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

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * Libav is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with Libav; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavutil/imgutils.h"
  29 #include "internal.h"
  30 #include "dsputil.h"
  31 #include "avcodec.h"
  32 #include "mpegvideo.h"
  33 #include "h264.h"
  34 #include "h264data.h"
  35 #include "h264_mvpred.h"
  36 #include "golomb.h"
  37 #include "mathops.h"
  38 #include "rectangle.h"
  39 #include "thread.h"
  40 #include "vdpau_internal.h"
  41 #include "libavutil/avassert.h"
  42
  43 #include "cabac.h"
  44
  45 //#undef NDEBUG
  46 #include <assert.h>
  47
  48 static const uint8_t rem6[QP_MAX_NUM+1]={
  49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[QP_MAX_NUM+1]={
  53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9,10,10,10,10,
  54 };
  55
  56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  57     PIX_FMT_DXVA2_VLD,
  58     PIX_FMT_VAAPI_VLD,
  59     PIX_FMT_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 /**
  64  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  65  */
  66 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  67     MpegEncContext * const s = &h->s;
  68     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  69     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  70     int i;
  71
  72     if(!(h->top_samples_available&0x8000)){
  73         for(i=0; i<4; i++){
  74             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  75             if(status<0){
  76                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  77                 return -1;
  78             } else if(status){
  79                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  80             }
  81         }
  82     }
  83
  84     if((h->left_samples_available&0x8888)!=0x8888){
  85         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  86         for(i=0; i<4; i++){
  87             if(!(h->left_samples_available&mask[i])){
  88                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  89                 if(status<0){
  90                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  91                     return -1;
  92                 } else if(status){
  93                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  94                 }
  95             }
  96         }
  97     }
  98
  99     return 0;
 100 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 101
 102 /**
 103  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 104  */
 105 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 106     MpegEncContext * const s = &h->s;
 107     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 108     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 109
 110     if(mode > 6U) {
 111         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 112         return -1;
 113     }
 114
 115     if(!(h->top_samples_available&0x8000)){
 116         mode= top[ mode ];
 117         if(mode<0){
 118             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 119             return -1;
 120         }
 121     }
 122
 123     if((h->left_samples_available&0x8080) != 0x8080){
 124         mode= left[ mode ];
 125         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 126             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 127         }
 128         if(mode<0){
 129             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 130             return -1;
 131         }
 132     }
 133
 134     return mode;
 135 }
 136
 137 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 138     int i, si, di;
 139     uint8_t *dst;
 140     int bufidx;
 141
 142 //    src[0]&0x80;                //forbidden bit
 143     h->nal_ref_idc= src[0]>>5;
 144     h->nal_unit_type= src[0]&0x1F;
 145
 146     src++; length--;
 147
 148 #if HAVE_FAST_UNALIGNED
 149 # if HAVE_FAST_64BIT
 150 #   define RS 7
 151     for(i=0; i+1<length; i+=9){
 152         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 153 # else
 154 #   define RS 3
 155     for(i=0; i+1<length; i+=5){
 156         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 157 # endif
 158             continue;
 159         if(i>0 && !src[i]) i--;
 160         while(src[i]) i++;
 161 #else
 162 #   define RS 0
 163     for(i=0; i+1<length; i+=2){
 164         if(src[i]) continue;
 165         if(i>0 && src[i-1]==0) i--;
 166 #endif
 167         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 168             if(src[i+2]!=3){
 169                 /* startcode, so we must be past the end */
 170                 length=i;
 171             }
 172             break;
 173         }
 174         i-= RS;
 175     }
 176
 177     if(i>=length-1){ //no escaped 0
 178         *dst_length= length;
 179         *consumed= length+1; //+1 for the header
 180         return src;
 181     }
 182
 183     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 184     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 185     dst= h->rbsp_buffer[bufidx];
 186
 187     if (dst == NULL){
 188         return NULL;
 189     }
 190
 191 //printf("decoding esc\n");
 192     memcpy(dst, src, i);
 193     si=di=i;
 194     while(si+2<length){
 195         //remove escapes (very rare 1:2^22)
 196         if(src[si+2]>3){
 197             dst[di++]= src[si++];
 198             dst[di++]= src[si++];
 199         }else if(src[si]==0 && src[si+1]==0){
 200             if(src[si+2]==3){ //escape
 201                 dst[di++]= 0;
 202                 dst[di++]= 0;
 203                 si+=3;
 204                 continue;
 205             }else //next start code
 206                 goto nsc;
 207         }
 208
 209         dst[di++]= src[si++];
 210     }
 211     while(si<length)
 212         dst[di++]= src[si++];
 213 nsc:
 214
 215     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 216
 217     *dst_length= di;
 218     *consumed= si + 1;//+1 for the header
 219 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 220     return dst;
 221 }
 222
 223 /**
 224  * Identify the exact end of the bitstream
 225  * @return the length of the trailing, or 0 if damaged
 226  */
 227 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 228     int v= *src;
 229     int r;
 230
 231     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 232
 233     for(r=1; r<9; r++){
 234         if(v&1) return r;
 235         v>>=1;
 236     }
 237     return 0;
 238 }
 239
 240 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 241                                  int y_offset, int list){
 242     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 243     int filter_height= (raw_my&3) ? 2 : 0;
 244     int full_my= (raw_my>>2) + y_offset;
 245     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 246
 247     return FFMAX(abs(top), bottom);
 248 }
 249
 250 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 251                                int y_offset, int list0, int list1, int *nrefs){
 252     MpegEncContext * const s = &h->s;
 253     int my;
 254
 255     y_offset += 16*(s->mb_y >> MB_FIELD);
 256
 257     if(list0){
 258         int ref_n = h->ref_cache[0][ scan8[n] ];
 259         Picture *ref= &h->ref_list[0][ref_n];
 260
 261         // Error resilience puts the current picture in the ref list.
 262         // Don't try to wait on these as it will cause a deadlock.
 263         // Fields can wait on each other, though.
 264         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 265            (ref->f.reference & 3) != s->picture_structure) {
 266             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 267             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 268             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 269         }
 270     }
 271
 272     if(list1){
 273         int ref_n = h->ref_cache[1][ scan8[n] ];
 274         Picture *ref= &h->ref_list[1][ref_n];
 275
 276         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 277            (ref->f.reference & 3) != s->picture_structure) {
 278             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 279             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 280             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 281         }
 282     }
 283 }
 284
 285 /**
 286  * Wait until all reference frames are available for MC operations.
 287  *
 288  * @param h the H264 context
 289  */
 290 static void await_references(H264Context *h){
 291     MpegEncContext * const s = &h->s;
 292     const int mb_xy= h->mb_xy;
 293     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 294     int refs[2][48];
 295     int nrefs[2] = {0};
 296     int ref, list;
 297
 298     memset(refs, -1, sizeof(refs));
 299
 300     if(IS_16X16(mb_type)){
 301         get_lowest_part_y(h, refs, 0, 16, 0,
 302                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 303     }else if(IS_16X8(mb_type)){
 304         get_lowest_part_y(h, refs, 0, 8, 0,
 305                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 306         get_lowest_part_y(h, refs, 8, 8, 8,
 307                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 308     }else if(IS_8X16(mb_type)){
 309         get_lowest_part_y(h, refs, 0, 16, 0,
 310                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 311         get_lowest_part_y(h, refs, 4, 16, 0,
 312                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 313     }else{
 314         int i;
 315
 316         assert(IS_8X8(mb_type));
 317
 318         for(i=0; i<4; i++){
 319             const int sub_mb_type= h->sub_mb_type[i];
 320             const int n= 4*i;
 321             int y_offset= (i&2)<<2;
 322
 323             if(IS_SUB_8X8(sub_mb_type)){
 324                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 325                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 326             }else if(IS_SUB_8X4(sub_mb_type)){
 327                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 328                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 329                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 330                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 331             }else if(IS_SUB_4X8(sub_mb_type)){
 332                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 333                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 334                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 335                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 336             }else{
 337                 int j;
 338                 assert(IS_SUB_4X4(sub_mb_type));
 339                 for(j=0; j<4; j++){
 340                     int sub_y_offset= y_offset + 2*(j&2);
 341                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 342                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 343                 }
 344             }
 345         }
 346     }
 347
 348     for(list=h->list_count-1; list>=0; list--){
 349         for(ref=0; ref<48 && nrefs[list]; ref++){
 350             int row = refs[list][ref];
 351             if(row >= 0){
 352                 Picture *ref_pic = &h->ref_list[list][ref];
 353                 int ref_field = ref_pic->f.reference - 1;
 354                 int ref_field_picture = ref_pic->field_picture;
 355                 int pic_height = 16*s->mb_height >> ref_field_picture;
 356
 357                 row <<= MB_MBAFF;
 358                 nrefs[list]--;
 359
 360                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 361                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 362                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 363                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 364                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 365                 }else if(FIELD_PICTURE){
 366                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 367                 }else{
 368                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 369                 }
 370             }
 371         }
 372     }
 373 }
 374
 375 #if 0
 376 /**
 377  * DCT transforms the 16 dc values.
 378  * @param qp quantization parameter ??? FIXME
 379  */
 380 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 381 //    const int qmul= dequant_coeff[qp][0];
 382     int i;
 383     int temp[16]; //FIXME check if this is a good idea
 384     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 385     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 386
 387     for(i=0; i<4; i++){
 388         const int offset= y_offset[i];
 389         const int z0= block[offset+stride*0] + block[offset+stride*4];
 390         const int z1= block[offset+stride*0] - block[offset+stride*4];
 391         const int z2= block[offset+stride*1] - block[offset+stride*5];
 392         const int z3= block[offset+stride*1] + block[offset+stride*5];
 393
 394         temp[4*i+0]= z0+z3;
 395         temp[4*i+1]= z1+z2;
 396         temp[4*i+2]= z1-z2;
 397         temp[4*i+3]= z0-z3;
 398     }
 399
 400     for(i=0; i<4; i++){
 401         const int offset= x_offset[i];
 402         const int z0= temp[4*0+i] + temp[4*2+i];
 403         const int z1= temp[4*0+i] - temp[4*2+i];
 404         const int z2= temp[4*1+i] - temp[4*3+i];
 405         const int z3= temp[4*1+i] + temp[4*3+i];
 406
 407         block[stride*0 +offset]= (z0 + z3)>>1;
 408         block[stride*2 +offset]= (z1 + z2)>>1;
 409         block[stride*8 +offset]= (z1 - z2)>>1;
 410         block[stride*10+offset]= (z0 - z3)>>1;
 411     }
 412 }
 413 #endif
 414
 415 #undef xStride
 416 #undef stride
 417
 418 #if 0
 419 static void chroma_dc_dct_c(DCTELEM *block){
 420     const int stride= 16*2;
 421     const int xStride= 16;
 422     int a,b,c,d,e;
 423
 424     a= block[stride*0 + xStride*0];
 425     b= block[stride*0 + xStride*1];
 426     c= block[stride*1 + xStride*0];
 427     d= block[stride*1 + xStride*1];
 428
 429     e= a-b;
 430     a= a+b;
 431     b= c-d;
 432     c= c+d;
 433
 434     block[stride*0 + xStride*0]= (a+c);
 435     block[stride*0 + xStride*1]= (e+b);
 436     block[stride*1 + xStride*0]= (a-c);
 437     block[stride*1 + xStride*1]= (e-b);
 438 }
 439 #endif
 440
 441 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 442                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 443                            int src_x_offset, int src_y_offset,
 444                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 445                            int pixel_shift, int chroma444){
 446     MpegEncContext * const s = &h->s;
 447     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 448     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 449     const int luma_xy= (mx&3) + ((my&3)<<2);
 450     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 451     uint8_t * src_y = pic->f.data[0] + offset;
 452     uint8_t * src_cb, * src_cr;
 453     int extra_width= h->emu_edge_width;
 454     int extra_height= h->emu_edge_height;
 455     int emu=0;
 456     const int full_mx= mx>>2;
 457     const int full_my= my>>2;
 458     const int pic_width  = 16*s->mb_width;
 459     const int pic_height = 16*s->mb_height >> MB_FIELD;
 460
 461     if(mx&7) extra_width -= 3;
 462     if(my&7) extra_height -= 3;
 463
 464     if(   full_mx < 0-extra_width
 465        || full_my < 0-extra_height
 466        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 467        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 468         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 469             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 470         emu=1;
 471     }
 472
 473     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 474     if(!square){
 475         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 476     }
 477
 478     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 479
 480     if(chroma444){
 481         src_cb = pic->f.data[1] + offset;
 482         if(emu){
 483             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 484                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 485             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 486         }
 487         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 488         if(!square){
 489             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 490         }
 491
 492         src_cr = pic->f.data[2] + offset;
 493         if(emu){
 494             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 495                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 496             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 497         }
 498         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 499         if(!square){
 500             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 501         }
 502         return;
 503     }
 504
 505     if(MB_FIELD){
 506         // chroma offset when predicting from a field of opposite parity
 507         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 508         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 509     }
 510     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 511     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> 3) * h->mb_uvlinesize;
 512
 513     if(emu){
 514         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 515             src_cb= s->edge_emu_buffer;
 516     }
 517     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 518
 519     if(emu){
 520         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 521             src_cr= s->edge_emu_buffer;
 522     }
 523     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 524 }
 525
 526 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 527                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 528                            int x_offset, int y_offset,
 529                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 530                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 531                            int list0, int list1, int pixel_shift, int chroma444){
 532     MpegEncContext * const s = &h->s;
 533     qpel_mc_func *qpix_op=  qpix_put;
 534     h264_chroma_mc_func chroma_op= chroma_put;
 535
 536     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 537     if(chroma444){
 538         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 539         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 540     }else{
 541         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 542         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 543     }
 544     x_offset += 8*s->mb_x;
 545     y_offset += 8*(s->mb_y >> MB_FIELD);
 546
 547     if(list0){
 548         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 549         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 550                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 551                            qpix_op, chroma_op, pixel_shift, chroma444);
 552
 553         qpix_op=  qpix_avg;
 554         chroma_op= chroma_avg;
 555     }
 556
 557     if(list1){
 558         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 559         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 560                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 561                            qpix_op, chroma_op, pixel_shift, chroma444);
 562     }
 563 }
 564
 565 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 566                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 567                            int x_offset, int y_offset,
 568                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 569                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 570                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 571                            int list0, int list1, int pixel_shift, int chroma444){
 572     MpegEncContext * const s = &h->s;
 573
 574     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 575     if(chroma444){
 576         chroma_weight_avg = luma_weight_avg;
 577         chroma_weight_op = luma_weight_op;
 578         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 579         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 580     }else{
 581         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 582         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 583     }
 584     x_offset += 8*s->mb_x;
 585     y_offset += 8*(s->mb_y >> MB_FIELD);
 586
 587     if(list0 && list1){
 588         /* don't optimize for luma-only case, since B-frames usually
 589          * use implicit weights => chroma too. */
 590         uint8_t *tmp_cb = s->obmc_scratchpad;
 591         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 592         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 593         int refn0 = h->ref_cache[0][ scan8[n] ];
 594         int refn1 = h->ref_cache[1][ scan8[n] ];
 595
 596         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 597                     dest_y, dest_cb, dest_cr,
 598                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 599         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 600                     tmp_y, tmp_cb, tmp_cr,
 601                     x_offset, y_offset, qpix_put, chroma_put, pixel_shift, chroma444);
 602
 603         if(h->use_weight == 2){
 604             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 605             int weight1 = 64 - weight0;
 606             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 607             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 608             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 609         }else{
 610             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 611                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 612                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 613             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 614                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 615                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 616             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 617                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 618                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 619         }
 620     }else{
 621         int list = list1 ? 1 : 0;
 622         int refn = h->ref_cache[list][ scan8[n] ];
 623         Picture *ref= &h->ref_list[list][refn];
 624         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 625                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 626                     qpix_put, chroma_put, pixel_shift, chroma444);
 627
 628         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 629                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 630         if(h->use_weight_chroma){
 631             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 632                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 633             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 634                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 635         }
 636     }
 637 }
 638
 639 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 640                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 641                            int x_offset, int y_offset,
 642                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 643                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 644                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 645                            int list0, int list1, int pixel_shift, int chroma444){
 646     if((h->use_weight==2 && list0 && list1
 647         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 648        || h->use_weight==1)
 649         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 650                          x_offset, y_offset, qpix_put, chroma_put,
 651                          weight_op[0], weight_op[3], weight_avg[0],
 652                          weight_avg[3], list0, list1, pixel_shift, chroma444);
 653     else
 654         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 655                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 656                     chroma_avg, list0, list1, pixel_shift, chroma444);
 657 }
 658
 659 static inline void prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma444){
 660     /* fetch pixels for estimated mv 4 macroblocks ahead
 661      * optimized for 64byte cache lines */
 662     MpegEncContext * const s = &h->s;
 663     const int refn = h->ref_cache[list][scan8[0]];
 664     if(refn >= 0){
 665         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 666         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 667         uint8_t **src = h->ref_list[list][refn].f.data;
 668         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 669         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 670         if(chroma444){
 671             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 672             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 673         }else{
 674             off= ((mx>>1) << pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + (64 << pixel_shift);
 675             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 676         }
 677     }
 678 }
 679
 680 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 681                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 682                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 683                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 684                       int pixel_shift, int chroma444){
 685     MpegEncContext * const s = &h->s;
 686     const int mb_xy= h->mb_xy;
 687     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 688
 689     assert(IS_INTER(mb_type));
 690
 691     if(HAVE_PTHREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 692         await_references(h);
 693     prefetch_motion(h, 0, pixel_shift, chroma444);
 694
 695     if(IS_16X16(mb_type)){
 696         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 697                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 698                 weight_op, weight_avg,
 699                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 700                 pixel_shift, chroma444);
 701     }else if(IS_16X8(mb_type)){
 702         mc_part(h, 0, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 703                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 704                 &weight_op[1], &weight_avg[1],
 705                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 706                 pixel_shift, chroma444);
 707         mc_part(h, 8, 0, 4, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 708                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 709                 &weight_op[1], &weight_avg[1],
 710                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 711                 pixel_shift, chroma444);
 712     }else if(IS_8X16(mb_type)){
 713         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 714                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 715                 &weight_op[2], &weight_avg[2],
 716                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 717                 pixel_shift, chroma444);
 718         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 719                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 720                 &weight_op[2], &weight_avg[2],
 721                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 722                 pixel_shift, chroma444);
 723     }else{
 724         int i;
 725
 726         assert(IS_8X8(mb_type));
 727
 728         for(i=0; i<4; i++){
 729             const int sub_mb_type= h->sub_mb_type[i];
 730             const int n= 4*i;
 731             int x_offset= (i&1)<<2;
 732             int y_offset= (i&2)<<1;
 733
 734             if(IS_SUB_8X8(sub_mb_type)){
 735                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 736                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 737                     &weight_op[3], &weight_avg[3],
 738                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 739                     pixel_shift, chroma444);
 740             }else if(IS_SUB_8X4(sub_mb_type)){
 741                 mc_part(h, n  , 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 742                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 743                     &weight_op[4], &weight_avg[4],
 744                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 745                     pixel_shift, chroma444);
 746                 mc_part(h, n+2, 0, 2, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 747                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 748                     &weight_op[4], &weight_avg[4],
 749                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 750                     pixel_shift, chroma444);
 751             }else if(IS_SUB_4X8(sub_mb_type)){
 752                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 753                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 754                     &weight_op[5], &weight_avg[5],
 755                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 756                     pixel_shift, chroma444);
 757                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 758                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 759                     &weight_op[5], &weight_avg[5],
 760                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 761                     pixel_shift, chroma444);
 762             }else{
 763                 int j;
 764                 assert(IS_SUB_4X4(sub_mb_type));
 765                 for(j=0; j<4; j++){
 766                     int sub_x_offset= x_offset + 2*(j&1);
 767                     int sub_y_offset= y_offset +   (j&2);
 768                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 769                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 770                         &weight_op[6], &weight_avg[6],
 771                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 772                         pixel_shift, chroma444);
 773                 }
 774             }
 775         }
 776     }
 777
 778     prefetch_motion(h, 1, pixel_shift, chroma444);
 779 }
 780
 781 static void free_tables(H264Context *h, int free_rbsp){
 782     int i;
 783     H264Context *hx;
 784
 785     av_freep(&h->intra4x4_pred_mode);
 786     av_freep(&h->chroma_pred_mode_table);
 787     av_freep(&h->cbp_table);
 788     av_freep(&h->mvd_table[0]);
 789     av_freep(&h->mvd_table[1]);
 790     av_freep(&h->direct_table);
 791     av_freep(&h->non_zero_count);
 792     av_freep(&h->slice_table_base);
 793     h->slice_table= NULL;
 794     av_freep(&h->list_counts);
 795
 796     av_freep(&h->mb2b_xy);
 797     av_freep(&h->mb2br_xy);
 798
 799     for(i = 0; i < MAX_THREADS; i++) {
 800         hx = h->thread_context[i];
 801         if(!hx) continue;
 802         av_freep(&hx->top_borders[1]);
 803         av_freep(&hx->top_borders[0]);
 804         av_freep(&hx->s.obmc_scratchpad);
 805         if (free_rbsp){
 806             av_freep(&hx->rbsp_buffer[1]);
 807             av_freep(&hx->rbsp_buffer[0]);
 808             hx->rbsp_buffer_size[0] = 0;
 809             hx->rbsp_buffer_size[1] = 0;
 810         }
 811         if (i) av_freep(&h->thread_context[i]);
 812     }
 813 }
 814
 815 static void init_dequant8_coeff_table(H264Context *h){
 816     int i,j,q,x;
 817     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 818
 819     for(i=0; i<6; i++ ){
 820         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 821         for(j=0; j<i; j++){
 822             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 823                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 824                 break;
 825             }
 826         }
 827         if(j<i)
 828             continue;
 829
 830         for(q=0; q<max_qp+1; q++){
 831             int shift = div6[q];
 832             int idx = rem6[q];
 833             for(x=0; x<64; x++)
 834                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 835                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 836                     h->pps.scaling_matrix8[i][x]) << shift;
 837         }
 838     }
 839 }
 840
 841 static void init_dequant4_coeff_table(H264Context *h){
 842     int i,j,q,x;
 843     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 844     for(i=0; i<6; i++ ){
 845         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 846         for(j=0; j<i; j++){
 847             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 848                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 849                 break;
 850             }
 851         }
 852         if(j<i)
 853             continue;
 854
 855         for(q=0; q<max_qp+1; q++){
 856             int shift = div6[q] + 2;
 857             int idx = rem6[q];
 858             for(x=0; x<16; x++)
 859                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 860                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 861                     h->pps.scaling_matrix4[i][x]) << shift;
 862         }
 863     }
 864 }
 865
 866 static void init_dequant_tables(H264Context *h){
 867     int i,x;
 868     init_dequant4_coeff_table(h);
 869     if(h->pps.transform_8x8_mode)
 870         init_dequant8_coeff_table(h);
 871     if(h->sps.transform_bypass){
 872         for(i=0; i<6; i++)
 873             for(x=0; x<16; x++)
 874                 h->dequant4_coeff[i][0][x] = 1<<6;
 875         if(h->pps.transform_8x8_mode)
 876             for(i=0; i<6; i++)
 877                 for(x=0; x<64; x++)
 878                     h->dequant8_coeff[i][0][x] = 1<<6;
 879     }
 880 }
 881
 882
 883 int ff_h264_alloc_tables(H264Context *h){
 884     MpegEncContext * const s = &h->s;
 885     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 886     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 887     int x,y;
 888
 889     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 890
 891     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 892     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 893     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 894
 895     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 896     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 897     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 898     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 899     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 900
 901     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 902     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 903
 904     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 905     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 906     for(y=0; y<s->mb_height; y++){
 907         for(x=0; x<s->mb_width; x++){
 908             const int mb_xy= x + y*s->mb_stride;
 909             const int b_xy = 4*x + 4*y*h->b_stride;
 910
 911             h->mb2b_xy [mb_xy]= b_xy;
 912             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 913         }
 914     }
 915
 916     s->obmc_scratchpad = NULL;
 917
 918     if(!h->dequant4_coeff[0])
 919         init_dequant_tables(h);
 920
 921     return 0;
 922 fail:
 923     free_tables(h, 1);
 924     return -1;
 925 }
 926
 927 /**
 928  * Mimic alloc_tables(), but for every context thread.
 929  */
 930 static void clone_tables(H264Context *dst, H264Context *src, int i){
 931     MpegEncContext * const s = &src->s;
 932     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 933     dst->non_zero_count           = src->non_zero_count;
 934     dst->slice_table              = src->slice_table;
 935     dst->cbp_table                = src->cbp_table;
 936     dst->mb2b_xy                  = src->mb2b_xy;
 937     dst->mb2br_xy                 = src->mb2br_xy;
 938     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 939     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 940     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 941     dst->direct_table             = src->direct_table;
 942     dst->list_counts              = src->list_counts;
 943
 944     dst->s.obmc_scratchpad = NULL;
 945     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma);
 946 }
 947
 948 /**
 949  * Init context
 950  * Allocate buffers which are not shared amongst multiple threads.
 951  */
 952 static int context_init(H264Context *h){
 953     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 954     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
 955
 956     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 957     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 958
 959     return 0;
 960 fail:
 961     return -1; // free_tables will clean up for us
 962 }
 963
 964 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 965
 966 static av_cold void common_init(H264Context *h){
 967     MpegEncContext * const s = &h->s;
 968
 969     s->width = s->avctx->width;
 970     s->height = s->avctx->height;
 971     s->codec_id= s->avctx->codec->id;
 972
 973     ff_h264dsp_init(&h->h264dsp, 8);
 974     ff_h264_pred_init(&h->hpc, s->codec_id, 8);
 975
 976     h->dequant_coeff_pps= -1;
 977     s->unrestricted_mv=1;
 978     s->decode=1; //FIXME
 979
 980     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 981
 982     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 983     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 984 }
 985
 986 int ff_h264_decode_extradata(H264Context *h)
 987 {
 988     AVCodecContext *avctx = h->s.avctx;
 989
 990     if(avctx->extradata[0] == 1){
 991         int i, cnt, nalsize;
 992         unsigned char *p = avctx->extradata;
 993
 994         h->is_avc = 1;
 995
 996         if(avctx->extradata_size < 7) {
 997             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
 998             return -1;
 999         }
1000         /* sps and pps in the avcC always have length coded with 2 bytes,
1001            so put a fake nal_length_size = 2 while parsing them */
1002         h->nal_length_size = 2;
1003         // Decode sps from avcC
1004         cnt = *(p+5) & 0x1f; // Number of sps
1005         p += 6;
1006         for (i = 0; i < cnt; i++) {
1007             nalsize = AV_RB16(p) + 2;
1008             if(decode_nal_units(h, p, nalsize) < 0) {
1009                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1010                 return -1;
1011             }
1012             p += nalsize;
1013         }
1014         // Decode pps from avcC
1015         cnt = *(p++); // Number of pps
1016         for (i = 0; i < cnt; i++) {
1017             nalsize = AV_RB16(p) + 2;
1018             if (decode_nal_units(h, p, nalsize) < 0) {
1019                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1020                 return -1;
1021             }
1022             p += nalsize;
1023         }
1024         // Now store right nal length size, that will be use to parse all other nals
1025         h->nal_length_size = (avctx->extradata[4] & 0x03) + 1;
1026     } else {
1027         h->is_avc = 0;
1028         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
1029             return -1;
1030     }
1031     return 0;
1032 }
1033
1034 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1035     H264Context *h= avctx->priv_data;
1036     MpegEncContext * const s = &h->s;
1037
1038     MPV_decode_defaults(s);
1039
1040     s->avctx = avctx;
1041     common_init(h);
1042
1043     s->out_format = FMT_H264;
1044     s->workaround_bugs= avctx->workaround_bugs;
1045
1046     // set defaults
1047 //    s->decode_mb= ff_h263_decode_mb;
1048     s->quarter_sample = 1;
1049     if(!avctx->has_b_frames)
1050     s->low_delay= 1;
1051
1052     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1053
1054     ff_h264_decode_init_vlc();
1055
1056     h->pixel_shift = 0;
1057     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1058
1059     h->thread_context[0] = h;
1060     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1061     h->prev_poc_msb= 1<<16;
1062     h->x264_build = -1;
1063     ff_h264_reset_sei(h);
1064     if(avctx->codec_id == CODEC_ID_H264){
1065         if(avctx->ticks_per_frame == 1){
1066             s->avctx->time_base.den *=2;
1067         }
1068         avctx->ticks_per_frame = 2;
1069     }
1070
1071     if(avctx->extradata_size > 0 && avctx->extradata &&
1072         ff_h264_decode_extradata(h))
1073         return -1;
1074
1075     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1076         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1077         s->low_delay = 0;
1078     }
1079
1080     return 0;
1081 }
1082
1083 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1084 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1085 {
1086     int i;
1087
1088     for (i=0; i<count; i++){
1089         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1090                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1091                 !from[i]));
1092         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1093     }
1094 }
1095
1096 static void copy_parameter_set(void **to, void **from, int count, int size)
1097 {
1098     int i;
1099
1100     for (i=0; i<count; i++){
1101         if (to[i] && !from[i]) av_freep(&to[i]);
1102         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1103
1104         if (from[i]) memcpy(to[i], from[i], size);
1105     }
1106 }
1107
1108 static int decode_init_thread_copy(AVCodecContext *avctx){
1109     H264Context *h= avctx->priv_data;
1110
1111     if (!avctx->is_copy) return 0;
1112     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1113     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1114
1115     return 0;
1116 }
1117
1118 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1119 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1120     H264Context *h= dst->priv_data, *h1= src->priv_data;
1121     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1122     int inited = s->context_initialized, err;
1123     int i;
1124
1125     if(dst == src || !s1->context_initialized) return 0;
1126
1127     err = ff_mpeg_update_thread_context(dst, src);
1128     if(err) return err;
1129
1130     //FIXME handle width/height changing
1131     if(!inited){
1132         for(i = 0; i < MAX_SPS_COUNT; i++)
1133             av_freep(h->sps_buffers + i);
1134
1135         for(i = 0; i < MAX_PPS_COUNT; i++)
1136             av_freep(h->pps_buffers + i);
1137
1138         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1139         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1140         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1141         if (ff_h264_alloc_tables(h) < 0) {
1142             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1143             return AVERROR(ENOMEM);
1144         }
1145         context_init(h);
1146
1147         for(i=0; i<2; i++){
1148             h->rbsp_buffer[i] = NULL;
1149             h->rbsp_buffer_size[i] = 0;
1150         }
1151
1152         h->thread_context[0] = h;
1153
1154         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1155         // so this has to be allocated here
1156         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1157
1158         s->dsp.clear_blocks(h->mb);
1159         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1160     }
1161
1162     //extradata/NAL handling
1163     h->is_avc          = h1->is_avc;
1164
1165     //SPS/PPS
1166     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1167     h->sps             = h1->sps;
1168     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1169     h->pps             = h1->pps;
1170
1171     //Dequantization matrices
1172     //FIXME these are big - can they be only copied when PPS changes?
1173     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1174
1175     for(i=0; i<6; i++)
1176         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1177
1178     for(i=0; i<6; i++)
1179         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1180
1181     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1182
1183     //POC timing
1184     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1185
1186     //reference lists
1187     copy_fields(h, h1, ref_count, list_count);
1188     copy_fields(h, h1, ref_list,  intra_gb);
1189     copy_fields(h, h1, short_ref, cabac_init_idc);
1190
1191     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1192     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1193     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1194
1195     h->last_slice_type = h1->last_slice_type;
1196
1197     if(!s->current_picture_ptr) return 0;
1198
1199     if(!s->dropable) {
1200         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1201         h->prev_poc_msb     = h->poc_msb;
1202         h->prev_poc_lsb     = h->poc_lsb;
1203     }
1204     h->prev_frame_num_offset= h->frame_num_offset;
1205     h->prev_frame_num       = h->frame_num;
1206     h->outputed_poc         = h->next_outputed_poc;
1207
1208     return err;
1209 }
1210
1211 int ff_h264_frame_start(H264Context *h){
1212     MpegEncContext * const s = &h->s;
1213     int i;
1214     const int pixel_shift = h->pixel_shift;
1215     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1216
1217     if(MPV_frame_start(s, s->avctx) < 0)
1218         return -1;
1219     ff_er_frame_start(s);
1220     /*
1221      * MPV_frame_start uses pict_type to derive key_frame.
1222      * This is incorrect for H.264; IDR markings must be used.
1223      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1224      * See decode_nal_units().
1225      */
1226     s->current_picture_ptr->f.key_frame = 0;
1227     s->current_picture_ptr->mmco_reset= 0;
1228
1229     assert(s->linesize && s->uvlinesize);
1230
1231     for(i=0; i<16; i++){
1232         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1233         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1234     }
1235     for(i=0; i<16; i++){
1236         h->block_offset[16+i]=
1237         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1238         h->block_offset[48+16+i]=
1239         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1240     }
1241
1242     /* can't be in alloc_tables because linesize isn't known there.
1243      * FIXME: redo bipred weight to not require extra buffer? */
1244     for(i = 0; i < thread_count; i++)
1245         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1246             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1247
1248     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1249     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1250
1251 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1252
1253     // We mark the current picture as non-reference after allocating it, so
1254     // that if we break out due to an error it can be released automatically
1255     // in the next MPV_frame_start().
1256     // SVQ3 as well as most other codecs have only last/next/current and thus
1257     // get released even with set reference, besides SVQ3 and others do not
1258     // mark frames as reference later "naturally".
1259     if(s->codec_id != CODEC_ID_SVQ3)
1260         s->current_picture_ptr->f.reference = 0;
1261
1262     s->current_picture_ptr->field_poc[0]=
1263     s->current_picture_ptr->field_poc[1]= INT_MAX;
1264
1265     h->next_output_pic = NULL;
1266
1267     assert(s->current_picture_ptr->long_ref==0);
1268
1269     return 0;
1270 }
1271
1272 /**
1273   * Run setup operations that must be run after slice header decoding.
1274   * This includes finding the next displayed frame.
1275   *
1276   * @param h h264 master context
1277   * @param setup_finished enough NALs have been read that we can call
1278   * ff_thread_finish_setup()
1279   */
1280 static void decode_postinit(H264Context *h, int setup_finished){
1281     MpegEncContext * const s = &h->s;
1282     Picture *out = s->current_picture_ptr;
1283     Picture *cur = s->current_picture_ptr;
1284     int i, pics, out_of_order, out_idx;
1285
1286     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1287     s->current_picture_ptr->f.pict_type   = s->pict_type;
1288
1289     if (h->next_output_pic) return;
1290
1291     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1292         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1293         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1294         //to find this yet, so we assume the worst for now.
1295         //if (setup_finished)
1296         //    ff_thread_finish_setup(s->avctx);
1297         return;
1298     }
1299
1300     cur->f.interlaced_frame = 0;
1301     cur->f.repeat_pict      = 0;
1302
1303     /* Signal interlacing information externally. */
1304     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1305
1306     if(h->sps.pic_struct_present_flag){
1307         switch (h->sei_pic_struct)
1308         {
1309         case SEI_PIC_STRUCT_FRAME:
1310             break;
1311         case SEI_PIC_STRUCT_TOP_FIELD:
1312         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1313             cur->f.interlaced_frame = 1;
1314             break;
1315         case SEI_PIC_STRUCT_TOP_BOTTOM:
1316         case SEI_PIC_STRUCT_BOTTOM_TOP:
1317             if (FIELD_OR_MBAFF_PICTURE)
1318                 cur->f.interlaced_frame = 1;
1319             else
1320                 // try to flag soft telecine progressive
1321                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1322             break;
1323         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1324         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1325             // Signal the possibility of telecined film externally (pic_struct 5,6)
1326             // From these hints, let the applications decide if they apply deinterlacing.
1327             cur->f.repeat_pict = 1;
1328             break;
1329         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1330             // Force progressive here, as doubling interlaced frame is a bad idea.
1331             cur->f.repeat_pict = 2;
1332             break;
1333         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1334             cur->f.repeat_pict = 4;
1335             break;
1336         }
1337
1338         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1339             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1340     }else{
1341         /* Derive interlacing flag from used decoding process. */
1342         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1343     }
1344     h->prev_interlaced_frame = cur->f.interlaced_frame;
1345
1346     if (cur->field_poc[0] != cur->field_poc[1]){
1347         /* Derive top_field_first from field pocs. */
1348         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1349     }else{
1350         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1351             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1352             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1353               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1354                 cur->f.top_field_first = 1;
1355             else
1356                 cur->f.top_field_first = 0;
1357         }else{
1358             /* Most likely progressive */
1359             cur->f.top_field_first = 0;
1360         }
1361     }
1362
1363     //FIXME do something with unavailable reference frames
1364
1365     /* Sort B-frames into display order */
1366
1367     if(h->sps.bitstream_restriction_flag
1368        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1369         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1370         s->low_delay = 0;
1371     }
1372
1373     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1374        && !h->sps.bitstream_restriction_flag){
1375         s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
1376         s->low_delay= 0;
1377     }
1378
1379     pics = 0;
1380     while(h->delayed_pic[pics]) pics++;
1381
1382     assert(pics <= MAX_DELAYED_PIC_COUNT);
1383
1384     h->delayed_pic[pics++] = cur;
1385     if (cur->f.reference == 0)
1386         cur->f.reference = DELAYED_PIC_REF;
1387
1388     out = h->delayed_pic[0];
1389     out_idx = 0;
1390     for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
1391         if(h->delayed_pic[i]->poc < out->poc){
1392             out = h->delayed_pic[i];
1393             out_idx = i;
1394         }
1395     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset))
1396         h->next_outputed_poc= INT_MIN;
1397     out_of_order = out->poc < h->next_outputed_poc;
1398
1399     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1400         { }
1401     else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
1402        || (s->low_delay &&
1403         ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2)
1404          || cur->f.pict_type == AV_PICTURE_TYPE_B)))
1405     {
1406         s->low_delay = 0;
1407         s->avctx->has_b_frames++;
1408     }
1409
1410     if(out_of_order || pics > s->avctx->has_b_frames){
1411         out->f.reference &= ~DELAYED_PIC_REF;
1412         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1413                          // or else the first thread can release the picture and reuse it unsafely
1414         for(i=out_idx; h->delayed_pic[i]; i++)
1415             h->delayed_pic[i] = h->delayed_pic[i+1];
1416     }
1417     if(!out_of_order && pics > s->avctx->has_b_frames){
1418         h->next_output_pic = out;
1419         if (out_idx == 0 && h->delayed_pic[0] && (h->delayed_pic[0]->f.key_frame || h->delayed_pic[0]->mmco_reset)) {
1420             h->next_outputed_poc = INT_MIN;
1421         } else
1422             h->next_outputed_poc = out->poc;
1423     }else{
1424         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1425     }
1426
1427     if (setup_finished)
1428         ff_thread_finish_setup(s->avctx);
1429 }
1430
1431 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int chroma444, int simple){
1432     MpegEncContext * const s = &h->s;
1433     uint8_t *top_border;
1434     int top_idx = 1;
1435     const int pixel_shift = h->pixel_shift;
1436
1437     src_y  -=   linesize;
1438     src_cb -= uvlinesize;
1439     src_cr -= uvlinesize;
1440
1441     if(!simple && FRAME_MBAFF){
1442         if(s->mb_y&1){
1443             if(!MB_MBAFF){
1444                 top_border = h->top_borders[0][s->mb_x];
1445                 AV_COPY128(top_border, src_y + 15*linesize);
1446                 if (pixel_shift)
1447                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1448                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1449                     if(chroma444){
1450                         if (pixel_shift){
1451                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1452                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1453                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1454                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1455                         } else {
1456                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1457                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1458                         }
1459                     } else {
1460                         if (pixel_shift) {
1461                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1462                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1463                         } else {
1464                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1465                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1466                         }
1467                     }
1468                 }
1469             }
1470         }else if(MB_MBAFF){
1471             top_idx = 0;
1472         }else
1473             return;
1474     }
1475
1476     top_border = h->top_borders[top_idx][s->mb_x];
1477     // There are two lines saved, the line above the the top macroblock of a pair,
1478     // and the line above the bottom macroblock
1479     AV_COPY128(top_border, src_y + 16*linesize);
1480     if (pixel_shift)
1481         AV_COPY128(top_border+16, src_y+16*linesize+16);
1482
1483     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1484         if(chroma444){
1485             if (pixel_shift){
1486                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1487                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1488                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1489                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1490             } else {
1491                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1492                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1493             }
1494         } else {
1495             if (pixel_shift) {
1496                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1497                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1498             } else {
1499                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1500                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1501             }
1502         }
1503     }
1504 }
1505
1506 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1507                                   uint8_t *src_cb, uint8_t *src_cr,
1508                                   int linesize, int uvlinesize,
1509                                   int xchg, int chroma444,
1510                                   int simple, int pixel_shift){
1511     MpegEncContext * const s = &h->s;
1512     int deblock_topleft;
1513     int deblock_top;
1514     int top_idx = 1;
1515     uint8_t *top_border_m1;
1516     uint8_t *top_border;
1517
1518     if(!simple && FRAME_MBAFF){
1519         if(s->mb_y&1){
1520             if(!MB_MBAFF)
1521                 return;
1522         }else{
1523             top_idx = MB_MBAFF ? 0 : 1;
1524         }
1525     }
1526
1527     if(h->deblocking_filter == 2) {
1528         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1529         deblock_top     = h->top_type;
1530     } else {
1531         deblock_topleft = (s->mb_x > 0);
1532         deblock_top     = (s->mb_y > !!MB_FIELD);
1533     }
1534
1535     src_y  -=   linesize + 1 + pixel_shift;
1536     src_cb -= uvlinesize + 1 + pixel_shift;
1537     src_cr -= uvlinesize + 1 + pixel_shift;
1538
1539     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1540     top_border    = h->top_borders[top_idx][s->mb_x];
1541
1542 #define XCHG(a,b,xchg)\
1543     if (pixel_shift) {\
1544         if (xchg) {\
1545             AV_SWAP64(b+0,a+0);\
1546             AV_SWAP64(b+8,a+8);\
1547         } else {\
1548             AV_COPY128(b,a); \
1549         }\
1550     } else \
1551 if (xchg) AV_SWAP64(b,a);\
1552 else      AV_COPY64(b,a);
1553
1554     if(deblock_top){
1555         if(deblock_topleft){
1556             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1557         }
1558         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1559         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1560         if(s->mb_x+1 < s->mb_width){
1561             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1562         }
1563     }
1564     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1565         if(chroma444){
1566             if(deblock_topleft){
1567                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1568                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1569             }
1570             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1571             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1572             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1573             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1574             if(s->mb_x+1 < s->mb_width){
1575                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1576                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1577             }
1578         } else {
1579             if(deblock_top){
1580                 if(deblock_topleft){
1581                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1582                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1583                 }
1584                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1585                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1586             }
1587         }
1588     }
1589 }
1590
1591 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1592     if (high_bit_depth) {
1593         return AV_RN32A(((int32_t*)mb) + index);
1594     } else
1595         return AV_RN16A(mb + index);
1596 }
1597
1598 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1599     if (high_bit_depth) {
1600         AV_WN32A(((int32_t*)mb) + index, value);
1601     } else
1602         AV_WN16A(mb + index, value);
1603 }
1604
1605 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1606                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1607 {
1608     MpegEncContext * const s = &h->s;
1609     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1610     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1611     int i;
1612     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1613     block_offset += 16*p;
1614     if(IS_INTRA4x4(mb_type)){
1615         if(simple || !s->encoding){
1616             if(IS_8x8DCT(mb_type)){
1617                 if(transform_bypass){
1618                     idct_dc_add =
1619                     idct_add    = s->dsp.add_pixels8;
1620                 }else{
1621                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1622                     idct_add    = h->h264dsp.h264_idct8_add;
1623                 }
1624                 for(i=0; i<16; i+=4){
1625                     uint8_t * const ptr= dest_y + block_offset[i];
1626                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1627                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1628                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1629                     }else{
1630                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1631                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1632                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1633                         if(nnz){
1634                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1635                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1636                             else
1637                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1638                         }
1639                     }
1640                 }
1641             }else{
1642                 if(transform_bypass){
1643                     idct_dc_add =
1644                     idct_add    = s->dsp.add_pixels4;
1645                 }else{
1646                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1647                     idct_add    = h->h264dsp.h264_idct_add;
1648                 }
1649                 for(i=0; i<16; i++){
1650                     uint8_t * const ptr= dest_y + block_offset[i];
1651                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1652
1653                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1654                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1655                     }else{
1656                         uint8_t *topright;
1657                         int nnz, tr;
1658                         uint64_t tr_high;
1659                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1660                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1661                             assert(s->mb_y || linesize <= block_offset[i]);
1662                             if(!topright_avail){
1663                                 if (pixel_shift) {
1664                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1665                                     topright= (uint8_t*) &tr_high;
1666                                 } else {
1667                                     tr= ptr[3 - linesize]*0x01010101;
1668                                     topright= (uint8_t*) &tr;
1669                                 }
1670                             }else
1671                                 topright= ptr + (4 << pixel_shift) - linesize;
1672                         }else
1673                             topright= NULL;
1674
1675                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1676                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1677                         if(nnz){
1678                             if(is_h264){
1679                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1680                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1681                                 else
1682                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1683                             }else
1684                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1685                         }
1686                     }
1687                 }
1688             }
1689         }
1690     }else{
1691         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1692         if(is_h264){
1693             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1694                 if(!transform_bypass)
1695                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1696                 else{
1697                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1698                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1699                     for(i = 0; i < 16; i++)
1700                         dctcoef_set(h->mb+p*256, pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1701                 }
1702             }
1703         }else
1704             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1705     }
1706 }
1707
1708 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1709                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1710 {
1711     MpegEncContext * const s = &h->s;
1712     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1713     int i;
1714     block_offset += 16*p;
1715     if(!IS_INTRA4x4(mb_type)){
1716         if(is_h264){
1717             if(IS_INTRA16x16(mb_type)){
1718                 if(transform_bypass){
1719                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1720                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1721                     }else{
1722                         for(i=0; i<16; i++){
1723                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1724                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1725                         }
1726                     }
1727                 }else{
1728                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1729                 }
1730             }else if(h->cbp&15){
1731                 if(transform_bypass){
1732                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1733                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1734                     for(i=0; i<16; i+=di){
1735                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1736                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1737                         }
1738                     }
1739                 }else{
1740                     if(IS_8x8DCT(mb_type)){
1741                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1742                     }else{
1743                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1744                     }
1745                 }
1746             }
1747         }else{
1748             for(i=0; i<16; i++){
1749                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1750                     uint8_t * const ptr= dest_y + block_offset[i];
1751                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1752                 }
1753             }
1754         }
1755     }
1756 }
1757
1758 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift){
1759     MpegEncContext * const s = &h->s;
1760     const int mb_x= s->mb_x;
1761     const int mb_y= s->mb_y;
1762     const int mb_xy= h->mb_xy;
1763     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1764     uint8_t  *dest_y, *dest_cb, *dest_cr;
1765     int linesize, uvlinesize /*dct_offset*/;
1766     int i, j;
1767     int *block_offset = &h->block_offset[0];
1768     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1769     /* is_h264 should always be true if SVQ3 is disabled. */
1770     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1771     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1772
1773     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1774     dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1775     dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) *  8;
1776
1777     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1778     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1779
1780     h->list_counts[mb_xy]= h->list_count;
1781
1782     if (!simple && MB_FIELD) {
1783         linesize   = h->mb_linesize   = s->linesize * 2;
1784         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1785         block_offset = &h->block_offset[48];
1786         if(mb_y&1){ //FIXME move out of this function?
1787             dest_y -= s->linesize*15;
1788             dest_cb-= s->uvlinesize*7;
1789             dest_cr-= s->uvlinesize*7;
1790         }
1791         if(FRAME_MBAFF) {
1792             int list;
1793             for(list=0; list<h->list_count; list++){
1794                 if(!USES_LIST(mb_type, list))
1795                     continue;
1796                 if(IS_16X16(mb_type)){
1797                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1798                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1799                 }else{
1800                     for(i=0; i<16; i+=4){
1801                         int ref = h->ref_cache[list][scan8[i]];
1802                         if(ref >= 0)
1803                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1804                     }
1805                 }
1806             }
1807         }
1808     } else {
1809         linesize   = h->mb_linesize   = s->linesize;
1810         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1811 //        dct_offset = s->linesize * 16;
1812     }
1813
1814     if (!simple && IS_INTRA_PCM(mb_type)) {
1815         if (pixel_shift) {
1816             const int bit_depth = h->sps.bit_depth_luma;
1817             int j;
1818             GetBitContext gb;
1819             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1820
1821             for (i = 0; i < 16; i++) {
1822                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1823                 for (j = 0; j < 16; j++)
1824                     tmp_y[j] = get_bits(&gb, bit_depth);
1825             }
1826             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1827                 if (!h->sps.chroma_format_idc) {
1828                     for (i = 0; i < 8; i++) {
1829                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1830                         for (j = 0; j < 8; j++) {
1831                             tmp_cb[j] = 1 << (bit_depth - 1);
1832                         }
1833                     }
1834                     for (i = 0; i < 8; i++) {
1835                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1836                         for (j = 0; j < 8; j++) {
1837                             tmp_cr[j] = 1 << (bit_depth - 1);
1838                         }
1839                     }
1840                 } else {
1841                     for (i = 0; i < 8; i++) {
1842                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
1843                         for (j = 0; j < 8; j++)
1844                             tmp_cb[j] = get_bits(&gb, bit_depth);
1845                     }
1846                     for (i = 0; i < 8; i++) {
1847                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
1848                         for (j = 0; j < 8; j++)
1849                             tmp_cr[j] = get_bits(&gb, bit_depth);
1850                     }
1851                 }
1852             }
1853         } else {
1854             for (i=0; i<16; i++) {
1855                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1856             }
1857             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1858                 if (!h->sps.chroma_format_idc) {
1859                     for (i = 0; i < 8; i++) {
1860                         memset(dest_cb + i*uvlinesize, 128, 8);
1861                         memset(dest_cr + i*uvlinesize, 128, 8);
1862                     }
1863                 } else {
1864                     for (i = 0; i < 8; i++) {
1865                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
1866                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
1867                     }
1868                 }
1869             }
1870         }
1871     } else {
1872         if(IS_INTRA(mb_type)){
1873             if(h->deblocking_filter)
1874                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
1875
1876             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1877                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1878                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1879             }
1880
1881             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1882
1883             if(h->deblocking_filter)
1884                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
1885         }else if(is_h264){
1886             hl_motion(h, dest_y, dest_cb, dest_cr,
1887                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1888                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1889                       h->h264dsp.weight_h264_pixels_tab,
1890                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 0);
1891         }
1892
1893         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
1894
1895         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1896             uint8_t *dest[2] = {dest_cb, dest_cr};
1897             if(transform_bypass){
1898                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1899                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
1900                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
1901                 }else{
1902                     idct_add = s->dsp.add_pixels4;
1903                     for(j=1; j<3; j++){
1904                         for(i=j*16; i<j*16+4; i++){
1905                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
1906                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
1907                         }
1908                     }
1909                 }
1910             }else{
1911                 if(is_h264){
1912                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
1913                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1914                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
1915                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1916                     h->h264dsp.h264_idct_add8(dest, block_offset,
1917                                               h->mb, uvlinesize,
1918                                               h->non_zero_count_cache);
1919                 }else{
1920                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*1, h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1921                     h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + 16*16*2, h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1922                     for(j=1; j<3; j++){
1923                         for(i=j*16; i<j*16+4; i++){
1924                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1925                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
1926                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
1927                             }
1928                         }
1929                     }
1930                 }
1931             }
1932         }
1933     }
1934     if(h->cbp || IS_INTRA(mb_type))
1935     {
1936         s->dsp.clear_blocks(h->mb);
1937         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
1938     }
1939 }
1940
1941 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
1942     MpegEncContext * const s = &h->s;
1943     const int mb_x= s->mb_x;
1944     const int mb_y= s->mb_y;
1945     const int mb_xy= h->mb_xy;
1946     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1947     uint8_t  *dest[3];
1948     int linesize;
1949     int i, j, p;
1950     int *block_offset = &h->block_offset[0];
1951     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1952     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
1953
1954     for (p = 0; p < plane_count; p++)
1955     {
1956         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
1957         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1958     }
1959
1960     h->list_counts[mb_xy]= h->list_count;
1961
1962     if (!simple && MB_FIELD) {
1963         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
1964         block_offset = &h->block_offset[48];
1965         if(mb_y&1) //FIXME move out of this function?
1966             for (p = 0; p < 3; p++)
1967                 dest[p] -= s->linesize*15;
1968         if(FRAME_MBAFF) {
1969             int list;
1970             for(list=0; list<h->list_count; list++){
1971                 if(!USES_LIST(mb_type, list))
1972                     continue;
1973                 if(IS_16X16(mb_type)){
1974                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1975                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1976                 }else{
1977                     for(i=0; i<16; i+=4){
1978                         int ref = h->ref_cache[list][scan8[i]];
1979                         if(ref >= 0)
1980                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1981                     }
1982                 }
1983             }
1984         }
1985     } else {
1986         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
1987     }
1988
1989     if (!simple && IS_INTRA_PCM(mb_type)) {
1990         if (pixel_shift) {
1991             const int bit_depth = h->sps.bit_depth_luma;
1992             GetBitContext gb;
1993             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
1994
1995             for (p = 0; p < plane_count; p++) {
1996                 for (i = 0; i < 16; i++) {
1997                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
1998                     for (j = 0; j < 16; j++)
1999                         tmp[j] = get_bits(&gb, bit_depth);
2000                 }
2001             }
2002         } else {
2003             for (p = 0; p < plane_count; p++) {
2004                 for (i = 0; i < 16; i++) {
2005                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2006                 }
2007             }
2008         }
2009     } else {
2010         if(IS_INTRA(mb_type)){
2011             if(h->deblocking_filter)
2012                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2013
2014             for (p = 0; p < plane_count; p++)
2015                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2016
2017             if(h->deblocking_filter)
2018                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2019         }else{
2020             hl_motion(h, dest[0], dest[1], dest[2],
2021                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2022                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2023                       h->h264dsp.weight_h264_pixels_tab,
2024                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 1);
2025         }
2026
2027         for (p = 0; p < plane_count; p++)
2028             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2029     }
2030     if(h->cbp || IS_INTRA(mb_type))
2031     {
2032         s->dsp.clear_blocks(h->mb);
2033         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2034     }
2035 }
2036
2037 /**
2038  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2039  */
2040 #define hl_decode_mb_simple(sh, bits) \
2041 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2042     hl_decode_mb_internal(h, 1, sh); \
2043 }
2044 hl_decode_mb_simple(0, 8);
2045 hl_decode_mb_simple(1, 16);
2046
2047 /**
2048  * Process a macroblock; this handles edge cases, such as interlacing.
2049  */
2050 static void av_noinline hl_decode_mb_complex(H264Context *h){
2051     hl_decode_mb_internal(h, 0, h->pixel_shift);
2052 }
2053
2054 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2055     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2056 }
2057
2058 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2059     hl_decode_mb_444_internal(h, 1, 0);
2060 }
2061
2062 void ff_h264_hl_decode_mb(H264Context *h){
2063     MpegEncContext * const s = &h->s;
2064     const int mb_xy= h->mb_xy;
2065     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2066     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2067
2068     if (CHROMA444) {
2069         if(is_complex || h->pixel_shift)
2070             hl_decode_mb_444_complex(h);
2071         else
2072             hl_decode_mb_444_simple(h);
2073     } else if (is_complex) {
2074         hl_decode_mb_complex(h);
2075     } else if (h->pixel_shift) {
2076         hl_decode_mb_simple_16(h);
2077     } else
2078         hl_decode_mb_simple_8(h);
2079 }
2080
2081 static int pred_weight_table(H264Context *h){
2082     MpegEncContext * const s = &h->s;
2083     int list, i;
2084     int luma_def, chroma_def;
2085
2086     h->use_weight= 0;
2087     h->use_weight_chroma= 0;
2088     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2089     if(h->sps.chroma_format_idc)
2090         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2091     luma_def = 1<<h->luma_log2_weight_denom;
2092     chroma_def = 1<<h->chroma_log2_weight_denom;
2093
2094     for(list=0; list<2; list++){
2095         h->luma_weight_flag[list]   = 0;
2096         h->chroma_weight_flag[list] = 0;
2097         for(i=0; i<h->ref_count[list]; i++){
2098             int luma_weight_flag, chroma_weight_flag;
2099
2100             luma_weight_flag= get_bits1(&s->gb);
2101             if(luma_weight_flag){
2102                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2103                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2104                 if(   h->luma_weight[i][list][0] != luma_def
2105                    || h->luma_weight[i][list][1] != 0) {
2106                     h->use_weight= 1;
2107                     h->luma_weight_flag[list]= 1;
2108                 }
2109             }else{
2110                 h->luma_weight[i][list][0]= luma_def;
2111                 h->luma_weight[i][list][1]= 0;
2112             }
2113
2114             if(h->sps.chroma_format_idc){
2115                 chroma_weight_flag= get_bits1(&s->gb);
2116                 if(chroma_weight_flag){
2117                     int j;
2118                     for(j=0; j<2; j++){
2119                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2120                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2121                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2122                            || h->chroma_weight[i][list][j][1] != 0) {
2123                             h->use_weight_chroma= 1;
2124                             h->chroma_weight_flag[list]= 1;
2125                         }
2126                     }
2127                 }else{
2128                     int j;
2129                     for(j=0; j<2; j++){
2130                         h->chroma_weight[i][list][j][0]= chroma_def;
2131                         h->chroma_weight[i][list][j][1]= 0;
2132                     }
2133                 }
2134             }
2135         }
2136         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2137     }
2138     h->use_weight= h->use_weight || h->use_weight_chroma;
2139     return 0;
2140 }
2141
2142 /**
2143  * Initialize implicit_weight table.
2144  * @param field  0/1 initialize the weight for interlaced MBAFF
2145  *                -1 initializes the rest
2146  */
2147 static void implicit_weight_table(H264Context *h, int field){
2148     MpegEncContext * const s = &h->s;
2149     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2150
2151     for (i = 0; i < 2; i++) {
2152         h->luma_weight_flag[i]   = 0;
2153         h->chroma_weight_flag[i] = 0;
2154     }
2155
2156     if(field < 0){
2157         cur_poc = s->current_picture_ptr->poc;
2158     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2159        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2160         h->use_weight= 0;
2161         h->use_weight_chroma= 0;
2162         return;
2163     }
2164         ref_start= 0;
2165         ref_count0= h->ref_count[0];
2166         ref_count1= h->ref_count[1];
2167     }else{
2168         cur_poc = s->current_picture_ptr->field_poc[field];
2169         ref_start= 16;
2170         ref_count0= 16+2*h->ref_count[0];
2171         ref_count1= 16+2*h->ref_count[1];
2172     }
2173
2174     h->use_weight= 2;
2175     h->use_weight_chroma= 2;
2176     h->luma_log2_weight_denom= 5;
2177     h->chroma_log2_weight_denom= 5;
2178
2179     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2180         int poc0 = h->ref_list[0][ref0].poc;
2181         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2182             int w = 32;
2183             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2184                 int poc1 = h->ref_list[1][ref1].poc;
2185                 int td = av_clip(poc1 - poc0, -128, 127);
2186                 if(td){
2187                     int tb = av_clip(cur_poc - poc0, -128, 127);
2188                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2189                     int dist_scale_factor = (tb*tx + 32) >> 8;
2190                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2191                         w = 64 - dist_scale_factor;
2192                 }
2193             }
2194             if(field<0){
2195                 h->implicit_weight[ref0][ref1][0]=
2196                 h->implicit_weight[ref0][ref1][1]= w;
2197             }else{
2198                 h->implicit_weight[ref0][ref1][field]=w;
2199             }
2200         }
2201     }
2202 }
2203
2204 /**
2205  * instantaneous decoder refresh.
2206  */
2207 static void idr(H264Context *h){
2208     ff_h264_remove_all_refs(h);
2209     h->prev_frame_num= 0;
2210     h->prev_frame_num_offset= 0;
2211     h->prev_poc_msb=
2212     h->prev_poc_lsb= 0;
2213 }
2214
2215 /* forget old pics after a seek */
2216 static void flush_dpb(AVCodecContext *avctx){
2217     H264Context *h= avctx->priv_data;
2218     int i;
2219     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
2220         if(h->delayed_pic[i])
2221             h->delayed_pic[i]->f.reference = 0;
2222         h->delayed_pic[i]= NULL;
2223     }
2224     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2225     h->prev_interlaced_frame = 1;
2226     idr(h);
2227     if(h->s.current_picture_ptr)
2228         h->s.current_picture_ptr->f.reference = 0;
2229     h->s.first_field= 0;
2230     ff_h264_reset_sei(h);
2231     ff_mpeg_flush(avctx);
2232 }
2233
2234 static int init_poc(H264Context *h){
2235     MpegEncContext * const s = &h->s;
2236     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2237     int field_poc[2];
2238     Picture *cur = s->current_picture_ptr;
2239
2240     h->frame_num_offset= h->prev_frame_num_offset;
2241     if(h->frame_num < h->prev_frame_num)
2242         h->frame_num_offset += max_frame_num;
2243
2244     if(h->sps.poc_type==0){
2245         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2246
2247         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2248             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2249         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2250             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2251         else
2252             h->poc_msb = h->prev_poc_msb;
2253 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2254         field_poc[0] =
2255         field_poc[1] = h->poc_msb + h->poc_lsb;
2256         if(s->picture_structure == PICT_FRAME)
2257             field_poc[1] += h->delta_poc_bottom;
2258     }else if(h->sps.poc_type==1){
2259         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2260         int i;
2261
2262         if(h->sps.poc_cycle_length != 0)
2263             abs_frame_num = h->frame_num_offset + h->frame_num;
2264         else
2265             abs_frame_num = 0;
2266
2267         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2268             abs_frame_num--;
2269
2270         expected_delta_per_poc_cycle = 0;
2271         for(i=0; i < h->sps.poc_cycle_length; i++)
2272             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2273
2274         if(abs_frame_num > 0){
2275             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2276             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2277
2278             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2279             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2280                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2281         } else
2282             expectedpoc = 0;
2283
2284         if(h->nal_ref_idc == 0)
2285             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2286
2287         field_poc[0] = expectedpoc + h->delta_poc[0];
2288         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2289
2290         if(s->picture_structure == PICT_FRAME)
2291             field_poc[1] += h->delta_poc[1];
2292     }else{
2293         int poc= 2*(h->frame_num_offset + h->frame_num);
2294
2295         if(!h->nal_ref_idc)
2296             poc--;
2297
2298         field_poc[0]= poc;
2299         field_poc[1]= poc;
2300     }
2301
2302     if(s->picture_structure != PICT_BOTTOM_FIELD)
2303         s->current_picture_ptr->field_poc[0]= field_poc[0];
2304     if(s->picture_structure != PICT_TOP_FIELD)
2305         s->current_picture_ptr->field_poc[1]= field_poc[1];
2306     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2307
2308     return 0;
2309 }
2310
2311
2312 /**
2313  * initialize scan tables
2314  */
2315 static void init_scan_tables(H264Context *h){
2316     int i;
2317     for(i=0; i<16; i++){
2318 #define T(x) (x>>2) | ((x<<2) & 0xF)
2319         h->zigzag_scan[i] = T(zigzag_scan[i]);
2320         h-> field_scan[i] = T( field_scan[i]);
2321 #undef T
2322     }
2323     for(i=0; i<64; i++){
2324 #define T(x) (x>>3) | ((x&7)<<3)
2325         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2326         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2327         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2328         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2329 #undef T
2330     }
2331     if(h->sps.transform_bypass){ //FIXME same ugly
2332         h->zigzag_scan_q0          = zigzag_scan;
2333         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2334         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2335         h->field_scan_q0           = field_scan;
2336         h->field_scan8x8_q0        = field_scan8x8;
2337         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2338     }else{
2339         h->zigzag_scan_q0          = h->zigzag_scan;
2340         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2341         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2342         h->field_scan_q0           = h->field_scan;
2343         h->field_scan8x8_q0        = h->field_scan8x8;
2344         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2345     }
2346 }
2347
2348 static int field_end(H264Context *h, int in_setup){
2349     MpegEncContext * const s = &h->s;
2350     AVCodecContext * const avctx= s->avctx;
2351     int err = 0;
2352     s->mb_y= 0;
2353
2354     if (!in_setup && !s->dropable)
2355         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2356                                  s->picture_structure==PICT_BOTTOM_FIELD);
2357
2358     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2359         ff_vdpau_h264_set_reference_frames(s);
2360
2361     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2362         if(!s->dropable) {
2363             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2364             h->prev_poc_msb= h->poc_msb;
2365             h->prev_poc_lsb= h->poc_lsb;
2366         }
2367         h->prev_frame_num_offset= h->frame_num_offset;
2368         h->prev_frame_num= h->frame_num;
2369         h->outputed_poc = h->next_outputed_poc;
2370     }
2371
2372     if (avctx->hwaccel) {
2373         if (avctx->hwaccel->end_frame(avctx) < 0)
2374             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2375     }
2376
2377     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2378         ff_vdpau_h264_picture_complete(s);
2379
2380     /*
2381      * FIXME: Error handling code does not seem to support interlaced
2382      * when slices span multiple rows
2383      * The ff_er_add_slice calls don't work right for bottom
2384      * fields; they cause massive erroneous error concealing
2385      * Error marking covers both fields (top and bottom).
2386      * This causes a mismatched s->error_count
2387      * and a bad error table. Further, the error count goes to
2388      * INT_MAX when called for bottom field, because mb_y is
2389      * past end by one (callers fault) and resync_mb_y != 0
2390      * causes problems for the first MB line, too.
2391      */
2392     if (!FIELD_PICTURE)
2393         ff_er_frame_end(s);
2394
2395     MPV_frame_end(s);
2396
2397     h->current_slice=0;
2398
2399     return err;
2400 }
2401
2402 /**
2403  * Replicate H264 "master" context to thread contexts.
2404  */
2405 static void clone_slice(H264Context *dst, H264Context *src)
2406 {
2407     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2408     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2409     dst->s.current_picture      = src->s.current_picture;
2410     dst->s.linesize             = src->s.linesize;
2411     dst->s.uvlinesize           = src->s.uvlinesize;
2412     dst->s.first_field          = src->s.first_field;
2413
2414     dst->prev_poc_msb           = src->prev_poc_msb;
2415     dst->prev_poc_lsb           = src->prev_poc_lsb;
2416     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2417     dst->prev_frame_num         = src->prev_frame_num;
2418     dst->short_ref_count        = src->short_ref_count;
2419
2420     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2421     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2422     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2423     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2424
2425     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2426     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2427 }
2428
2429 /**
2430  * computes profile from profile_idc and constraint_set?_flags
2431  *
2432  * @param sps SPS
2433  *
2434  * @return profile as defined by FF_PROFILE_H264_*
2435  */
2436 int ff_h264_get_profile(SPS *sps)
2437 {
2438     int profile = sps->profile_idc;
2439
2440     switch(sps->profile_idc) {
2441     case FF_PROFILE_H264_BASELINE:
2442         // constraint_set1_flag set to 1
2443         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2444         break;
2445     case FF_PROFILE_H264_HIGH_10:
2446     case FF_PROFILE_H264_HIGH_422:
2447     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2448         // constraint_set3_flag set to 1
2449         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2450         break;
2451     }
2452
2453     return profile;
2454 }
2455
2456 /**
2457  * decodes a slice header.
2458  * This will also call MPV_common_init() and frame_start() as needed.
2459  *
2460  * @param h h264context
2461  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2462  *
2463  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2464  */
2465 static int decode_slice_header(H264Context *h, H264Context *h0){
2466     MpegEncContext * const s = &h->s;
2467     MpegEncContext * const s0 = &h0->s;
2468     unsigned int first_mb_in_slice;
2469     unsigned int pps_id;
2470     int num_ref_idx_active_override_flag;
2471     unsigned int slice_type, tmp, i, j;
2472     int default_ref_list_done = 0;
2473     int last_pic_structure;
2474
2475     s->dropable= h->nal_ref_idc == 0;
2476
2477     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2478     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2479         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2480         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2481     }else{
2482         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2483         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2484     }
2485
2486     first_mb_in_slice= get_ue_golomb(&s->gb);
2487
2488     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2489         if(h0->current_slice && FIELD_PICTURE){
2490             field_end(h, 1);
2491         }
2492
2493         h0->current_slice = 0;
2494         if (!s0->first_field)
2495             s->current_picture_ptr= NULL;
2496     }
2497
2498     slice_type= get_ue_golomb_31(&s->gb);
2499     if(slice_type > 9){
2500         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);
2501         return -1;
2502     }
2503     if(slice_type > 4){
2504         slice_type -= 5;
2505         h->slice_type_fixed=1;
2506     }else
2507         h->slice_type_fixed=0;
2508
2509     slice_type= golomb_to_pict_type[ slice_type ];
2510     if (slice_type == AV_PICTURE_TYPE_I
2511         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2512         default_ref_list_done = 1;
2513     }
2514     h->slice_type= slice_type;
2515     h->slice_type_nos= slice_type & 3;
2516
2517     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2518
2519     pps_id= get_ue_golomb(&s->gb);
2520     if(pps_id>=MAX_PPS_COUNT){
2521         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2522         return -1;
2523     }
2524     if(!h0->pps_buffers[pps_id]) {
2525         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2526         return -1;
2527     }
2528     h->pps= *h0->pps_buffers[pps_id];
2529
2530     if(!h0->sps_buffers[h->pps.sps_id]) {
2531         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2532         return -1;
2533     }
2534     h->sps = *h0->sps_buffers[h->pps.sps_id];
2535
2536     s->avctx->profile = ff_h264_get_profile(&h->sps);
2537     s->avctx->level   = h->sps.level_idc;
2538     s->avctx->refs    = h->sps.ref_frame_count;
2539
2540     if(h == h0 && h->dequant_coeff_pps != pps_id){
2541         h->dequant_coeff_pps = pps_id;
2542         init_dequant_tables(h);
2543     }
2544
2545     s->mb_width= h->sps.mb_width;
2546     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2547
2548     h->b_stride=  s->mb_width*4;
2549
2550     s->width = 16*s->mb_width - (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2551     if(h->sps.frame_mbs_only_flag)
2552         s->height= 16*s->mb_height - (2>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2553     else
2554         s->height= 16*s->mb_height - (4>>CHROMA444)*FFMIN(h->sps.crop_bottom, (8<<CHROMA444)-1);
2555
2556     if (s->context_initialized
2557         && (   s->width != s->avctx->width || s->height != s->avctx->height
2558             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2559         if(h != h0) {
2560             av_log_missing_feature(s->avctx, "Width/height changing with threads is", 0);
2561             return -1;   // width / height changed during parallelized decoding
2562         }
2563         free_tables(h, 0);
2564         flush_dpb(s->avctx);
2565         MPV_common_end(s);
2566     }
2567     if (!s->context_initialized) {
2568         if (h != h0) {
2569             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2570             return -1;
2571         }
2572
2573         avcodec_set_dimensions(s->avctx, s->width, s->height);
2574         s->avctx->sample_aspect_ratio= h->sps.sar;
2575         av_assert0(s->avctx->sample_aspect_ratio.den);
2576
2577         h->s.avctx->coded_width = 16*s->mb_width;
2578         h->s.avctx->coded_height = 16*s->mb_height;
2579
2580         if(h->sps.video_signal_type_present_flag){
2581             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2582             if(h->sps.colour_description_present_flag){
2583                 s->avctx->color_primaries = h->sps.color_primaries;
2584                 s->avctx->color_trc       = h->sps.color_trc;
2585                 s->avctx->colorspace      = h->sps.colorspace;
2586             }
2587         }
2588
2589         if(h->sps.timing_info_present_flag){
2590             int64_t den= h->sps.time_scale;
2591             if(h->x264_build < 44U)
2592                 den *= 2;
2593             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2594                       h->sps.num_units_in_tick, den, 1<<30);
2595         }
2596
2597         switch (h->sps.bit_depth_luma) {
2598             case 9 :
2599                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P9 : PIX_FMT_YUV420P9;
2600                 break;
2601             case 10 :
2602                 s->avctx->pix_fmt = CHROMA444 ? PIX_FMT_YUV444P10 : PIX_FMT_YUV420P10;
2603                 break;
2604             default:
2605                 if (CHROMA444){
2606                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2607                 }else{
2608                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2609                                                              s->avctx->codec->pix_fmts ?
2610                                                              s->avctx->codec->pix_fmts :
2611                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2612                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2613                                                              ff_hwaccel_pixfmt_list_420);
2614                 }
2615         }
2616
2617         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2618
2619         if (MPV_common_init(s) < 0) {
2620             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2621             return -1;
2622         }
2623         s->first_field = 0;
2624         h->prev_interlaced_frame = 1;
2625
2626         init_scan_tables(h);
2627         if (ff_h264_alloc_tables(h) < 0) {
2628             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2629             return AVERROR(ENOMEM);
2630         }
2631
2632         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2633             if (context_init(h) < 0) {
2634                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2635                 return -1;
2636             }
2637         } else {
2638             for(i = 1; i < s->avctx->thread_count; i++) {
2639                 H264Context *c;
2640                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2641                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2642                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2643                 c->h264dsp = h->h264dsp;
2644                 c->sps = h->sps;
2645                 c->pps = h->pps;
2646                 c->pixel_shift = h->pixel_shift;
2647                 init_scan_tables(c);
2648                 clone_tables(c, h, i);
2649             }
2650
2651             for(i = 0; i < s->avctx->thread_count; i++)
2652                 if (context_init(h->thread_context[i]) < 0) {
2653                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2654                     return -1;
2655                 }
2656         }
2657     }
2658
2659     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2660
2661     h->mb_mbaff = 0;
2662     h->mb_aff_frame = 0;
2663     last_pic_structure = s0->picture_structure;
2664     if(h->sps.frame_mbs_only_flag){
2665         s->picture_structure= PICT_FRAME;
2666     }else{
2667         if(get_bits1(&s->gb)) { //field_pic_flag
2668             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2669         } else {
2670             s->picture_structure= PICT_FRAME;
2671             h->mb_aff_frame = h->sps.mb_aff;
2672         }
2673     }
2674     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2675
2676     if(h0->current_slice == 0){
2677         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2678         if(h->frame_num != h->prev_frame_num) {
2679             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2680
2681             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2682
2683             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2684                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2685                 if (unwrap_prev_frame_num < 0)
2686                     unwrap_prev_frame_num += max_frame_num;
2687
2688                 h->prev_frame_num = unwrap_prev_frame_num;
2689             }
2690         }
2691
2692         while(h->frame_num !=  h->prev_frame_num &&
2693               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2694             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2695             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2696             if (ff_h264_frame_start(h) < 0)
2697                 return -1;
2698             h->prev_frame_num++;
2699             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2700             s->current_picture_ptr->frame_num= h->prev_frame_num;
2701             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2702             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2703             ff_generate_sliding_window_mmcos(h);
2704             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2705                 s->avctx->error_recognition >= FF_ER_EXPLODE)
2706                 return AVERROR_INVALIDDATA;
2707             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2708              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2709              * about there being no actual duplicates.
2710              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2711              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2712              * be fixed. */
2713             if (h->short_ref_count) {
2714                 if (prev) {
2715                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2716                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2717                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2718                     h->short_ref[0]->poc = prev->poc+2;
2719                 }
2720                 h->short_ref[0]->frame_num = h->prev_frame_num;
2721             }
2722         }
2723
2724         /* See if we have a decoded first field looking for a pair... */
2725         if (s0->first_field) {
2726             assert(s0->current_picture_ptr);
2727             assert(s0->current_picture_ptr->f.data[0]);
2728             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
2729
2730             /* figure out if we have a complementary field pair */
2731             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2732                 /*
2733                  * Previous field is unmatched. Don't display it, but let it
2734                  * remain for reference if marked as such.
2735                  */
2736                 s0->current_picture_ptr = NULL;
2737                 s0->first_field = FIELD_PICTURE;
2738
2739             } else {
2740                 if (h->nal_ref_idc &&
2741                         s0->current_picture_ptr->f.reference &&
2742                         s0->current_picture_ptr->frame_num != h->frame_num) {
2743                     /*
2744                      * This and previous field were reference, but had
2745                      * different frame_nums. Consider this field first in
2746                      * pair. Throw away previous field except for reference
2747                      * purposes.
2748                      */
2749                     s0->first_field = 1;
2750                     s0->current_picture_ptr = NULL;
2751
2752                 } else {
2753                     /* Second field in complementary pair */
2754                     s0->first_field = 0;
2755                 }
2756             }
2757
2758         } else {
2759             /* Frame or first field in a potentially complementary pair */
2760             assert(!s0->current_picture_ptr);
2761             s0->first_field = FIELD_PICTURE;
2762         }
2763
2764         if(!FIELD_PICTURE || s0->first_field) {
2765             if (ff_h264_frame_start(h) < 0) {
2766                 s0->first_field = 0;
2767                 return -1;
2768             }
2769         } else {
2770             ff_release_unused_pictures(s, 0);
2771         }
2772     }
2773     if(h != h0)
2774         clone_slice(h, h0);
2775
2776     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
2777
2778     assert(s->mb_num == s->mb_width * s->mb_height);
2779     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
2780        first_mb_in_slice                    >= s->mb_num){
2781         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
2782         return -1;
2783     }
2784     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
2785     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
2786     if (s->picture_structure == PICT_BOTTOM_FIELD)
2787         s->resync_mb_y = s->mb_y = s->mb_y + 1;
2788     assert(s->mb_y < s->mb_height);
2789
2790     if(s->picture_structure==PICT_FRAME){
2791         h->curr_pic_num=   h->frame_num;
2792         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
2793     }else{
2794         h->curr_pic_num= 2*h->frame_num + 1;
2795         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
2796     }
2797
2798     if(h->nal_unit_type == NAL_IDR_SLICE){
2799         get_ue_golomb(&s->gb); /* idr_pic_id */
2800     }
2801
2802     if(h->sps.poc_type==0){
2803         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2804
2805         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2806             h->delta_poc_bottom= get_se_golomb(&s->gb);
2807         }
2808     }
2809
2810     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2811         h->delta_poc[0]= get_se_golomb(&s->gb);
2812
2813         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2814             h->delta_poc[1]= get_se_golomb(&s->gb);
2815     }
2816
2817     init_poc(h);
2818
2819     if(h->pps.redundant_pic_cnt_present){
2820         h->redundant_pic_count= get_ue_golomb(&s->gb);
2821     }
2822
2823     //set defaults, might be overridden a few lines later
2824     h->ref_count[0]= h->pps.ref_count[0];
2825     h->ref_count[1]= h->pps.ref_count[1];
2826
2827     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
2828         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
2829             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2830         }
2831         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2832
2833         if(num_ref_idx_active_override_flag){
2834             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2835             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
2836                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2837
2838             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2839                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2840                 h->ref_count[0]= h->ref_count[1]= 1;
2841                 return -1;
2842             }
2843         }
2844         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
2845             h->list_count= 2;
2846         else
2847             h->list_count= 1;
2848     }else
2849         h->list_count= 0;
2850
2851     if(!default_ref_list_done){
2852         ff_h264_fill_default_ref_list(h);
2853     }
2854
2855     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2856         return -1;
2857
2858     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
2859         s->last_picture_ptr= &h->ref_list[0][0];
2860         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2861     }
2862     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
2863         s->next_picture_ptr= &h->ref_list[1][0];
2864         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2865     }
2866
2867     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
2868        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
2869         pred_weight_table(h);
2870     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2871         implicit_weight_table(h, -1);
2872     }else {
2873         h->use_weight = 0;
2874         for (i = 0; i < 2; i++) {
2875             h->luma_weight_flag[i]   = 0;
2876             h->chroma_weight_flag[i] = 0;
2877         }
2878     }
2879
2880     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
2881        s->avctx->error_recognition >= FF_ER_EXPLODE)
2882         return AVERROR_INVALIDDATA;
2883
2884     if(FRAME_MBAFF){
2885         ff_h264_fill_mbaff_ref_list(h);
2886
2887         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
2888             implicit_weight_table(h, 0);
2889             implicit_weight_table(h, 1);
2890         }
2891     }
2892
2893     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
2894         ff_h264_direct_dist_scale_factor(h);
2895     ff_h264_direct_ref_list_init(h);
2896
2897     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
2898         tmp = get_ue_golomb_31(&s->gb);
2899         if(tmp > 2){
2900             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2901             return -1;
2902         }
2903         h->cabac_init_idc= tmp;
2904     }
2905
2906     h->last_qscale_diff = 0;
2907     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2908     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
2909         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2910         return -1;
2911     }
2912     s->qscale= tmp;
2913     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2914     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2915     //FIXME qscale / qp ... stuff
2916     if(h->slice_type == AV_PICTURE_TYPE_SP){
2917         get_bits1(&s->gb); /* sp_for_switch_flag */
2918     }
2919     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
2920         get_se_golomb(&s->gb); /* slice_qs_delta */
2921     }
2922
2923     h->deblocking_filter = 1;
2924     h->slice_alpha_c0_offset = 52;
2925     h->slice_beta_offset = 52;
2926     if( h->pps.deblocking_filter_parameters_present ) {
2927         tmp= get_ue_golomb_31(&s->gb);
2928         if(tmp > 2){
2929             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2930             return -1;
2931         }
2932         h->deblocking_filter= tmp;
2933         if(h->deblocking_filter < 2)
2934             h->deblocking_filter^= 1; // 1<->0
2935
2936         if( h->deblocking_filter ) {
2937             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2938             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2939             if(   h->slice_alpha_c0_offset > 104U
2940                || h->slice_beta_offset     > 104U){
2941                 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);
2942                 return -1;
2943             }
2944         }
2945     }
2946
2947     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2948        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
2949        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
2950        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2951         h->deblocking_filter= 0;
2952
2953     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2954         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2955             /* Cheat slightly for speed:
2956                Do not bother to deblock across slices. */
2957             h->deblocking_filter = 2;
2958         } else {
2959             h0->max_contexts = 1;
2960             if(!h0->single_decode_warning) {
2961                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2962                 h0->single_decode_warning = 1;
2963             }
2964             if (h != h0) {
2965                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
2966                 return 1;
2967             }
2968         }
2969     }
2970     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
2971                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
2972                  + 6 * (h->sps.bit_depth_luma - 8);
2973
2974 #if 0 //FMO
2975     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2976         slice_group_change_cycle= get_bits(&s->gb, ?);
2977 #endif
2978
2979     h0->last_slice_type = slice_type;
2980     h->slice_num = ++h0->current_slice;
2981     if(h->slice_num >= MAX_SLICES){
2982         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2983     }
2984
2985     for(j=0; j<2; j++){
2986         int id_list[16];
2987         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2988         for(i=0; i<16; i++){
2989             id_list[i]= 60;
2990             if (h->ref_list[j][i].f.data[0]) {
2991                 int k;
2992                 uint8_t *base = h->ref_list[j][i].f.base[0];
2993                 for(k=0; k<h->short_ref_count; k++)
2994                     if (h->short_ref[k]->f.base[0] == base) {
2995                         id_list[i]= k;
2996                         break;
2997                     }
2998                 for(k=0; k<h->long_ref_count; k++)
2999                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3000                         id_list[i]= h->short_ref_count + k;
3001                         break;
3002                     }
3003             }
3004         }
3005
3006         ref2frm[0]=
3007         ref2frm[1]= -1;
3008         for(i=0; i<16; i++)
3009             ref2frm[i+2]= 4*id_list[i]
3010                           + (h->ref_list[j][i].f.reference & 3);
3011         ref2frm[18+0]=
3012         ref2frm[18+1]= -1;
3013         for(i=16; i<48; i++)
3014             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3015                           + (h->ref_list[j][i].f.reference & 3);
3016     }
3017
3018     //FIXME: fix draw_edges+PAFF+frame threads
3019     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3020     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3021
3022     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3023         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",
3024                h->slice_num,
3025                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3026                first_mb_in_slice,
3027                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3028                pps_id, h->frame_num,
3029                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3030                h->ref_count[0], h->ref_count[1],
3031                s->qscale,
3032                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3033                h->use_weight,
3034                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3035                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3036                );
3037     }
3038
3039     return 0;
3040 }
3041
3042 int ff_h264_get_slice_type(const H264Context *h)
3043 {
3044     switch (h->slice_type) {
3045     case AV_PICTURE_TYPE_P:  return 0;
3046     case AV_PICTURE_TYPE_B:  return 1;
3047     case AV_PICTURE_TYPE_I:  return 2;
3048     case AV_PICTURE_TYPE_SP: return 3;
3049     case AV_PICTURE_TYPE_SI: return 4;
3050     default:         return -1;
3051     }
3052 }
3053
3054 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3055                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3056 {
3057     int b_stride = h->b_stride;
3058     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3059     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3060     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3061         if(USES_LIST(top_type, list)){
3062             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3063             const int b8_xy= 4*top_xy + 2;
3064             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3065             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3066             ref_cache[0 - 1*8]=
3067             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3068             ref_cache[2 - 1*8]=
3069             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3070         }else{
3071             AV_ZERO128(mv_dst - 1*8);
3072             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3073         }
3074
3075         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3076             if(USES_LIST(left_type[LTOP], list)){
3077                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3078                 const int b8_xy= 4*left_xy[LTOP] + 1;
3079                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3080                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3081                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3082                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3083                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3084                 ref_cache[-1 +  0]=
3085                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3086                 ref_cache[-1 + 16]=
3087                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3088             }else{
3089                 AV_ZERO32(mv_dst - 1 + 0);
3090                 AV_ZERO32(mv_dst - 1 + 8);
3091                 AV_ZERO32(mv_dst - 1 +16);
3092                 AV_ZERO32(mv_dst - 1 +24);
3093                 ref_cache[-1 +  0]=
3094                 ref_cache[-1 +  8]=
3095                 ref_cache[-1 + 16]=
3096                 ref_cache[-1 + 24]= LIST_NOT_USED;
3097             }
3098         }
3099     }
3100
3101     if(!USES_LIST(mb_type, list)){
3102         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3103         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3104         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3105         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3106         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3107         return;
3108     }
3109
3110     {
3111         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3112         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3113         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3114         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3115         AV_WN32A(&ref_cache[0*8], ref01);
3116         AV_WN32A(&ref_cache[1*8], ref01);
3117         AV_WN32A(&ref_cache[2*8], ref23);
3118         AV_WN32A(&ref_cache[3*8], ref23);
3119     }
3120
3121     {
3122         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3123         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3124         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3125         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3126         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3127     }
3128 }
3129
3130 /**
3131  *
3132  * @return non zero if the loop filter can be skiped
3133  */
3134 static int fill_filter_caches(H264Context *h, int mb_type){
3135     MpegEncContext * const s = &h->s;
3136     const int mb_xy= h->mb_xy;
3137     int top_xy, left_xy[LEFT_MBS];
3138     int top_type, left_type[LEFT_MBS];
3139     uint8_t *nnz;
3140     uint8_t *nnz_cache;
3141
3142     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3143
3144     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3145      * stuff, I can't imagine that these complex rules are worth it. */
3146
3147     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3148     if(FRAME_MBAFF){
3149         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3150         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3151         if(s->mb_y&1){
3152             if (left_mb_field_flag != curr_mb_field_flag) {
3153                 left_xy[LTOP] -= s->mb_stride;
3154             }
3155         }else{
3156             if(curr_mb_field_flag){
3157                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3158             }
3159             if (left_mb_field_flag != curr_mb_field_flag) {
3160                 left_xy[LBOT] += s->mb_stride;
3161             }
3162         }
3163     }
3164
3165     h->top_mb_xy = top_xy;
3166     h->left_mb_xy[LTOP] = left_xy[LTOP];
3167     h->left_mb_xy[LBOT] = left_xy[LBOT];
3168     {
3169         //for sufficiently low qp, filtering wouldn't do anything
3170         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3171         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3172         int qp = s->current_picture.f.qscale_table[mb_xy];
3173         if(qp <= qp_thresh
3174            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3175            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3176             if(!FRAME_MBAFF)
3177                 return 1;
3178             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3179                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3180                 return 1;
3181         }
3182     }
3183
3184     top_type        = s->current_picture.f.mb_type[top_xy];
3185     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3186     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3187     if(h->deblocking_filter == 2){
3188         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3189         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3190     }else{
3191         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3192         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3193     }
3194     h->top_type       = top_type;
3195     h->left_type[LTOP]= left_type[LTOP];
3196     h->left_type[LBOT]= left_type[LBOT];
3197
3198     if(IS_INTRA(mb_type))
3199         return 0;
3200
3201     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3202     if(h->list_count == 2)
3203         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3204
3205     nnz = h->non_zero_count[mb_xy];
3206     nnz_cache = h->non_zero_count_cache;
3207     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3208     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3209     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3210     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3211     h->cbp= h->cbp_table[mb_xy];
3212
3213     if(top_type){
3214         nnz = h->non_zero_count[top_xy];
3215         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3216     }
3217
3218     if(left_type[LTOP]){
3219         nnz = h->non_zero_count[left_xy[LTOP]];
3220         nnz_cache[3+8*1]= nnz[3+0*4];
3221         nnz_cache[3+8*2]= nnz[3+1*4];
3222         nnz_cache[3+8*3]= nnz[3+2*4];
3223         nnz_cache[3+8*4]= nnz[3+3*4];
3224     }
3225
3226     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3227     if(!CABAC && h->pps.transform_8x8_mode){
3228         if(IS_8x8DCT(top_type)){
3229             nnz_cache[4+8*0]=
3230             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3231             nnz_cache[6+8*0]=
3232             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3233         }
3234         if(IS_8x8DCT(left_type[LTOP])){
3235             nnz_cache[3+8*1]=
3236             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3237         }
3238         if(IS_8x8DCT(left_type[LBOT])){
3239             nnz_cache[3+8*3]=
3240             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3241         }
3242
3243         if(IS_8x8DCT(mb_type)){
3244             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3245             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3246
3247             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3248             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3249
3250             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3251             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3252
3253             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3254             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3255         }
3256     }
3257
3258     return 0;
3259 }
3260
3261 static void loop_filter(H264Context *h, int start_x, int end_x){
3262     MpegEncContext * const s = &h->s;
3263     uint8_t  *dest_y, *dest_cb, *dest_cr;
3264     int linesize, uvlinesize, mb_x, mb_y;
3265     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3266     const int old_slice_type= h->slice_type;
3267     const int pixel_shift = h->pixel_shift;
3268
3269     if(h->deblocking_filter) {
3270         for(mb_x= start_x; mb_x<end_x; mb_x++){
3271             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3272                 int mb_xy, mb_type;
3273                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3274                 h->slice_num= h->slice_table[mb_xy];
3275                 mb_type = s->current_picture.f.mb_type[mb_xy];
3276                 h->list_count= h->list_counts[mb_xy];
3277
3278                 if(FRAME_MBAFF)
3279                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3280
3281                 s->mb_x= mb_x;
3282                 s->mb_y= mb_y;
3283                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3284                 dest_cb = s->current_picture.f.data[1] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3285                 dest_cr = s->current_picture.f.data[2] + ((mb_x << pixel_shift) + mb_y * s->uvlinesize) * (8 << CHROMA444);
3286                     //FIXME simplify above
3287
3288                 if (MB_FIELD) {
3289                     linesize   = h->mb_linesize   = s->linesize * 2;
3290                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3291                     if(mb_y&1){ //FIXME move out of this function?
3292                         dest_y -= s->linesize*15;
3293                         dest_cb-= s->uvlinesize*((8 << CHROMA444)-1);
3294                         dest_cr-= s->uvlinesize*((8 << CHROMA444)-1);
3295                     }
3296                 } else {
3297                     linesize   = h->mb_linesize   = s->linesize;
3298                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3299                 }
3300                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, CHROMA444, 0);
3301                 if(fill_filter_caches(h, mb_type))
3302                     continue;
3303                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3304                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3305
3306                 if (FRAME_MBAFF) {
3307                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3308                 } else {
3309                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3310                 }
3311             }
3312         }
3313     }
3314     h->slice_type= old_slice_type;
3315     s->mb_x= end_x;
3316     s->mb_y= end_mb_y - FRAME_MBAFF;
3317     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3318     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3319 }
3320
3321 static void predict_field_decoding_flag(H264Context *h){
3322     MpegEncContext * const s = &h->s;
3323     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3324     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3325                 ? s->current_picture.f.mb_type[mb_xy - 1]
3326                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3327                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3328                 : 0;
3329     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3330 }
3331
3332 /**
3333  * Draw edges and report progress for the last MB row.
3334  */
3335 static void decode_finish_row(H264Context *h){
3336     MpegEncContext * const s = &h->s;
3337     int top = 16*(s->mb_y >> FIELD_PICTURE);
3338     int height = 16 << FRAME_MBAFF;
3339     int deblock_border = (16 + 4) << FRAME_MBAFF;
3340     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3341
3342     if (h->deblocking_filter) {
3343         if((top + height) >= pic_height)
3344             height += deblock_border;
3345
3346         top -= deblock_border;
3347     }
3348
3349     if (top >= pic_height || (top + height) < h->emu_edge_height)
3350         return;
3351
3352     height = FFMIN(height, pic_height - top);
3353     if (top < h->emu_edge_height) {
3354         height = top+height;
3355         top = 0;
3356     }
3357
3358     ff_draw_horiz_band(s, top, height);
3359
3360     if (s->dropable) return;
3361
3362     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3363                              s->picture_structure==PICT_BOTTOM_FIELD);
3364 }
3365
3366 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3367     H264Context *h = *(void**)arg;
3368     MpegEncContext * const s = &h->s;
3369     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3370     int lf_x_start = s->mb_x;
3371
3372     s->mb_skip_run= -1;
3373
3374     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3375                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3376
3377     if( h->pps.cabac ) {
3378         /* realign */
3379         align_get_bits( &s->gb );
3380
3381         /* init cabac */
3382         ff_init_cabac_states( &h->cabac);
3383         ff_init_cabac_decoder( &h->cabac,
3384                                s->gb.buffer + get_bits_count(&s->gb)/8,
3385                                (get_bits_left(&s->gb) + 7)/8);
3386
3387         ff_h264_init_cabac_states(h);
3388
3389         for(;;){
3390 //START_TIMER
3391             int ret = ff_h264_decode_mb_cabac(h);
3392             int eos;
3393 //STOP_TIMER("decode_mb_cabac")
3394
3395             if(ret>=0) ff_h264_hl_decode_mb(h);
3396
3397             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3398                 s->mb_y++;
3399
3400                 ret = ff_h264_decode_mb_cabac(h);
3401
3402                 if(ret>=0) ff_h264_hl_decode_mb(h);
3403                 s->mb_y--;
3404             }
3405             eos = get_cabac_terminate( &h->cabac );
3406
3407             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3408                 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);
3409                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3410                 return 0;
3411             }
3412             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3413                 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);
3414                 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);
3415                 return -1;
3416             }
3417
3418             if( ++s->mb_x >= s->mb_width ) {
3419                 loop_filter(h, lf_x_start, s->mb_x);
3420                 s->mb_x = lf_x_start = 0;
3421                 decode_finish_row(h);
3422                 ++s->mb_y;
3423                 if(FIELD_OR_MBAFF_PICTURE) {
3424                     ++s->mb_y;
3425                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3426                         predict_field_decoding_flag(h);
3427                 }
3428             }
3429
3430             if( eos || s->mb_y >= s->mb_height ) {
3431                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3432                 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);
3433                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3434                 return 0;
3435             }
3436         }
3437
3438     } else {
3439         for(;;){
3440             int ret = ff_h264_decode_mb_cavlc(h);
3441
3442             if(ret>=0) ff_h264_hl_decode_mb(h);
3443
3444             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3445                 s->mb_y++;
3446                 ret = ff_h264_decode_mb_cavlc(h);
3447
3448                 if(ret>=0) ff_h264_hl_decode_mb(h);
3449                 s->mb_y--;
3450             }
3451
3452             if(ret<0){
3453                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3454                 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);
3455                 return -1;
3456             }
3457
3458             if(++s->mb_x >= s->mb_width){
3459                 loop_filter(h, lf_x_start, s->mb_x);
3460                 s->mb_x = lf_x_start = 0;
3461                 decode_finish_row(h);
3462                 ++s->mb_y;
3463                 if(FIELD_OR_MBAFF_PICTURE) {
3464                     ++s->mb_y;
3465                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3466                         predict_field_decoding_flag(h);
3467                 }
3468                 if(s->mb_y >= s->mb_height){
3469                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3470
3471                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
3472                         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);
3473
3474                         return 0;
3475                     }else{
3476                         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);
3477
3478                         return -1;
3479                     }
3480                 }
3481             }
3482
3483             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3484                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3485                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3486                     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);
3487                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3488
3489                     return 0;
3490                 }else{
3491                     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);
3492
3493                     return -1;
3494                 }
3495             }
3496         }
3497     }
3498 }
3499
3500 /**
3501  * Call decode_slice() for each context.
3502  *
3503  * @param h h264 master context
3504  * @param context_count number of contexts to execute
3505  */
3506 static int execute_decode_slices(H264Context *h, int context_count){
3507     MpegEncContext * const s = &h->s;
3508     AVCodecContext * const avctx= s->avctx;
3509     H264Context *hx;
3510     int i;
3511
3512     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3513         return 0;
3514     if(context_count == 1) {
3515         return decode_slice(avctx, &h);
3516     } else {
3517         for(i = 1; i < context_count; i++) {
3518             hx = h->thread_context[i];
3519             hx->s.error_recognition = avctx->error_recognition;
3520             hx->s.error_count = 0;
3521         }
3522
3523         avctx->execute(avctx, (void *)decode_slice,
3524                        h->thread_context, NULL, context_count, sizeof(void*));
3525
3526         /* pull back stuff from slices to master context */
3527         hx = h->thread_context[context_count - 1];
3528         s->mb_x = hx->s.mb_x;
3529         s->mb_y = hx->s.mb_y;
3530         s->dropable = hx->s.dropable;
3531         s->picture_structure = hx->s.picture_structure;
3532         for(i = 1; i < context_count; i++)
3533             h->s.error_count += h->thread_context[i]->s.error_count;
3534     }
3535
3536     return 0;
3537 }
3538
3539
3540 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3541     MpegEncContext * const s = &h->s;
3542     AVCodecContext * const avctx= s->avctx;
3543     H264Context *hx; ///< thread context
3544     int buf_index;
3545     int context_count;
3546     int next_avc;
3547     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3548     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3549     int nal_index;
3550
3551     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3552     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3553         h->current_slice = 0;
3554         if (!s->first_field)
3555             s->current_picture_ptr= NULL;
3556         ff_h264_reset_sei(h);
3557     }
3558
3559     for(;pass <= 1;pass++){
3560         buf_index = 0;
3561         context_count = 0;
3562         next_avc = h->is_avc ? 0 : buf_size;
3563         nal_index = 0;
3564     for(;;){
3565         int consumed;
3566         int dst_length;
3567         int bit_length;
3568         const uint8_t *ptr;
3569         int i, nalsize = 0;
3570         int err;
3571
3572         if(buf_index >= next_avc) {
3573             if(buf_index >= buf_size) break;
3574             nalsize = 0;
3575             for(i = 0; i < h->nal_length_size; i++)
3576                 nalsize = (nalsize << 8) | buf[buf_index++];
3577             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3578                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3579                 break;
3580             }
3581             next_avc= buf_index + nalsize;
3582         } else {
3583             // start code prefix search
3584             for(; buf_index + 3 < next_avc; buf_index++){
3585                 // This should always succeed in the first iteration.
3586                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3587                     break;
3588             }
3589
3590             if(buf_index+3 >= buf_size) break;
3591
3592             buf_index+=3;
3593             if(buf_index >= next_avc) continue;
3594         }
3595
3596         hx = h->thread_context[context_count];
3597
3598         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3599         if (ptr==NULL || dst_length < 0){
3600             return -1;
3601         }
3602         i= buf_index + consumed;
3603         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3604            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3605             s->workaround_bugs |= FF_BUG_TRUNCATED;
3606
3607         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3608         while(ptr[dst_length - 1] == 0 && dst_length > 0)
3609             dst_length--;
3610         }
3611         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3612
3613         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3614             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);
3615         }
3616
3617         if (h->is_avc && (nalsize != consumed) && nalsize){
3618             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3619         }
3620
3621         buf_index += consumed;
3622         nal_index++;
3623
3624         if(pass == 0) {
3625             // packets can sometimes contain multiple PPS/SPS
3626             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3627             // if so, when frame threading we can't start the next thread until we've read all of them
3628             switch (hx->nal_unit_type) {
3629                 case NAL_SPS:
3630                 case NAL_PPS:
3631                 case NAL_IDR_SLICE:
3632                 case NAL_SLICE:
3633                     nals_needed = nal_index;
3634             }
3635             continue;
3636         }
3637
3638         //FIXME do not discard SEI id
3639         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3640             continue;
3641
3642       again:
3643         err = 0;
3644         switch(hx->nal_unit_type){
3645         case NAL_IDR_SLICE:
3646             if (h->nal_unit_type != NAL_IDR_SLICE) {
3647                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3648                 return -1;
3649             }
3650             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3651         case NAL_SLICE:
3652             init_get_bits(&hx->s.gb, ptr, bit_length);
3653             hx->intra_gb_ptr=
3654             hx->inter_gb_ptr= &hx->s.gb;
3655             hx->s.data_partitioning = 0;
3656
3657             if((err = decode_slice_header(hx, h)))
3658                break;
3659
3660             s->current_picture_ptr->f.key_frame |=
3661                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
3662                     (h->sei_recovery_frame_cnt >= 0);
3663
3664             if (h->current_slice == 1) {
3665                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3666                     decode_postinit(h, nal_index >= nals_needed);
3667                 }
3668
3669                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3670                     return -1;
3671                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3672                     ff_vdpau_h264_picture_start(s);
3673             }
3674
3675             if(hx->redundant_pic_count==0
3676                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3677                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3678                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3679                && avctx->skip_frame < AVDISCARD_ALL){
3680                 if(avctx->hwaccel) {
3681                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3682                         return -1;
3683                 }else
3684                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3685                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3686                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3687                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3688                 }else
3689                     context_count++;
3690             }
3691             break;
3692         case NAL_DPA:
3693             init_get_bits(&hx->s.gb, ptr, bit_length);
3694             hx->intra_gb_ptr=
3695             hx->inter_gb_ptr= NULL;
3696
3697             if ((err = decode_slice_header(hx, h)) < 0)
3698                 break;
3699
3700             hx->s.data_partitioning = 1;
3701
3702             break;
3703         case NAL_DPB:
3704             init_get_bits(&hx->intra_gb, ptr, bit_length);
3705             hx->intra_gb_ptr= &hx->intra_gb;
3706             break;
3707         case NAL_DPC:
3708             init_get_bits(&hx->inter_gb, ptr, bit_length);
3709             hx->inter_gb_ptr= &hx->inter_gb;
3710
3711             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3712                && s->context_initialized
3713                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3714                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3715                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3716                && avctx->skip_frame < AVDISCARD_ALL)
3717                 context_count++;
3718             break;
3719         case NAL_SEI:
3720             init_get_bits(&s->gb, ptr, bit_length);
3721             ff_h264_decode_sei(h);
3722             break;
3723         case NAL_SPS:
3724             init_get_bits(&s->gb, ptr, bit_length);
3725             ff_h264_decode_seq_parameter_set(h);
3726
3727             if (s->flags& CODEC_FLAG_LOW_DELAY ||
3728                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
3729                 s->low_delay=1;
3730
3731             if(avctx->has_b_frames < 2)
3732                 avctx->has_b_frames= !s->low_delay;
3733
3734             if (avctx->bits_per_raw_sample != h->sps.bit_depth_luma) {
3735                 if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10) {
3736                     avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
3737                     h->pixel_shift = h->sps.bit_depth_luma > 8;
3738
3739                     ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma);
3740                     ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma);
3741                     s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
3742                     dsputil_init(&s->dsp, s->avctx);
3743                 } else {
3744                     av_log(avctx, AV_LOG_DEBUG, "Unsupported bit depth: %d\n", h->sps.bit_depth_luma);
3745                     return -1;
3746                 }
3747             }
3748             break;
3749         case NAL_PPS:
3750             init_get_bits(&s->gb, ptr, bit_length);
3751
3752             ff_h264_decode_picture_parameter_set(h, bit_length);
3753
3754             break;
3755         case NAL_AUD:
3756         case NAL_END_SEQUENCE:
3757         case NAL_END_STREAM:
3758         case NAL_FILLER_DATA:
3759         case NAL_SPS_EXT:
3760         case NAL_AUXILIARY_SLICE:
3761             break;
3762         default:
3763             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
3764         }
3765
3766         if(context_count == h->max_contexts) {
3767             execute_decode_slices(h, context_count);
3768             context_count = 0;
3769         }
3770
3771         if (err < 0)
3772             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
3773         else if(err == 1) {
3774             /* Slice could not be decoded in parallel mode, copy down
3775              * NAL unit stuff to context 0 and restart. Note that
3776              * rbsp_buffer is not transferred, but since we no longer
3777              * run in parallel mode this should not be an issue. */
3778             h->nal_unit_type = hx->nal_unit_type;
3779             h->nal_ref_idc   = hx->nal_ref_idc;
3780             hx = h;
3781             goto again;
3782         }
3783     }
3784     }
3785     if(context_count)
3786         execute_decode_slices(h, context_count);
3787     return buf_index;
3788 }
3789
3790 /**
3791  * returns the number of bytes consumed for building the current frame
3792  */
3793 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
3794         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
3795         if(pos+10>buf_size) pos=buf_size; // oops ;)
3796
3797         return pos;
3798 }
3799
3800 static int decode_frame(AVCodecContext *avctx,
3801                              void *data, int *data_size,
3802                              AVPacket *avpkt)
3803 {
3804     const uint8_t *buf = avpkt->data;
3805     int buf_size = avpkt->size;
3806     H264Context *h = avctx->priv_data;
3807     MpegEncContext *s = &h->s;
3808     AVFrame *pict = data;
3809     int buf_index;
3810
3811     s->flags= avctx->flags;
3812     s->flags2= avctx->flags2;
3813
3814    /* end of stream, output what is still in the buffers */
3815  out:
3816     if (buf_size == 0) {
3817         Picture *out;
3818         int i, out_idx;
3819
3820         s->current_picture_ptr = NULL;
3821
3822 //FIXME factorize this with the output code below
3823         out = h->delayed_pic[0];
3824         out_idx = 0;
3825         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3826             if(h->delayed_pic[i]->poc < out->poc){
3827                 out = h->delayed_pic[i];
3828                 out_idx = i;
3829             }
3830
3831         for(i=out_idx; h->delayed_pic[i]; i++)
3832             h->delayed_pic[i] = h->delayed_pic[i+1];
3833
3834         if(out){
3835             *data_size = sizeof(AVFrame);
3836             *pict= *(AVFrame*)out;
3837         }
3838
3839         return 0;
3840     }
3841
3842     buf_index=decode_nal_units(h, buf, buf_size);
3843     if(buf_index < 0)
3844         return -1;
3845
3846     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
3847         buf_size = 0;
3848         goto out;
3849     }
3850
3851     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3852         if (avctx->skip_frame >= AVDISCARD_NONREF)
3853             return 0;
3854         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3855         return -1;
3856     }
3857
3858     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3859
3860         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
3861
3862         field_end(h, 0);
3863
3864         if (!h->next_output_pic) {
3865             /* Wait for second field. */
3866             *data_size = 0;
3867
3868         } else {
3869             *data_size = sizeof(AVFrame);
3870             *pict = *(AVFrame*)h->next_output_pic;
3871         }
3872     }
3873
3874     assert(pict->data[0] || !*data_size);
3875     ff_print_debug_info(s, pict);
3876 //printf("out %d\n", (int)pict->data[0]);
3877
3878     return get_consumed_bytes(s, buf_index, buf_size);
3879 }
3880 #if 0
3881 static inline void fill_mb_avail(H264Context *h){
3882     MpegEncContext * const s = &h->s;
3883     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3884
3885     if(s->mb_y){
3886         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3887         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3888         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3889     }else{
3890         h->mb_avail[0]=
3891         h->mb_avail[1]=
3892         h->mb_avail[2]= 0;
3893     }
3894     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3895     h->mb_avail[4]= 1; //FIXME move out
3896     h->mb_avail[5]= 0; //FIXME move out
3897 }
3898 #endif
3899
3900 #ifdef TEST
3901 #undef printf
3902 #undef random
3903 #define COUNT 8000
3904 #define SIZE (COUNT*40)
3905 int main(void){
3906     int i;
3907     uint8_t temp[SIZE];
3908     PutBitContext pb;
3909     GetBitContext gb;
3910 //    int int_temp[10000];
3911     DSPContext dsp;
3912     AVCodecContext avctx;
3913
3914     dsputil_init(&dsp, &avctx);
3915
3916     init_put_bits(&pb, temp, SIZE);
3917     printf("testing unsigned exp golomb\n");
3918     for(i=0; i<COUNT; i++){
3919         START_TIMER
3920         set_ue_golomb(&pb, i);
3921         STOP_TIMER("set_ue_golomb");
3922     }
3923     flush_put_bits(&pb);
3924
3925     init_get_bits(&gb, temp, 8*SIZE);
3926     for(i=0; i<COUNT; i++){
3927         int j, s;
3928
3929         s= show_bits(&gb, 24);
3930
3931         START_TIMER
3932         j= get_ue_golomb(&gb);
3933         if(j != i){
3934             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3935 //            return -1;
3936         }
3937         STOP_TIMER("get_ue_golomb");
3938     }
3939
3940
3941     init_put_bits(&pb, temp, SIZE);
3942     printf("testing signed exp golomb\n");
3943     for(i=0; i<COUNT; i++){
3944         START_TIMER
3945         set_se_golomb(&pb, i - COUNT/2);
3946         STOP_TIMER("set_se_golomb");
3947     }
3948     flush_put_bits(&pb);
3949
3950     init_get_bits(&gb, temp, 8*SIZE);
3951     for(i=0; i<COUNT; i++){
3952         int j, s;
3953
3954         s= show_bits(&gb, 24);
3955
3956         START_TIMER
3957         j= get_se_golomb(&gb);
3958         if(j != i - COUNT/2){
3959             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3960 //            return -1;
3961         }
3962         STOP_TIMER("get_se_golomb");
3963     }
3964
3965     printf("Testing RBSP\n");
3966
3967
3968     return 0;
3969 }
3970 #endif /* TEST */
3971
3972
3973 av_cold void ff_h264_free_context(H264Context *h)
3974 {
3975     int i;
3976
3977     free_tables(h, 1); //FIXME cleanup init stuff perhaps
3978
3979     for(i = 0; i < MAX_SPS_COUNT; i++)
3980         av_freep(h->sps_buffers + i);
3981
3982     for(i = 0; i < MAX_PPS_COUNT; i++)
3983         av_freep(h->pps_buffers + i);
3984 }
3985
3986 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
3987 {
3988     H264Context *h = avctx->priv_data;
3989     MpegEncContext *s = &h->s;
3990
3991     ff_h264_free_context(h);
3992
3993     MPV_common_end(s);
3994
3995 //    memset(h, 0, sizeof(H264Context));
3996
3997     return 0;
3998 }
3999
4000 static const AVProfile profiles[] = {
4001     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4002     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4003     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4004     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4005     { FF_PROFILE_H264_HIGH,                 "High"                  },
4006     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4007     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4008     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4009     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4010     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4011     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4012     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4013     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4014     { FF_PROFILE_UNKNOWN },
4015 };
4016
4017 AVCodec ff_h264_decoder = {
4018     .name           = "h264",
4019     .type           = AVMEDIA_TYPE_VIDEO,
4020     .id             = CODEC_ID_H264,
4021     .priv_data_size = sizeof(H264Context),
4022     .init           = ff_h264_decode_init,
4023     .close          = ff_h264_decode_end,
4024     .decode         = decode_frame,
4025     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4026                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4027     .flush= flush_dpb,
4028     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4029     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4030     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4031     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4032 };
4033
4034 #if CONFIG_H264_VDPAU_DECODER
4035 AVCodec ff_h264_vdpau_decoder = {
4036     .name           = "h264_vdpau",
4037     .type           = AVMEDIA_TYPE_VIDEO,
4038     .id             = CODEC_ID_H264,
4039     .priv_data_size = sizeof(H264Context),
4040     .init           = ff_h264_decode_init,
4041     .close          = ff_h264_decode_end,
4042     .decode         = decode_frame,
4043     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4044     .flush= flush_dpb,
4045     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4046     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4047     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4048 };
4049 #endif