata_size > 0 && avctx->extradata &&
1052         ff_h264_decode_extradata(h))
1053         return -1;
1054
1055     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1056         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1057         s->low_delay = 0;
1058     }
1059
1060     return 0;
1061 }
1062
1063 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1064 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1065 {
1066     int i;
1067
1068     for (i=0; i<count; i++){
1069         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1070                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1071                 !from[i]));
1072         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1073     }
1074 }
1075
1076 static void copy_parameter_set(void **to, void **from, int count, int size)
1077 {
1078     int i;
1079
1080     for (i=0; i<count; i++){
1081         if (to[i] && !from[i]) av_freep(&to[i]);
1082         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1083
1084         if (from[i]) memcpy(to[i], from[i], size);
1085     }
1086 }
1087
1088 static int decode_init_thread_copy(AVCodecContext *avctx){
1089     H264Context *h= avctx->priv_data;
1090
1091     if (!avctx->is_copy) return 0;
1092     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1093     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1094
1095     return 0;
1096 }
1097
1098 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1099 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1100     H264Context *h= dst->priv_data, *h1= src->priv_data;
1101     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1102     int inited = s->context_initialized, err;
1103     int i;
1104
1105     if(dst == src || !s1->context_initialized) return 0;
1106
1107     err = ff_mpeg_update_thread_context(dst, src);
1108     if(err) return err;
1109
1110     //FIXME handle width/height changing
1111     if(!inited){
1112         for(i = 0; i < MAX_SPS_COUNT; i++)
1113             av_freep(h->sps_buffers + i);
1114
1115         for(i = 0; i < MAX_PPS_COUNT; i++)
1116             av_freep(h->pps_buffers + i);
1117
1118         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1119         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1120         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1121         ff_h264_alloc_tables(h);
1122         context_init(h);
1123
1124         for(i=0; i<2; i++){
1125             h->rbsp_buffer[i] = NULL;
1126             h->rbsp_buffer_size[i] = 0;
1127         }
1128
1129         h->thread_context[0] = h;
1130
1131         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1132         // so this has to be allocated here
1133         h->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
1134
1135         s->dsp.clear_blocks(h->mb);
1136     }
1137
1138     //extradata/NAL handling
1139     h->is_avc          = h1->is_avc;
1140
1141     //SPS/PPS
1142     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1143     h->sps             = h1->sps;
1144     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1145     h->pps             = h1->pps;
1146
1147     //Dequantization matrices
1148     //FIXME these are big - can they be only copied when PPS changes?
1149     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1150
1151     for(i=0; i<6; i++)
1152         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1153
1154     for(i=0; i<2; i++)
1155         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1156
1157     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1158
1159     //POC timing
1160     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1161
1162     //reference lists
1163     copy_fields(h, h1, ref_count, list_count);
1164     copy_fields(h, h1, ref_list,  intra_gb);
1165     copy_fields(h, h1, short_ref, cabac_init_idc);
1166
1167     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1168     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1169     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1170
1171     h->last_slice_type = h1->last_slice_type;
1172
1173     if(!s->current_picture_ptr) return 0;
1174
1175     if(!s->dropable) {
1176         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1177         h->prev_poc_msb     = h->poc_msb;
1178         h->prev_poc_lsb     = h->poc_lsb;
1179     }
1180     h->prev_frame_num_offset= h->frame_num_offset;
1181     h->prev_frame_num       = h->frame_num;
1182     h->outputed_poc         = h->next_outputed_poc;
1183
1184     return 0;
1185 }
1186
1187 int ff_h264_frame_start(H264Context *h){
1188     MpegEncContext * const s = &h->s;
1189     int i;
1190     const int pixel_shift = h->pixel_shift;
1191     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1192
1193     if(MPV_frame_start(s, s->avctx) < 0)
1194         return -1;
1195     ff_er_frame_start(s);
1196     /*
1197      * MPV_frame_start uses pict_type to derive key_frame.
1198      * This is incorrect for H.264; IDR markings must be used.
1199      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1200      * See decode_nal_units().
1201      */
1202     s->current_picture_ptr->key_frame= 0;
1203     s->current_picture_ptr->mmco_reset= 0;
1204
1205     assert(s->linesize && s->uvlinesize);
1206
1207     for(i=0; i<16; i++){
1208         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1209         h->block_offset[24+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1210     }
1211     for(i=0; i<4; i++){
1212         h->block_offset[16+i]=
1213         h->block_offset[20+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1214         h->block_offset[24+16+i]=
1215         h->block_offset[24+20+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1216     }
1217
1218     /* can't be in alloc_tables because linesize isn't known there.
1219      * FIXME: redo bipred weight to not require extra buffer? */
1220     for(i = 0; i < thread_count; i++)
1221         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1222             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
1223
1224     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1225     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1226
1227 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
1228
1229     // We mark the current picture as non-reference after allocating it, so
1230     // that if we break out due to an error it can be released automatically
1231     // in the next MPV_frame_start().
1232     // SVQ3 as well as most other codecs have only last/next/current and thus
1233     // get released even with set reference, besides SVQ3 and others do not
1234     // mark frames as reference later "naturally".
1235     if(s->codec_id != CODEC_ID_SVQ3)
1236         s->current_picture_ptr->reference= 0;
1237
1238     s->current_picture_ptr->field_poc[0]=
1239     s->current_picture_ptr->field_poc[1]= INT_MAX;
1240
1241     h->next_output_pic = NULL;
1242
1243     assert(s->current_picture_ptr->long_ref==0);
1244
1245     return 0;
1246 }
1247
1248 /**
1249   * Run setup operations that must be run after slice header decoding.
1250   * This includes finding the next displayed frame.
1251   *
1252   * @param h h264 master context
1253   * @param setup_finished enough NALs have been read that we can call
1254   * ff_thread_finish_setup()
1255   */
1256 static void decode_postinit(H264Context *h, int setup_finished){
1257     MpegEncContext * const s = &h->s;
1258     Picture *out = s->current_picture_ptr;
1259     Picture *cur = s->current_picture_ptr;
1260     int i, pics, out_of_order, out_idx;
1261
1262     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1263     s->current_picture_ptr->pict_type= s->pict_type;
1264
1265     if (h->next_output_pic) return;
1266
1267     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1268         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1269         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1270         //to find this yet, so we assume the worst for now.
1271         //if (setup_finished)
1272         //    ff_thread_finish_setup(s->avctx);
1273         return;
1274     }
1275
1276     cur->interlaced_frame = 0;
1277     cur->repeat_pict = 0;
1278
1279     /* Signal interlacing information externally. */
1280     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1281
1282     if(h->sps.pic_struct_present_flag){
1283         switch (h->sei_pic_struct)
1284         {
1285         case SEI_PIC_STRUCT_FRAME:
1286             break;
1287         case SEI_PIC_STRUCT_TOP_FIELD:
1288         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1289             cur->interlaced_frame = 1;
1290             break;
1291         case SEI_PIC_STRUCT_TOP_BOTTOM:
1292         case SEI_PIC_STRUCT_BOTTOM_TOP:
1293             if (FIELD_OR_MBAFF_PICTURE)
1294                 cur->interlaced_frame = 1;
1295             else
1296                 // try to flag soft telecine progressive
1297                 cur->interlaced_frame = h->prev_interlaced_frame;
1298             break;
1299         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1300         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1301             // Signal the possibility of telecined film externally (pic_struct 5,6)
1302             // From these hints, let the applications decide if they apply deinterlacing.
1303             cur->repeat_pict = 1;
1304             break;
1305         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1306             // Force progressive here, as doubling interlaced frame is a bad idea.
1307             cur->repeat_pict = 2;
1308             break;
1309         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1310             cur->repeat_pict = 4;
1311             break;
1312         }
1313
1314         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1315             cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
1316     }else{
1317         /* Derive interlacing flag from used decoding process. */
1318         cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1319     }
1320     h->prev_interlaced_frame = cur->interlaced_frame;
1321
1322     if (cur->field_poc[0] != cur->field_poc[1]){
1323         /* Derive top_field_first from field pocs. */
1324         cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
1325     }else{
1326         if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
1327             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1328             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1329               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1330                 cur->top_field_first = 1;
1331             else
1332                 cur->top_field_first = 0;
1333         }else{
1334             /* Most likely progressive */
1335             cur->top_field_first = 0;
1336         }
1337     }
1338
1339     //FIXME do something with unavailable reference frames
1340
1341     /* Sort B-frames into display order */
1342
1343     if(h->sps.bitstream_restriction_flag
1344        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1345         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1346         s->low_delay = 0;
1347     }
1348
1349     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1350        && !h->sps.bitstream_restriction_flag){
1351         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1352         s->low_delay= 0;
1353     }
1354
1355     pics = 0;
1356     while(h->delayed_pic[pics]) pics++;
1357
1358     assert(pics <= MAX_DELAYED_PIC_COUNT);
1359
1360     h->delayed_pic[pics++] = cur;
1361     if(cur->reference == 0)
1362         cur->reference = DELAYED_PIC_REF;
1363
1364     out = h->delayed_pic[0];
1365     out_idx = 0;
1366     for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1367         if(h->delayed_pic[i]->poc < out->poc){
1368             out = h->delayed_pic[i];
1369             out_idx = i;
1370         }
1371     if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
1372         h->next_outputed_poc= INT_MIN;
1373     out_of_order = out->poc < h->next_outputed_poc;
1374
1375     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1376         { }
1377     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1378        || (s->low_delay &&
1379         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1380          || cur->pict_type == AV_PICTURE_TYPE_B)))
1381     {
1382         s->low_delay = 0;
1383         s->avctx->has_b_frames++;
1384     }
1385
1386     if(out_of_order || pics > s->avctx->has_b_frames){
1387         out->reference &= ~DELAYED_PIC_REF;
1388         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1389                          // or else the first thread can release the picture and reuse it unsafely
1390         for(i=out_idx; h->delayed_pic[i]; i++)
1391             h->delayed_pic[i] = h->delayed_pic[i+1];
1392     }
1393     if(!out_of_order && pics > s->avctx->has_b_frames){
1394         h->next_output_pic = out;
1395         if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
1396             h->next_outputed_poc = INT_MIN;
1397         } else
1398             h->next_outputed_poc = out->poc;
1399     }else{
1400         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1401     }
1402
1403     if (setup_finished)
1404         ff_thread_finish_setup(s->avctx);
1405 }
1406
1407 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
1408     MpegEncContext * const s = &h->s;
1409     uint8_t *top_border;
1410     int top_idx = 1;
1411     const int pixel_shift = h->pixel_shift;
1412
1413     src_y  -=   linesize;
1414     src_cb -= uvlinesize;
1415     src_cr -= uvlinesize;
1416
1417     if(!simple && FRAME_MBAFF){
1418         if(s->mb_y&1){
1419             if(!MB_MBAFF){
1420                 top_border = h->top_borders[0][s->mb_x];
1421                 AV_COPY128(top_border, src_y + 15*linesize);
1422                 if (pixel_shift)
1423                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1424                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1425                     if (pixel_shift) {
1426                         AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1427                         AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1428                     } else {
1429                     AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1430                     AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1431                     }
1432                 }
1433             }
1434         }else if(MB_MBAFF){
1435             top_idx = 0;
1436         }else
1437             return;
1438     }
1439
1440     top_border = h->top_borders[top_idx][s->mb_x];
1441     // There are two lines saved, the line above the the top macroblock of a pair,
1442     // and the line above the bottom macroblock
1443     AV_COPY128(top_border, src_y + 16*linesize);
1444     if (pixel_shift)
1445         AV_COPY128(top_border+16, src_y+16*linesize+16);
1446
1447     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1448         if (pixel_shift) {
1449             AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1450             AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1451         } else {
1452         AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1453         AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1454         }
1455     }
1456 }
1457
1458 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1459                                   uint8_t *src_cb, uint8_t *src_cr,
1460                                   int linesize, int uvlinesize,
1461                                   int xchg, int simple, int pixel_shift){
1462     MpegEncContext * const s = &h->s;
1463     int deblock_topleft;
1464     int deblock_top;
1465     int top_idx = 1;
1466     uint8_t *top_border_m1;
1467     uint8_t *top_border;
1468
1469     if(!simple && FRAME_MBAFF){
1470         if(s->mb_y&1){
1471             if(!MB_MBAFF)
1472                 return;
1473         }else{
1474             top_idx = MB_MBAFF ? 0 : 1;
1475         }
1476     }
1477
1478     if(h->deblocking_filter == 2) {
1479         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1480         deblock_top     = h->top_type;
1481     } else {
1482         deblock_topleft = (s->mb_x > 0);
1483         deblock_top     = (s->mb_y > !!MB_FIELD);
1484     }
1485
1486     src_y  -=   linesize + 1 + pixel_shift;
1487     src_cb -= uvlinesize + 1 + pixel_shift;
1488     src_cr -= uvlinesize + 1 + pixel_shift;
1489
1490     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1491     top_border    = h->top_borders[top_idx][s->mb_x];
1492
1493 #define XCHG(a,b,xchg)\
1494     if (pixel_shift) {\
1495         if (xchg) {\
1496             AV_SWAP64(b+0,a+0);\
1497             AV_SWAP64(b+8,a+8);\
1498         } else {\
1499             AV_COPY128(b,a); \
1500         }\
1501     } else \
1502 if (xchg) AV_SWAP64(b,a);\
1503 else      AV_COPY64(b,a);
1504
1505     if(deblock_top){
1506         if(deblock_topleft){
1507             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1508         }
1509         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1510         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1511         if(s->mb_x+1 < s->mb_width){
1512             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1513         }
1514     }
1515     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1516         if(deblock_top){
1517             if(deblock_topleft){
1518                 XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1519                 XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1520             }
1521             XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1522             XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1523         }
1524     }
1525 }
1526
1527 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1528     if (high_bit_depth) {
1529         return AV_RN32A(((int32_t*)mb) + index);
1530     } else
1531         return AV_RN16A(mb + index);
1532 }
1533
1534 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1535     if (high_bit_depth) {
1536         AV_WN32A(((int32_t*)mb) + index, value);
1537     } else
1538         AV_WN16A(mb + index, value);
1539 }
1540
1541 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1542     MpegEncContext * const s = &h->s;
1543     const int mb_x= s->mb_x;
1544     const int mb_y= s->mb_y;
1545     const int mb_xy= h->mb_xy;
1546     const int mb_type= s->current_picture.mb_type[mb_xy];
1547     uint8_t  *dest_y, *dest_cb, *dest_cr;
1548     int linesize, uvlinesize /*dct_offset*/;
1549     int i;
1550     int *block_offset = &h->block_offset[0];
1551     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1552     /* is_h264 should always be true if SVQ3 is disabled. */
1553     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1554     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1555     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1556
1557     dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1558     dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1559     dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
1560
1561     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1562     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1563
1564     h->list_counts[mb_xy]= h->list_count;
1565
1566     if (!simple && MB_FIELD) {
1567         linesize   = h->mb_linesize   = s->linesize * 2;
1568         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1569         block_offset = &h->block_offset[24];
1570         if(mb_y&1){ //FIXME move out of this function?
1571             dest_y -= s->linesize*15;
1572             dest_cb-= s->uvlinesize*7;
1573             dest_cr-= s->uvlinesize*7;
1574         }
1575         if(FRAME_MBAFF) {
1576             int list;
1577             for(list=0; list<h->list_count; list++){
1578                 if(!USES_LIST(mb_type, list))
1579                     continue;
1580                 if(IS_16X16(mb_type)){
1581                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1582                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1583                 }else{
1584                     for(i=0; i<16; i+=4){
1585                         int ref = h->ref_cache[list][scan8[i]];
1586                         if(ref >= 0)
1587                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1588                     }
1589                 }
1590             }
1591         }
1592     } else {
1593         linesize   = h->mb_linesize   = s->linesize;
1594         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1595 //        dct_offset = s->linesize * 16;
1596     }
1597
1598     if (!simple && IS_INTRA_PCM(mb_type)) {
1599         if (pixel_shift) {
1600             const int bit_depth = h->sps.bit_depth_luma;
1601             int j;
1602             GetBitContext gb;
1603             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1604
1605             for (i = 0; i < 16; i++) {
1606                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1607                 for (j = 0; j < 16; j++)
1608                     tmp_y[j] = get_bits(&gb, bit_depth);
1609             }
1610             for (i = 0; i < 8; i++) {
1611                 uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1612                 for (j = 0; j < 8; j++)
1613                     tmp_cb[j] = get_bits(&gb, bit_depth);
1614             }
1615             for (i = 0; i < 8; i++) {
1616                 uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1617                 for (j = 0; j < 8; j++)
1618                     tmp_cr[j] = get_bits(&gb, bit_depth);
1619             }
1620         } else {
1621         for (i=0; i<16; i++) {
1622             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1623         }
1624         for (i=0; i<8; i++) {
1625             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1626             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1627         }
1628         }
1629     } else {
1630         if(IS_INTRA(mb_type)){
1631             if(h->deblocking_filter)
1632                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple, pixel_shift);
1633
1634             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1635                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1636                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1637             }
1638
1639             if(IS_INTRA4x4(mb_type)){
1640                 if(simple || !s->encoding){
1641                     if(IS_8x8DCT(mb_type)){
1642                         if(transform_bypass){
1643                             idct_dc_add =
1644                             idct_add    = s->dsp.add_pixels8;
1645                         }else{
1646                             idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1647                             idct_add    = h->h264dsp.h264_idct8_add;
1648                         }
1649                         for(i=0; i<16; i+=4){
1650                             uint8_t * const ptr= dest_y + block_offset[i];
1651                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1652                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1653                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16 << pixel_shift), linesize);
1654                             }else{
1655                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
1656                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1657                                                             (h->topright_samples_available<<i)&0x4000, linesize);
1658                                 if(nnz){
1659                                     if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16))
1660                                         idct_dc_add(ptr, h->mb + (i*16 << pixel_shift), linesize);
1661                                     else
1662                                         idct_add   (ptr, h->mb + (i*16 << pixel_shift), linesize);
1663                                 }
1664                             }
1665                         }
1666                     }else{
1667                         if(transform_bypass){
1668                             idct_dc_add =
1669                             idct_add    = s->dsp.add_pixels4;
1670                         }else{
1671                             idct_dc_add = h->h264dsp.h264_idct_dc_add;
1672                             idct_add    = h->h264dsp.h264_idct_add;
1673                         }
1674                         for(i=0; i<16; i++){
1675                             uint8_t * const ptr= dest_y + block_offset[i];
1676                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1677
1678                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1679                                 h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16 << pixel_shift), linesize);
1680                             }else{
1681                                 uint8_t *topright;
1682                                 int nnz, tr;
1683                                 uint64_t tr_high;
1684                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1685                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1686                                     assert(mb_y || linesize <= block_offset[i]);
1687                                     if(!topright_avail){
1688                                         if (pixel_shift) {
1689                                             tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1690                                             topright= (uint8_t*) &tr_high;
1691                                         } else {
1692                                         tr= ptr[3 - linesize]*0x01010101;
1693                                         topright= (uint8_t*) &tr;
1694                                         }
1695                                     }else
1696                                         topright= ptr + (4 << pixel_shift) - linesize;
1697                                 }else
1698                                     topright= NULL;
1699
1700                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1701                                 nnz = h->non_zero_count_cache[ scan8[i] ];
1702                                 if(nnz){
1703                                     if(is_h264){
1704                                         if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16))
1705                                             idct_dc_add(ptr, h->mb + (i*16 << pixel_shift), linesize);
1706                                         else
1707                                             idct_add   (ptr, h->mb + (i*16 << pixel_shift), linesize);
1708                                     }else
1709                                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
1710                                 }
1711                             }
1712                         }
1713                     }
1714                 }
1715             }else{
1716                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1717                 if(is_h264){
1718                     if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX] ]){
1719                         if(!transform_bypass)
1720                             h->h264dsp.h264_luma_dc_dequant_idct(h->mb, h->mb_luma_dc, h->dequant4_coeff[0][s->qscale][0]);
1721                         else{
1722                             static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1723                                                                     8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1724                             for(i = 0; i < 16; i++)
1725                                 dctcoef_set(h->mb, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc, pixel_shift, i));
1726                         }
1727                     }
1728                 }else
1729                     ff_svq3_luma_dc_dequant_idct_c(h->mb, h->mb_luma_dc, s->qscale);
1730             }
1731             if(h->deblocking_filter)
1732                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple, pixel_shift);
1733         }else if(is_h264){
1734             if (pixel_shift) {
1735                 hl_motion_16(h, dest_y, dest_cb, dest_cr,
1736                              s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1737                              s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1738                              h->h264dsp.weight_h264_pixels_tab,
1739                              h->h264dsp.biweight_h264_pixels_tab);
1740             } else
1741                 hl_motion_8(h, dest_y, dest_cb, dest_cr,
1742                             s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1743                             s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1744                             h->h264dsp.weight_h264_pixels_tab,
1745                             h->h264dsp.biweight_h264_pixels_tab);
1746         }
1747
1748
1749         if(!IS_INTRA4x4(mb_type)){
1750             if(is_h264){
1751                 if(IS_INTRA16x16(mb_type)){
1752                     if(transform_bypass){
1753                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1754                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
1755                         }else{
1756                             for(i=0; i<16; i++){
1757                                 if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1758                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16 << pixel_shift), linesize);
1759                             }
1760                         }
1761                     }else{
1762                          h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1763                     }
1764                 }else if(h->cbp&15){
1765                     if(transform_bypass){
1766                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1767                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1768                         for(i=0; i<16; i+=di){
1769                             if(h->non_zero_count_cache[ scan8[i] ]){
1770                                 idct_add(dest_y + block_offset[i], h->mb + (i*16 << pixel_shift), linesize);
1771                             }
1772                         }
1773                     }else{
1774                         if(IS_8x8DCT(mb_type)){
1775                             h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1776                         }else{
1777                             h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1778                         }
1779                     }
1780                 }
1781             }else{
1782                 for(i=0; i<16; i++){
1783                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
1784                         uint8_t * const ptr= dest_y + block_offset[i];
1785                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1786                     }
1787                 }
1788             }
1789         }
1790
1791         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1792             uint8_t *dest[2] = {dest_cb, dest_cr};
1793             if(transform_bypass){
1794                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1795                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16 << pixel_shift), uvlinesize);
1796                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + (20*16 << pixel_shift), uvlinesize);
1797                 }else{
1798                     idct_add = s->dsp.add_pixels4;
1799                     for(i=16; i<16+8; i++){
1800                         if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1801                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1802                     }
1803                 }
1804             }else{
1805                 if(is_h264){
1806                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1807                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16 << pixel_shift)       , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1808                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1809                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + ((16*16+4*16) << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1810                     h->h264dsp.h264_idct_add8(dest, block_offset,
1811                                               h->mb, uvlinesize,
1812                                               h->non_zero_count_cache);
1813                 }else{
1814                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16     , h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1815                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16+4*16, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1816                     for(i=16; i<16+8; i++){
1817                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1818                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
1819                             ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1820                         }
1821                     }
1822                 }
1823             }
1824         }
1825     }
1826     if(h->cbp || IS_INTRA(mb_type))
1827         s->dsp.clear_blocks(h->mb);
1828 }
1829
1830 /**
1831  * Process a macroblock; this case avoids checks for expensive uncommon cases.
1832  */
1833 #define hl_decode_mb_simple(sh, bits) \
1834 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
1835     hl_decode_mb_internal(h, 1, sh); \
1836 }
1837 hl_decode_mb_simple(0, 8);
1838 hl_decode_mb_simple(1, 16);
1839
1840 /**
1841  * Process a macroblock; this handles edge cases, such as interlacing.
1842  */
1843 static void av_noinline hl_decode_mb_complex(H264Context *h){
1844     hl_decode_mb_internal(h, 0, h->pixel_shift);
1845 }
1846
1847 void ff_h264_hl_decode_mb(H264Context *h){
1848     MpegEncContext * const s = &h->s;
1849     const int mb_xy= h->mb_xy;
1850     const int mb_type= s->current_picture.mb_type[mb_xy];
1851     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1852
1853     if (is_complex) {
1854         hl_decode_mb_complex(h);
1855     } else if (h->pixel_shift) {
1856         hl_decode_mb_simple_16(h);
1857     } else
1858         hl_decode_mb_simple_8(h);
1859 }
1860
1861 static int pred_weight_table(H264Context *h){
1862     MpegEncContext * const s = &h->s;
1863     int list, i;
1864     int luma_def, chroma_def;
1865
1866     h->use_weight= 0;
1867     h->use_weight_chroma= 0;
1868     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
1869     if(CHROMA)
1870         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
1871     luma_def = 1<<h->luma_log2_weight_denom;
1872     chroma_def = 1<<h->chroma_log2_weight_denom;
1873
1874     for(list=0; list<2; list++){
1875         h->luma_weight_flag[list]   = 0;
1876         h->chroma_weight_flag[list] = 0;
1877         for(i=0; i<h->ref_count[list]; i++){
1878             int luma_weight_flag, chroma_weight_flag;
1879
1880             luma_weight_flag= get_bits1(&s->gb);
1881             if(luma_weight_flag){
1882                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
1883                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
1884                 if(   h->luma_weight[i][list][0] != luma_def
1885                    || h->luma_weight[i][list][1] != 0) {
1886                     h->use_weight= 1;
1887                     h->luma_weight_flag[list]= 1;
1888                 }
1889             }else{
1890                 h->luma_weight[i][list][0]= luma_def;
1891                 h->luma_weight[i][list][1]= 0;
1892             }
1893
1894             if(CHROMA){
1895                 chroma_weight_flag= get_bits1(&s->gb);
1896                 if(chroma_weight_flag){
1897                     int j;
1898                     for(j=0; j<2; j++){
1899                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
1900                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
1901                         if(   h->chroma_weight[i][list][j][0] != chroma_def
1902                            || h->chroma_weight[i][list][j][1] != 0) {
1903                             h->use_weight_chroma= 1;
1904                             h->chroma_weight_flag[list]= 1;
1905                         }
1906                     }
1907                 }else{
1908                     int j;
1909                     for(j=0; j<2; j++){
1910                         h->chroma_weight[i][list][j][0]= chroma_def;
1911                         h->chroma_weight[i][list][j][1]= 0;
1912                     }
1913                 }
1914             }
1915         }
1916         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
1917     }
1918     h->use_weight= h->use_weight || h->use_weight_chroma;
1919     return 0;
1920 }
1921
1922 /**
1923  * Initialize implicit_weight table.
1924  * @param field  0/1 initialize the weight for interlaced MBAFF
1925  *                -1 initializes the rest
1926  */
1927 static void implicit_weight_table(H264Context *h, int field){
1928     MpegEncContext * const s = &h->s;
1929     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
1930
1931     for (i = 0; i < 2; i++) {
1932         h->luma_weight_flag[i]   = 0;
1933         h->chroma_weight_flag[i] = 0;
1934     }
1935
1936     if(field < 0){
1937         cur_poc = s->current_picture_ptr->poc;
1938     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
1939        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
1940         h->use_weight= 0;
1941         h->use_weight_chroma= 0;
1942         return;
1943     }
1944         ref_start= 0;
1945         ref_count0= h->ref_count[0];
1946         ref_count1= h->ref_count[1];
1947     }else{
1948         cur_poc = s->current_picture_ptr->field_poc[field];
1949         ref_start= 16;
1950         ref_count0= 16+2*h->ref_count[0];
1951         ref_count1= 16+2*h->ref_count[1];
1952     }
1953
1954     h->use_weight= 2;
1955     h->use_weight_chroma= 2;
1956     h->luma_log2_weight_denom= 5;
1957     h->chroma_log2_weight_denom= 5;
1958
1959     for(ref0=ref_start; ref0 < ref_count0; ref0++){
1960         int poc0 = h->ref_list[0][ref0].poc;
1961         for(ref1=ref_start; ref1 < ref_count1; ref1++){
1962             int poc1 = h->ref_list[1][ref1].poc;
1963             int td = av_clip(poc1 - poc0, -128, 127);
1964             int w= 32;
1965             if(td){
1966                 int tb = av_clip(cur_poc - poc0, -128, 127);
1967                 int tx = (16384 + (FFABS(td) >> 1)) / td;
1968                 int dist_scale_factor = (tb*tx + 32) >> 8;
1969                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
1970                     w = 64 - dist_scale_factor;
1971             }
1972             if(field<0){
1973                 h->implicit_weight[ref0][ref1][0]=
1974                 h->implicit_weight[ref0][ref1][1]= w;
1975             }else{
1976                 h->implicit_weight[ref0][ref1][field]=w;
1977             }
1978         }
1979     }
1980 }
1981
1982 /**
1983  * instantaneous decoder refresh.
1984  */
1985 static void idr(H264Context *h){
1986     ff_h264_remove_all_refs(h);
1987     h->prev_frame_num= 0;
1988     h->prev_frame_num_offset= 0;
1989     h->prev_poc_msb=
1990     h->prev_poc_lsb= 0;
1991 }
1992
1993 /* forget old pics after a seek */
1994 static void flush_dpb(AVCodecContext *avctx){
1995     H264Context *h= avctx->priv_data;
1996     int i;
1997     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
1998         if(h->delayed_pic[i])
1999             h->delayed_pic[i]->reference= 0;
2000         h->delayed_pic[i]= NULL;
2001     }
2002     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2003     h->prev_interlaced_frame = 1;
2004     idr(h);
2005     if(h->s.current_picture_ptr)
2006         h->s.current_picture_ptr->reference= 0;
2007     h->s.first_field= 0;
2008     ff_h264_reset_sei(h);
2009     ff_mpeg_flush(avctx);
2010 }
2011
2012 static int init_poc(H264Context *h){
2013     MpegEncContext * const s = &h->s;
2014     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2015     int field_poc[2];
2016     Picture *cur = s->current_picture_ptr;
2017
2018     h->frame_num_offset= h->prev_frame_num_offset;
2019     if(h->frame_num < h->prev_frame_num)
2020         h->frame_num_offset += max_frame_num;
2021
2022     if(h->sps.poc_type==0){
2023         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2024
2025         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2026             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2027         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2028             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2029         else
2030             h->poc_msb = h->prev_poc_msb;
2031 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2032         field_poc[0] =
2033         field_poc[1] = h->poc_msb + h->poc_lsb;
2034         if(s->picture_structure == PICT_FRAME)
2035             field_poc[1] += h->delta_poc_bottom;
2036     }else if(h->sps.poc_type==1){
2037         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2038         int i;
2039
2040         if(h->sps.poc_cycle_length != 0)
2041             abs_frame_num = h->frame_num_offset + h->frame_num;
2042         else
2043             abs_frame_num = 0;
2044
2045         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2046             abs_frame_num--;
2047
2048         expected_delta_per_poc_cycle = 0;
2049         for(i=0; i < h->sps.poc_cycle_length; i++)
2050             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2051
2052         if(abs_frame_num > 0){
2053             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2054             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2055
2056             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2057             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2058                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2059         } else
2060             expectedpoc = 0;
2061
2062         if(h->nal_ref_idc == 0)
2063             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2064
2065         field_poc[0] = expectedpoc + h->delta_poc[0];
2066         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2067
2068         if(s->picture_structure == PICT_FRAME)
2069             field_poc[1] += h->delta_poc[1];
2070     }else{
2071         int poc= 2*(h->frame_num_offset + h->frame_num);
2072
2073         if(!h->nal_ref_idc)
2074             poc--;
2075
2076         field_poc[0]= poc;
2077         field_poc[1]= poc;
2078     }
2079
2080     if(s->picture_structure != PICT_BOTTOM_FIELD)
2081         s->current_picture_ptr->field_poc[0]= field_poc[0];
2082     if(s->picture_structure != PICT_TOP_FIELD)
2083         s->current_picture_ptr->field_poc[1]= field_poc[1];
2084     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2085
2086     return 0;
2087 }
2088
2089
2090 /**
2091  * initialize scan tables
2092  */
2093 static void init_scan_tables(H264Context *h){
2094     int i;
2095     for(i=0; i<16; i++){
2096 #define T(x) (x>>2) | ((x<<2) & 0xF)
2097         h->zigzag_scan[i] = T(zigzag_scan[i]);
2098         h-> field_scan[i] = T( field_scan[i]);
2099 #undef T
2100     }
2101     for(i=0; i<64; i++){
2102 #define T(x) (x>>3) | ((x&7)<<3)
2103         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2104         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2105         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2106         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2107 #undef T
2108     }
2109     if(h->sps.transform_bypass){ //FIXME same ugly
2110         h->zigzag_scan_q0          = zigzag_scan;
2111         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2112         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2113         h->field_scan_q0           = field_scan;
2114         h->field_scan8x8_q0        = field_scan8x8;
2115         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2116     }else{
2117         h->zigzag_scan_q0          = h->zigzag_scan;
2118         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2119         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2120         h->field_scan_q0           = h->field_scan;
2121         h->field_scan8x8_q0        = h->field_scan8x8;
2122         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2123     }
2124 }
2125
2126 static void field_end(H264Context *h, int in_setup){
2127     MpegEncContext * const s = &h->s;
2128     AVCodecContext * const avctx= s->avctx;
2129     s->mb_y= 0;
2130
2131     if (!in_setup && !s->dropable)
2132         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2133                                  s->picture_structure==PICT_BOTTOM_FIELD);
2134
2135     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2136         ff_vdpau_h264_set_reference_frames(s);
2137
2138     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2139         if(!s->dropable) {
2140             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2141             h->prev_poc_msb= h->poc_msb;
2142             h->prev_poc_lsb= h->poc_lsb;
2143         }
2144         h->prev_frame_num_offset= h->frame_num_offset;
2145         h->prev_frame_num= h->frame_num;
2146         h->outputed_poc = h->next_outputed_poc;
2147     }
2148
2149     if (avctx->hwaccel) {
2150         if (avctx->hwaccel->end_frame(avctx) < 0)
2151             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2152     }
2153
2154     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2155         ff_vdpau_h264_picture_complete(s);
2156
2157     /*
2158      * FIXME: Error handling code does not seem to support interlaced
2159      * when slices span multiple rows
2160      * The ff_er_add_slice calls don't work right for bottom
2161      * fields; they cause massive erroneous error concealing
2162      * Error marking covers both fields (top and bottom).
2163      * This causes a mismatched s->error_count
2164      * and a bad error table. Further, the error count goes to
2165      * INT_MAX when called for bottom field, because mb_y is
2166      * past end by one (callers fault) and resync_mb_y != 0
2167      * causes problems for the first MB line, too.
2168      */
2169     if (!FIELD_PICTURE)
2170         ff_er_frame_end(s);
2171
2172     MPV_frame_end(s);
2173
2174     h->current_slice=0;
2175 }
2176
2177 /**
2178  * Replicate H264 "master" context to thread contexts.
2179  */
2180 static void clone_slice(H264Context *dst, H264Context *src)
2181 {
2182     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2183     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2184     dst->s.current_picture      = src->s.current_picture;
2185     dst->s.linesize             = src->s.linesize;
2186     dst->s.uvlinesize           = src->s.uvlinesize;
2187     dst->s.first_field          = src->s.first_field;
2188
2189     dst->prev_poc_msb           = src->prev_poc_msb;
2190     dst->prev_poc_lsb           = src->prev_poc_lsb;
2191     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2192     dst->prev_frame_num         = src->prev_frame_num;
2193     dst->short_ref_count        = src->short_ref_count;
2194
2195     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2196     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2197     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2198     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2199
2200     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2201     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2202 }
2203
2204 /**
2205  * computes profile from profile_idc and constraint_set?_flags
2206  *
2207  * @param sps SPS
2208  *
2209  * @return profile as defined by FF_PROFILE_H264_*
2210  */
2211 int ff_h264_get_profile(SPS *sps)
2212 {
2213     int profile = sps->profile_idc;
2214
2215     switch(sps->profile_idc) {
2216     case FF_PROFILE_H264_BASELINE:
2217         // constraint_set1_flag set to 1
2218         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2219         break;
2220     case FF_PROFILE_H264_HIGH_10:
2221     case FF_PROFILE_H264_HIGH_422:
2222     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2223         // constraint_set3_flag set to 1
2224         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2225         break;
2226     }
2227
2228     return profile;
2229 }
2230
2231 /**
2232  * decodes a slice header.
2233  * This will also call MPV_common_init() and frame_start() as needed.
2234  *
2235  * @param h h264context
2236  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2237  *
2238  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2239  */
2240 static int decode_slice_header(H264Context *h, H264Context *h0){
2241     MpegEncContext * const s = &h->s;
2242     MpegEncContext * const s0 = &h0->s;
2243     unsigned int first_mb_in_slice;
2244     unsigned int pps_id;
2245     int num_ref_idx_active_override_flag;
2246     unsigned int slice_type, tmp, i, j;
2247     int default_ref_list_done = 0;
2248     int last_pic_structure;
2249
2250     s->dropable= h->nal_ref_idc == 0;
2251
2252     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
2253         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2254         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2255     }else{
2256         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2257         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2258     }
2259
2260     first_mb_in_slice= get_ue_golomb(&s->gb);
2261
2262     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2263         if(h0->current_slice && FIELD_PICTURE){
2264             field_end(h, 1);
2265         }
2266
2267         h0->current_slice = 0;
2268         if (!s0->first_field)
2269             s->current_picture_ptr= NULL;
2270     }
2271
2272     slice_type= get_ue_golomb_31(&s->gb);
2273     if(slice_type > 9){
2274         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);
2275         return -1;
2276     }
2277     if(slice_type > 4){
2278         slice_type -= 5;
2279         h->slice_type_fixed=1;
2280     }else
2281         h->slice_type_fixed=0;
2282
2283     slice_type= golomb_to_pict_type[ slice_type ];
2284     if (slice_type == AV_PICTURE_TYPE_I
2285         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2286         default_ref_list_done = 1;
2287     }
2288     h->slice_type= slice_type;
2289     h->slice_type_nos= slice_type & 3;
2290
2291     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2292
2293     pps_id= get_ue_golomb(&s->gb);
2294     if(pps_id>=MAX_PPS_COUNT){
2295         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2296         return -1;
2297     }
2298     if(!h0->pps_buffers[pps_id]) {
2299         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2300         return -1;
2301     }
2302     h->pps= *h0->pps_buffers[pps_id];
2303
2304     if(!h0->sps_buffers[h->pps.sps_id]) {
2305         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2306         return -1;
2307     }
2308     h->sps = *h0->sps_buffers[h->pps.sps_id];
2309
2310     s->avctx->profile = ff_h264_get_profile(&h->sps);
2311     s->avctx->level   = h->sps.level_idc;
2312     s->avctx->refs    = h->sps.ref_frame_count;
2313
2314     if(h == h0 && h->dequant_coeff_pps != pps_id){
2315         h->dequant_coeff_pps = pps_id;
2316         init_dequant_tables(h);
2317     }
2318
2319     s->mb_width= h->sps.mb_width;
2320     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2321
2322     h->b_stride=  s->mb_width*4;
2323
2324     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
2325     if(h->sps.frame_mbs_only_flag)
2326         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
2327     else
2328         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 7);
2329
2330     if (s->context_initialized
2331         && (   s->width != s->avctx->width || s->height != s->avctx->height
2332             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2333         if(h != h0) {
2334             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2335             return -1;   // width / height changed during parallelized decoding
2336         }
2337         free_tables(h, 0);
2338         flush_dpb(s->avctx);
2339         MPV_common_end(s);
2340     }
2341     if (!s->context_initialized) {
2342         if(h != h0)
2343             return -1;  // we cant (re-)initialize context during parallel decoding
2344
2345         avcodec_set_dimensions(s->avctx, s->width, s->height);
2346         s->avctx->sample_aspect_ratio= h->sps.sar;
2347         av_assert0(s->avctx->sample_aspect_ratio.den);
2348
2349         h->s.avctx->coded_width = 16*s->mb_width;
2350         h->s.avctx->coded_height = 16*s->mb_height;
2351
2352         if(h->sps.video_signal_type_present_flag){
2353             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2354             if(h->sps.colour_description_present_flag){
2355                 s->avctx->color_primaries = h->sps.color_primaries;
2356                 s->avctx->color_trc       = h->sps.color_trc;
2357                 s->avctx->colorspace      = h->sps.colorspace;
2358             }
2359         }
2360
2361         if(h->sps.timing_info_present_flag){
2362             int64_t den= h->sps.time_scale;
2363             if(h->x264_build < 44U)
2364                 den *= 2;
2365             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2366                       h->sps.num_units_in_tick, den, 1<<30);
2367         }
2368
2369         switch (h->sps.bit_depth_luma) {
2370             case 9 :
2371                 s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2372                 break;
2373             case 10 :
2374                 s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2375                 break;
2376             default:
2377         s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2378                                                  s->avctx->codec->pix_fmts ?
2379                                                  s->avctx->codec->pix_fmts :
2380                                                  s->avctx->color_range == AVCOL_RANGE_JPEG ?
2381                                                  hwaccel_pixfmt_list_h264_jpeg_420 :
2382                                                  ff_hwaccel_pixfmt_list_420);
2383         }
2384
2385         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2386
2387         if (MPV_common_init(s) < 0)
2388             return -1;
2389         s->first_field = 0;
2390         h->prev_interlaced_frame = 1;
2391
2392         init_scan_tables(h);
2393         ff_h264_alloc_tables(h);
2394
2395         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2396             if (context_init(h) < 0)
2397                 return -1;
2398         } else {
2399             for(i = 1; i < s->avctx->thread_count; i++) {
2400                 H264Context *c;
2401                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2402                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2403                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2404                 c->h264dsp = h->h264dsp;
2405                 c->sps = h->sps;
2406                 c->pps = h->pps;
2407                 c->pixel_shift = h->pixel_shift;
2408                 init_scan_tables(c);
2409                 clone_tables(c, h, i);
2410             }
2411
2412             for(i = 0; i < s->avctx->thread_count; i++)
2413                 if(context_init(h->thread_context[i]) < 0)
2414                     return -1;
2415         }
2416     }
2417
2418     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2419
2420     h->mb_mbaff = 0;
2421     h->mb_aff_frame = 0;
2422     last_pic_structure = s0->picture_structure;
2423     if(h->sps.frame_mbs_only_flag){
2424         s->picture_structure= PICT_FRAME;
2425     }else{
2426         if(get_bits1(&s->gb)) { //field_pic_flag
2427             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2428         } else {
2429             s->picture_structure= PICT_FRAME;
2430             h->mb_aff_frame = h->sps.mb_aff;
2431         }
2432     }
2433     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2434
2435     if(h0->current_slice == 0){
2436         if(h->frame_num != h->prev_frame_num &&
2437           (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num) < (h->frame_num - h->sps.ref_frame_count))
2438             h->prev_frame_num = h->frame_num - h->sps.ref_frame_count - 1;
2439
2440         while(h->frame_num !=  h->prev_frame_num &&
2441               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2442             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2443             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2444             if (ff_h264_frame_start(h) < 0)
2445                 return -1;
2446             h->prev_frame_num++;
2447             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2448             s->current_picture_ptr->frame_num= h->prev_frame_num;
2449             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2450             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2451             ff_generate_sliding_window_mmcos(h);
2452             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2453             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2454              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2455              * about there being no actual duplicates.
2456              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2457              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2458              * be fixed. */
2459             if (h->short_ref_count) {
2460                 if (prev) {
2461                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
2462                                   (const uint8_t**)prev->data, prev->linesize,
2463                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2464                     h->short_ref[0]->poc = prev->poc+2;
2465                 }
2466                 h->short_ref[0]->frame_num = h->prev_frame_num;
2467             }
2468         }
2469
2470         /* See if we have a decoded first field looking for a pair... */
2471         if (s0->first_field) {
2472             assert(s0->current_picture_ptr);
2473             assert(s0->current_picture_ptr->data[0]);
2474             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2475
2476             /* figure out if we have a complementary field pair */
2477             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2478                 /*
2479                  * Previous field is unmatched. Don't display it, but let it
2480                  * remain for reference if marked as such.
2481                  */
2482                 s0->current_picture_ptr = NULL;
2483                 s0->first_field = FIELD_PICTURE;
2484
2485             } else {
2486                 if (h->nal_ref_idc &&
2487                         s0->current_picture_ptr->reference &&
2488                         s0->current_picture_ptr->frame_num != h->frame_num) {
2489                     /*
2490                      * This and previous field were reference, but had
2491                      * different frame_nums. Consider this field first in
2492                      * pair. Throw away previous field except for reference
2493                      * purposes.
2494                      */
2495                     s0->first_field = 1;
2496                     s0->current_picture_ptr = NULL;
2497
2498                 } else {
2499                     /* Second field in complementary pair */
2500                     s0->first_field = 0;
2501                 }
2502             }
2503
2504         } else {
2505             /* Frame or first field in a potentially complementary pair */
2506             assert(!s0->current_picture_ptr);
2507             s0->first_field = FIELD_PICTURE;
2508         }
2509
2510         if(!FIELD_PICTURE || s0->first_field) {
2511             if (ff_h264_frame_start(h) < 0) {
2512                 s0->first_field = 0;
2513                 return -1;
2514             }
2515         } else {
2516             ff_release_unused_pictures(s, 0);
2517         }
2518     }
2519     if(h != h0)
2520         clone_slice(h, h0);
2521
2522     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2523
2524     assert(s->mb_num == s->mb_width * s->mb_height);
2525     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2526        first_mb_in_slice                    >= s->mb_num){
2527         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2528         return -1;
2529     }
2530     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2531     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2532     if (s->picture_structure == PICT_BOTTOM_FIELD)
2533         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2534     assert(s->mb_y < s->mb_height);
2535
2536     if(s->picture_structure==PICT_FRAME){
2537         h->curr_pic_num=   h->frame_num;
2538         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2539     }else{
2540         h->curr_pic_num= 2*h->frame_num + 1;
2541         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2542     }
2543
2544     if(h->nal_unit_type == NAL_IDR_SLICE){
2545         get_ue_golomb(&s->gb); /* idr_pic_id */
2546     }
2547
2548     if(h->sps.poc_type==0){
2549         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2550
2551         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2552             h->delta_poc_bottom= get_se_golomb(&s->gb);
2553         }
2554     }
2555
2556     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2557         h->delta_poc[0]= get_se_golomb(&s->gb);
2558
2559         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2560             h->delta_poc[1]= get_se_golomb(&s->gb);
2561     }
2562
2563     init_poc(h);
2564
2565     if(h->pps.redundant_pic_cnt_present){
2566         h->redundant_pic_count= get_ue_golomb(&s->gb);
2567     }
2568
2569     //set defaults, might be overridden a few lines later
2570     h->ref_count[0]= h->pps.ref_count[0];
2571     h->ref_count[1]= h->pps.ref_count[1];
2572
2573     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2574         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2575             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2576         }
2577         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2578
2579         if(num_ref_idx_active_override_flag){
2580             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2581             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2582                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2583
2584             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2585                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2586                 h->ref_count[0]= h->ref_count[1]= 1;
2587                 return -1;
2588             }
2589         }
2590         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2591             h->list_count= 2;
2592         else
2593             h->list_count= 1;
2594     }else
2595         h->list_count= 0;
2596
2597     if(!default_ref_list_done){
2598         ff_h264_fill_default_ref_list(h);
2599     }
2600
2601     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2602         return -1;
2603
2604     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2605         s->last_picture_ptr= &h->ref_list[0][0];
2606         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2607     }
2608     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2609         s->next_picture_ptr= &h->ref_list[1][0];
2610         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2611     }
2612
2613     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2614        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2615         pred_weight_table(h);
2616     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2617         implicit_weight_table(h, -1);
2618     }else {
2619         h->use_weight = 0;
2620         for (i = 0; i < 2; i++) {
2621             h->luma_weight_flag[i]   = 0;
2622             h->chroma_weight_flag[i] = 0;
2623         }
2624     }
2625
2626     if(h->nal_ref_idc)
2627         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2628
2629     if(FRAME_MBAFF){
2630         ff_h264_fill_mbaff_ref_list(h);
2631
2632         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2633             implicit_weight_table(h, 0);
2634             implicit_weight_table(h, 1);
2635         }
2636     }
2637
2638     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2639         ff_h264_direct_dist_scale_factor(h);
2640     ff_h264_direct_ref_list_init(h);
2641
2642     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2643         tmp = get_ue_golomb_31(&s->gb);
2644         if(tmp > 2){
2645             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2646             return -1;
2647         }
2648         h->cabac_init_idc= tmp;
2649     }
2650
2651     h->last_qscale_diff = 0;
2652     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2653     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2654         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2655         return -1;
2656     }
2657     s->qscale= tmp;
2658     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2659     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2660     //FIXME qscale / qp ... stuff
2661     if(h->slice_type == AV_PICTURE_TYPE_SP){
2662         get_bits1(&s->gb); /* sp_for_switch_flag */
2663     }
2664     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2665         get_se_golomb(&s->gb); /* slice_qs_delta */
2666     }
2667
2668     h->deblocking_filter = 1;
2669     h->slice_alpha_c0_offset = 52;
2670     h->slice_beta_offset = 52;
2671     if( h->pps.deblocking_filter_parameters_present ) {
2672         tmp= get_ue_golomb_31(&s->gb);
2673         if(tmp > 2){
2674             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2675             return -1;
2676         }
2677         h->deblocking_filter= tmp;
2678         if(h->deblocking_filter < 2)
2679             h->deblocking_filter^= 1; // 1<->0
2680
2681         if( h->deblocking_filter ) {
2682             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2683             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2684             if(   h->slice_alpha_c0_offset > 104U
2685                || h->slice_beta_offset     > 104U){
2686                 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);
2687                 return -1;
2688             }
2689         }
2690     }
2691
2692     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2693        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2694        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2695        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2696         h->deblocking_filter= 0;
2697
2698     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2699         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2700             /* Cheat slightly for speed:
2701                Do not bother to deblock across slices. */
2702             h->deblocking_filter = 2;
2703         } else {
2704             h0->max_contexts = 1;
2705             if(!h0->single_decode_warning) {
2706                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2707                 h0->single_decode_warning = 1;
2708             }
2709             if(h != h0)
2710                 return 1; // deblocking switched inside frame
2711         }
2712     }
2713     h->qp_thresh= 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1]);
2714
2715 #if 0 //FMO
2716     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2717         slice_group_change_cycle= get_bits(&s->gb, ?);
2718 #endif
2719
2720     h0->last_slice_type = slice_type;
2721     h->slice_num = ++h0->current_slice;
2722     if(h->slice_num >= MAX_SLICES){
2723         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2724     }
2725
2726     for(j=0; j<2; j++){
2727         int id_list[16];
2728         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2729         for(i=0; i<16; i++){
2730             id_list[i]= 60;
2731             if(h->ref_list[j][i].data[0]){
2732                 int k;
2733                 uint8_t *base= h->ref_list[j][i].base[0];
2734                 for(k=0; k<h->short_ref_count; k++)
2735                     if(h->short_ref[k]->base[0] == base){
2736                         id_list[i]= k;
2737                         break;
2738                     }
2739                 for(k=0; k<h->long_ref_count; k++)
2740                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2741                         id_list[i]= h->short_ref_count + k;
2742                         break;
2743                     }
2744             }
2745         }
2746
2747         ref2frm[0]=
2748         ref2frm[1]= -1;
2749         for(i=0; i<16; i++)
2750             ref2frm[i+2]= 4*id_list[i]
2751                           +(h->ref_list[j][i].reference&3);
2752         ref2frm[18+0]=
2753         ref2frm[18+1]= -1;
2754         for(i=16; i<48; i++)
2755             ref2frm[i+4]= 4*id_list[(i-16)>>1]
2756                           +(h->ref_list[j][i].reference&3);
2757     }
2758
2759     //FIXME: fix draw_edges+PAFF+frame threads
2760     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
2761     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
2762
2763     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
2764         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",
2765                h->slice_num,
2766                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
2767                first_mb_in_slice,
2768                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
2769                pps_id, h->frame_num,
2770                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
2771                h->ref_count[0], h->ref_count[1],
2772                s->qscale,
2773                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
2774                h->use_weight,
2775                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
2776                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
2777                );
2778     }
2779
2780     return 0;
2781 }
2782
2783 int ff_h264_get_slice_type(const H264Context *h)
2784 {
2785     switch (h->slice_type) {
2786     case AV_PICTURE_TYPE_P:  return 0;
2787     case AV_PICTURE_TYPE_B:  return 1;
2788     case AV_PICTURE_TYPE_I:  return 2;
2789     case AV_PICTURE_TYPE_SP: return 3;
2790     case AV_PICTURE_TYPE_SI: return 4;
2791     default:         return -1;
2792     }
2793 }
2794
2795 /**
2796  *
2797  * @return non zero if the loop filter can be skiped
2798  */
2799 static int fill_filter_caches(H264Context *h, int mb_type){
2800     MpegEncContext * const s = &h->s;
2801     const int mb_xy= h->mb_xy;
2802     int top_xy, left_xy[2];
2803     int top_type, left_type[2];
2804
2805     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
2806
2807     //FIXME deblocking could skip the intra and nnz parts.
2808
2809     /* Wow, what a mess, why didn't they simplify the interlacing & intra
2810      * stuff, I can't imagine that these complex rules are worth it. */
2811
2812     left_xy[1] = left_xy[0] = mb_xy-1;
2813     if(FRAME_MBAFF){
2814         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
2815         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
2816         if(s->mb_y&1){
2817             if (left_mb_field_flag != curr_mb_field_flag) {
2818                 left_xy[0] -= s->mb_stride;
2819             }
2820         }else{
2821             if(curr_mb_field_flag){
2822                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
2823             }
2824             if (left_mb_field_flag != curr_mb_field_flag) {
2825                 left_xy[1] += s->mb_stride;
2826             }
2827         }
2828     }
2829
2830     h->top_mb_xy = top_xy;
2831     h->left_mb_xy[0] = left_xy[0];
2832     h->left_mb_xy[1] = left_xy[1];
2833     {
2834         //for sufficiently low qp, filtering wouldn't do anything
2835         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
2836         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
2837         int qp = s->current_picture.qscale_table[mb_xy];
2838         if(qp <= qp_thresh
2839            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
2840            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
2841             if(!FRAME_MBAFF)
2842                 return 1;
2843             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
2844                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
2845                 return 1;
2846         }
2847     }
2848
2849     top_type     = s->current_picture.mb_type[top_xy]    ;
2850     left_type[0] = s->current_picture.mb_type[left_xy[0]];
2851     left_type[1] = s->current_picture.mb_type[left_xy[1]];
2852     if(h->deblocking_filter == 2){
2853         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
2854         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
2855     }else{
2856         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
2857         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
2858     }
2859     h->top_type    = top_type    ;
2860     h->left_type[0]= left_type[0];
2861     h->left_type[1]= left_type[1];
2862
2863     if(IS_INTRA(mb_type))
2864         return 0;
2865
2866     AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
2867     AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
2868     AV_COPY32(&h->non_zero_count_cache[0+8*5], &h->non_zero_count[mb_xy][16]);
2869     AV_COPY32(&h->non_zero_count_cache[4+8*3], &h->non_zero_count[mb_xy][20]);
2870     AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
2871
2872     h->cbp= h->cbp_table[mb_xy];
2873
2874     {
2875         int list;
2876         for(list=0; list<h->list_count; list++){
2877             int8_t *ref;
2878             int y, b_stride;
2879             int16_t (*mv_dst)[2];
2880             int16_t (*mv_src)[2];
2881
2882             if(!USES_LIST(mb_type, list)){
2883                 fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
2884                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2885                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2886                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2887                 AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2888                 continue;
2889             }
2890
2891             ref = &s->current_picture.ref_index[list][4*mb_xy];
2892             {
2893                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2894                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2895                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2896                 ref += 2;
2897                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2898                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2899             }
2900
2901             b_stride = h->b_stride;
2902             mv_dst   = &h->mv_cache[list][scan8[0]];
2903             mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
2904             for(y=0; y<4; y++){
2905                 AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
2906             }
2907
2908         }
2909     }
2910
2911
2912 /*
2913 0 . T T. T T T T
2914 1 L . .L . . . .
2915 2 L . .L . . . .
2916 3 . T TL . . . .
2917 4 L . .L . . . .
2918 5 L . .. . . . .
2919 */
2920 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
2921     if(top_type){
2922         AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
2923     }
2924
2925     if(left_type[0]){
2926         h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][7+0*8];
2927         h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][7+1*8];
2928         h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][7+2*8];
2929         h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][7+3*8];
2930     }
2931
2932     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
2933     if(!CABAC && h->pps.transform_8x8_mode){
2934         if(IS_8x8DCT(top_type)){
2935             h->non_zero_count_cache[4+8*0]=
2936             h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
2937             h->non_zero_count_cache[6+8*0]=
2938             h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
2939         }
2940         if(IS_8x8DCT(left_type[0])){
2941             h->non_zero_count_cache[3+8*1]=
2942             h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
2943         }
2944         if(IS_8x8DCT(left_type[1])){
2945             h->non_zero_count_cache[3+8*3]=
2946             h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
2947         }
2948
2949         if(IS_8x8DCT(mb_type)){
2950             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
2951             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
2952
2953             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
2954             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
2955
2956             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
2957             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
2958
2959             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
2960             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
2961         }
2962     }
2963
2964     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
2965         int list;
2966         for(list=0; list<h->list_count; list++){
2967             if(USES_LIST(top_type, list)){
2968                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
2969                 const int b8_xy= 4*top_xy + 2;
2970                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2971                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
2972                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
2973                 h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
2974                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
2975                 h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
2976             }else{
2977                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
2978                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2979             }
2980
2981             if(!IS_INTERLACED(mb_type^left_type[0])){
2982                 if(USES_LIST(left_type[0], list)){
2983                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
2984                     const int b8_xy= 4*left_xy[0] + 1;
2985                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2986                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
2987                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
2988                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
2989                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
2990                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
2991                     h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
2992                     h->ref_cache[list][scan8[0] - 1 +16 ]=
2993                     h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
2994                 }else{
2995                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
2996                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
2997                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
2998                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
2999                     h->ref_cache[list][scan8[0] - 1 + 0  ]=
3000                     h->ref_cache[list][scan8[0] - 1 + 8  ]=
3001                     h->ref_cache[list][scan8[0] - 1 + 16 ]=
3002                     h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
3003                 }
3004             }
3005         }
3006     }
3007
3008     return 0;
3009 }
3010
3011 static void loop_filter(H264Context *h, int start_x, int end_x){
3012     MpegEncContext * const s = &h->s;
3013     uint8_t  *dest_y, *dest_cb, *dest_cr;
3014     int linesize, uvlinesize, mb_x, mb_y;
3015     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3016     const int old_slice_type= h->slice_type;
3017     const int pixel_shift = h->pixel_shift;
3018
3019     if(h->deblocking_filter) {
3020         for(mb_x= start_x; mb_x<end_x; mb_x++){
3021             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3022                 int mb_xy, mb_type;
3023                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3024                 h->slice_num= h->slice_table[mb_xy];
3025                 mb_type= s->current_picture.mb_type[mb_xy];
3026                 h->list_count= h->list_counts[mb_xy];
3027
3028                 if(FRAME_MBAFF)
3029                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3030
3031                 s->mb_x= mb_x;
3032                 s->mb_y= mb_y;
3033                 dest_y  = s->current_picture.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3034                 dest_cb = s->current_picture.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
3035                 dest_cr = s->current_picture.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * 8;
3036                     //FIXME simplify above
3037
3038                 if (MB_FIELD) {
3039                     linesize   = h->mb_linesize   = s->linesize * 2;
3040                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3041                     if(mb_y&1){ //FIXME move out of this function?
3042                         dest_y -= s->linesize*15;
3043                         dest_cb-= s->uvlinesize*7;
3044                         dest_cr-= s->uvlinesize*7;
3045                     }
3046                 } else {
3047                     linesize   = h->mb_linesize   = s->linesize;
3048                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3049                 }
3050                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3051                 if(fill_filter_caches(h, mb_type))
3052                     continue;
3053                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
3054                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
3055
3056                 if (FRAME_MBAFF) {
3057                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3058                 } else {
3059                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3060                 }
3061             }
3062         }
3063     }
3064     h->slice_type= old_slice_type;
3065     s->mb_x= end_x;
3066     s->mb_y= end_mb_y - FRAME_MBAFF;
3067     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3068     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3069 }
3070
3071 static void predict_field_decoding_flag(H264Context *h){
3072     MpegEncContext * const s = &h->s;
3073     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3074     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3075                 ? s->current_picture.mb_type[mb_xy-1]
3076                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3077                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
3078                 : 0;
3079     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3080 }
3081
3082 /**
3083  * Draw edges and report progress for the last MB row.
3084  */
3085 static void decode_finish_row(H264Context *h){
3086     MpegEncContext * const s = &h->s;
3087     int top = 16*(s->mb_y >> FIELD_PICTURE);
3088     int height = 16 << FRAME_MBAFF;
3089     int deblock_border = (16 + 4) << FRAME_MBAFF;
3090     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3091
3092     if (h->deblocking_filter) {
3093         if((top + height) >= pic_height)
3094             height += deblock_border;
3095
3096         top -= deblock_border;
3097     }
3098
3099     if (top >= pic_height || (top + height) < h->emu_edge_height)
3100         return;
3101
3102     height = FFMIN(height, pic_height - top);
3103     if (top < h->emu_edge_height) {
3104         height = top+height;
3105         top = 0;
3106     }
3107
3108     ff_draw_horiz_band(s, top, height);
3109
3110     if (s->dropable) return;
3111
3112     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3113                              s->picture_structure==PICT_BOTTOM_FIELD);
3114 }
3115
3116 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3117     H264Context *h = *(void**)arg;
3118     MpegEncContext * const s = &h->s;
3119     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3120     int lf_x_start = s->mb_x;
3121
3122     s->mb_skip_run= -1;
3123
3124     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3125                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3126
3127     if( h->pps.cabac ) {
3128         /* realign */
3129         align_get_bits( &s->gb );
3130
3131         /* init cabac */
3132         ff_init_cabac_states( &h->cabac);
3133         ff_init_cabac_decoder( &h->cabac,
3134                                s->gb.buffer + get_bits_count(&s->gb)/8,
3135                                (get_bits_left(&s->gb) + 7)/8);
3136
3137         ff_h264_init_cabac_states(h);
3138
3139         for(;;){
3140 //START_TIMER
3141             int ret = ff_h264_decode_mb_cabac(h);
3142             int eos;
3143 //STOP_TIMER("decode_mb_cabac")
3144
3145             if(ret>=0) ff_h264_hl_decode_mb(h);
3146
3147             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3148                 s->mb_y++;
3149
3150                 ret = ff_h264_decode_mb_cabac(h);
3151
3152                 if(ret>=0) ff_h264_hl_decode_mb(h);
3153                 s->mb_y--;
3154             }
3155             eos = get_cabac_terminate( &h->cabac );
3156
3157             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3158                 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);
3159                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3160                 return 0;
3161             }
3162             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3163                 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);
3164                 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);
3165                 return -1;
3166             }
3167
3168             if( ++s->mb_x >= s->mb_width ) {
3169                 loop_filter(h, lf_x_start, s->mb_x);
3170                 s->mb_x = lf_x_start = 0;
3171                 decode_finish_row(h);
3172                 ++s->mb_y;
3173                 if(FIELD_OR_MBAFF_PICTURE) {
3174                     ++s->mb_y;
3175                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3176                         predict_field_decoding_flag(h);
3177                 }
3178             }
3179
3180             if( eos || s->mb_y >= s->mb_height ) {
3181                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3182                 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);
3183                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3184                 return 0;
3185             }
3186         }
3187
3188     } else {
3189         for(;;){
3190             int ret = ff_h264_decode_mb_cavlc(h);
3191
3192             if(ret>=0) ff_h264_hl_decode_mb(h);
3193
3194             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3195                 s->mb_y++;
3196                 ret = ff_h264_decode_mb_cavlc(h);
3197
3198                 if(ret>=0) ff_h264_hl_decode_mb(h);
3199                 s->mb_y--;
3200             }
3201
3202             if(ret<0){
3203                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3204                 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);
3205                 return -1;
3206             }
3207
3208             if(++s->mb_x >= s->mb_width){
3209                 loop_filter(h, lf_x_start, s->mb_x);
3210                 s->mb_x = lf_x_start = 0;
3211                 decode_finish_row(h);
3212                 ++s->mb_y;
3213                 if(FIELD_OR_MBAFF_PICTURE) {
3214                     ++s->mb_y;
3215                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3216                         predict_field_decoding_flag(h);
3217                 }
3218                 if(s->mb_y >= s->mb_height){
3219                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3220
3221                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3222                         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);
3223
3224                         return 0;
3225                     }else{
3226                         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);
3227
3228                         return -1;
3229                     }
3230                 }
3231             }
3232
3233             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3234                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3235                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3236                     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);
3237                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3238
3239                     return 0;
3240                 }else{
3241                     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);
3242
3243                     return -1;
3244                 }
3245             }
3246         }
3247     }
3248
3249 #if 0
3250     for(;s->mb_y < s->mb_height; s->mb_y++){
3251         for(;s->mb_x < s->mb_width; s->mb_x++){
3252             int ret= decode_mb(h);
3253
3254             ff_h264_hl_decode_mb(h);
3255
3256             if(ret<0){
3257                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3258                 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);
3259
3260                 return -1;
3261             }
3262
3263             if(++s->mb_x >= s->mb_width){
3264                 s->mb_x=0;
3265                 if(++s->mb_y >= s->mb_height){
3266                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
3267                         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);
3268
3269                         return 0;
3270                     }else{
3271                         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);
3272
3273                         return -1;
3274                     }
3275                 }
3276             }
3277
3278             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
3279                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
3280                     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);
3281
3282                     return 0;
3283                 }else{
3284                     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);
3285
3286                     return -1;
3287                 }
3288             }
3289         }
3290         s->mb_x=0;
3291         ff_draw_horiz_band(s, 16*s->mb_y, 16);
3292     }
3293 #endif
3294     return -1; //not reached
3295 }
3296
3297 /**
3298  * Call decode_slice() for each context.
3299  *
3300  * @param h h264 master context
3301  * @param context_count number of contexts to execute
3302  */
3303 static void execute_decode_slices(H264Context *h, int context_count){
3304     MpegEncContext * const s = &h->s;
3305     AVCodecContext * const avctx= s->avctx;
3306     H264Context *hx;
3307     int i;
3308
3309     if (s->avctx->hwaccel)
3310         return;
3311     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3312         return;
3313     if(context_count == 1) {
3314         decode_slice(avctx, &h);
3315     } else {
3316         for(i = 1; i < context_count; i++) {
3317             hx = h->thread_context[i];
3318             hx->s.error_recognition = avctx->error_recognition;
3319             hx->s.error_count = 0;
3320         }
3321
3322         avctx->execute(avctx, (void *)decode_slice,
3323                        h->thread_context, NULL, context_count, sizeof(void*));
3324
3325         /* pull back stuff from slices to master context */
3326         hx = h->thread_context[context_count - 1];
3327         s->mb_x = hx->s.mb_x;
3328         s->mb_y = hx->s.mb_y;
3329         s->dropable = hx->s.dropable;
3330         s->picture_structure = hx->s.picture_structure;
3331         for(i = 1; i < context_count; i++)
3332             h->s.error_count += h->thread_context[i]->s.error_count;
3333     }
3334 }
3335
3336
3337 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3338     MpegEncContext * const s = &h->s;
3339     AVCodecContext * const avctx= s->avctx;
3340     H264Context *hx; ///< thread context
3341     int buf_index;
3342     int context_count;
3343     int next_avc;
3344     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3345     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3346     int nal_index;
3347
3348     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3349     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3350         h->current_slice = 0;
3351         if (!s->first_field)
3352             s->current_picture_ptr= NULL;
3353         ff_h264_reset_sei(h);
3354     }
3355
3356     for(;pass <= 1;pass++){
3357         buf_index = 0;
3358         context_count = 0;
3359         next_avc = h->is_avc ? 0 : buf_size;
3360         nal_index = 0;
3361     for(;;){
3362         int consumed;
3363         int dst_length;
3364         int bit_length;
3365         const uint8_t *ptr;
3366         int i, nalsize = 0;
3367         int err;
3368
3369         if(buf_index >= next_avc) {
3370             if(buf_index >= buf_size) break;
3371             nalsize = 0;
3372             for(i = 0; i < h->nal_length_size; i++)
3373                 nalsize = (nalsize << 8) | buf[buf_index++];
3374             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3375                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3376                 break;
3377             }
3378             next_avc= buf_index + nalsize;
3379         } else {
3380             // start code prefix search
3381             for(; buf_index + 3 < next_avc; buf_index++){
3382                 // This should always succeed in the first iteration.
3383                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3384                     break;
3385             }
3386
3387             if(buf_index+3 >= buf_size) break;
3388
3389             buf_index+=3;
3390             if(buf_index >= next_avc) continue;
3391         }
3392
3393         hx = h->thread_context[context_count];
3394
3395         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3396         if (ptr==NULL || dst_length < 0){
3397             return -1;
3398         }
3399         i= buf_index + consumed;
3400         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3401            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3402             s->workaround_bugs |= FF_BUG_TRUNCATED;
3403
3404         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3405         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3406             dst_length--;
3407         }
3408         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3409
3410         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3411             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
3412         }
3413
3414         if (h->is_avc && (nalsize != consumed) && nalsize){
3415             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3416         }
3417
3418         buf_index += consumed;
3419         nal_index++;
3420
3421         if(pass == 0) {
3422             // packets can sometimes contain multiple PPS/SPS
3423             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3424             // if so, when frame threading we can't start the next thread until we've read all of them
3425             switch (hx->nal_unit_type) {
3426                 case NAL_SPS:
3427                 case NAL_PPS:
3428                     nals_needed = nal_index;
3429             }
3430             continue;
3431         }
3432
3433         //FIXME do not discard SEI id
3434         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3435             continue;
3436
3437       again:
3438         err = 0;
3439         switch(hx->nal_unit_type){
3440         case NAL_IDR_SLICE:
3441             if (h->nal_unit_type != NAL_IDR_SLICE) {
3442                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3443                 return -1;
3444             }
3445             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3446         case NAL_SLICE:
3447             init_get_bits(&hx->s.gb, ptr, bit_length);
3448             hx->intra_gb_ptr=
3449             hx->inter_gb_ptr= &hx->s.gb;
3450             hx->s.data_partitioning = 0;
3451
3452             if((err = decode_slice_header(hx, h)))
3453                break;
3454
3455             s->current_picture_ptr->key_frame |=
3456                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3457                     (h->sei_recovery_frame_cnt >= 0);
3458
3459             if (h->current_slice == 1) {
3460                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3461                     decode_postinit(h, nal_index >= nals_needed);
3462                 }
3463
3464                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3465                     return -1;
3466                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3467                     ff_vdpau_h264_picture_start(s);
3468             }
3469
3470             if(hx->redundant_pic_count==0
3471                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3472                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3473                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3474                && avctx->skip_frame < AVDISCARD_ALL){
3475                 if(avctx->hwaccel) {
3476                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3477                         return -1;
3478                 }else
3479                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3480                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3481                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3482                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3483                 }else
3484                     context_count++;
3485             }
3486             break;
3487         case NAL_DPA:
3488             init_get_bits(&hx->s.gb, ptr, bit_length);
3489             hx->intra_gb_ptr=
3490             hx->inter_gb_ptr= NULL;
3491
3492             if ((err = decode_slice_header(hx, h)) < 0)
3493                 break;
3494
3495             hx->s.data_partitioning = 1;
3496
3497             break;
3498         case NAL_DPB:
3499             init_get_bits(&hx->intra_gb, ptr, bit_length);
3500             hx->intra_gb_ptr= &hx->intra_gb;
3501             break;
3502         case NAL_DPC:
3503             init_get_bits(&hx->inter_gb, ptr, bit_length);
3504             hx->inter_gb_ptr= &hx->inter_gb;
3505
3506             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3507                && s->context_initialized
3508                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3509                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3510                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3511                && avctx->skip_frame < AVDISCARD_ALL)
3512                 context_count++;
3513             break;
3514         case NAL_SEI:
3515             init_get_bits(&s->gb, ptr, bit_length);
3516             ff_h264_decode_sei(h);
3517             break;
3518         case NAL_SPS:
3519             init_get_bits(&s->gb, ptr, bit_length);
3520             ff_h264_decode_seq_parameter_set(h);
3521
3522             if(s->flags& CODEC_FLAG_LOW_DELAY)
3523                 s->low_delay=1;
3524
3525             if(avctx->has_b_frames < 2)
3526                 avctx->has_b_frames= !s->low_delay;
3527
3528             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3529                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3530                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3531                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3532
3533                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3534                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3535                     dsputil_init(&s->dsp, s->avctx);
3536                 } else {
3537                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3538                     return -1;
3539                 }
3540             }
3541             break;
3542         case NAL_PPS:
3543             init_get_bits(&s->gb, ptr, bit_length);
3544
3545             ff_h264_decode_picture_parameter_set(h, bit_length);
3546
3547             break;
3548         case NAL_AUD:
3549         case NAL_END_SEQUENCE:
3550         case NAL_END_STREAM:
3551         case NAL_FILLER_DATA:
3552         case NAL_SPS_EXT:
3553         case NAL_AUXILIARY_SLICE:
3554             break;
3555         default:
3556             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3557         }
3558
3559         if(context_count == h->max_contexts) {
3560             execute_decode_slices(h, context_count);
3561             context_count = 0;
3562         }
3563
3564         if (err < 0)
3565             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3566         else if(err == 1) {
3567             /* Slice could not be decoded in parallel mode, copy down
3568              * NAL unit stuff to context 0 and restart. Note that
3569              * rbsp_buffer is not transferred, but since we no longer
3570              * run in parallel mode this should not be an issue. */
3571             h->nal_unit_type = hx->nal_unit_type;
3572             h->nal_ref_idc   = hx->nal_ref_idc;
3573             hx = h;
3574             goto again;
3575         }
3576     }
3577     }
3578     if(context_count)
3579         execute_decode_slices(h, context_count);
3580     return buf_index;
3581 }
3582
3583 /**
3584  * returns the number of bytes consumed for building the current frame
3585  */
3586 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3587         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3588         if(pos+10>buf_size) pos=buf_size; // oops ;)
3589
3590         return pos;
3591 }
3592
3593 static int decode_frame(AVCodecContext *avctx,
3594                              void *data, int *data_size,
3595                              AVPacket *avpkt)
3596 {
3597     const uint8_t *buf = avpkt->data;
3598     int buf_size = avpkt->size;
3599     H264Context *h = avctx->priv_data;
3600     MpegEncContext *s = &h->s;
3601     AVFrame *pict = data;
3602     int buf_index;
3603
3604     s->flags= avctx->flags;
3605     s->flags2= avctx->flags2;
3606
3607    /* end of stream, output what is still in the buffers */
3608  out:
3609     if (buf_size == 0) {
3610         Picture *out;
3611         int i, out_idx;
3612
3613         s->current_picture_ptr = NULL;
3614
3615 //FIXME factorize this with the output code below
3616         out = h->delayed_pic[0];
3617         out_idx = 0;
3618         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3619             if(h->delayed_pic[i]->poc < out->poc){
3620                 out = h->delayed_pic[i];
3621                 out_idx = i;
3622             }
3623
3624         for(i=out_idx; h->delayed_pic[i]; i++)
3625             h->delayed_pic[i] = h->delayed_pic[i+1];
3626
3627         if(out){
3628             *data_size = sizeof(AVFrame);
3629             *pict= *(AVFrame*)out;
3630         }
3631
3632         return 0;
3633     }
3634
3635     buf_index=decode_nal_units(h, buf, buf_size);
3636     if(buf_index < 0)
3637         return -1;
3638
3639     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3640         buf_size = 0;
3641         goto out;
3642     }
3643
3644     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3645         if (avctx->skip_frame >= AVDISCARD_NONREF)
3646             return 0;
3647         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3648         return -1;
3649     }
3650
3651     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3652
3653         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3654
3655         field_end(h, 0);
3656
3657         if (!h->next_output_pic) {
3658             /* Wait for second field. */
3659             *data_size = 0;
3660
3661         } else {
3662             *data_size = sizeof(AVFrame);
3663             *pict = *(AVFrame*)h->next_output_pic;
3664         }
3665     }
3666
3667     assert(pict->data[0] || !*data_size);
3668     ff_print_debug_info(s, pict);
3669 //printf("out %d\n", (int)pict->data[0]);
3670
3671     return get_consumed_bytes(s, buf_index, buf_size);
3672 }
3673 #if 0
3674 static inline void fill_mb_avail(H264Context *h){
3675     MpegEncContext * const s = &h->s;
3676     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3677
3678     if(s->mb_y){
3679         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3680         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3681         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3682     }else{
3683         h->mb_avail[0]=
3684         h->mb_avail[1]=
3685         h->mb_avail[2]= 0;
3686     }
3687     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3688     h->mb_avail[4]= 1; //FIXME move out
3689     h->mb_avail[5]= 0; //FIXME move out
3690 }
3691 #endif
3692
3693 #ifdef TEST
3694 #undef printf
3695 #undef random
3696 #define COUNT 8000
3697 #define SIZE (COUNT*40)
3698 int main(void){
3699     int i;
3700     uint8_t temp[SIZE];
3701     PutBitContext pb;
3702     GetBitContext gb;
3703 //    int int_temp[10000];
3704     DSPContext dsp;
3705     AVCodecContext avctx;
3706
3707     dsputil_init(&dsp, &avctx);
3708
3709     init_put_bits(&pb, temp, SIZE);
3710     printf("testing unsigned exp golomb\n");
3711     for(i=0; i<COUNT; i++){
3712         START_TIMER
3713         set_ue_golomb(&pb, i);
3714         STOP_TIMER("set_ue_golomb");
3715     }
3716     flush_put_bits(&pb);
3717
3718     init_get_bits(&gb, temp, 8*SIZE);
3719     for(i=0; i<COUNT; i++){
3720         int j, s;
3721
3722         s= show_bits(&gb, 24);
3723
3724         START_TIMER
3725         j= get_ue_golomb(&gb);
3726         if(j != i){
3727             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3728 //            return -1;
3729         }
3730         STOP_TIMER("get_ue_golomb");
3731     }
3732
3733
3734     init_put_bits(&pb, temp, SIZE);
3735     printf("testing signed exp golomb\n");
3736     for(i=0; i<COUNT; i++){
3737         START_TIMER
3738         set_se_golomb(&pb, i - COUNT/2);
3739         STOP_TIMER("set_se_golomb");
3740     }
3741     flush_put_bits(&pb);
3742
3743     init_get_bits(&gb, temp, 8*SIZE);
3744     for(i=0; i<COUNT; i++){
3745         int j, s;
3746
3747         s= show_bits(&gb, 24);
3748
3749         START_TIMER
3750         j= get_se_golomb(&gb);
3751         if(j != i - COUNT/2){
3752             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3753 //            return -1;
3754         }
3755         STOP_TIMER("get_se_golomb");
3756     }
3757
3758 #if 0
3759     printf("testing 4x4 (I)DCT\n");
3760
3761     DCTELEM block[16];
3762     uint8_t src[16], ref[16];
3763     uint64_t error= 0, max_error=0;
3764
3765     for(i=0; i<COUNT; i++){
3766         int j;
3767 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
3768         for(j=0; j<16; j++){
3769             ref[j]= random()%255;
3770             src[j]= random()%255;
3771         }
3772
3773         h264_diff_dct_c(block, src, ref, 4);
3774
3775         //normalize
3776         for(j=0; j<16; j++){
3777 //            printf("%d ", block[j]);
3778             block[j]= block[j]*4;
3779             if(j&1) block[j]= (block[j]*4 + 2)/5;
3780             if(j&4) block[j]= (block[j]*4 + 2)/5;
3781         }
3782 //        printf("\n");
3783
3784         h->h264dsp.h264_idct_add(ref, block, 4);
3785 /*        for(j=0; j<16; j++){
3786             printf("%d ", ref[j]);
3787         }
3788         printf("\n");*/
3789
3790         for(j=0; j<16; j++){
3791             int diff= FFABS(src[j] - ref[j]);
3792
3793             error+= diff*diff;
3794             max_error= FFMAX(max_error, diff);
3795         }
3796     }
3797     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
3798     printf("testing quantizer\n");
3799     for(qp=0; qp<52; qp++){
3800         for(i=0; i<16; i++)
3801             src1_block[i]= src2_block[i]= random()%255;
3802
3803     }
3804     printf("Testing NAL layer\n");
3805
3806     uint8_t bitstream[COUNT];
3807     uint8_t nal[COUNT*2];
3808     H264Context h;
3809     memset(&h, 0, sizeof(H264Context));
3810
3811     for(i=0; i<COUNT; i++){
3812         int zeros= i;
3813         int nal_length;
3814         int consumed;
3815         int out_length;
3816         uint8_t *out;
3817         int j;
3818
3819         for(j=0; j<COUNT; j++){
3820             bitstream[j]= (random() % 255) + 1;
3821         }
3822
3823         for(j=0; j<zeros; j++){
3824             int pos= random() % COUNT;
3825             while(bitstream[pos] == 0){
3826                 pos++;
3827                 pos %= COUNT;
3828             }
3829             bitstream[pos]=0;
3830         }
3831
3832         START_TIMER
3833
3834         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
3835         if(nal_length<0){
3836             printf("encoding failed\n");
3837             return -1;
3838         }
3839
3840         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
3841
3842         STOP_TIMER("NAL")
3843
3844         if(out_length != COUNT){
3845             printf("incorrect length %d %d\n", out_length, COUNT);
3846             return -1;
3847         }
3848
3849         if(consumed != nal_length){
3850             printf("incorrect consumed length %d %d\n", nal_length, consumed);
3851             return -1;
3852         }
3853
3854         if(memcmp(bitstream, out, COUNT)){
3855             printf("mismatch\n");
3856             return -1;
3857         }
3858     }
3859 #endif
3860
3861     printf("Testing RBSP\n");
3862
3863
3864     return 0;
3865 }
3866 #endif /* TEST */
3867
3868
3869 av_cold void ff_h264_free_context(H264Context *h)
3870 {
3871     int i;
3872
3873     free_tables(h, 1); //FIXME cleanup init stuff perhaps
3874
3875     for(i = 0; i < MAX_SPS_COUNT; i++)
3876         av_freep(h->sps_buffers + i);
3877
3878     for(i = 0; i < MAX_PPS_COUNT; i++)
3879         av_freep(h->pps_buffers + i);
3880 }
3881
3882 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
3883 {
3884     H264Context *h = avctx->priv_data;
3885     MpegEncContext *s = &h->s;
3886
3887     ff_h264_free_context(h);
3888
3889     MPV_common_end(s);
3890
3891 //    memset(h, 0, sizeof(H264Context));
3892
3893     return 0;
3894 }
3895
3896 static const AVProfile profiles[] = {
3897     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
3898     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
3899     { FF_PROFILE_H264_MAIN,                 "Main"                  },
3900     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
3901     { FF_PROFILE_H264_HIGH,                 "High"                  },
3902     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
3903     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
3904     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
3905     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
3906     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
3907     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
3908     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
3909     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
3910     { FF_PROFILE_UNKNOWN },
3911 };
3912
3913 AVCodec ff_h264_decoder = {
3914     "h264",
3915     AVMEDIA_TYPE_VIDEO,
3916     CODEC_ID_H264,
3917     sizeof(H264Context),
3918     ff_h264_decode_init,
3919     NULL,
3920     ff_h264_decode_end,
3921     decode_frame,
3922     /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
3923         CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
3924     .flush= flush_dpb,
3925     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
3926     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
3927     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
3928     .profiles = NULL_IF_CONFIG_SMALL(profiles),
3929 };
3930
3931 #if CONFIG_H264_VDPAU_DECODER
3932 AVCodec ff_h264_vdpau_decoder = {
3933     "h264_vdpau",
3934     AVMEDIA_TYPE_VIDEO,
3935     CODEC_ID_H264,
3936     sizeof(H264Context),
3937     ff_h264_decode_init,
3938     NULL,
3939     ff_h264_decode_end,
3940     decode_frame,
3941     CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
3942     .flush= flush_dpb,
3943     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
3944     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
3945     .profiles = NULL_IF_CONFIG_SMALL(profiles),
3946 };
3947 #endif