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

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "libavutil/imgutils.h"
  29 #include "libavutil/opt.h"
  30 #include "internal.h"
  31 #include "dsputil.h"
  32 #include "avcodec.h"
  33 #include "mpegvideo.h"
  34 #include "h264.h"
  35 #include "h264data.h"
  36 #include "h264_mvpred.h"
  37 #include "golomb.h"
  38 #include "mathops.h"
  39 #include "rectangle.h"
  40 #include "thread.h"
  41 #include "vdpau_internal.h"
  42 #include "libavutil/avassert.h"
  43
  44 #include "cabac.h"
  45
  46 //#undef NDEBUG
  47 #include <assert.h>
  48
  49 static const uint8_t rem6[QP_MAX_NUM+1]={
  50 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,
  51 };
  52
  53 static const uint8_t div6[QP_MAX_NUM+1]={
  54 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,
  55 };
  56
  57 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  58     PIX_FMT_DXVA2_VLD,
  59     PIX_FMT_VAAPI_VLD,
  60     PIX_FMT_VDA_VLD,
  61     PIX_FMT_YUVJ420P,
  62     PIX_FMT_NONE
  63 };
  64
  65 /**
  66  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  67  */
  68 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  69     MpegEncContext * const s = &h->s;
  70     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  71     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  72     int i;
  73
  74     if(!(h->top_samples_available&0x8000)){
  75         for(i=0; i<4; i++){
  76             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  77             if(status<0){
  78                 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);
  79                 return -1;
  80             } else if(status){
  81                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  82             }
  83         }
  84     }
  85
  86     if((h->left_samples_available&0x8888)!=0x8888){
  87         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  88         for(i=0; i<4; i++){
  89             if(!(h->left_samples_available&mask[i])){
  90                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  91                 if(status<0){
  92                     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);
  93                     return -1;
  94                 } else if(status){
  95                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
  96                 }
  97             }
  98         }
  99     }
 100
 101     return 0;
 102 } //FIXME cleanup like check_intra_pred_mode
 103
 104 static int check_intra_pred_mode(H264Context *h, int mode, int is_chroma){
 105     MpegEncContext * const s = &h->s;
 106     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 107     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 108
 109     if(mode > 6U) {
 110         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);
 111         return -1;
 112     }
 113
 114     if(!(h->top_samples_available&0x8000)){
 115         mode= top[ mode ];
 116         if(mode<0){
 117             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);
 118             return -1;
 119         }
 120     }
 121
 122     if((h->left_samples_available&0x8080) != 0x8080){
 123         mode= left[ mode ];
 124         if(is_chroma && (h->left_samples_available&0x8080)){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 125             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 126         }
 127         if(mode<0){
 128             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);
 129             return -1;
 130         }
 131     }
 132
 133     return mode;
 134 }
 135
 136 /**
 137  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 138  */
 139 int ff_h264_check_intra16x16_pred_mode(H264Context *h, int mode)
 140 {
 141     return check_intra_pred_mode(h, mode, 0);
 142 }
 143
 144 /**
 145  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 146  */
 147 int ff_h264_check_intra_chroma_pred_mode(H264Context *h, int mode)
 148 {
 149     return check_intra_pred_mode(h, mode, 1);
 150 }
 151
 152
 153 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 154     int i, si, di;
 155     uint8_t *dst;
 156     int bufidx;
 157
 158 //    src[0]&0x80;                //forbidden bit
 159     h->nal_ref_idc= src[0]>>5;
 160     h->nal_unit_type= src[0]&0x1F;
 161
 162     src++; length--;
 163
 164 #if HAVE_FAST_UNALIGNED
 165 # if HAVE_FAST_64BIT
 166 #   define RS 7
 167     for(i=0; i+1<length; i+=9){
 168         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 169 # else
 170 #   define RS 3
 171     for(i=0; i+1<length; i+=5){
 172         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 173 # endif
 174             continue;
 175         if(i>0 && !src[i]) i--;
 176         while(src[i]) i++;
 177 #else
 178 #   define RS 0
 179     for(i=0; i+1<length; i+=2){
 180         if(src[i]) continue;
 181         if(i>0 && src[i-1]==0) i--;
 182 #endif
 183         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 184             if(src[i+2]!=3){
 185                 /* startcode, so we must be past the end */
 186                 length=i;
 187             }
 188             break;
 189         }
 190         i-= RS;
 191     }
 192
 193     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 194     si=h->rbsp_buffer_size[bufidx];
 195     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 196     dst= h->rbsp_buffer[bufidx];
 197     if(si != h->rbsp_buffer_size[bufidx])
 198         memset(dst + length, 0, FF_INPUT_BUFFER_PADDING_SIZE+MAX_MBPAIR_SIZE);
 199
 200     if (dst == NULL){
 201         return NULL;
 202     }
 203
 204     if(i>=length-1){ //no escaped 0
 205         *dst_length= length;
 206         *consumed= length+1; //+1 for the header
 207         if(h->s.avctx->flags2 & CODEC_FLAG2_FAST){
 208             return src;
 209         }else{
 210             memcpy(dst, src, length);
 211             return dst;
 212         }
 213     }
 214
 215 //printf("decoding esc\n");
 216     memcpy(dst, src, i);
 217     si=di=i;
 218     while(si+2<length){
 219         //remove escapes (very rare 1:2^22)
 220         if(src[si+2]>3){
 221             dst[di++]= src[si++];
 222             dst[di++]= src[si++];
 223         }else if(src[si]==0 && src[si+1]==0){
 224             if(src[si+2]==3){ //escape
 225                 dst[di++]= 0;
 226                 dst[di++]= 0;
 227                 si+=3;
 228                 continue;
 229             }else //next start code
 230                 goto nsc;
 231         }
 232
 233         dst[di++]= src[si++];
 234     }
 235     while(si<length)
 236         dst[di++]= src[si++];
 237 nsc:
 238
 239     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 240
 241     *dst_length= di;
 242     *consumed= si + 1;//+1 for the header
 243 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 244     return dst;
 245 }
 246
 247 /**
 248  * Identify the exact end of the bitstream
 249  * @return the length of the trailing, or 0 if damaged
 250  */
 251 static int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 252     int v= *src;
 253     int r;
 254
 255     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 256
 257     for(r=1; r<9; r++){
 258         if(v&1) return r;
 259         v>>=1;
 260     }
 261     return 0;
 262 }
 263
 264 static inline int get_lowest_part_list_y(H264Context *h, Picture *pic, int n, int height,
 265                                  int y_offset, int list){
 266     int raw_my= h->mv_cache[list][ scan8[n] ][1];
 267     int filter_height= (raw_my&3) ? 2 : 0;
 268     int full_my= (raw_my>>2) + y_offset;
 269     int top = full_my - filter_height, bottom = full_my + height + filter_height;
 270
 271     return FFMAX(abs(top), bottom);
 272 }
 273
 274 static inline void get_lowest_part_y(H264Context *h, int refs[2][48], int n, int height,
 275                                int y_offset, int list0, int list1, int *nrefs){
 276     MpegEncContext * const s = &h->s;
 277     int my;
 278
 279     y_offset += 16*(s->mb_y >> MB_FIELD);
 280
 281     if(list0){
 282         int ref_n = h->ref_cache[0][ scan8[n] ];
 283         Picture *ref= &h->ref_list[0][ref_n];
 284
 285         // Error resilience puts the current picture in the ref list.
 286         // Don't try to wait on these as it will cause a deadlock.
 287         // Fields can wait on each other, though.
 288         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 289            (ref->f.reference & 3) != s->picture_structure) {
 290             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 0);
 291             if (refs[0][ref_n] < 0) nrefs[0] += 1;
 292             refs[0][ref_n] = FFMAX(refs[0][ref_n], my);
 293         }
 294     }
 295
 296     if(list1){
 297         int ref_n = h->ref_cache[1][ scan8[n] ];
 298         Picture *ref= &h->ref_list[1][ref_n];
 299
 300         if (ref->f.thread_opaque != s->current_picture.f.thread_opaque ||
 301            (ref->f.reference & 3) != s->picture_structure) {
 302             my = get_lowest_part_list_y(h, ref, n, height, y_offset, 1);
 303             if (refs[1][ref_n] < 0) nrefs[1] += 1;
 304             refs[1][ref_n] = FFMAX(refs[1][ref_n], my);
 305         }
 306     }
 307 }
 308
 309 /**
 310  * Wait until all reference frames are available for MC operations.
 311  *
 312  * @param h the H264 context
 313  */
 314 static void await_references(H264Context *h){
 315     MpegEncContext * const s = &h->s;
 316     const int mb_xy= h->mb_xy;
 317     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 318     int refs[2][48];
 319     int nrefs[2] = {0};
 320     int ref, list;
 321
 322     memset(refs, -1, sizeof(refs));
 323
 324     if(IS_16X16(mb_type)){
 325         get_lowest_part_y(h, refs, 0, 16, 0,
 326                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 327     }else if(IS_16X8(mb_type)){
 328         get_lowest_part_y(h, refs, 0, 8, 0,
 329                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 330         get_lowest_part_y(h, refs, 8, 8, 8,
 331                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 332     }else if(IS_8X16(mb_type)){
 333         get_lowest_part_y(h, refs, 0, 16, 0,
 334                   IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1), nrefs);
 335         get_lowest_part_y(h, refs, 4, 16, 0,
 336                   IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1), nrefs);
 337     }else{
 338         int i;
 339
 340         assert(IS_8X8(mb_type));
 341
 342         for(i=0; i<4; i++){
 343             const int sub_mb_type= h->sub_mb_type[i];
 344             const int n= 4*i;
 345             int y_offset= (i&2)<<2;
 346
 347             if(IS_SUB_8X8(sub_mb_type)){
 348                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 349                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 350             }else if(IS_SUB_8X4(sub_mb_type)){
 351                 get_lowest_part_y(h, refs, n  , 4, y_offset,
 352                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 353                 get_lowest_part_y(h, refs, n+2, 4, y_offset+4,
 354                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 355             }else if(IS_SUB_4X8(sub_mb_type)){
 356                 get_lowest_part_y(h, refs, n  , 8, y_offset,
 357                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 358                 get_lowest_part_y(h, refs, n+1, 8, y_offset,
 359                           IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 360             }else{
 361                 int j;
 362                 assert(IS_SUB_4X4(sub_mb_type));
 363                 for(j=0; j<4; j++){
 364                     int sub_y_offset= y_offset + 2*(j&2);
 365                     get_lowest_part_y(h, refs, n+j, 4, sub_y_offset,
 366                               IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1), nrefs);
 367                 }
 368             }
 369         }
 370     }
 371
 372     for(list=h->list_count-1; list>=0; list--){
 373         for(ref=0; ref<48 && nrefs[list]; ref++){
 374             int row = refs[list][ref];
 375             if(row >= 0){
 376                 Picture *ref_pic = &h->ref_list[list][ref];
 377                 int ref_field = ref_pic->f.reference - 1;
 378                 int ref_field_picture = ref_pic->field_picture;
 379                 int pic_height = 16*s->mb_height >> ref_field_picture;
 380
 381                 row <<= MB_MBAFF;
 382                 nrefs[list]--;
 383
 384                 if(!FIELD_PICTURE && ref_field_picture){ // frame referencing two fields
 385                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1) - !(row&1), pic_height-1), 1);
 386                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN((row >> 1)           , pic_height-1), 0);
 387                 }else if(FIELD_PICTURE && !ref_field_picture){ // field referencing one field of a frame
 388                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row*2 + ref_field    , pic_height-1), 0);
 389                 }else if(FIELD_PICTURE){
 390                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), ref_field);
 391                 }else{
 392                     ff_thread_await_progress((AVFrame*)ref_pic, FFMIN(row, pic_height-1), 0);
 393                 }
 394             }
 395         }
 396     }
 397 }
 398
 399 #if 0
 400 /**
 401  * DCT transforms the 16 dc values.
 402  * @param qp quantization parameter ??? FIXME
 403  */
 404 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 405 //    const int qmul= dequant_coeff[qp][0];
 406     int i;
 407     int temp[16]; //FIXME check if this is a good idea
 408     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 409     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 410
 411     for(i=0; i<4; i++){
 412         const int offset= y_offset[i];
 413         const int z0= block[offset+stride*0] + block[offset+stride*4];
 414         const int z1= block[offset+stride*0] - block[offset+stride*4];
 415         const int z2= block[offset+stride*1] - block[offset+stride*5];
 416         const int z3= block[offset+stride*1] + block[offset+stride*5];
 417
 418         temp[4*i+0]= z0+z3;
 419         temp[4*i+1]= z1+z2;
 420         temp[4*i+2]= z1-z2;
 421         temp[4*i+3]= z0-z3;
 422     }
 423
 424     for(i=0; i<4; i++){
 425         const int offset= x_offset[i];
 426         const int z0= temp[4*0+i] + temp[4*2+i];
 427         const int z1= temp[4*0+i] - temp[4*2+i];
 428         const int z2= temp[4*1+i] - temp[4*3+i];
 429         const int z3= temp[4*1+i] + temp[4*3+i];
 430
 431         block[stride*0 +offset]= (z0 + z3)>>1;
 432         block[stride*2 +offset]= (z1 + z2)>>1;
 433         block[stride*8 +offset]= (z1 - z2)>>1;
 434         block[stride*10+offset]= (z0 - z3)>>1;
 435     }
 436 }
 437 #endif
 438
 439 #undef xStride
 440 #undef stride
 441
 442 #if 0
 443 static void chroma_dc_dct_c(DCTELEM *block){
 444     const int stride= 16*2;
 445     const int xStride= 16;
 446     int a,b,c,d,e;
 447
 448     a= block[stride*0 + xStride*0];
 449     b= block[stride*0 + xStride*1];
 450     c= block[stride*1 + xStride*0];
 451     d= block[stride*1 + xStride*1];
 452
 453     e= a-b;
 454     a= a+b;
 455     b= c-d;
 456     c= c+d;
 457
 458     block[stride*0 + xStride*0]= (a+c);
 459     block[stride*0 + xStride*1]= (e+b);
 460     block[stride*1 + xStride*0]= (a-c);
 461     block[stride*1 + xStride*1]= (e-b);
 462 }
 463 #endif
 464
 465 static av_always_inline void
 466 mc_dir_part(H264Context *h, Picture *pic, int n, int square,
 467             int height, int delta, int list,
 468             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 469             int src_x_offset, int src_y_offset,
 470             qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op,
 471             int pixel_shift, int chroma_idc)
 472 {
 473     MpegEncContext * const s = &h->s;
 474     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 475     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 476     const int luma_xy= (mx&3) + ((my&3)<<2);
 477     int offset = ((mx>>2) << pixel_shift) + (my>>2)*h->mb_linesize;
 478     uint8_t * src_y = pic->f.data[0] + offset;
 479     uint8_t * src_cb, * src_cr;
 480     int extra_width= h->emu_edge_width;
 481     int extra_height= h->emu_edge_height;
 482     int emu=0;
 483     const int full_mx= mx>>2;
 484     const int full_my= my>>2;
 485     const int pic_width  = 16*s->mb_width;
 486     const int pic_height = 16*s->mb_height >> MB_FIELD;
 487     int ysh;
 488
 489     if(mx&7) extra_width -= 3;
 490     if(my&7) extra_height -= 3;
 491
 492     if(   full_mx < 0-extra_width
 493        || full_my < 0-extra_height
 494        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 495        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 496         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_y - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 497                                 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 498             src_y= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 499         emu=1;
 500     }
 501
 502     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 503     if(!square){
 504         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 505     }
 506
 507     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 508
 509     if(chroma_idc == 3 /* yuv444 */){
 510         src_cb = pic->f.data[1] + offset;
 511         if(emu){
 512             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 513                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 514             src_cb= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 515         }
 516         qpix_op[luma_xy](dest_cb, src_cb, h->mb_linesize); //FIXME try variable height perhaps?
 517         if(!square){
 518             qpix_op[luma_xy](dest_cb + delta, src_cb + delta, h->mb_linesize);
 519         }
 520
 521         src_cr = pic->f.data[2] + offset;
 522         if(emu){
 523             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr - (2 << pixel_shift) - 2*h->mb_linesize, h->mb_linesize,
 524                                     16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 525             src_cr= s->edge_emu_buffer + (2 << pixel_shift) + 2*h->mb_linesize;
 526         }
 527         qpix_op[luma_xy](dest_cr, src_cr, h->mb_linesize); //FIXME try variable height perhaps?
 528         if(!square){
 529             qpix_op[luma_xy](dest_cr + delta, src_cr + delta, h->mb_linesize);
 530         }
 531         return;
 532     }
 533
 534     ysh = 3 - (chroma_idc == 2 /* yuv422 */);
 535     if(chroma_idc == 1 /* yuv420 */ && MB_FIELD){
 536         // chroma offset when predicting from a field of opposite parity
 537         my += 2 * ((s->mb_y & 1) - (pic->f.reference - 1));
 538         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 539     }
 540
 541     src_cb = pic->f.data[1] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 542     src_cr = pic->f.data[2] + ((mx >> 3) << pixel_shift) + (my >> ysh) * h->mb_uvlinesize;
 543
 544     if(emu){
 545         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize,
 546                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 547                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 548             src_cb= s->edge_emu_buffer;
 549     }
 550     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 551               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 552
 553     if(emu){
 554         s->dsp.emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize,
 555                                 9, 8 * chroma_idc + 1, (mx >> 3), (my >> ysh),
 556                                 pic_width >> 1, pic_height >> (chroma_idc == 1 /* yuv420 */));
 557             src_cr= s->edge_emu_buffer;
 558     }
 559     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, height >> (chroma_idc == 1 /* yuv420 */),
 560               mx&7, (my << (chroma_idc == 2 /* yuv422 */)) &7);
 561 }
 562
 563 static av_always_inline void
 564 mc_part_std(H264Context *h, int n, int square, int height, int delta,
 565             uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 566             int x_offset, int y_offset,
 567             qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 568             qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 569             int list0, int list1, int pixel_shift, int chroma_idc)
 570 {
 571     MpegEncContext * const s = &h->s;
 572     qpel_mc_func *qpix_op=  qpix_put;
 573     h264_chroma_mc_func chroma_op= chroma_put;
 574
 575     dest_y  += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 576     if (chroma_idc == 3 /* yuv444 */) {
 577         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 578         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 579     } else if (chroma_idc == 2 /* yuv422 */) {
 580         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 581         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 582     } else /* yuv420 */ {
 583         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 584         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 585     }
 586     x_offset += 8*s->mb_x;
 587     y_offset += 8*(s->mb_y >> MB_FIELD);
 588
 589     if(list0){
 590         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 591         mc_dir_part(h, ref, n, square, height, delta, 0,
 592                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 593                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 594
 595         qpix_op=  qpix_avg;
 596         chroma_op= chroma_avg;
 597     }
 598
 599     if(list1){
 600         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 601         mc_dir_part(h, ref, n, square, height, delta, 1,
 602                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 603                            qpix_op, chroma_op, pixel_shift, chroma_idc);
 604     }
 605 }
 606
 607 static av_always_inline void
 608 mc_part_weighted(H264Context *h, int n, int square, int height, int delta,
 609                  uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 610                  int x_offset, int y_offset,
 611                  qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 612                  h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 613                  h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 614                  int list0, int list1, int pixel_shift, int chroma_idc){
 615     MpegEncContext * const s = &h->s;
 616     int chroma_height;
 617
 618     dest_y += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 619     if (chroma_idc == 3 /* yuv444 */) {
 620         chroma_height = height;
 621         chroma_weight_avg = luma_weight_avg;
 622         chroma_weight_op = luma_weight_op;
 623         dest_cb += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 624         dest_cr += (2*x_offset << pixel_shift) + 2*y_offset*h->mb_linesize;
 625     } else if (chroma_idc == 2 /* yuv422 */) {
 626         chroma_height = height;
 627         dest_cb += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 628         dest_cr += (  x_offset << pixel_shift) + 2*y_offset*h->mb_uvlinesize;
 629     } else /* yuv420 */ {
 630         chroma_height = height >> 1;
 631         dest_cb += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 632         dest_cr += (  x_offset << pixel_shift) +   y_offset*h->mb_uvlinesize;
 633     }
 634     x_offset += 8*s->mb_x;
 635     y_offset += 8*(s->mb_y >> MB_FIELD);
 636
 637     if(list0 && list1){
 638         /* don't optimize for luma-only case, since B-frames usually
 639          * use implicit weights => chroma too. */
 640         uint8_t *tmp_cb = s->obmc_scratchpad;
 641         uint8_t *tmp_cr = s->obmc_scratchpad + (16 << pixel_shift);
 642         uint8_t *tmp_y  = s->obmc_scratchpad + 16*h->mb_uvlinesize;
 643         int refn0 = h->ref_cache[0][ scan8[n] ];
 644         int refn1 = h->ref_cache[1][ scan8[n] ];
 645
 646         mc_dir_part(h, &h->ref_list[0][refn0], n, square, height, delta, 0,
 647                     dest_y, dest_cb, dest_cr,
 648                     x_offset, y_offset, qpix_put, chroma_put,
 649                     pixel_shift, chroma_idc);
 650         mc_dir_part(h, &h->ref_list[1][refn1], n, square, height, delta, 1,
 651                     tmp_y, tmp_cb, tmp_cr,
 652                     x_offset, y_offset, qpix_put, chroma_put,
 653                     pixel_shift, chroma_idc);
 654
 655         if(h->use_weight == 2){
 656             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 657             int weight1 = 64 - weight0;
 658             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize,
 659                               height,        5, weight0, weight1, 0);
 660             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize,
 661                               chroma_height, 5, weight0, weight1, 0);
 662             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize,
 663                               chroma_height, 5, weight0, weight1, 0);
 664         }else{
 665             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 666                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 667                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 668             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 669                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 670                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 671             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 672                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 673                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 674         }
 675     }else{
 676         int list = list1 ? 1 : 0;
 677         int refn = h->ref_cache[list][ scan8[n] ];
 678         Picture *ref= &h->ref_list[list][refn];
 679         mc_dir_part(h, ref, n, square, height, delta, list,
 680                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 681                     qpix_put, chroma_put, pixel_shift, chroma_idc);
 682
 683         luma_weight_op(dest_y, h->mb_linesize, height, h->luma_log2_weight_denom,
 684                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 685         if(h->use_weight_chroma){
 686             chroma_weight_op(dest_cb, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 687                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 688             chroma_weight_op(dest_cr, h->mb_uvlinesize, chroma_height, h->chroma_log2_weight_denom,
 689                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 690         }
 691     }
 692 }
 693
 694 static av_always_inline void
 695 mc_part(H264Context *h, int n, int square, int height, int delta,
 696         uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 697         int x_offset, int y_offset,
 698         qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 699         qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 700         h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 701         int list0, int list1, int pixel_shift, int chroma_idc)
 702 {
 703     if((h->use_weight==2 && list0 && list1
 704         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 705        || h->use_weight==1)
 706         mc_part_weighted(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 707                          x_offset, y_offset, qpix_put, chroma_put,
 708                          weight_op[0], weight_op[1], weight_avg[0],
 709                          weight_avg[1], list0, list1, pixel_shift, chroma_idc);
 710     else
 711         mc_part_std(h, n, square, height, delta, dest_y, dest_cb, dest_cr,
 712                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg,
 713                     chroma_avg, list0, list1, pixel_shift, chroma_idc);
 714 }
 715
 716 static av_always_inline void
 717 prefetch_motion(H264Context *h, int list, int pixel_shift, int chroma_idc)
 718 {
 719     /* fetch pixels for estimated mv 4 macroblocks ahead
 720      * optimized for 64byte cache lines */
 721     MpegEncContext * const s = &h->s;
 722     const int refn = h->ref_cache[list][scan8[0]];
 723     if(refn >= 0){
 724         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 725         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 726         uint8_t **src = h->ref_list[list][refn].f.data;
 727         int off= (mx << pixel_shift) + (my + (s->mb_x&3)*4)*h->mb_linesize + (64 << pixel_shift);
 728         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 729         if (chroma_idc == 3 /* yuv444 */) {
 730             s->dsp.prefetch(src[1]+off, s->linesize, 4);
 731             s->dsp.prefetch(src[2]+off, s->linesize, 4);
 732         }else{
 733             off= (((mx>>1)+64)<<pixel_shift) + ((my>>1) + (s->mb_x&7))*s->uvlinesize;
 734             s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 735         }
 736     }
 737 }
 738
 739 static av_always_inline void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 740                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 741                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 742                       h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 743                       int pixel_shift, int chroma_idc)
 744 {
 745     MpegEncContext * const s = &h->s;
 746     const int mb_xy= h->mb_xy;
 747     const int mb_type = s->current_picture.f.mb_type[mb_xy];
 748
 749     assert(IS_INTER(mb_type));
 750
 751     if(HAVE_THREADS && (s->avctx->active_thread_type & FF_THREAD_FRAME))
 752         await_references(h);
 753     prefetch_motion(h, 0, pixel_shift, chroma_idc);
 754
 755     if(IS_16X16(mb_type)){
 756         mc_part(h, 0, 1, 16, 0, dest_y, dest_cb, dest_cr, 0, 0,
 757                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 758                 weight_op, weight_avg,
 759                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 760                 pixel_shift, chroma_idc);
 761     }else if(IS_16X8(mb_type)){
 762         mc_part(h, 0, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 0,
 763                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 764                 weight_op, weight_avg,
 765                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 766                 pixel_shift, chroma_idc);
 767         mc_part(h, 8, 0, 8, 8 << pixel_shift, dest_y, dest_cb, dest_cr, 0, 4,
 768                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 769                 weight_op, weight_avg,
 770                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 771                 pixel_shift, chroma_idc);
 772     }else if(IS_8X16(mb_type)){
 773         mc_part(h, 0, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 774                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 775                 &weight_op[1], &weight_avg[1],
 776                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1),
 777                 pixel_shift, chroma_idc);
 778         mc_part(h, 4, 0, 16, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 779                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 780                 &weight_op[1], &weight_avg[1],
 781                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1),
 782                 pixel_shift, chroma_idc);
 783     }else{
 784         int i;
 785
 786         assert(IS_8X8(mb_type));
 787
 788         for(i=0; i<4; i++){
 789             const int sub_mb_type= h->sub_mb_type[i];
 790             const int n= 4*i;
 791             int x_offset= (i&1)<<2;
 792             int y_offset= (i&2)<<1;
 793
 794             if(IS_SUB_8X8(sub_mb_type)){
 795                 mc_part(h, n, 1, 8, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 796                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 797                     &weight_op[1], &weight_avg[1],
 798                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 799                     pixel_shift, chroma_idc);
 800             }else if(IS_SUB_8X4(sub_mb_type)){
 801                 mc_part(h, n  , 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 802                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 803                     &weight_op[1], &weight_avg[1],
 804                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 805                     pixel_shift, chroma_idc);
 806                 mc_part(h, n+2, 0, 4, 4 << pixel_shift, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 807                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 808                     &weight_op[1], &weight_avg[1],
 809                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 810                     pixel_shift, chroma_idc);
 811             }else if(IS_SUB_4X8(sub_mb_type)){
 812                 mc_part(h, n  , 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 813                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 814                     &weight_op[2], &weight_avg[2],
 815                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 816                     pixel_shift, chroma_idc);
 817                 mc_part(h, n+1, 0, 8, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 818                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 819                     &weight_op[2], &weight_avg[2],
 820                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 821                     pixel_shift, chroma_idc);
 822             }else{
 823                 int j;
 824                 assert(IS_SUB_4X4(sub_mb_type));
 825                 for(j=0; j<4; j++){
 826                     int sub_x_offset= x_offset + 2*(j&1);
 827                     int sub_y_offset= y_offset +   (j&2);
 828                     mc_part(h, n+j, 1, 4, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 829                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 830                         &weight_op[2], &weight_avg[2],
 831                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1),
 832                         pixel_shift, chroma_idc);
 833                 }
 834             }
 835         }
 836     }
 837
 838     prefetch_motion(h, 1, pixel_shift, chroma_idc);
 839 }
 840
 841 static av_always_inline void
 842 hl_motion_420(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 843               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 844               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 845               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 846               int pixel_shift)
 847 {
 848     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 849               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 1);
 850 }
 851
 852 static av_always_inline void
 853 hl_motion_422(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 854               qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 855               qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 856               h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 857               int pixel_shift)
 858 {
 859     hl_motion(h, dest_y, dest_cb, dest_cr, qpix_put, chroma_put,
 860               qpix_avg, chroma_avg, weight_op, weight_avg, pixel_shift, 2);
 861 }
 862
 863 static void free_tables(H264Context *h, int free_rbsp){
 864     int i;
 865     H264Context *hx;
 866
 867     av_freep(&h->intra4x4_pred_mode);
 868     av_freep(&h->chroma_pred_mode_table);
 869     av_freep(&h->cbp_table);
 870     av_freep(&h->mvd_table[0]);
 871     av_freep(&h->mvd_table[1]);
 872     av_freep(&h->direct_table);
 873     av_freep(&h->non_zero_count);
 874     av_freep(&h->slice_table_base);
 875     h->slice_table= NULL;
 876     av_freep(&h->list_counts);
 877
 878     av_freep(&h->mb2b_xy);
 879     av_freep(&h->mb2br_xy);
 880
 881     for(i = 0; i < MAX_THREADS; i++) {
 882         hx = h->thread_context[i];
 883         if(!hx) continue;
 884         av_freep(&hx->top_borders[1]);
 885         av_freep(&hx->top_borders[0]);
 886         av_freep(&hx->s.obmc_scratchpad);
 887         if (free_rbsp){
 888             av_freep(&hx->rbsp_buffer[1]);
 889             av_freep(&hx->rbsp_buffer[0]);
 890             hx->rbsp_buffer_size[0] = 0;
 891             hx->rbsp_buffer_size[1] = 0;
 892         }
 893         if (i) av_freep(&h->thread_context[i]);
 894     }
 895 }
 896
 897 static void init_dequant8_coeff_table(H264Context *h){
 898     int i,j,q,x;
 899     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 900
 901     for(i=0; i<6; i++ ){
 902         h->dequant8_coeff[i] = h->dequant8_buffer[i];
 903         for(j=0; j<i; j++){
 904             if(!memcmp(h->pps.scaling_matrix8[j], h->pps.scaling_matrix8[i], 64*sizeof(uint8_t))){
 905                 h->dequant8_coeff[i] = h->dequant8_buffer[j];
 906                 break;
 907             }
 908         }
 909         if(j<i)
 910             continue;
 911
 912         for(q=0; q<max_qp+1; q++){
 913             int shift = div6[q];
 914             int idx = rem6[q];
 915             for(x=0; x<64; x++)
 916                 h->dequant8_coeff[i][q][(x>>3)|((x&7)<<3)] =
 917                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 918                     h->pps.scaling_matrix8[i][x]) << shift;
 919         }
 920     }
 921 }
 922
 923 static void init_dequant4_coeff_table(H264Context *h){
 924     int i,j,q,x;
 925     const int max_qp = 51 + 6*(h->sps.bit_depth_luma-8);
 926     for(i=0; i<6; i++ ){
 927         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 928         for(j=0; j<i; j++){
 929             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 930                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 931                 break;
 932             }
 933         }
 934         if(j<i)
 935             continue;
 936
 937         for(q=0; q<max_qp+1; q++){
 938             int shift = div6[q] + 2;
 939             int idx = rem6[q];
 940             for(x=0; x<16; x++)
 941                 h->dequant4_coeff[i][q][(x>>2)|((x<<2)&0xF)] =
 942                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 943                     h->pps.scaling_matrix4[i][x]) << shift;
 944         }
 945     }
 946 }
 947
 948 static void init_dequant_tables(H264Context *h){
 949     int i,x;
 950     init_dequant4_coeff_table(h);
 951     if(h->pps.transform_8x8_mode)
 952         init_dequant8_coeff_table(h);
 953     if(h->sps.transform_bypass){
 954         for(i=0; i<6; i++)
 955             for(x=0; x<16; x++)
 956                 h->dequant4_coeff[i][0][x] = 1<<6;
 957         if(h->pps.transform_8x8_mode)
 958             for(i=0; i<6; i++)
 959                 for(x=0; x<64; x++)
 960                     h->dequant8_coeff[i][0][x] = 1<<6;
 961     }
 962 }
 963
 964
 965 int ff_h264_alloc_tables(H264Context *h){
 966     MpegEncContext * const s = &h->s;
 967     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 968     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 969     int x,y;
 970
 971     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 972
 973     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 48 * sizeof(uint8_t), fail)
 974     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 975     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 976
 977     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 978     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 979     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 980     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 981     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 982
 983     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 984     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 985
 986     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 987     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 988     for(y=0; y<s->mb_height; y++){
 989         for(x=0; x<s->mb_width; x++){
 990             const int mb_xy= x + y*s->mb_stride;
 991             const int b_xy = 4*x + 4*y*h->b_stride;
 992
 993             h->mb2b_xy [mb_xy]= b_xy;
 994             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 995         }
 996     }
 997
 998     s->obmc_scratchpad = NULL;
 999
1000     if(!h->dequant4_coeff[0])
1001         init_dequant_tables(h);
1002
1003     return 0;
1004 fail:
1005     free_tables(h, 1);
1006     return -1;
1007 }
1008
1009 /**
1010  * Mimic alloc_tables(), but for every context thread.
1011  */
1012 static void clone_tables(H264Context *dst, H264Context *src, int i){
1013     MpegEncContext * const s = &src->s;
1014     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
1015     dst->non_zero_count           = src->non_zero_count;
1016     dst->slice_table              = src->slice_table;
1017     dst->cbp_table                = src->cbp_table;
1018     dst->mb2b_xy                  = src->mb2b_xy;
1019     dst->mb2br_xy                 = src->mb2br_xy;
1020     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
1021     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
1022     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
1023     dst->direct_table             = src->direct_table;
1024     dst->list_counts              = src->list_counts;
1025
1026     dst->s.obmc_scratchpad = NULL;
1027     ff_h264_pred_init(&dst->hpc, src->s.codec_id, src->sps.bit_depth_luma, src->sps.chroma_format_idc);
1028 }
1029
1030 /**
1031  * Init context
1032  * Allocate buffers which are not shared amongst multiple threads.
1033  */
1034 static int context_init(H264Context *h){
1035     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1036     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * 16*3 * sizeof(uint8_t)*2, fail)
1037
1038     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
1039     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
1040
1041     return 0;
1042 fail:
1043     return -1; // free_tables will clean up for us
1044 }
1045
1046 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
1047
1048 static av_cold void common_init(H264Context *h){
1049     MpegEncContext * const s = &h->s;
1050
1051     s->width = s->avctx->width;
1052     s->height = s->avctx->height;
1053     s->codec_id= s->avctx->codec->id;
1054
1055     s->avctx->bits_per_raw_sample = 8;
1056     h->cur_chroma_format_idc = 1;
1057
1058     ff_h264dsp_init(&h->h264dsp,
1059                     s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1060     ff_h264_pred_init(&h->hpc, s->codec_id,
1061                       s->avctx->bits_per_raw_sample, h->cur_chroma_format_idc);
1062
1063     h->dequant_coeff_pps= -1;
1064     s->unrestricted_mv=1;
1065
1066     s->dsp.dct_bits = 16;
1067     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
1068
1069     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
1070     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
1071 }
1072
1073 int ff_h264_decode_extradata(H264Context *h, const uint8_t *buf, int size)
1074 {
1075     AVCodecContext *avctx = h->s.avctx;
1076
1077     if(!buf || size <= 0)
1078         return -1;
1079
1080     if(buf[0] == 1){
1081         int i, cnt, nalsize;
1082         const unsigned char *p = buf;
1083
1084         h->is_avc = 1;
1085
1086         if(size < 7) {
1087             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
1088             return -1;
1089         }
1090         /* sps and pps in the avcC always have length coded with 2 bytes,
1091            so put a fake nal_length_size = 2 while parsing them */
1092         h->nal_length_size = 2;
1093         // Decode sps from avcC
1094         cnt = *(p+5) & 0x1f; // Number of sps
1095         p += 6;
1096         for (i = 0; i < cnt; i++) {
1097             nalsize = AV_RB16(p) + 2;
1098             if(nalsize > size - (p-buf))
1099                 return -1;
1100             if(decode_nal_units(h, p, nalsize) < 0) {
1101                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
1102                 return -1;
1103             }
1104             p += nalsize;
1105         }
1106         // Decode pps from avcC
1107         cnt = *(p++); // Number of pps
1108         for (i = 0; i < cnt; i++) {
1109             nalsize = AV_RB16(p) + 2;
1110             if(nalsize > size - (p-buf))
1111                 return -1;
1112             if (decode_nal_units(h, p, nalsize) < 0) {
1113                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
1114                 return -1;
1115             }
1116             p += nalsize;
1117         }
1118         // Now store right nal length size, that will be use to parse all other nals
1119         h->nal_length_size = (buf[4] & 0x03) + 1;
1120     } else {
1121         h->is_avc = 0;
1122         if(decode_nal_units(h, buf, size) < 0)
1123             return -1;
1124     }
1125     return 0;
1126 }
1127
1128 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
1129     H264Context *h= avctx->priv_data;
1130     MpegEncContext * const s = &h->s;
1131     int i;
1132
1133     MPV_decode_defaults(s);
1134
1135     s->avctx = avctx;
1136     common_init(h);
1137
1138     s->out_format = FMT_H264;
1139     s->workaround_bugs= avctx->workaround_bugs;
1140
1141     // set defaults
1142 //    s->decode_mb= ff_h263_decode_mb;
1143     s->quarter_sample = 1;
1144     if(!avctx->has_b_frames)
1145     s->low_delay= 1;
1146
1147     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
1148
1149     ff_h264_decode_init_vlc();
1150
1151     h->pixel_shift = 0;
1152     h->sps.bit_depth_luma = avctx->bits_per_raw_sample = 8;
1153
1154     h->thread_context[0] = h;
1155     h->outputed_poc = h->next_outputed_poc = INT_MIN;
1156     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1157         h->last_pocs[i] = INT_MIN;
1158     h->prev_poc_msb= 1<<16;
1159     h->prev_frame_num= -1;
1160     h->x264_build = -1;
1161     ff_h264_reset_sei(h);
1162     if(avctx->codec_id == CODEC_ID_H264){
1163         if(avctx->ticks_per_frame == 1){
1164             s->avctx->time_base.den *=2;
1165         }
1166         avctx->ticks_per_frame = 2;
1167     }
1168
1169     if(avctx->extradata_size > 0 && avctx->extradata &&
1170         ff_h264_decode_extradata(h, avctx->extradata, avctx->extradata_size))
1171         return -1;
1172
1173     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1174         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1175         s->low_delay = 0;
1176     }
1177
1178     return 0;
1179 }
1180
1181 #define IN_RANGE(a, b, size) (((a) >= (b)) && ((a) < ((b)+(size))))
1182 static void copy_picture_range(Picture **to, Picture **from, int count, MpegEncContext *new_base, MpegEncContext *old_base)
1183 {
1184     int i;
1185
1186     for (i=0; i<count; i++){
1187         assert((IN_RANGE(from[i], old_base, sizeof(*old_base)) ||
1188                 IN_RANGE(from[i], old_base->picture, sizeof(Picture) * old_base->picture_count) ||
1189                 !from[i]));
1190         to[i] = REBASE_PICTURE(from[i], new_base, old_base);
1191     }
1192 }
1193
1194 static void copy_parameter_set(void **to, void **from, int count, int size)
1195 {
1196     int i;
1197
1198     for (i=0; i<count; i++){
1199         if (to[i] && !from[i]) av_freep(&to[i]);
1200         else if (from[i] && !to[i]) to[i] = av_malloc(size);
1201
1202         if (from[i]) memcpy(to[i], from[i], size);
1203     }
1204 }
1205
1206 static int decode_init_thread_copy(AVCodecContext *avctx){
1207     H264Context *h= avctx->priv_data;
1208
1209     if (!avctx->internal->is_copy)
1210         return 0;
1211     memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1212     memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1213
1214     return 0;
1215 }
1216
1217 #define copy_fields(to, from, start_field, end_field) memcpy(&to->start_field, &from->start_field, (char*)&to->end_field - (char*)&to->start_field)
1218 static int decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src){
1219     H264Context *h= dst->priv_data, *h1= src->priv_data;
1220     MpegEncContext * const s = &h->s, * const s1 = &h1->s;
1221     int inited = s->context_initialized, err;
1222     int i;
1223
1224     if(dst == src || !s1->context_initialized) return 0;
1225
1226     err = ff_mpeg_update_thread_context(dst, src);
1227     if(err) return err;
1228
1229     //FIXME handle width/height changing
1230     if(!inited){
1231         for(i = 0; i < MAX_SPS_COUNT; i++)
1232             av_freep(h->sps_buffers + i);
1233
1234         for(i = 0; i < MAX_PPS_COUNT; i++)
1235             av_freep(h->pps_buffers + i);
1236
1237         memcpy(&h->s + 1, &h1->s + 1, sizeof(H264Context) - sizeof(MpegEncContext)); //copy all fields after MpegEnc
1238         memset(h->sps_buffers, 0, sizeof(h->sps_buffers));
1239         memset(h->pps_buffers, 0, sizeof(h->pps_buffers));
1240         if (ff_h264_alloc_tables(h) < 0) {
1241             av_log(dst, AV_LOG_ERROR, "Could not allocate memory for h264\n");
1242             return AVERROR(ENOMEM);
1243         }
1244         context_init(h);
1245
1246         for(i=0; i<2; i++){
1247             h->rbsp_buffer[i] = NULL;
1248             h->rbsp_buffer_size[i] = 0;
1249         }
1250
1251         h->thread_context[0] = h;
1252
1253         // frame_start may not be called for the next thread (if it's decoding a bottom field)
1254         // so this has to be allocated here
1255         h->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1256
1257         s->dsp.clear_blocks(h->mb);
1258         s->dsp.clear_blocks(h->mb+(24*16<<h->pixel_shift));
1259     }
1260
1261     //extradata/NAL handling
1262     h->is_avc          = h1->is_avc;
1263
1264     //SPS/PPS
1265     copy_parameter_set((void**)h->sps_buffers, (void**)h1->sps_buffers, MAX_SPS_COUNT, sizeof(SPS));
1266     h->sps             = h1->sps;
1267     copy_parameter_set((void**)h->pps_buffers, (void**)h1->pps_buffers, MAX_PPS_COUNT, sizeof(PPS));
1268     h->pps             = h1->pps;
1269
1270     //Dequantization matrices
1271     //FIXME these are big - can they be only copied when PPS changes?
1272     copy_fields(h, h1, dequant4_buffer, dequant4_coeff);
1273
1274     for(i=0; i<6; i++)
1275         h->dequant4_coeff[i] = h->dequant4_buffer[0] + (h1->dequant4_coeff[i] - h1->dequant4_buffer[0]);
1276
1277     for(i=0; i<6; i++)
1278         h->dequant8_coeff[i] = h->dequant8_buffer[0] + (h1->dequant8_coeff[i] - h1->dequant8_buffer[0]);
1279
1280     h->dequant_coeff_pps = h1->dequant_coeff_pps;
1281
1282     //POC timing
1283     copy_fields(h, h1, poc_lsb, redundant_pic_count);
1284
1285     //reference lists
1286     copy_fields(h, h1, ref_count, list_count);
1287     copy_fields(h, h1, ref_list,  intra_gb);
1288     copy_fields(h, h1, short_ref, cabac_init_idc);
1289
1290     copy_picture_range(h->short_ref,   h1->short_ref,   32, s, s1);
1291     copy_picture_range(h->long_ref,    h1->long_ref,    32, s, s1);
1292     copy_picture_range(h->delayed_pic, h1->delayed_pic, MAX_DELAYED_PIC_COUNT+2, s, s1);
1293
1294     h->last_slice_type = h1->last_slice_type;
1295     h->sync            = h1->sync;
1296
1297     if(!s->current_picture_ptr) return 0;
1298
1299     if(!s->dropable) {
1300         err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1301         h->prev_poc_msb     = h->poc_msb;
1302         h->prev_poc_lsb     = h->poc_lsb;
1303     }
1304     h->prev_frame_num_offset= h->frame_num_offset;
1305     h->prev_frame_num       = h->frame_num;
1306     h->outputed_poc         = h->next_outputed_poc;
1307
1308     return err;
1309 }
1310
1311 int ff_h264_frame_start(H264Context *h){
1312     MpegEncContext * const s = &h->s;
1313     int i;
1314     const int pixel_shift = h->pixel_shift;
1315     int thread_count = (s->avctx->active_thread_type & FF_THREAD_SLICE) ? s->avctx->thread_count : 1;
1316
1317     if(MPV_frame_start(s, s->avctx) < 0)
1318         return -1;
1319     ff_er_frame_start(s);
1320     /*
1321      * MPV_frame_start uses pict_type to derive key_frame.
1322      * This is incorrect for H.264; IDR markings must be used.
1323      * Zero here; IDR markings per slice in frame or fields are ORed in later.
1324      * See decode_nal_units().
1325      */
1326     s->current_picture_ptr->f.key_frame = 0;
1327     s->current_picture_ptr->mmco_reset= 0;
1328
1329     assert(s->linesize && s->uvlinesize);
1330
1331     for(i=0; i<16; i++){
1332         h->block_offset[i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
1333         h->block_offset[48+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
1334     }
1335     for(i=0; i<16; i++){
1336         h->block_offset[16+i]=
1337         h->block_offset[32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1338         h->block_offset[48+16+i]=
1339         h->block_offset[48+32+i]= (4*((scan8[i] - scan8[0])&7) << pixel_shift) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
1340     }
1341
1342     /* can't be in alloc_tables because linesize isn't known there.
1343      * FIXME: redo bipred weight to not require extra buffer? */
1344     for(i = 0; i < thread_count; i++)
1345         if(h->thread_context[i] && !h->thread_context[i]->s.obmc_scratchpad)
1346             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*6*s->linesize);
1347
1348     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
1349     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
1350
1351 //    s->decode = (s->flags & CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.f.reference /*|| h->contains_intra*/ || 1;
1352
1353     // We mark the current picture as non-reference after allocating it, so
1354     // that if we break out due to an error it can be released automatically
1355     // in the next MPV_frame_start().
1356     // SVQ3 as well as most other codecs have only last/next/current and thus
1357     // get released even with set reference, besides SVQ3 and others do not
1358     // mark frames as reference later "naturally".
1359     if(s->codec_id != CODEC_ID_SVQ3)
1360         s->current_picture_ptr->f.reference = 0;
1361
1362     s->current_picture_ptr->field_poc[0]=
1363     s->current_picture_ptr->field_poc[1]= INT_MAX;
1364
1365     h->next_output_pic = NULL;
1366
1367     assert(s->current_picture_ptr->long_ref==0);
1368
1369     return 0;
1370 }
1371
1372 /**
1373   * Run setup operations that must be run after slice header decoding.
1374   * This includes finding the next displayed frame.
1375   *
1376   * @param h h264 master context
1377   * @param setup_finished enough NALs have been read that we can call
1378   * ff_thread_finish_setup()
1379   */
1380 static void decode_postinit(H264Context *h, int setup_finished){
1381     MpegEncContext * const s = &h->s;
1382     Picture *out = s->current_picture_ptr;
1383     Picture *cur = s->current_picture_ptr;
1384     int i, pics, out_of_order, out_idx;
1385     int invalid = 0, cnt = 0;
1386
1387     s->current_picture_ptr->f.qscale_type = FF_QSCALE_TYPE_H264;
1388     s->current_picture_ptr->f.pict_type   = s->pict_type;
1389
1390     if (h->next_output_pic) return;
1391
1392     if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
1393         //FIXME: if we have two PAFF fields in one packet, we can't start the next thread here.
1394         //If we have one field per packet, we can. The check in decode_nal_units() is not good enough
1395         //to find this yet, so we assume the worst for now.
1396         //if (setup_finished)
1397         //    ff_thread_finish_setup(s->avctx);
1398         return;
1399     }
1400
1401     cur->f.interlaced_frame = 0;
1402     cur->f.repeat_pict      = 0;
1403
1404     /* Signal interlacing information externally. */
1405     /* Prioritize picture timing SEI information over used decoding process if it exists. */
1406
1407     if(h->sps.pic_struct_present_flag){
1408         switch (h->sei_pic_struct)
1409         {
1410         case SEI_PIC_STRUCT_FRAME:
1411             break;
1412         case SEI_PIC_STRUCT_TOP_FIELD:
1413         case SEI_PIC_STRUCT_BOTTOM_FIELD:
1414             cur->f.interlaced_frame = 1;
1415             break;
1416         case SEI_PIC_STRUCT_TOP_BOTTOM:
1417         case SEI_PIC_STRUCT_BOTTOM_TOP:
1418             if (FIELD_OR_MBAFF_PICTURE)
1419                 cur->f.interlaced_frame = 1;
1420             else
1421                 // try to flag soft telecine progressive
1422                 cur->f.interlaced_frame = h->prev_interlaced_frame;
1423             break;
1424         case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
1425         case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
1426             // Signal the possibility of telecined film externally (pic_struct 5,6)
1427             // From these hints, let the applications decide if they apply deinterlacing.
1428             cur->f.repeat_pict = 1;
1429             break;
1430         case SEI_PIC_STRUCT_FRAME_DOUBLING:
1431             // Force progressive here, as doubling interlaced frame is a bad idea.
1432             cur->f.repeat_pict = 2;
1433             break;
1434         case SEI_PIC_STRUCT_FRAME_TRIPLING:
1435             cur->f.repeat_pict = 4;
1436             break;
1437         }
1438
1439         if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
1440             cur->f.interlaced_frame = (h->sei_ct_type & (1 << 1)) != 0;
1441     }else{
1442         /* Derive interlacing flag from used decoding process. */
1443         cur->f.interlaced_frame = FIELD_OR_MBAFF_PICTURE;
1444     }
1445     h->prev_interlaced_frame = cur->f.interlaced_frame;
1446
1447     if (cur->field_poc[0] != cur->field_poc[1]){
1448         /* Derive top_field_first from field pocs. */
1449         cur->f.top_field_first = cur->field_poc[0] < cur->field_poc[1];
1450     }else{
1451         if (cur->f.interlaced_frame || h->sps.pic_struct_present_flag) {
1452             /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
1453             if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
1454               || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
1455                 cur->f.top_field_first = 1;
1456             else
1457                 cur->f.top_field_first = 0;
1458         }else{
1459             /* Most likely progressive */
1460             cur->f.top_field_first = 0;
1461         }
1462     }
1463
1464     //FIXME do something with unavailable reference frames
1465
1466     /* Sort B-frames into display order */
1467
1468     if(h->sps.bitstream_restriction_flag
1469        && s->avctx->has_b_frames < h->sps.num_reorder_frames){
1470         s->avctx->has_b_frames = h->sps.num_reorder_frames;
1471         s->low_delay = 0;
1472     }
1473
1474     if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
1475        && !h->sps.bitstream_restriction_flag){
1476         s->avctx->has_b_frames = MAX_DELAYED_PIC_COUNT - 1;
1477         s->low_delay= 0;
1478     }
1479
1480     pics = 0;
1481     while(h->delayed_pic[pics]) pics++;
1482
1483     av_assert0(pics <= MAX_DELAYED_PIC_COUNT);
1484
1485     h->delayed_pic[pics++] = cur;
1486     if (cur->f.reference == 0)
1487         cur->f.reference = DELAYED_PIC_REF;
1488
1489     /* Frame reordering. This code takes pictures from coding order and sorts
1490      * them by their incremental POC value into display order. It supports POC
1491      * gaps, MMCO reset codes and random resets.
1492      * A "display group" can start either with a IDR frame (f.key_frame = 1),
1493      * and/or can be closed down with a MMCO reset code. In sequences where
1494      * there is no delay, we can't detect that (since the frame was already
1495      * output to the user), so we also set h->mmco_reset to detect the MMCO
1496      * reset code.
1497      * FIXME: if we detect insufficient delays (as per s->avctx->has_b_frames),
1498      * we increase the delay between input and output. All frames affected by
1499      * the lag (e.g. those that should have been output before another frame
1500      * that we already returned to the user) will be dropped. This is a bug
1501      * that we will fix later. */
1502     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++) {
1503         cnt     += out->poc < h->last_pocs[i];
1504         invalid += out->poc == INT_MIN;
1505     }
1506     if (!h->mmco_reset && !cur->f.key_frame && cnt + invalid == MAX_DELAYED_PIC_COUNT && cnt > 0) {
1507         h->mmco_reset = 2;
1508         if (pics > 1)
1509             h->delayed_pic[pics - 2]->mmco_reset = 2;
1510     }
1511     if (h->mmco_reset || cur->f.key_frame) {
1512         for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
1513             h->last_pocs[i] = INT_MIN;
1514         cnt     = 0;
1515         invalid = MAX_DELAYED_PIC_COUNT;
1516     }
1517     out = h->delayed_pic[0];
1518     out_idx = 0;
1519     for (i = 1; i < MAX_DELAYED_PIC_COUNT && h->delayed_pic[i] &&
1520          !h->delayed_pic[i-1]->mmco_reset && !h->delayed_pic[i]->f.key_frame; i++)
1521     {
1522         if(h->delayed_pic[i]->poc < out->poc){
1523             out = h->delayed_pic[i];
1524             out_idx = i;
1525         }
1526     }
1527     if (s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->f.key_frame || h->mmco_reset))
1528         h->next_outputed_poc = INT_MIN;
1529     out_of_order = !out->f.key_frame && !h->mmco_reset && (out->poc < h->next_outputed_poc);
1530
1531     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
1532         { }
1533     else if (out_of_order && pics-1 == s->avctx->has_b_frames &&
1534              s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT) {
1535         if (invalid + cnt < MAX_DELAYED_PIC_COUNT) {
1536             if(cnt > s->avctx->has_b_frames)
1537                 av_log(s->avctx, AV_LOG_WARNING, "Increasing reorder buffer to %d\n", cnt);
1538             s->avctx->has_b_frames = FFMAX(s->avctx->has_b_frames, cnt);
1539         }
1540         s->low_delay = 0;
1541     } else if (s->low_delay &&
1542                ((h->next_outputed_poc != INT_MIN && out->poc > h->next_outputed_poc + 2) ||
1543                 cur->f.pict_type == AV_PICTURE_TYPE_B)) {
1544         s->low_delay = 0;
1545         s->avctx->has_b_frames++;
1546     }
1547
1548     if(pics > s->avctx->has_b_frames){
1549         out->f.reference &= ~DELAYED_PIC_REF;
1550         out->owner2 = s; // for frame threading, the owner must be the second field's thread
1551                          // or else the first thread can release the picture and reuse it unsafely
1552         for(i=out_idx; h->delayed_pic[i]; i++)
1553             h->delayed_pic[i] = h->delayed_pic[i+1];
1554     }
1555     memmove(h->last_pocs, &h->last_pocs[1], sizeof(*h->last_pocs) * (MAX_DELAYED_PIC_COUNT - 1));
1556     h->last_pocs[MAX_DELAYED_PIC_COUNT - 1] = cur->poc;
1557     if(!out_of_order && pics > s->avctx->has_b_frames){
1558         h->next_output_pic = out;
1559         if (out->mmco_reset) {
1560             if (out_idx > 0) {
1561                 h->next_outputed_poc = out->poc;
1562                 h->delayed_pic[out_idx - 1]->mmco_reset = out->mmco_reset;
1563             } else {
1564                 h->next_outputed_poc = INT_MIN;
1565             }
1566         } else {
1567             h->next_outputed_poc = out->poc;
1568         }
1569         h->mmco_reset = 0;
1570     }else{
1571         av_log(s->avctx, AV_LOG_DEBUG, "no picture\n");
1572     }
1573
1574     if (h->next_output_pic && h->next_output_pic->sync) {
1575         h->sync |= 2;
1576     }
1577
1578     if (setup_finished)
1579         ff_thread_finish_setup(s->avctx);
1580 }
1581
1582 static av_always_inline void backup_mb_border(H264Context *h, uint8_t *src_y,
1583                                               uint8_t *src_cb, uint8_t *src_cr,
1584                                               int linesize, int uvlinesize, int simple)
1585 {
1586     MpegEncContext * const s = &h->s;
1587     uint8_t *top_border;
1588     int top_idx = 1;
1589     const int pixel_shift = h->pixel_shift;
1590     int chroma444 = CHROMA444;
1591     int chroma422 = CHROMA422;
1592
1593     src_y  -=   linesize;
1594     src_cb -= uvlinesize;
1595     src_cr -= uvlinesize;
1596
1597     if(!simple && FRAME_MBAFF){
1598         if(s->mb_y&1){
1599             if(!MB_MBAFF){
1600                 top_border = h->top_borders[0][s->mb_x];
1601                 AV_COPY128(top_border, src_y + 15*linesize);
1602                 if (pixel_shift)
1603                     AV_COPY128(top_border+16, src_y+15*linesize+16);
1604                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1605                     if(chroma444){
1606                         if (pixel_shift){
1607                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1608                             AV_COPY128(top_border+48, src_cb + 15*uvlinesize+16);
1609                             AV_COPY128(top_border+64, src_cr + 15*uvlinesize);
1610                             AV_COPY128(top_border+80, src_cr + 15*uvlinesize+16);
1611                         } else {
1612                             AV_COPY128(top_border+16, src_cb + 15*uvlinesize);
1613                             AV_COPY128(top_border+32, src_cr + 15*uvlinesize);
1614                         }
1615                     } else if(chroma422){
1616                         if (pixel_shift) {
1617                             AV_COPY128(top_border+32, src_cb + 15*uvlinesize);
1618                             AV_COPY128(top_border+48, src_cr + 15*uvlinesize);
1619                         } else {
1620                             AV_COPY64(top_border+16, src_cb +  15*uvlinesize);
1621                             AV_COPY64(top_border+24, src_cr +  15*uvlinesize);
1622                         }
1623                     } else {
1624                         if (pixel_shift) {
1625                             AV_COPY128(top_border+32, src_cb+7*uvlinesize);
1626                             AV_COPY128(top_border+48, src_cr+7*uvlinesize);
1627                         } else {
1628                             AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1629                             AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1630                         }
1631                     }
1632                 }
1633             }
1634         }else if(MB_MBAFF){
1635             top_idx = 0;
1636         }else
1637             return;
1638     }
1639
1640     top_border = h->top_borders[top_idx][s->mb_x];
1641     // There are two lines saved, the line above the the top macroblock of a pair,
1642     // and the line above the bottom macroblock
1643     AV_COPY128(top_border, src_y + 16*linesize);
1644     if (pixel_shift)
1645         AV_COPY128(top_border+16, src_y+16*linesize+16);
1646
1647     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1648         if(chroma444){
1649             if (pixel_shift){
1650                 AV_COPY128(top_border+32, src_cb + 16*linesize);
1651                 AV_COPY128(top_border+48, src_cb + 16*linesize+16);
1652                 AV_COPY128(top_border+64, src_cr + 16*linesize);
1653                 AV_COPY128(top_border+80, src_cr + 16*linesize+16);
1654             } else {
1655                 AV_COPY128(top_border+16, src_cb + 16*linesize);
1656                 AV_COPY128(top_border+32, src_cr + 16*linesize);
1657             }
1658         } else if(chroma422) {
1659             if (pixel_shift) {
1660                 AV_COPY128(top_border+32, src_cb+16*uvlinesize);
1661                 AV_COPY128(top_border+48, src_cr+16*uvlinesize);
1662             } else {
1663                 AV_COPY64(top_border+16, src_cb+16*uvlinesize);
1664                 AV_COPY64(top_border+24, src_cr+16*uvlinesize);
1665             }
1666         } else {
1667             if (pixel_shift) {
1668                 AV_COPY128(top_border+32, src_cb+8*uvlinesize);
1669                 AV_COPY128(top_border+48, src_cr+8*uvlinesize);
1670             } else {
1671                 AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1672                 AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1673             }
1674         }
1675     }
1676 }
1677
1678 static av_always_inline void xchg_mb_border(H264Context *h, uint8_t *src_y,
1679                                   uint8_t *src_cb, uint8_t *src_cr,
1680                                   int linesize, int uvlinesize,
1681                                   int xchg, int chroma444,
1682                                   int simple, int pixel_shift){
1683     MpegEncContext * const s = &h->s;
1684     int deblock_topleft;
1685     int deblock_top;
1686     int top_idx = 1;
1687     uint8_t *top_border_m1;
1688     uint8_t *top_border;
1689
1690     if(!simple && FRAME_MBAFF){
1691         if(s->mb_y&1){
1692             if(!MB_MBAFF)
1693                 return;
1694         }else{
1695             top_idx = MB_MBAFF ? 0 : 1;
1696         }
1697     }
1698
1699     if(h->deblocking_filter == 2) {
1700         deblock_topleft = h->slice_table[h->mb_xy - 1 - s->mb_stride] == h->slice_num;
1701         deblock_top     = h->top_type;
1702     } else {
1703         deblock_topleft = (s->mb_x > 0);
1704         deblock_top     = (s->mb_y > !!MB_FIELD);
1705     }
1706
1707     src_y  -=   linesize + 1 + pixel_shift;
1708     src_cb -= uvlinesize + 1 + pixel_shift;
1709     src_cr -= uvlinesize + 1 + pixel_shift;
1710
1711     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1712     top_border    = h->top_borders[top_idx][s->mb_x];
1713
1714 #define XCHG(a,b,xchg)\
1715     if (pixel_shift) {\
1716         if (xchg) {\
1717             AV_SWAP64(b+0,a+0);\
1718             AV_SWAP64(b+8,a+8);\
1719         } else {\
1720             AV_COPY128(b,a); \
1721         }\
1722     } else \
1723 if (xchg) AV_SWAP64(b,a);\
1724 else      AV_COPY64(b,a);
1725
1726     if(deblock_top){
1727         if(deblock_topleft){
1728             XCHG(top_border_m1 + (8 << pixel_shift), src_y - (7 << pixel_shift), 1);
1729         }
1730         XCHG(top_border + (0 << pixel_shift), src_y + (1 << pixel_shift), xchg);
1731         XCHG(top_border + (8 << pixel_shift), src_y + (9 << pixel_shift), 1);
1732         if(s->mb_x+1 < s->mb_width){
1733             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y + (17 << pixel_shift), 1);
1734         }
1735     }
1736     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1737         if(chroma444){
1738             if(deblock_topleft){
1739                 XCHG(top_border_m1 + (24 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1740                 XCHG(top_border_m1 + (40 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1741             }
1742             XCHG(top_border + (16 << pixel_shift), src_cb + (1 << pixel_shift), xchg);
1743             XCHG(top_border + (24 << pixel_shift), src_cb + (9 << pixel_shift), 1);
1744             XCHG(top_border + (32 << pixel_shift), src_cr + (1 << pixel_shift), xchg);
1745             XCHG(top_border + (40 << pixel_shift), src_cr + (9 << pixel_shift), 1);
1746             if(s->mb_x+1 < s->mb_width){
1747                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (16 << pixel_shift), src_cb + (17 << pixel_shift), 1);
1748                 XCHG(h->top_borders[top_idx][s->mb_x+1] + (32 << pixel_shift), src_cr + (17 << pixel_shift), 1);
1749             }
1750         } else {
1751             if(deblock_top){
1752                 if(deblock_topleft){
1753                     XCHG(top_border_m1 + (16 << pixel_shift), src_cb - (7 << pixel_shift), 1);
1754                     XCHG(top_border_m1 + (24 << pixel_shift), src_cr - (7 << pixel_shift), 1);
1755                 }
1756                 XCHG(top_border + (16 << pixel_shift), src_cb+1+pixel_shift, 1);
1757                 XCHG(top_border + (24 << pixel_shift), src_cr+1+pixel_shift, 1);
1758             }
1759         }
1760     }
1761 }
1762
1763 static av_always_inline int dctcoef_get(DCTELEM *mb, int high_bit_depth, int index) {
1764     if (high_bit_depth) {
1765         return AV_RN32A(((int32_t*)mb) + index);
1766     } else
1767         return AV_RN16A(mb + index);
1768 }
1769
1770 static av_always_inline void dctcoef_set(DCTELEM *mb, int high_bit_depth, int index, int value) {
1771     if (high_bit_depth) {
1772         AV_WN32A(((int32_t*)mb) + index, value);
1773     } else
1774         AV_WN16A(mb + index, value);
1775 }
1776
1777 static av_always_inline void hl_decode_mb_predict_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1778                                                        int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1779 {
1780     MpegEncContext * const s = &h->s;
1781     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1782     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1783     int i;
1784     int qscale = p == 0 ? s->qscale : h->chroma_qp[p-1];
1785     block_offset += 16*p;
1786     if(IS_INTRA4x4(mb_type)){
1787         if(simple || !s->encoding){
1788             if(IS_8x8DCT(mb_type)){
1789                 if(transform_bypass){
1790                     idct_dc_add =
1791                     idct_add    = s->dsp.add_pixels8;
1792                 }else{
1793                     idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1794                     idct_add    = h->h264dsp.h264_idct8_add;
1795                 }
1796                 for(i=0; i<16; i+=4){
1797                     uint8_t * const ptr= dest_y + block_offset[i];
1798                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1799                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1800                         h->hpc.pred8x8l_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1801                     }else{
1802                         const int nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1803                         h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1804                                                     (h->topright_samples_available<<i)&0x4000, linesize);
1805                         if(nnz){
1806                             if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1807                                 idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1808                             else
1809                                 idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1810                         }
1811                     }
1812                 }
1813             }else{
1814                 if(transform_bypass){
1815                     idct_dc_add =
1816                     idct_add    = s->dsp.add_pixels4;
1817                 }else{
1818                     idct_dc_add = h->h264dsp.h264_idct_dc_add;
1819                     idct_add    = h->h264dsp.h264_idct_add;
1820                 }
1821                 for(i=0; i<16; i++){
1822                     uint8_t * const ptr= dest_y + block_offset[i];
1823                     const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1824
1825                     if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1826                         h->hpc.pred4x4_add[dir](ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1827                     }else{
1828                         uint8_t *topright;
1829                         int nnz, tr;
1830                         uint64_t tr_high;
1831                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1832                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1833                             assert(s->mb_y || linesize <= block_offset[i]);
1834                             if(!topright_avail){
1835                                 if (pixel_shift) {
1836                                     tr_high= ((uint16_t*)ptr)[3 - linesize/2]*0x0001000100010001ULL;
1837                                     topright= (uint8_t*) &tr_high;
1838                                 } else {
1839                                     tr= ptr[3 - linesize]*0x01010101u;
1840                                     topright= (uint8_t*) &tr;
1841                                 }
1842                             }else
1843                                 topright= ptr + (4 << pixel_shift) - linesize;
1844                         }else
1845                             topright= NULL;
1846
1847                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1848                         nnz = h->non_zero_count_cache[ scan8[i+p*16] ];
1849                         if(nnz){
1850                             if(is_h264){
1851                                 if(nnz == 1 && dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1852                                     idct_dc_add(ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1853                                 else
1854                                     idct_add   (ptr, h->mb + (i*16+p*256 << pixel_shift), linesize);
1855                             }else
1856                                 ff_svq3_add_idct_c(ptr, h->mb + i*16+p*256, linesize, qscale, 0);
1857                         }
1858                     }
1859                 }
1860             }
1861         }
1862     }else{
1863         h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1864         if(is_h264){
1865             if(h->non_zero_count_cache[ scan8[LUMA_DC_BLOCK_INDEX+p] ]){
1866                 if(!transform_bypass)
1867                     h->h264dsp.h264_luma_dc_dequant_idct(h->mb+(p*256 << pixel_shift), h->mb_luma_dc[p], h->dequant4_coeff[p][qscale][0]);
1868                 else{
1869                     static const uint8_t dc_mapping[16] = { 0*16, 1*16, 4*16, 5*16, 2*16, 3*16, 6*16, 7*16,
1870                                                             8*16, 9*16,12*16,13*16,10*16,11*16,14*16,15*16};
1871                     for(i = 0; i < 16; i++)
1872                         dctcoef_set(h->mb+(p*256 << pixel_shift), pixel_shift, dc_mapping[i], dctcoef_get(h->mb_luma_dc[p], pixel_shift, i));
1873                 }
1874             }
1875         }else
1876             ff_svq3_luma_dc_dequant_idct_c(h->mb+p*256, h->mb_luma_dc[p], qscale);
1877     }
1878 }
1879
1880 static av_always_inline void hl_decode_mb_idct_luma(H264Context *h, int mb_type, int is_h264, int simple, int transform_bypass,
1881                                                     int pixel_shift, int *block_offset, int linesize, uint8_t *dest_y, int p)
1882 {
1883     MpegEncContext * const s = &h->s;
1884     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1885     int i;
1886     block_offset += 16*p;
1887     if(!IS_INTRA4x4(mb_type)){
1888         if(is_h264){
1889             if(IS_INTRA16x16(mb_type)){
1890                 if(transform_bypass){
1891                     if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1892                         h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize);
1893                     }else{
1894                         for(i=0; i<16; i++){
1895                             if(h->non_zero_count_cache[ scan8[i+p*16] ] || dctcoef_get(h->mb, pixel_shift, i*16+p*256))
1896                                 s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1897                         }
1898                     }
1899                 }else{
1900                     h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1901                 }
1902             }else if(h->cbp&15){
1903                 if(transform_bypass){
1904                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1905                     idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1906                     for(i=0; i<16; i+=di){
1907                         if(h->non_zero_count_cache[ scan8[i+p*16] ]){
1908                             idct_add(dest_y + block_offset[i], h->mb + (i*16+p*256 << pixel_shift), linesize);
1909                         }
1910                     }
1911                 }else{
1912                     if(IS_8x8DCT(mb_type)){
1913                         h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1914                     }else{
1915                         h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb + (p*256 << pixel_shift), linesize, h->non_zero_count_cache+p*5*8);
1916                     }
1917                 }
1918             }
1919         }else{
1920             for(i=0; i<16; i++){
1921                 if(h->non_zero_count_cache[ scan8[i+p*16] ] || h->mb[i*16+p*256]){ //FIXME benchmark weird rule, & below
1922                     uint8_t * const ptr= dest_y + block_offset[i];
1923                     ff_svq3_add_idct_c(ptr, h->mb + i*16 + p*256, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1924                 }
1925             }
1926         }
1927     }
1928 }
1929
1930 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple, int pixel_shift)
1931 {
1932     MpegEncContext * const s = &h->s;
1933     const int mb_x= s->mb_x;
1934     const int mb_y= s->mb_y;
1935     const int mb_xy= h->mb_xy;
1936     const int mb_type = s->current_picture.f.mb_type[mb_xy];
1937     uint8_t  *dest_y, *dest_cb, *dest_cr;
1938     int linesize, uvlinesize /*dct_offset*/;
1939     int i, j;
1940     int *block_offset = &h->block_offset[0];
1941     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1942     /* is_h264 should always be true if SVQ3 is disabled. */
1943     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1944     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1945     const int block_h = 16 >> s->chroma_y_shift;
1946     const int chroma422 = CHROMA422;
1947
1948     dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
1949     dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1950     dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift)*8 + mb_y * s->uvlinesize * block_h;
1951
1952     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
1953     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + (64 << pixel_shift), dest_cr - dest_cb, 2);
1954
1955     h->list_counts[mb_xy]= h->list_count;
1956
1957     if (!simple && MB_FIELD) {
1958         linesize   = h->mb_linesize   = s->linesize * 2;
1959         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1960         block_offset = &h->block_offset[48];
1961         if(mb_y&1){ //FIXME move out of this function?
1962             dest_y -= s->linesize*15;
1963             dest_cb-= s->uvlinesize * (block_h - 1);
1964             dest_cr-= s->uvlinesize * (block_h - 1);
1965         }
1966         if(FRAME_MBAFF) {
1967             int list;
1968             for(list=0; list<h->list_count; list++){
1969                 if(!USES_LIST(mb_type, list))
1970                     continue;
1971                 if(IS_16X16(mb_type)){
1972                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1973                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1974                 }else{
1975                     for(i=0; i<16; i+=4){
1976                         int ref = h->ref_cache[list][scan8[i]];
1977                         if(ref >= 0)
1978                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1979                     }
1980                 }
1981             }
1982         }
1983     } else {
1984         linesize   = h->mb_linesize   = s->linesize;
1985         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1986 //        dct_offset = s->linesize * 16;
1987     }
1988
1989     if (!simple && IS_INTRA_PCM(mb_type)) {
1990         const int bit_depth = h->sps.bit_depth_luma;
1991         if (pixel_shift) {
1992             int j;
1993             GetBitContext gb;
1994             init_get_bits(&gb, (uint8_t*)h->mb, 384*bit_depth);
1995
1996             for (i = 0; i < 16; i++) {
1997                 uint16_t *tmp_y  = (uint16_t*)(dest_y  + i*linesize);
1998                 for (j = 0; j < 16; j++)
1999                     tmp_y[j] = get_bits(&gb, bit_depth);
2000             }
2001             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2002                 if (!h->sps.chroma_format_idc) {
2003                     for (i = 0; i < block_h; i++) {
2004                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
2005                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
2006                         for (j = 0; j < 8; j++) {
2007                             tmp_cb[j] = tmp_cr[j] = 1 << (bit_depth - 1);
2008                         }
2009                     }
2010                 } else {
2011                     for (i = 0; i < block_h; i++) {
2012                         uint16_t *tmp_cb = (uint16_t*)(dest_cb + i*uvlinesize);
2013                         for (j = 0; j < 8; j++)
2014                             tmp_cb[j] = get_bits(&gb, bit_depth);
2015                     }
2016                     for (i = 0; i < block_h; i++) {
2017                         uint16_t *tmp_cr = (uint16_t*)(dest_cr + i*uvlinesize);
2018                         for (j = 0; j < 8; j++)
2019                             tmp_cr[j] = get_bits(&gb, bit_depth);
2020                     }
2021                 }
2022             }
2023         } else {
2024             for (i=0; i<16; i++) {
2025                 memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
2026             }
2027             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2028                 if (!h->sps.chroma_format_idc) {
2029                     for (i=0; i<8; i++) {
2030                         memset(dest_cb + i*uvlinesize, 1 << (bit_depth - 1), 8);
2031                         memset(dest_cr + i*uvlinesize, 1 << (bit_depth - 1), 8);
2032                     }
2033                 } else {
2034                     for (i=0; i<block_h; i++) {
2035                         memcpy(dest_cb + i*uvlinesize, h->mb + 128 + i*4,  8);
2036                         memcpy(dest_cr + i*uvlinesize, h->mb + 160 + i*4,  8);
2037                     }
2038                 }
2039             }
2040         }
2041     } else {
2042         if(IS_INTRA(mb_type)){
2043             if(h->deblocking_filter)
2044                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, 0, simple, pixel_shift);
2045
2046             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2047                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2048                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2049             }
2050
2051             hl_decode_mb_predict_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2052
2053             if(h->deblocking_filter)
2054                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, 0, simple, pixel_shift);
2055         }else if(is_h264){
2056             if (chroma422) {
2057                 hl_motion_422(h, dest_y, dest_cb, dest_cr,
2058                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2059                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2060                               h->h264dsp.weight_h264_pixels_tab,
2061                               h->h264dsp.biweight_h264_pixels_tab,
2062                               pixel_shift);
2063             } else {
2064                 hl_motion_420(h, dest_y, dest_cb, dest_cr,
2065                               s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2066                               s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2067                               h->h264dsp.weight_h264_pixels_tab,
2068                               h->h264dsp.biweight_h264_pixels_tab,
2069                               pixel_shift);
2070             }
2071         }
2072
2073         hl_decode_mb_idct_luma(h, mb_type, is_h264, simple, transform_bypass, pixel_shift, block_offset, linesize, dest_y, 0);
2074
2075         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
2076             uint8_t *dest[2] = {dest_cb, dest_cr};
2077             if(transform_bypass){
2078                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
2079                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + (16*16*1 << pixel_shift), uvlinesize);
2080                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 32, h->mb + (16*16*2 << pixel_shift), uvlinesize);
2081                 }else{
2082                     idct_add = s->dsp.add_pixels4;
2083                     for(j=1; j<3; j++){
2084                         for(i=j*16; i<j*16+4; i++){
2085                             if(h->non_zero_count_cache[ scan8[i] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2086                                 idct_add   (dest[j-1] + block_offset[i], h->mb + (i*16 << pixel_shift), uvlinesize);
2087                         }
2088                         if (chroma422) {
2089                             for(i=j*16+4; i<j*16+8; i++){
2090                                 if(h->non_zero_count_cache[ scan8[i+4] ] || dctcoef_get(h->mb, pixel_shift, i*16))
2091                                     idct_add   (dest[j-1] + block_offset[i+4], h->mb + (i*16 << pixel_shift), uvlinesize);
2092                             }
2093                         }
2094                     }
2095                 }
2096             }else{
2097                 if(is_h264){
2098                     int qp[2];
2099                     if (chroma422) {
2100                         qp[0] = h->chroma_qp[0] + 3;
2101                         qp[1] = h->chroma_qp[1] + 3;
2102                     } else {
2103                         qp[0] = h->chroma_qp[0];
2104                         qp[1] = h->chroma_qp[1];
2105                     }
2106                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+0] ])
2107                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*1 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][qp[0]][0]);
2108                     if(h->non_zero_count_cache[ scan8[CHROMA_DC_BLOCK_INDEX+1] ])
2109                         h->h264dsp.h264_chroma_dc_dequant_idct(h->mb + (16*16*2 << pixel_shift), h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][qp[1]][0]);
2110                     h->h264dsp.h264_idct_add8(dest, block_offset,
2111                                               h->mb, uvlinesize,
2112                                               h->non_zero_count_cache);
2113                 }
2114 #if CONFIG_SVQ3_DECODER
2115                 else{
2116                     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]);
2117                     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]);
2118                     for(j=1; j<3; j++){
2119                         for(i=j*16; i<j*16+4; i++){
2120                             if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2121                                 uint8_t * const ptr= dest[j-1] + block_offset[i];
2122                                 ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[0][s->qscale + 12] - 12, 2);
2123                             }
2124                         }
2125                     }
2126                 }
2127 #endif
2128             }
2129         }
2130     }
2131     if(h->cbp || IS_INTRA(mb_type))
2132     {
2133         s->dsp.clear_blocks(h->mb);
2134         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2135     }
2136 }
2137
2138 static av_always_inline void hl_decode_mb_444_internal(H264Context *h, int simple, int pixel_shift){
2139     MpegEncContext * const s = &h->s;
2140     const int mb_x= s->mb_x;
2141     const int mb_y= s->mb_y;
2142     const int mb_xy= h->mb_xy;
2143     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2144     uint8_t  *dest[3];
2145     int linesize;
2146     int i, j, p;
2147     int *block_offset = &h->block_offset[0];
2148     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
2149     const int plane_count = (simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) ? 3 : 1;
2150
2151     for (p = 0; p < plane_count; p++)
2152     {
2153         dest[p] = s->current_picture.f.data[p] + ((mb_x << pixel_shift) + mb_y * s->linesize) * 16;
2154         s->dsp.prefetch(dest[p] + (s->mb_x&3)*4*s->linesize + (64 << pixel_shift), s->linesize, 4);
2155     }
2156
2157     h->list_counts[mb_xy]= h->list_count;
2158
2159     if (!simple && MB_FIELD) {
2160         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize * 2;
2161         block_offset = &h->block_offset[48];
2162         if(mb_y&1) //FIXME move out of this function?
2163             for (p = 0; p < 3; p++)
2164                 dest[p] -= s->linesize*15;
2165         if(FRAME_MBAFF) {
2166             int list;
2167             for(list=0; list<h->list_count; list++){
2168                 if(!USES_LIST(mb_type, list))
2169                     continue;
2170                 if(IS_16X16(mb_type)){
2171                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2172                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2173                 }else{
2174                     for(i=0; i<16; i+=4){
2175                         int ref = h->ref_cache[list][scan8[i]];
2176                         if(ref >= 0)
2177                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2178                     }
2179                 }
2180             }
2181         }
2182     } else {
2183         linesize   = h->mb_linesize = h->mb_uvlinesize = s->linesize;
2184     }
2185
2186     if (!simple && IS_INTRA_PCM(mb_type)) {
2187         if (pixel_shift) {
2188             const int bit_depth = h->sps.bit_depth_luma;
2189             GetBitContext gb;
2190             init_get_bits(&gb, (uint8_t*)h->mb, 768*bit_depth);
2191
2192             for (p = 0; p < plane_count; p++) {
2193                 for (i = 0; i < 16; i++) {
2194                     uint16_t *tmp = (uint16_t*)(dest[p] + i*linesize);
2195                     for (j = 0; j < 16; j++)
2196                         tmp[j] = get_bits(&gb, bit_depth);
2197                 }
2198             }
2199         } else {
2200             for (p = 0; p < plane_count; p++) {
2201                 for (i = 0; i < 16; i++) {
2202                     memcpy(dest[p] + i*linesize, h->mb + p*128 + i*8, 16);
2203                 }
2204             }
2205         }
2206     } else {
2207         if(IS_INTRA(mb_type)){
2208             if(h->deblocking_filter)
2209                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 1, 1, simple, pixel_shift);
2210
2211             for (p = 0; p < plane_count; p++)
2212                 hl_decode_mb_predict_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2213
2214             if(h->deblocking_filter)
2215                 xchg_mb_border(h, dest[0], dest[1], dest[2], linesize, linesize, 0, 1, simple, pixel_shift);
2216         }else{
2217             hl_motion(h, dest[0], dest[1], dest[2],
2218                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2219                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2220                       h->h264dsp.weight_h264_pixels_tab,
2221                       h->h264dsp.biweight_h264_pixels_tab, pixel_shift, 3);
2222         }
2223
2224         for (p = 0; p < plane_count; p++)
2225             hl_decode_mb_idct_luma(h, mb_type, 1, simple, transform_bypass, pixel_shift, block_offset, linesize, dest[p], p);
2226     }
2227     if(h->cbp || IS_INTRA(mb_type))
2228     {
2229         s->dsp.clear_blocks(h->mb);
2230         s->dsp.clear_blocks(h->mb+(24*16<<pixel_shift));
2231     }
2232 }
2233
2234 /**
2235  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2236  */
2237 #define hl_decode_mb_simple(sh, bits) \
2238 static void hl_decode_mb_simple_ ## bits(H264Context *h){ \
2239     hl_decode_mb_internal(h, 1, sh); \
2240 }
2241 hl_decode_mb_simple(0, 8);
2242 hl_decode_mb_simple(1, 16);
2243
2244 /**
2245  * Process a macroblock; this handles edge cases, such as interlacing.
2246  */
2247 static void av_noinline hl_decode_mb_complex(H264Context *h){
2248     hl_decode_mb_internal(h, 0, h->pixel_shift);
2249 }
2250
2251 static void av_noinline hl_decode_mb_444_complex(H264Context *h){
2252     hl_decode_mb_444_internal(h, 0, h->pixel_shift);
2253 }
2254
2255 static void av_noinline hl_decode_mb_444_simple(H264Context *h){
2256     hl_decode_mb_444_internal(h, 1, 0);
2257 }
2258
2259 void ff_h264_hl_decode_mb(H264Context *h){
2260     MpegEncContext * const s = &h->s;
2261     const int mb_xy= h->mb_xy;
2262     const int mb_type = s->current_picture.f.mb_type[mb_xy];
2263     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
2264
2265     if (CHROMA444) {
2266         if(is_complex || h->pixel_shift)
2267             hl_decode_mb_444_complex(h);
2268         else
2269             hl_decode_mb_444_simple(h);
2270     } else if (is_complex) {
2271         hl_decode_mb_complex(h);
2272     } else if (h->pixel_shift) {
2273         hl_decode_mb_simple_16(h);
2274     } else
2275         hl_decode_mb_simple_8(h);
2276 }
2277
2278 static int pred_weight_table(H264Context *h){
2279     MpegEncContext * const s = &h->s;
2280     int list, i;
2281     int luma_def, chroma_def;
2282
2283     h->use_weight= 0;
2284     h->use_weight_chroma= 0;
2285     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
2286     if(h->sps.chroma_format_idc)
2287         h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
2288     luma_def = 1<<h->luma_log2_weight_denom;
2289     chroma_def = 1<<h->chroma_log2_weight_denom;
2290
2291     for(list=0; list<2; list++){
2292         h->luma_weight_flag[list]   = 0;
2293         h->chroma_weight_flag[list] = 0;
2294         for(i=0; i<h->ref_count[list]; i++){
2295             int luma_weight_flag, chroma_weight_flag;
2296
2297             luma_weight_flag= get_bits1(&s->gb);
2298             if(luma_weight_flag){
2299                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
2300                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
2301                 if(   h->luma_weight[i][list][0] != luma_def
2302                    || h->luma_weight[i][list][1] != 0) {
2303                     h->use_weight= 1;
2304                     h->luma_weight_flag[list]= 1;
2305                 }
2306             }else{
2307                 h->luma_weight[i][list][0]= luma_def;
2308                 h->luma_weight[i][list][1]= 0;
2309             }
2310
2311             if(h->sps.chroma_format_idc){
2312                 chroma_weight_flag= get_bits1(&s->gb);
2313                 if(chroma_weight_flag){
2314                     int j;
2315                     for(j=0; j<2; j++){
2316                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
2317                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
2318                         if(   h->chroma_weight[i][list][j][0] != chroma_def
2319                            || h->chroma_weight[i][list][j][1] != 0) {
2320                             h->use_weight_chroma= 1;
2321                             h->chroma_weight_flag[list]= 1;
2322                         }
2323                     }
2324                 }else{
2325                     int j;
2326                     for(j=0; j<2; j++){
2327                         h->chroma_weight[i][list][j][0]= chroma_def;
2328                         h->chroma_weight[i][list][j][1]= 0;
2329                     }
2330                 }
2331             }
2332         }
2333         if(h->slice_type_nos != AV_PICTURE_TYPE_B) break;
2334     }
2335     h->use_weight= h->use_weight || h->use_weight_chroma;
2336     return 0;
2337 }
2338
2339 /**
2340  * Initialize implicit_weight table.
2341  * @param field  0/1 initialize the weight for interlaced MBAFF
2342  *                -1 initializes the rest
2343  */
2344 static void implicit_weight_table(H264Context *h, int field){
2345     MpegEncContext * const s = &h->s;
2346     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
2347
2348     for (i = 0; i < 2; i++) {
2349         h->luma_weight_flag[i]   = 0;
2350         h->chroma_weight_flag[i] = 0;
2351     }
2352
2353     if(field < 0){
2354         if (s->picture_structure == PICT_FRAME) {
2355             cur_poc = s->current_picture_ptr->poc;
2356         } else {
2357             cur_poc = s->current_picture_ptr->field_poc[s->picture_structure - 1];
2358         }
2359     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
2360        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
2361         h->use_weight= 0;
2362         h->use_weight_chroma= 0;
2363         return;
2364     }
2365         ref_start= 0;
2366         ref_count0= h->ref_count[0];
2367         ref_count1= h->ref_count[1];
2368     }else{
2369         cur_poc = s->current_picture_ptr->field_poc[field];
2370         ref_start= 16;
2371         ref_count0= 16+2*h->ref_count[0];
2372         ref_count1= 16+2*h->ref_count[1];
2373     }
2374
2375     h->use_weight= 2;
2376     h->use_weight_chroma= 2;
2377     h->luma_log2_weight_denom= 5;
2378     h->chroma_log2_weight_denom= 5;
2379
2380     for(ref0=ref_start; ref0 < ref_count0; ref0++){
2381         int poc0 = h->ref_list[0][ref0].poc;
2382         for(ref1=ref_start; ref1 < ref_count1; ref1++){
2383             int w = 32;
2384             if (!h->ref_list[0][ref0].long_ref && !h->ref_list[1][ref1].long_ref) {
2385                 int poc1 = h->ref_list[1][ref1].poc;
2386                 int td = av_clip(poc1 - poc0, -128, 127);
2387                 if(td){
2388                     int tb = av_clip(cur_poc - poc0, -128, 127);
2389                     int tx = (16384 + (FFABS(td) >> 1)) / td;
2390                     int dist_scale_factor = (tb*tx + 32) >> 8;
2391                     if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
2392                         w = 64 - dist_scale_factor;
2393                 }
2394             }
2395             if(field<0){
2396                 h->implicit_weight[ref0][ref1][0]=
2397                 h->implicit_weight[ref0][ref1][1]= w;
2398             }else{
2399                 h->implicit_weight[ref0][ref1][field]=w;
2400             }
2401         }
2402     }
2403 }
2404
2405 /**
2406  * instantaneous decoder refresh.
2407  */
2408 static void idr(H264Context *h){
2409     int i;
2410     ff_h264_remove_all_refs(h);
2411     h->prev_frame_num= -1;
2412     h->prev_frame_num_offset= 0;
2413     h->prev_poc_msb=
2414     h->prev_poc_lsb= 0;
2415     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2416         h->last_pocs[i] = INT_MIN;
2417 }
2418
2419 /* forget old pics after a seek */
2420 static void flush_dpb(AVCodecContext *avctx){
2421     H264Context *h= avctx->priv_data;
2422     int i;
2423     for(i=0; i<=MAX_DELAYED_PIC_COUNT; i++) {
2424         if(h->delayed_pic[i])
2425             h->delayed_pic[i]->f.reference = 0;
2426         h->delayed_pic[i]= NULL;
2427     }
2428     for (i = 0; i < MAX_DELAYED_PIC_COUNT; i++)
2429         h->last_pocs[i] = INT_MIN;
2430     h->outputed_poc=h->next_outputed_poc= INT_MIN;
2431     h->prev_interlaced_frame = 1;
2432     idr(h);
2433     if(h->s.current_picture_ptr)
2434         h->s.current_picture_ptr->f.reference = 0;
2435     h->s.first_field= 0;
2436     ff_h264_reset_sei(h);
2437     ff_mpeg_flush(avctx);
2438     h->recovery_frame= -1;
2439     h->sync= 0;
2440 }
2441
2442 static int init_poc(H264Context *h){
2443     MpegEncContext * const s = &h->s;
2444     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
2445     int field_poc[2];
2446     Picture *cur = s->current_picture_ptr;
2447
2448     h->frame_num_offset= h->prev_frame_num_offset;
2449     if(h->frame_num < h->prev_frame_num)
2450         h->frame_num_offset += max_frame_num;
2451
2452     if(h->sps.poc_type==0){
2453         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
2454
2455         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
2456             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
2457         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
2458             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
2459         else
2460             h->poc_msb = h->prev_poc_msb;
2461 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
2462         field_poc[0] =
2463         field_poc[1] = h->poc_msb + h->poc_lsb;
2464         if(s->picture_structure == PICT_FRAME)
2465             field_poc[1] += h->delta_poc_bottom;
2466     }else if(h->sps.poc_type==1){
2467         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
2468         int i;
2469
2470         if(h->sps.poc_cycle_length != 0)
2471             abs_frame_num = h->frame_num_offset + h->frame_num;
2472         else
2473             abs_frame_num = 0;
2474
2475         if(h->nal_ref_idc==0 && abs_frame_num > 0)
2476             abs_frame_num--;
2477
2478         expected_delta_per_poc_cycle = 0;
2479         for(i=0; i < h->sps.poc_cycle_length; i++)
2480             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
2481
2482         if(abs_frame_num > 0){
2483             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
2484             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
2485
2486             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
2487             for(i = 0; i <= frame_num_in_poc_cycle; i++)
2488                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
2489         } else
2490             expectedpoc = 0;
2491
2492         if(h->nal_ref_idc == 0)
2493             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
2494
2495         field_poc[0] = expectedpoc + h->delta_poc[0];
2496         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
2497
2498         if(s->picture_structure == PICT_FRAME)
2499             field_poc[1] += h->delta_poc[1];
2500     }else{
2501         int poc= 2*(h->frame_num_offset + h->frame_num);
2502
2503         if(!h->nal_ref_idc)
2504             poc--;
2505
2506         field_poc[0]= poc;
2507         field_poc[1]= poc;
2508     }
2509
2510     if(s->picture_structure != PICT_BOTTOM_FIELD)
2511         s->current_picture_ptr->field_poc[0]= field_poc[0];
2512     if(s->picture_structure != PICT_TOP_FIELD)
2513         s->current_picture_ptr->field_poc[1]= field_poc[1];
2514     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
2515
2516     return 0;
2517 }
2518
2519
2520 /**
2521  * initialize scan tables
2522  */
2523 static void init_scan_tables(H264Context *h){
2524     int i;
2525     for(i=0; i<16; i++){
2526 #define T(x) (x>>2) | ((x<<2) & 0xF)
2527         h->zigzag_scan[i] = T(zigzag_scan[i]);
2528         h-> field_scan[i] = T( field_scan[i]);
2529 #undef T
2530     }
2531     for(i=0; i<64; i++){
2532 #define T(x) (x>>3) | ((x&7)<<3)
2533         h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
2534         h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
2535         h->field_scan8x8[i]        = T(field_scan8x8[i]);
2536         h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
2537 #undef T
2538     }
2539     if(h->sps.transform_bypass){ //FIXME same ugly
2540         h->zigzag_scan_q0          = zigzag_scan;
2541         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
2542         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
2543         h->field_scan_q0           = field_scan;
2544         h->field_scan8x8_q0        = field_scan8x8;
2545         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
2546     }else{
2547         h->zigzag_scan_q0          = h->zigzag_scan;
2548         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
2549         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
2550         h->field_scan_q0           = h->field_scan;
2551         h->field_scan8x8_q0        = h->field_scan8x8;
2552         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
2553     }
2554 }
2555
2556 static int field_end(H264Context *h, int in_setup){
2557     MpegEncContext * const s = &h->s;
2558     AVCodecContext * const avctx= s->avctx;
2559     int err = 0;
2560     s->mb_y= 0;
2561
2562     if (!in_setup && !s->dropable)
2563         ff_thread_report_progress((AVFrame*)s->current_picture_ptr, (16*s->mb_height >> FIELD_PICTURE) - 1,
2564                                  s->picture_structure==PICT_BOTTOM_FIELD);
2565
2566     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2567         ff_vdpau_h264_set_reference_frames(s);
2568
2569     if(in_setup || !(avctx->active_thread_type&FF_THREAD_FRAME)){
2570         if(!s->dropable) {
2571             err = ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
2572             h->prev_poc_msb= h->poc_msb;
2573             h->prev_poc_lsb= h->poc_lsb;
2574         }
2575         h->prev_frame_num_offset= h->frame_num_offset;
2576         h->prev_frame_num= h->frame_num;
2577         h->outputed_poc = h->next_outputed_poc;
2578     }
2579
2580     if (avctx->hwaccel) {
2581         if (avctx->hwaccel->end_frame(avctx) < 0)
2582             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
2583     }
2584
2585     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2586         ff_vdpau_h264_picture_complete(s);
2587
2588     /*
2589      * FIXME: Error handling code does not seem to support interlaced
2590      * when slices span multiple rows
2591      * The ff_er_add_slice calls don't work right for bottom
2592      * fields; they cause massive erroneous error concealing
2593      * Error marking covers both fields (top and bottom).
2594      * This causes a mismatched s->error_count
2595      * and a bad error table. Further, the error count goes to
2596      * INT_MAX when called for bottom field, because mb_y is
2597      * past end by one (callers fault) and resync_mb_y != 0
2598      * causes problems for the first MB line, too.
2599      */
2600     if (!FIELD_PICTURE)
2601         ff_er_frame_end(s);
2602
2603     MPV_frame_end(s);
2604
2605     h->current_slice=0;
2606
2607     return err;
2608 }
2609
2610 /**
2611  * Replicate H264 "master" context to thread contexts.
2612  */
2613 static void clone_slice(H264Context *dst, H264Context *src)
2614 {
2615     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
2616     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
2617     dst->s.current_picture      = src->s.current_picture;
2618     dst->s.linesize             = src->s.linesize;
2619     dst->s.uvlinesize           = src->s.uvlinesize;
2620     dst->s.first_field          = src->s.first_field;
2621
2622     dst->prev_poc_msb           = src->prev_poc_msb;
2623     dst->prev_poc_lsb           = src->prev_poc_lsb;
2624     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
2625     dst->prev_frame_num         = src->prev_frame_num;
2626     dst->short_ref_count        = src->short_ref_count;
2627
2628     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
2629     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
2630     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
2631     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
2632
2633     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
2634     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
2635 }
2636
2637 /**
2638  * computes profile from profile_idc and constraint_set?_flags
2639  *
2640  * @param sps SPS
2641  *
2642  * @return profile as defined by FF_PROFILE_H264_*
2643  */
2644 int ff_h264_get_profile(SPS *sps)
2645 {
2646     int profile = sps->profile_idc;
2647
2648     switch(sps->profile_idc) {
2649     case FF_PROFILE_H264_BASELINE:
2650         // constraint_set1_flag set to 1
2651         profile |= (sps->constraint_set_flags & 1<<1) ? FF_PROFILE_H264_CONSTRAINED : 0;
2652         break;
2653     case FF_PROFILE_H264_HIGH_10:
2654     case FF_PROFILE_H264_HIGH_422:
2655     case FF_PROFILE_H264_HIGH_444_PREDICTIVE:
2656         // constraint_set3_flag set to 1
2657         profile |= (sps->constraint_set_flags & 1<<3) ? FF_PROFILE_H264_INTRA : 0;
2658         break;
2659     }
2660
2661     return profile;
2662 }
2663
2664 /**
2665  * decodes a slice header.
2666  * This will also call MPV_common_init() and frame_start() as needed.
2667  *
2668  * @param h h264context
2669  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
2670  *
2671  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
2672  */
2673 static int decode_slice_header(H264Context *h, H264Context *h0){
2674     MpegEncContext * const s = &h->s;
2675     MpegEncContext * const s0 = &h0->s;
2676     unsigned int first_mb_in_slice;
2677     unsigned int pps_id;
2678     int num_ref_idx_active_override_flag;
2679     unsigned int slice_type, tmp, i, j;
2680     int default_ref_list_done = 0;
2681     int last_pic_structure;
2682
2683     s->dropable= h->nal_ref_idc == 0;
2684
2685     /* FIXME: 2tap qpel isn't implemented for high bit depth. */
2686     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc && !h->pixel_shift){
2687         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
2688         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
2689     }else{
2690         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
2691         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
2692     }
2693
2694     first_mb_in_slice= get_ue_golomb_long(&s->gb);
2695
2696     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
2697         if(h0->current_slice && FIELD_PICTURE){
2698             field_end(h, 1);
2699         }
2700
2701         h0->current_slice = 0;
2702         if (!s0->first_field)
2703             s->current_picture_ptr= NULL;
2704     }
2705
2706     slice_type= get_ue_golomb_31(&s->gb);
2707     if(slice_type > 9){
2708         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);
2709         return -1;
2710     }
2711     if(slice_type > 4){
2712         slice_type -= 5;
2713         h->slice_type_fixed=1;
2714     }else
2715         h->slice_type_fixed=0;
2716
2717     slice_type= golomb_to_pict_type[ slice_type ];
2718     if (slice_type == AV_PICTURE_TYPE_I
2719         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
2720         default_ref_list_done = 1;
2721     }
2722     h->slice_type= slice_type;
2723     h->slice_type_nos= slice_type & 3;
2724
2725     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
2726
2727     pps_id= get_ue_golomb(&s->gb);
2728     if(pps_id>=MAX_PPS_COUNT){
2729         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
2730         return -1;
2731     }
2732     if(!h0->pps_buffers[pps_id]) {
2733         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
2734         return -1;
2735     }
2736     h->pps= *h0->pps_buffers[pps_id];
2737
2738     if(!h0->sps_buffers[h->pps.sps_id]) {
2739         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
2740         return -1;
2741     }
2742     h->sps = *h0->sps_buffers[h->pps.sps_id];
2743
2744     s->avctx->profile = ff_h264_get_profile(&h->sps);
2745     s->avctx->level   = h->sps.level_idc;
2746     s->avctx->refs    = h->sps.ref_frame_count;
2747
2748     if(h == h0 && h->dequant_coeff_pps != pps_id){
2749         h->dequant_coeff_pps = pps_id;
2750         init_dequant_tables(h);
2751     }
2752
2753     s->mb_width= h->sps.mb_width;
2754     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
2755
2756     h->b_stride=  s->mb_width*4;
2757
2758     s->chroma_y_shift = h->sps.chroma_format_idc <= 1; // 400 uses yuv420p
2759
2760     s->width = 16*s->mb_width;
2761     s->height= 16*s->mb_height;
2762
2763     if (s->context_initialized
2764         && (   s->width != s->avctx->coded_width || s->height != s->avctx->coded_height
2765             || s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma
2766             || h->cur_chroma_format_idc != h->sps.chroma_format_idc
2767             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
2768         if(h != h0) {
2769             av_log_missing_feature(s->avctx, "Width/height/bit depth/chroma idc changing with threads is", 0);
2770             return -1;   // width / height changed during parallelized decoding
2771         }
2772         free_tables(h, 0);
2773         flush_dpb(s->avctx);
2774         MPV_common_end(s);
2775         h->list_count = 0;
2776     }
2777     if (!s->context_initialized) {
2778         if (h != h0) {
2779             av_log(h->s.avctx, AV_LOG_ERROR, "Cannot (re-)initialize context during parallel decoding.\n");
2780             return -1;
2781         }
2782         avcodec_set_dimensions(s->avctx, s->width, s->height);
2783         s->avctx->width  -= (2>>CHROMA444)*FFMIN(h->sps.crop_right, (8<<CHROMA444)-1);
2784         s->avctx->height -= (1<<s->chroma_y_shift)*FFMIN(h->sps.crop_bottom, (16>>s->chroma_y_shift)-1) * (2 - h->sps.frame_mbs_only_flag);
2785         s->avctx->sample_aspect_ratio= h->sps.sar;
2786         av_assert0(s->avctx->sample_aspect_ratio.den);
2787
2788         if (s->avctx->bits_per_raw_sample != h->sps.bit_depth_luma ||
2789             h->cur_chroma_format_idc != h->sps.chroma_format_idc) {
2790             if (h->sps.bit_depth_luma >= 8 && h->sps.bit_depth_luma <= 10 &&
2791                 (h->sps.bit_depth_luma != 9 || !CHROMA422)) {
2792                 s->avctx->bits_per_raw_sample = h->sps.bit_depth_luma;
2793                 h->cur_chroma_format_idc = h->sps.chroma_format_idc;
2794                 h->pixel_shift = h->sps.bit_depth_luma > 8;
2795
2796                 ff_h264dsp_init(&h->h264dsp, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2797                 ff_h264_pred_init(&h->hpc, s->codec_id, h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2798                 s->dsp.dct_bits = h->sps.bit_depth_luma > 8 ? 32 : 16;
2799                 dsputil_init(&s->dsp, s->avctx);
2800             } else {
2801                 av_log(s->avctx, AV_LOG_ERROR, "Unsupported bit depth: %d chroma_idc: %d\n",
2802                        h->sps.bit_depth_luma, h->sps.chroma_format_idc);
2803                 return -1;
2804             }
2805         }
2806
2807         if(h->sps.video_signal_type_present_flag){
2808             s->avctx->color_range = h->sps.full_range>0 ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
2809             if(h->sps.colour_description_present_flag){
2810                 s->avctx->color_primaries = h->sps.color_primaries;
2811                 s->avctx->color_trc       = h->sps.color_trc;
2812                 s->avctx->colorspace      = h->sps.colorspace;
2813             }
2814         }
2815
2816         if(h->sps.timing_info_present_flag){
2817             int64_t den= h->sps.time_scale;
2818             if(h->x264_build < 44U)
2819                 den *= 2;
2820             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
2821                       h->sps.num_units_in_tick, den, 1<<30);
2822         }
2823
2824         switch (h->sps.bit_depth_luma) {
2825             case 9 :
2826                 if (CHROMA444) {
2827                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2828                         s->avctx->pix_fmt = PIX_FMT_GBRP9;
2829                     } else
2830                         s->avctx->pix_fmt = PIX_FMT_YUV444P9;
2831                 } else if (CHROMA422)
2832                     s->avctx->pix_fmt = PIX_FMT_YUV422P9;
2833                 else
2834                     s->avctx->pix_fmt = PIX_FMT_YUV420P9;
2835                 break;
2836             case 10 :
2837                 if (CHROMA444) {
2838                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2839                         s->avctx->pix_fmt = PIX_FMT_GBRP10;
2840                     } else
2841                         s->avctx->pix_fmt = PIX_FMT_YUV444P10;
2842                 } else if (CHROMA422)
2843                     s->avctx->pix_fmt = PIX_FMT_YUV422P10;
2844                 else
2845                     s->avctx->pix_fmt = PIX_FMT_YUV420P10;
2846                 break;
2847             default:
2848                 if (CHROMA444){
2849                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ444P : PIX_FMT_YUV444P;
2850                     if (s->avctx->colorspace == AVCOL_SPC_RGB) {
2851                        s->avctx->pix_fmt = PIX_FMT_GBR24P;
2852                        av_log(h->s.avctx, AV_LOG_DEBUG, "Detected GBR colorspace.\n");
2853                     } else if (s->avctx->colorspace == AVCOL_SPC_YCGCO) {
2854                         av_log(h->s.avctx, AV_LOG_WARNING, "Detected unsupported YCgCo colorspace.\n");
2855                     }
2856                 } else if (CHROMA422) {
2857                     s->avctx->pix_fmt = s->avctx->color_range == AVCOL_RANGE_JPEG ? PIX_FMT_YUVJ422P : PIX_FMT_YUV422P;
2858                 }else{
2859                     s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
2860                                                              s->avctx->codec->pix_fmts ?
2861                                                              s->avctx->codec->pix_fmts :
2862                                                              s->avctx->color_range == AVCOL_RANGE_JPEG ?
2863                                                              hwaccel_pixfmt_list_h264_jpeg_420 :
2864                                                              ff_hwaccel_pixfmt_list_420);
2865                 }
2866         }
2867
2868         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
2869
2870         if (MPV_common_init(s) < 0) {
2871             av_log(h->s.avctx, AV_LOG_ERROR, "MPV_common_init() failed.\n");
2872             return -1;
2873         }
2874         s->first_field = 0;
2875         h->prev_interlaced_frame = 1;
2876
2877         init_scan_tables(h);
2878         if (ff_h264_alloc_tables(h) < 0) {
2879             av_log(h->s.avctx, AV_LOG_ERROR, "Could not allocate memory for h264\n");
2880             return AVERROR(ENOMEM);
2881         }
2882
2883         if (!HAVE_THREADS || !(s->avctx->active_thread_type&FF_THREAD_SLICE)) {
2884             if (context_init(h) < 0) {
2885                 av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2886                 return -1;
2887             }
2888         } else {
2889             for(i = 1; i < s->avctx->thread_count; i++) {
2890                 H264Context *c;
2891                 c = h->thread_context[i] = av_malloc(sizeof(H264Context));
2892                 memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
2893                 memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
2894                 c->h264dsp = h->h264dsp;
2895                 c->sps = h->sps;
2896                 c->pps = h->pps;
2897                 c->pixel_shift = h->pixel_shift;
2898                 c->cur_chroma_format_idc = h->cur_chroma_format_idc;
2899                 init_scan_tables(c);
2900                 clone_tables(c, h, i);
2901             }
2902
2903             for(i = 0; i < s->avctx->thread_count; i++)
2904                 if (context_init(h->thread_context[i]) < 0) {
2905                     av_log(h->s.avctx, AV_LOG_ERROR, "context_init() failed.\n");
2906                     return -1;
2907                 }
2908         }
2909     }
2910
2911     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
2912
2913     h->mb_mbaff = 0;
2914     h->mb_aff_frame = 0;
2915     last_pic_structure = s0->picture_structure;
2916     if(h->sps.frame_mbs_only_flag){
2917         s->picture_structure= PICT_FRAME;
2918     }else{
2919         if(!h->sps.direct_8x8_inference_flag && slice_type == AV_PICTURE_TYPE_B){
2920             av_log(h->s.avctx, AV_LOG_ERROR, "This stream was generated by a broken encoder, invalid 8x8 inference\n");
2921             return -1;
2922         }
2923         if(get_bits1(&s->gb)) { //field_pic_flag
2924             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
2925         } else {
2926             s->picture_structure= PICT_FRAME;
2927             h->mb_aff_frame = h->sps.mb_aff;
2928         }
2929     }
2930     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
2931
2932     if(h0->current_slice == 0){
2933         // Shorten frame num gaps so we don't have to allocate reference frames just to throw them away
2934         if(h->frame_num != h->prev_frame_num && h->prev_frame_num >= 0) {
2935             int unwrap_prev_frame_num = h->prev_frame_num, max_frame_num = 1<<h->sps.log2_max_frame_num;
2936
2937             if (unwrap_prev_frame_num > h->frame_num) unwrap_prev_frame_num -= max_frame_num;
2938
2939             if ((h->frame_num - unwrap_prev_frame_num) > h->sps.ref_frame_count) {
2940                 unwrap_prev_frame_num = (h->frame_num - h->sps.ref_frame_count) - 1;
2941                 if (unwrap_prev_frame_num < 0)
2942                     unwrap_prev_frame_num += max_frame_num;
2943
2944                 h->prev_frame_num = unwrap_prev_frame_num;
2945             }
2946         }
2947
2948         while(h->frame_num !=  h->prev_frame_num && h->prev_frame_num >= 0 &&
2949               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
2950             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
2951             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
2952             if (ff_h264_frame_start(h) < 0)
2953                 return -1;
2954             h->prev_frame_num++;
2955             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
2956             s->current_picture_ptr->frame_num= h->prev_frame_num;
2957             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 0);
2958             ff_thread_report_progress((AVFrame*)s->current_picture_ptr, INT_MAX, 1);
2959             ff_generate_sliding_window_mmcos(h);
2960             if (ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index) < 0 &&
2961                 (s->avctx->err_recognition & AV_EF_EXPLODE))
2962                 return AVERROR_INVALIDDATA;
2963             /* Error concealment: if a ref is missing, copy the previous ref in its place.
2964              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
2965              * about there being no actual duplicates.
2966              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
2967              * concealing a lost frame, this probably isn't noticable by comparison, but it should
2968              * be fixed. */
2969             if (h->short_ref_count) {
2970                 if (prev) {
2971                     av_image_copy(h->short_ref[0]->f.data, h->short_ref[0]->f.linesize,
2972                                   (const uint8_t**)prev->f.data, prev->f.linesize,
2973                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
2974                     h->short_ref[0]->poc = prev->poc+2;
2975                 }
2976                 h->short_ref[0]->frame_num = h->prev_frame_num;
2977             }
2978         }
2979
2980         /* See if we have a decoded first field looking for a pair... */
2981         if (s0->first_field) {
2982             assert(s0->current_picture_ptr);
2983             assert(s0->current_picture_ptr->f.data[0]);
2984             assert(s0->current_picture_ptr->f.reference != DELAYED_PIC_REF);
2985
2986             /* figure out if we have a complementary field pair */
2987             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
2988                 /*
2989                  * Previous field is unmatched. Don't display it, but let it
2990                  * remain for reference if marked as such.
2991                  */
2992                 s0->current_picture_ptr = NULL;
2993                 s0->first_field = FIELD_PICTURE;
2994
2995             } else {
2996                 if (s0->current_picture_ptr->frame_num != h->frame_num) {
2997                     /*
2998                      * This and previous field had
2999                      * different frame_nums. Consider this field first in
3000                      * pair. Throw away previous field except for reference
3001                      * purposes.
3002                      */
3003                     s0->first_field = 1;
3004                     s0->current_picture_ptr = NULL;
3005
3006                 } else {
3007                     /* Second field in complementary pair */
3008                     s0->first_field = 0;
3009                 }
3010             }
3011
3012         } else {
3013             /* Frame or first field in a potentially complementary pair */
3014             assert(!s0->current_picture_ptr);
3015             s0->first_field = FIELD_PICTURE;
3016         }
3017
3018         if(!FIELD_PICTURE || s0->first_field) {
3019             if (ff_h264_frame_start(h) < 0) {
3020                 s0->first_field = 0;
3021                 return -1;
3022             }
3023         } else {
3024             ff_release_unused_pictures(s, 0);
3025         }
3026     }
3027     if(h != h0)
3028         clone_slice(h, h0);
3029
3030     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
3031
3032     assert(s->mb_num == s->mb_width * s->mb_height);
3033     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
3034        first_mb_in_slice                    >= s->mb_num){
3035         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
3036         return -1;
3037     }
3038     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
3039     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
3040     if (s->picture_structure == PICT_BOTTOM_FIELD)
3041         s->resync_mb_y = s->mb_y = s->mb_y + 1;
3042     assert(s->mb_y < s->mb_height);
3043
3044     if(s->picture_structure==PICT_FRAME){
3045         h->curr_pic_num=   h->frame_num;
3046         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
3047     }else{
3048         h->curr_pic_num= 2*h->frame_num + 1;
3049         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
3050     }
3051
3052     if(h->nal_unit_type == NAL_IDR_SLICE){
3053         get_ue_golomb(&s->gb); /* idr_pic_id */
3054     }
3055
3056     if(h->sps.poc_type==0){
3057         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
3058
3059         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
3060             h->delta_poc_bottom= get_se_golomb(&s->gb);
3061         }
3062     }
3063
3064     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
3065         h->delta_poc[0]= get_se_golomb(&s->gb);
3066
3067         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
3068             h->delta_poc[1]= get_se_golomb(&s->gb);
3069     }
3070
3071     init_poc(h);
3072
3073     if(h->pps.redundant_pic_cnt_present){
3074         h->redundant_pic_count= get_ue_golomb(&s->gb);
3075     }
3076
3077     //set defaults, might be overridden a few lines later
3078     h->ref_count[0]= h->pps.ref_count[0];
3079     h->ref_count[1]= h->pps.ref_count[1];
3080
3081     if(h->slice_type_nos != AV_PICTURE_TYPE_I){
3082         unsigned max= (16<<(s->picture_structure != PICT_FRAME))-1;
3083         if(h->slice_type_nos == AV_PICTURE_TYPE_B){
3084             h->direct_spatial_mv_pred= get_bits1(&s->gb);
3085         }
3086         num_ref_idx_active_override_flag= get_bits1(&s->gb);
3087
3088         if(num_ref_idx_active_override_flag){
3089             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
3090             if(h->slice_type_nos==AV_PICTURE_TYPE_B)
3091                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
3092
3093         }
3094         if(h->ref_count[0]-1 > max || h->ref_count[1]-1 > max){
3095             av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
3096             h->ref_count[0]= h->ref_count[1]= 1;
3097             return -1;
3098         }
3099         if(h->slice_type_nos == AV_PICTURE_TYPE_B)
3100             h->list_count= 2;
3101         else
3102             h->list_count= 1;
3103     }else
3104         h->ref_count[1]= h->ref_count[0]= h->list_count= 0;
3105
3106     if(!default_ref_list_done){
3107         ff_h264_fill_default_ref_list(h);
3108     }
3109
3110     if(h->slice_type_nos!=AV_PICTURE_TYPE_I && ff_h264_decode_ref_pic_list_reordering(h) < 0) {
3111         h->ref_count[1]= h->ref_count[0]= 0;
3112         return -1;
3113     }
3114
3115     if(h->slice_type_nos!=AV_PICTURE_TYPE_I){
3116         s->last_picture_ptr= &h->ref_list[0][0];
3117         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
3118     }
3119     if(h->slice_type_nos==AV_PICTURE_TYPE_B){
3120         s->next_picture_ptr= &h->ref_list[1][0];
3121         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
3122     }
3123
3124     if(   (h->pps.weighted_pred          && h->slice_type_nos == AV_PICTURE_TYPE_P )
3125        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== AV_PICTURE_TYPE_B ) )
3126         pred_weight_table(h);
3127     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3128         implicit_weight_table(h, -1);
3129     }else {
3130         h->use_weight = 0;
3131         for (i = 0; i < 2; i++) {
3132             h->luma_weight_flag[i]   = 0;
3133             h->chroma_weight_flag[i] = 0;
3134         }
3135     }
3136
3137     if(h->nal_ref_idc && ff_h264_decode_ref_pic_marking(h0, &s->gb) < 0 &&
3138        (s->avctx->err_recognition & AV_EF_EXPLODE))
3139         return AVERROR_INVALIDDATA;
3140
3141     if(FRAME_MBAFF){
3142         ff_h264_fill_mbaff_ref_list(h);
3143
3144         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== AV_PICTURE_TYPE_B){
3145             implicit_weight_table(h, 0);
3146             implicit_weight_table(h, 1);
3147         }
3148     }
3149
3150     if(h->slice_type_nos==AV_PICTURE_TYPE_B && !h->direct_spatial_mv_pred)
3151         ff_h264_direct_dist_scale_factor(h);
3152     ff_h264_direct_ref_list_init(h);
3153
3154     if( h->slice_type_nos != AV_PICTURE_TYPE_I && h->pps.cabac ){
3155         tmp = get_ue_golomb_31(&s->gb);
3156         if(tmp > 2){
3157             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
3158             return -1;
3159         }
3160         h->cabac_init_idc= tmp;
3161     }
3162
3163     h->last_qscale_diff = 0;
3164     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
3165     if(tmp>51+6*(h->sps.bit_depth_luma-8)){
3166         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
3167         return -1;
3168     }
3169     s->qscale= tmp;
3170     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3171     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3172     //FIXME qscale / qp ... stuff
3173     if(h->slice_type == AV_PICTURE_TYPE_SP){
3174         get_bits1(&s->gb); /* sp_for_switch_flag */
3175     }
3176     if(h->slice_type==AV_PICTURE_TYPE_SP || h->slice_type == AV_PICTURE_TYPE_SI){
3177         get_se_golomb(&s->gb); /* slice_qs_delta */
3178     }
3179
3180     h->deblocking_filter = 1;
3181     h->slice_alpha_c0_offset = 52;
3182     h->slice_beta_offset = 52;
3183     if( h->pps.deblocking_filter_parameters_present ) {
3184         tmp= get_ue_golomb_31(&s->gb);
3185         if(tmp > 2){
3186             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
3187             return -1;
3188         }
3189         h->deblocking_filter= tmp;
3190         if(h->deblocking_filter < 2)
3191             h->deblocking_filter^= 1; // 1<->0
3192
3193         if( h->deblocking_filter ) {
3194             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
3195             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
3196             if(   h->slice_alpha_c0_offset > 104U
3197                || h->slice_beta_offset     > 104U){
3198                 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);
3199                 return -1;
3200             }
3201         }
3202     }
3203
3204     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
3205        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != AV_PICTURE_TYPE_I)
3206        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == AV_PICTURE_TYPE_B)
3207        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
3208         h->deblocking_filter= 0;
3209
3210     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
3211         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
3212             /* Cheat slightly for speed:
3213                Do not bother to deblock across slices. */
3214             h->deblocking_filter = 2;
3215         } else {
3216             h0->max_contexts = 1;
3217             if(!h0->single_decode_warning) {
3218                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
3219                 h0->single_decode_warning = 1;
3220             }
3221             if (h != h0) {
3222                 av_log(h->s.avctx, AV_LOG_ERROR, "Deblocking switched inside frame.\n");
3223                 return 1;
3224             }
3225         }
3226     }
3227     h->qp_thresh = 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset)
3228                  - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1])
3229                  + 6 * (h->sps.bit_depth_luma - 8);
3230
3231 #if 0 //FMO
3232     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
3233         slice_group_change_cycle= get_bits(&s->gb, ?);
3234 #endif
3235
3236     h0->last_slice_type = slice_type;
3237     h->slice_num = ++h0->current_slice;
3238
3239     if(h->slice_num)
3240         h0->slice_row[(h->slice_num-1)&(MAX_SLICES-1)]= s->resync_mb_y;
3241     if (   h0->slice_row[h->slice_num&(MAX_SLICES-1)] + 3 >= s->resync_mb_y
3242         && h0->slice_row[h->slice_num&(MAX_SLICES-1)] <= s->resync_mb_y
3243         && h->slice_num >= MAX_SLICES) {
3244         //in case of ASO this check needs to be updated depending on how we decide to assign slice numbers in this case
3245         av_log(s->avctx, AV_LOG_WARNING, "Possibly too many slices (%d >= %d), increase MAX_SLICES and recompile if there are artifacts\n", h->slice_num, MAX_SLICES);
3246     }
3247
3248     for(j=0; j<2; j++){
3249         int id_list[16];
3250         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
3251         for(i=0; i<16; i++){
3252             id_list[i]= 60;
3253             if (h->ref_list[j][i].f.data[0]) {
3254                 int k;
3255                 uint8_t *base = h->ref_list[j][i].f.base[0];
3256                 for(k=0; k<h->short_ref_count; k++)
3257                     if (h->short_ref[k]->f.base[0] == base) {
3258                         id_list[i]= k;
3259                         break;
3260                     }
3261                 for(k=0; k<h->long_ref_count; k++)
3262                     if (h->long_ref[k] && h->long_ref[k]->f.base[0] == base) {
3263                         id_list[i]= h->short_ref_count + k;
3264                         break;
3265                     }
3266             }
3267         }
3268
3269         ref2frm[0]=
3270         ref2frm[1]= -1;
3271         for(i=0; i<16; i++)
3272             ref2frm[i+2]= 4*id_list[i]
3273                           + (h->ref_list[j][i].f.reference & 3);
3274         ref2frm[18+0]=
3275         ref2frm[18+1]= -1;
3276         for(i=16; i<48; i++)
3277             ref2frm[i+4]= 4*id_list[(i-16)>>1]
3278                           + (h->ref_list[j][i].f.reference & 3);
3279     }
3280
3281     //FIXME: fix draw_edges+PAFF+frame threads
3282     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE || (!h->sps.frame_mbs_only_flag && s->avctx->active_thread_type)) ? 0 : 16;
3283     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
3284
3285     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
3286         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",
3287                h->slice_num,
3288                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
3289                first_mb_in_slice,
3290                av_get_picture_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
3291                pps_id, h->frame_num,
3292                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
3293                h->ref_count[0], h->ref_count[1],
3294                s->qscale,
3295                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
3296                h->use_weight,
3297                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
3298                h->slice_type == AV_PICTURE_TYPE_B ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
3299                );
3300     }
3301
3302     return 0;
3303 }
3304
3305 int ff_h264_get_slice_type(const H264Context *h)
3306 {
3307     switch (h->slice_type) {
3308     case AV_PICTURE_TYPE_P:  return 0;
3309     case AV_PICTURE_TYPE_B:  return 1;
3310     case AV_PICTURE_TYPE_I:  return 2;
3311     case AV_PICTURE_TYPE_SP: return 3;
3312     case AV_PICTURE_TYPE_SI: return 4;
3313     default:         return -1;
3314     }
3315 }
3316
3317 static av_always_inline void fill_filter_caches_inter(H264Context *h, MpegEncContext * const s, int mb_type, int top_xy,
3318                                                       int left_xy[LEFT_MBS], int top_type, int left_type[LEFT_MBS], int mb_xy, int list)
3319 {
3320     int b_stride = h->b_stride;
3321     int16_t (*mv_dst)[2] = &h->mv_cache[list][scan8[0]];
3322     int8_t *ref_cache = &h->ref_cache[list][scan8[0]];
3323     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
3324         if(USES_LIST(top_type, list)){
3325             const int b_xy= h->mb2b_xy[top_xy] + 3*b_stride;
3326             const int b8_xy= 4*top_xy + 2;
3327             int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3328             AV_COPY128(mv_dst - 1*8, s->current_picture.f.motion_val[list][b_xy + 0]);
3329             ref_cache[0 - 1*8]=
3330             ref_cache[1 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 0]];
3331             ref_cache[2 - 1*8]=
3332             ref_cache[3 - 1*8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 1]];
3333         }else{
3334             AV_ZERO128(mv_dst - 1*8);
3335             AV_WN32A(&ref_cache[0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3336         }
3337
3338         if(!IS_INTERLACED(mb_type^left_type[LTOP])){
3339             if(USES_LIST(left_type[LTOP], list)){
3340                 const int b_xy= h->mb2b_xy[left_xy[LTOP]] + 3;
3341                 const int b8_xy= 4*left_xy[LTOP] + 1;
3342                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[LTOP]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3343                 AV_COPY32(mv_dst - 1 +  0, s->current_picture.f.motion_val[list][b_xy + b_stride*0]);
3344                 AV_COPY32(mv_dst - 1 +  8, s->current_picture.f.motion_val[list][b_xy + b_stride*1]);
3345                 AV_COPY32(mv_dst - 1 + 16, s->current_picture.f.motion_val[list][b_xy + b_stride*2]);
3346                 AV_COPY32(mv_dst - 1 + 24, s->current_picture.f.motion_val[list][b_xy + b_stride*3]);
3347                 ref_cache[-1 +  0]=
3348                 ref_cache[-1 +  8]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*0]];
3349                 ref_cache[-1 + 16]=
3350                 ref_cache[-1 + 24]= ref2frm[list][s->current_picture.f.ref_index[list][b8_xy + 2*1]];
3351             }else{
3352                 AV_ZERO32(mv_dst - 1 + 0);
3353                 AV_ZERO32(mv_dst - 1 + 8);
3354                 AV_ZERO32(mv_dst - 1 +16);
3355                 AV_ZERO32(mv_dst - 1 +24);
3356                 ref_cache[-1 +  0]=
3357                 ref_cache[-1 +  8]=
3358                 ref_cache[-1 + 16]=
3359                 ref_cache[-1 + 24]= LIST_NOT_USED;
3360             }
3361         }
3362     }
3363
3364     if(!USES_LIST(mb_type, list)){
3365         fill_rectangle(mv_dst, 4, 4, 8, pack16to32(0,0), 4);
3366         AV_WN32A(&ref_cache[0*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3367         AV_WN32A(&ref_cache[1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3368         AV_WN32A(&ref_cache[2*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3369         AV_WN32A(&ref_cache[3*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
3370         return;
3371     }
3372
3373     {
3374         int8_t *ref = &s->current_picture.f.ref_index[list][4*mb_xy];
3375         int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
3376         uint32_t ref01 = (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101;
3377         uint32_t ref23 = (pack16to32(ref2frm[list][ref[2]],ref2frm[list][ref[3]])&0x00FF00FF)*0x0101;
3378         AV_WN32A(&ref_cache[0*8], ref01);
3379         AV_WN32A(&ref_cache[1*8], ref01);
3380         AV_WN32A(&ref_cache[2*8], ref23);
3381         AV_WN32A(&ref_cache[3*8], ref23);
3382     }
3383
3384     {
3385         int16_t (*mv_src)[2] = &s->current_picture.f.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
3386         AV_COPY128(mv_dst + 8*0, mv_src + 0*b_stride);
3387         AV_COPY128(mv_dst + 8*1, mv_src + 1*b_stride);
3388         AV_COPY128(mv_dst + 8*2, mv_src + 2*b_stride);
3389         AV_COPY128(mv_dst + 8*3, mv_src + 3*b_stride);
3390     }
3391 }
3392
3393 /**
3394  *
3395  * @return non zero if the loop filter can be skiped
3396  */
3397 static int fill_filter_caches(H264Context *h, int mb_type){
3398     MpegEncContext * const s = &h->s;
3399     const int mb_xy= h->mb_xy;
3400     int top_xy, left_xy[LEFT_MBS];
3401     int top_type, left_type[LEFT_MBS];
3402     uint8_t *nnz;
3403     uint8_t *nnz_cache;
3404
3405     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
3406
3407     /* Wow, what a mess, why didn't they simplify the interlacing & intra
3408      * stuff, I can't imagine that these complex rules are worth it. */
3409
3410     left_xy[LBOT] = left_xy[LTOP] = mb_xy-1;
3411     if(FRAME_MBAFF){
3412         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.f.mb_type[mb_xy - 1]);
3413         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
3414         if(s->mb_y&1){
3415             if (left_mb_field_flag != curr_mb_field_flag) {
3416                 left_xy[LTOP] -= s->mb_stride;
3417             }
3418         }else{
3419             if(curr_mb_field_flag){
3420                 top_xy += s->mb_stride & (((s->current_picture.f.mb_type[top_xy] >> 7) & 1) - 1);
3421             }
3422             if (left_mb_field_flag != curr_mb_field_flag) {
3423                 left_xy[LBOT] += s->mb_stride;
3424             }
3425         }
3426     }
3427
3428     h->top_mb_xy = top_xy;
3429     h->left_mb_xy[LTOP] = left_xy[LTOP];
3430     h->left_mb_xy[LBOT] = left_xy[LBOT];
3431     {
3432         //for sufficiently low qp, filtering wouldn't do anything
3433         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
3434         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
3435         int qp = s->current_picture.f.qscale_table[mb_xy];
3436         if(qp <= qp_thresh
3437            && (left_xy[LTOP] < 0 || ((qp + s->current_picture.f.qscale_table[left_xy[LTOP]] + 1) >> 1) <= qp_thresh)
3438            && (top_xy        < 0 || ((qp + s->current_picture.f.qscale_table[top_xy       ] + 1) >> 1) <= qp_thresh)) {
3439             if(!FRAME_MBAFF)
3440                 return 1;
3441             if ((left_xy[LTOP] < 0            || ((qp + s->current_picture.f.qscale_table[left_xy[LBOT]        ] + 1) >> 1) <= qp_thresh) &&
3442                 (top_xy        < s->mb_stride || ((qp + s->current_picture.f.qscale_table[top_xy - s->mb_stride] + 1) >> 1) <= qp_thresh))
3443                 return 1;
3444         }
3445     }
3446
3447     top_type        = s->current_picture.f.mb_type[top_xy];
3448     left_type[LTOP] = s->current_picture.f.mb_type[left_xy[LTOP]];
3449     left_type[LBOT] = s->current_picture.f.mb_type[left_xy[LBOT]];
3450     if(h->deblocking_filter == 2){
3451         if(h->slice_table[top_xy       ] != h->slice_num) top_type= 0;
3452         if(h->slice_table[left_xy[LBOT]] != h->slice_num) left_type[LTOP]= left_type[LBOT]= 0;
3453     }else{
3454         if(h->slice_table[top_xy       ] == 0xFFFF) top_type= 0;
3455         if(h->slice_table[left_xy[LBOT]] == 0xFFFF) left_type[LTOP]= left_type[LBOT] =0;
3456     }
3457     h->top_type       = top_type;
3458     h->left_type[LTOP]= left_type[LTOP];
3459     h->left_type[LBOT]= left_type[LBOT];
3460
3461     if(IS_INTRA(mb_type))
3462         return 0;
3463
3464     fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 0);
3465     if(h->list_count == 2)
3466         fill_filter_caches_inter(h, s, mb_type, top_xy, left_xy, top_type, left_type, mb_xy, 1);
3467
3468     nnz = h->non_zero_count[mb_xy];
3469     nnz_cache = h->non_zero_count_cache;
3470     AV_COPY32(&nnz_cache[4+8*1], &nnz[ 0]);
3471     AV_COPY32(&nnz_cache[4+8*2], &nnz[ 4]);
3472     AV_COPY32(&nnz_cache[4+8*3], &nnz[ 8]);
3473     AV_COPY32(&nnz_cache[4+8*4], &nnz[12]);
3474     h->cbp= h->cbp_table[mb_xy];
3475
3476     if(top_type){
3477         nnz = h->non_zero_count[top_xy];
3478         AV_COPY32(&nnz_cache[4+8*0], &nnz[3*4]);
3479     }
3480
3481     if(left_type[LTOP]){
3482         nnz = h->non_zero_count[left_xy[LTOP]];
3483         nnz_cache[3+8*1]= nnz[3+0*4];
3484         nnz_cache[3+8*2]= nnz[3+1*4];
3485         nnz_cache[3+8*3]= nnz[3+2*4];
3486         nnz_cache[3+8*4]= nnz[3+3*4];
3487     }
3488
3489     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
3490     if(!CABAC && h->pps.transform_8x8_mode){
3491         if(IS_8x8DCT(top_type)){
3492             nnz_cache[4+8*0]=
3493             nnz_cache[5+8*0]= (h->cbp_table[top_xy] & 0x4000) >> 12;
3494             nnz_cache[6+8*0]=
3495             nnz_cache[7+8*0]= (h->cbp_table[top_xy] & 0x8000) >> 12;
3496         }
3497         if(IS_8x8DCT(left_type[LTOP])){
3498             nnz_cache[3+8*1]=
3499             nnz_cache[3+8*2]= (h->cbp_table[left_xy[LTOP]]&0x2000) >> 12; //FIXME check MBAFF
3500         }
3501         if(IS_8x8DCT(left_type[LBOT])){
3502             nnz_cache[3+8*3]=
3503             nnz_cache[3+8*4]= (h->cbp_table[left_xy[LBOT]]&0x8000) >> 12; //FIXME check MBAFF
3504         }
3505
3506         if(IS_8x8DCT(mb_type)){
3507             nnz_cache[scan8[0   ]]= nnz_cache[scan8[1   ]]=
3508             nnz_cache[scan8[2   ]]= nnz_cache[scan8[3   ]]= (h->cbp & 0x1000) >> 12;
3509
3510             nnz_cache[scan8[0+ 4]]= nnz_cache[scan8[1+ 4]]=
3511             nnz_cache[scan8[2+ 4]]= nnz_cache[scan8[3+ 4]]= (h->cbp & 0x2000) >> 12;
3512
3513             nnz_cache[scan8[0+ 8]]= nnz_cache[scan8[1+ 8]]=
3514             nnz_cache[scan8[2+ 8]]= nnz_cache[scan8[3+ 8]]= (h->cbp & 0x4000) >> 12;
3515
3516             nnz_cache[scan8[0+12]]= nnz_cache[scan8[1+12]]=
3517             nnz_cache[scan8[2+12]]= nnz_cache[scan8[3+12]]= (h->cbp & 0x8000) >> 12;
3518         }
3519     }
3520
3521     return 0;
3522 }
3523
3524 static void loop_filter(H264Context *h, int start_x, int end_x){
3525     MpegEncContext * const s = &h->s;
3526     uint8_t  *dest_y, *dest_cb, *dest_cr;
3527     int linesize, uvlinesize, mb_x, mb_y;
3528     const int end_mb_y= s->mb_y + FRAME_MBAFF;
3529     const int old_slice_type= h->slice_type;
3530     const int pixel_shift = h->pixel_shift;
3531     const int block_h = 16 >> s->chroma_y_shift;
3532
3533     if(h->deblocking_filter) {
3534         for(mb_x= start_x; mb_x<end_x; mb_x++){
3535             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
3536                 int mb_xy, mb_type;
3537                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
3538                 h->slice_num= h->slice_table[mb_xy];
3539                 mb_type = s->current_picture.f.mb_type[mb_xy];
3540                 h->list_count= h->list_counts[mb_xy];
3541
3542                 if(FRAME_MBAFF)
3543                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
3544
3545                 s->mb_x= mb_x;
3546                 s->mb_y= mb_y;
3547                 dest_y  = s->current_picture.f.data[0] + ((mb_x << pixel_shift) + mb_y * s->linesize  ) * 16;
3548                 dest_cb = s->current_picture.f.data[1] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3549                 dest_cr = s->current_picture.f.data[2] + (mb_x << pixel_shift) * (8 << CHROMA444) + mb_y * s->uvlinesize * block_h;
3550                     //FIXME simplify above
3551
3552                 if (MB_FIELD) {
3553                     linesize   = h->mb_linesize   = s->linesize * 2;
3554                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
3555                     if(mb_y&1){ //FIXME move out of this function?
3556                         dest_y -= s->linesize*15;
3557                         dest_cb-= s->uvlinesize * (block_h - 1);
3558                         dest_cr-= s->uvlinesize * (block_h - 1);
3559                     }
3560                 } else {
3561                     linesize   = h->mb_linesize   = s->linesize;
3562                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
3563                 }
3564                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
3565                 if(fill_filter_caches(h, mb_type))
3566                     continue;
3567                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.f.qscale_table[mb_xy]);
3568                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.f.qscale_table[mb_xy]);
3569
3570                 if (FRAME_MBAFF) {
3571                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3572                 } else {
3573                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
3574                 }
3575             }
3576         }
3577     }
3578     h->slice_type= old_slice_type;
3579     s->mb_x= end_x;
3580     s->mb_y= end_mb_y - FRAME_MBAFF;
3581     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
3582     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
3583 }
3584
3585 static void predict_field_decoding_flag(H264Context *h){
3586     MpegEncContext * const s = &h->s;
3587     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3588     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
3589                 ? s->current_picture.f.mb_type[mb_xy - 1]
3590                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
3591                 ? s->current_picture.f.mb_type[mb_xy - s->mb_stride]
3592                 : 0;
3593     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
3594 }
3595
3596 /**
3597  * Draw edges and report progress for the last MB row.
3598  */
3599 static void decode_finish_row(H264Context *h){
3600     MpegEncContext * const s = &h->s;
3601     int top = 16*(s->mb_y >> FIELD_PICTURE);
3602     int height = 16 << FRAME_MBAFF;
3603     int deblock_border = (16 + 4) << FRAME_MBAFF;
3604     int pic_height = 16*s->mb_height >> FIELD_PICTURE;
3605
3606     if (h->deblocking_filter) {
3607         if((top + height) >= pic_height)
3608             height += deblock_border;
3609
3610         top -= deblock_border;
3611     }
3612
3613     if (top >= pic_height || (top + height) < h->emu_edge_height)
3614         return;
3615
3616     height = FFMIN(height, pic_height - top);
3617     if (top < h->emu_edge_height) {
3618         height = top+height;
3619         top = 0;
3620     }
3621
3622     ff_draw_horiz_band(s, top, height);
3623
3624     if (s->dropable) return;
3625
3626     ff_thread_report_progress((AVFrame*)s->current_picture_ptr, top + height - 1,
3627                              s->picture_structure==PICT_BOTTOM_FIELD);
3628 }
3629
3630 static int decode_slice(struct AVCodecContext *avctx, void *arg){
3631     H264Context *h = *(void**)arg;
3632     MpegEncContext * const s = &h->s;
3633     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
3634     int lf_x_start = s->mb_x;
3635
3636     s->mb_skip_run= -1;
3637
3638     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
3639                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
3640
3641     if( h->pps.cabac ) {
3642         /* realign */
3643         align_get_bits( &s->gb );
3644
3645         /* init cabac */
3646         ff_init_cabac_states( &h->cabac);
3647         ff_init_cabac_decoder( &h->cabac,
3648                                s->gb.buffer + get_bits_count(&s->gb)/8,
3649                                (get_bits_left(&s->gb) + 7)/8);
3650
3651         ff_h264_init_cabac_states(h);
3652
3653         for(;;){
3654 //START_TIMER
3655             int ret = ff_h264_decode_mb_cabac(h);
3656             int eos;
3657 //STOP_TIMER("decode_mb_cabac")
3658
3659             if(ret>=0) ff_h264_hl_decode_mb(h);
3660
3661             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
3662                 s->mb_y++;
3663
3664                 ret = ff_h264_decode_mb_cabac(h);
3665
3666                 if(ret>=0) ff_h264_hl_decode_mb(h);
3667                 s->mb_y--;
3668             }
3669             eos = get_cabac_terminate( &h->cabac );
3670
3671             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
3672                 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);
3673                 if (s->mb_x >= lf_x_start) loop_filter(h, lf_x_start, s->mb_x + 1);
3674                 return 0;
3675             }
3676             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
3677                 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);
3678                 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);
3679                 return -1;
3680             }
3681
3682             if( ++s->mb_x >= s->mb_width ) {
3683                 loop_filter(h, lf_x_start, s->mb_x);
3684                 s->mb_x = lf_x_start = 0;
3685                 decode_finish_row(h);
3686                 ++s->mb_y;
3687                 if(FIELD_OR_MBAFF_PICTURE) {
3688                     ++s->mb_y;
3689                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3690                         predict_field_decoding_flag(h);
3691                 }
3692             }
3693
3694             if( eos || s->mb_y >= s->mb_height ) {
3695                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3696                 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);
3697                 if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3698                 return 0;
3699             }
3700         }
3701
3702     } else {
3703         for(;;){
3704             int ret = ff_h264_decode_mb_cavlc(h);
3705
3706             if(ret>=0) ff_h264_hl_decode_mb(h);
3707
3708             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
3709                 s->mb_y++;
3710                 ret = ff_h264_decode_mb_cavlc(h);
3711
3712                 if(ret>=0) ff_h264_hl_decode_mb(h);
3713                 s->mb_y--;
3714             }
3715
3716             if(ret<0){
3717                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
3718                 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);
3719                 return -1;
3720             }
3721
3722             if(++s->mb_x >= s->mb_width){
3723                 loop_filter(h, lf_x_start, s->mb_x);
3724                 s->mb_x = lf_x_start = 0;
3725                 decode_finish_row(h);
3726                 ++s->mb_y;
3727                 if(FIELD_OR_MBAFF_PICTURE) {
3728                     ++s->mb_y;
3729                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
3730                         predict_field_decoding_flag(h);
3731                 }
3732                 if(s->mb_y >= s->mb_height){
3733                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3734
3735                     if(   get_bits_count(&s->gb) == s->gb.size_in_bits
3736                        || get_bits_count(&s->gb) <  s->gb.size_in_bits && s->avctx->error_recognition < FF_ER_AGGRESSIVE) {
3737                         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);
3738
3739                         return 0;
3740                     }else{
3741                         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);
3742
3743                         return -1;
3744                     }
3745                 }
3746             }
3747
3748             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
3749                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
3750                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
3751                     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);
3752                     if (s->mb_x > lf_x_start) loop_filter(h, lf_x_start, s->mb_x);
3753
3754                     return 0;
3755                 }else{
3756                     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);
3757
3758                     return -1;
3759                 }
3760             }
3761         }
3762     }
3763 }
3764
3765 /**
3766  * Call decode_slice() for each context.
3767  *
3768  * @param h h264 master context
3769  * @param context_count number of contexts to execute
3770  */
3771 static int execute_decode_slices(H264Context *h, int context_count){
3772     MpegEncContext * const s = &h->s;
3773     AVCodecContext * const avctx= s->avctx;
3774     H264Context *hx;
3775     int i;
3776
3777     if (s->avctx->hwaccel || s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3778         return 0;
3779     if(context_count == 1) {
3780         return decode_slice(avctx, &h);
3781     } else {
3782         for(i = 1; i < context_count; i++) {
3783             hx = h->thread_context[i];
3784             hx->s.error_recognition = avctx->error_recognition;
3785             hx->s.error_count = 0;
3786             hx->x264_build= h->x264_build;
3787         }
3788
3789         avctx->execute(avctx, (void *)decode_slice,
3790                        h->thread_context, NULL, context_count, sizeof(void*));
3791
3792         /* pull back stuff from slices to master context */
3793         hx = h->thread_context[context_count - 1];
3794         s->mb_x = hx->s.mb_x;
3795         s->mb_y = hx->s.mb_y;
3796         s->dropable = hx->s.dropable;
3797         s->picture_structure = hx->s.picture_structure;
3798         for(i = 1; i < context_count; i++)
3799             h->s.error_count += h->thread_context[i]->s.error_count;
3800     }
3801
3802     return 0;
3803 }
3804
3805
3806 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
3807     MpegEncContext * const s = &h->s;
3808     AVCodecContext * const avctx= s->avctx;
3809     H264Context *hx; ///< thread context
3810     int buf_index;
3811     int context_count;
3812     int next_avc;
3813     int pass = !(avctx->active_thread_type & FF_THREAD_FRAME);
3814     int nals_needed=0; ///< number of NALs that need decoding before the next frame thread starts
3815     int nal_index;
3816
3817     h->max_contexts = (HAVE_THREADS && (s->avctx->active_thread_type&FF_THREAD_SLICE)) ? avctx->thread_count : 1;
3818     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
3819         h->current_slice = 0;
3820         if (!s->first_field)
3821             s->current_picture_ptr= NULL;
3822         ff_h264_reset_sei(h);
3823     }
3824
3825     for(;pass <= 1;pass++){
3826         buf_index = 0;
3827         context_count = 0;
3828         next_avc = h->is_avc ? 0 : buf_size;
3829         nal_index = 0;
3830     for(;;){
3831         int consumed;
3832         int dst_length;
3833         int bit_length;
3834         const uint8_t *ptr;
3835         int i, nalsize = 0;
3836         int err;
3837
3838         if(buf_index >= next_avc) {
3839             if(buf_index >= buf_size) break;
3840             nalsize = 0;
3841             for(i = 0; i < h->nal_length_size; i++)
3842                 nalsize = (nalsize << 8) | buf[buf_index++];
3843             if(nalsize <= 0 || nalsize > buf_size - buf_index){
3844                 av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
3845                 break;
3846             }
3847             next_avc= buf_index + nalsize;
3848         } else {
3849             // start code prefix search
3850             for(; buf_index + 3 < next_avc; buf_index++){
3851                 // This should always succeed in the first iteration.
3852                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
3853                     break;
3854             }
3855
3856             if(buf_index+3 >= buf_size) break;
3857
3858             buf_index+=3;
3859             if(buf_index >= next_avc) continue;
3860         }
3861
3862         hx = h->thread_context[context_count];
3863
3864         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
3865         if (ptr==NULL || dst_length < 0){
3866             return -1;
3867         }
3868         i= buf_index + consumed;
3869         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
3870            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
3871             s->workaround_bugs |= FF_BUG_TRUNCATED;
3872
3873         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
3874         while(dst_length > 0 && ptr[dst_length - 1] == 0)
3875             dst_length--;
3876         }
3877         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
3878
3879         if(s->avctx->debug&FF_DEBUG_STARTCODE){
3880             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d/%d at %d/%d length %d\n", hx->nal_unit_type, hx->nal_ref_idc, buf_index, buf_size, dst_length);
3881         }
3882
3883         if (h->is_avc && (nalsize != consumed) && nalsize){
3884             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
3885         }
3886
3887         buf_index += consumed;
3888         nal_index++;
3889
3890         if(pass == 0) {
3891             // packets can sometimes contain multiple PPS/SPS
3892             // e.g. two PAFF field pictures in one packet, or a demuxer which splits NALs strangely
3893             // if so, when frame threading we can't start the next thread until we've read all of them
3894             switch (hx->nal_unit_type) {
3895                 case NAL_SPS:
3896                 case NAL_PPS:
3897                     nals_needed = nal_index;
3898                     break;
3899                 case NAL_IDR_SLICE:
3900                 case NAL_SLICE:
3901                     init_get_bits(&hx->s.gb, ptr, bit_length);
3902                     if (!get_ue_golomb(&hx->s.gb))
3903                         nals_needed = nal_index;
3904             }
3905             continue;
3906         }
3907
3908         //FIXME do not discard SEI id
3909         if(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0)
3910             continue;
3911
3912       again:
3913         err = 0;
3914         switch(hx->nal_unit_type){
3915         case NAL_IDR_SLICE:
3916             if (h->nal_unit_type != NAL_IDR_SLICE) {
3917                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
3918                 return -1;
3919             }
3920             idr(h); //FIXME ensure we don't loose some frames if there is reordering
3921         case NAL_SLICE:
3922             init_get_bits(&hx->s.gb, ptr, bit_length);
3923             hx->intra_gb_ptr=
3924             hx->inter_gb_ptr= &hx->s.gb;
3925             hx->s.data_partitioning = 0;
3926
3927             if((err = decode_slice_header(hx, h)))
3928                break;
3929
3930             if (   h->sei_recovery_frame_cnt >= 0
3931                 && ((h->recovery_frame - h->frame_num) & ((1 << h->sps.log2_max_frame_num)-1)) > h->sei_recovery_frame_cnt) {
3932                 h->recovery_frame = (h->frame_num + h->sei_recovery_frame_cnt) %
3933                                     (1 << h->sps.log2_max_frame_num);
3934             }
3935
3936             s->current_picture_ptr->f.key_frame |=
3937                     (hx->nal_unit_type == NAL_IDR_SLICE);
3938
3939             if (h->recovery_frame == h->frame_num) {
3940                 h->sync |= 1;
3941                 h->recovery_frame = -1;
3942             }
3943
3944             h->sync |= !!s->current_picture_ptr->f.key_frame;
3945             h->sync |= 3*!!(s->flags2 & CODEC_FLAG2_SHOW_ALL);
3946             s->current_picture_ptr->sync = h->sync;
3947
3948             if (h->current_slice == 1) {
3949                 if(!(s->flags2 & CODEC_FLAG2_CHUNKS)) {
3950                     decode_postinit(h, nal_index >= nals_needed);
3951                 }
3952
3953                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
3954                     return -1;
3955                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
3956                     ff_vdpau_h264_picture_start(s);
3957             }
3958
3959             if(hx->redundant_pic_count==0
3960                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3961                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3962                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
3963                && avctx->skip_frame < AVDISCARD_ALL){
3964                 if(avctx->hwaccel) {
3965                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
3966                         return -1;
3967                 }else
3968                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
3969                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
3970                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
3971                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
3972                 }else
3973                     context_count++;
3974             }
3975             break;
3976         case NAL_DPA:
3977             init_get_bits(&hx->s.gb, ptr, bit_length);
3978             hx->intra_gb_ptr=
3979             hx->inter_gb_ptr= NULL;
3980
3981             if ((err = decode_slice_header(hx, h)) < 0)
3982                 break;
3983
3984             hx->s.data_partitioning = 1;
3985
3986             break;
3987         case NAL_DPB:
3988             init_get_bits(&hx->intra_gb, ptr, bit_length);
3989             hx->intra_gb_ptr= &hx->intra_gb;
3990             break;
3991         case NAL_DPC:
3992             init_get_bits(&hx->inter_gb, ptr, bit_length);
3993             hx->inter_gb_ptr= &hx->inter_gb;
3994
3995             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
3996                && s->context_initialized
3997                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
3998                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=AV_PICTURE_TYPE_B)
3999                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==AV_PICTURE_TYPE_I)
4000                && avctx->skip_frame < AVDISCARD_ALL)
4001                 context_count++;
4002             break;
4003         case NAL_SEI:
4004             init_get_bits(&s->gb, ptr, bit_length);
4005             ff_h264_decode_sei(h);
4006             break;
4007         case NAL_SPS:
4008             init_get_bits(&s->gb, ptr, bit_length);
4009             if(ff_h264_decode_seq_parameter_set(h) < 0 && (h->is_avc ? (nalsize != consumed) && nalsize : 1)){
4010                 av_log(h->s.avctx, AV_LOG_DEBUG, "SPS decoding failure, trying alternative mode\n");
4011                 if(h->is_avc) av_assert0(next_avc - buf_index + consumed == nalsize);
4012                 init_get_bits(&s->gb, &buf[buf_index + 1 - consumed], 8*(next_avc - buf_index + consumed));
4013                 ff_h264_decode_seq_parameter_set(h);
4014             }
4015
4016             if (s->flags& CODEC_FLAG_LOW_DELAY ||
4017                 (h->sps.bitstream_restriction_flag && !h->sps.num_reorder_frames))
4018                 s->low_delay=1;
4019
4020             if(avctx->has_b_frames < 2)
4021                 avctx->has_b_frames= !s->low_delay;
4022             break;
4023         case NAL_PPS:
4024             init_get_bits(&s->gb, ptr, bit_length);
4025
4026             ff_h264_decode_picture_parameter_set(h, bit_length);
4027
4028             break;
4029         case NAL_AUD:
4030         case NAL_END_SEQUENCE:
4031         case NAL_END_STREAM:
4032         case NAL_FILLER_DATA:
4033         case NAL_SPS_EXT:
4034         case NAL_AUXILIARY_SLICE:
4035             break;
4036         default:
4037             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
4038         }
4039
4040         if(context_count == h->max_contexts) {
4041             execute_decode_slices(h, context_count);
4042             context_count = 0;
4043         }
4044
4045         if (err < 0)
4046             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
4047         else if(err == 1) {
4048             /* Slice could not be decoded in parallel mode, copy down
4049              * NAL unit stuff to context 0 and restart. Note that
4050              * rbsp_buffer is not transferred, but since we no longer
4051              * run in parallel mode this should not be an issue. */
4052             h->nal_unit_type = hx->nal_unit_type;
4053             h->nal_ref_idc   = hx->nal_ref_idc;
4054             hx = h;
4055             goto again;
4056         }
4057     }
4058     }
4059     if(context_count)
4060         execute_decode_slices(h, context_count);
4061     return buf_index;
4062 }
4063
4064 /**
4065  * returns the number of bytes consumed for building the current frame
4066  */
4067 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
4068         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
4069         if(pos+10>buf_size) pos=buf_size; // oops ;)
4070
4071         return pos;
4072 }
4073
4074 static int decode_frame(AVCodecContext *avctx,
4075                              void *data, int *data_size,
4076                              AVPacket *avpkt)
4077 {
4078     const uint8_t *buf = avpkt->data;
4079     int buf_size = avpkt->size;
4080     H264Context *h = avctx->priv_data;
4081     MpegEncContext *s = &h->s;
4082     AVFrame *pict = data;
4083     int buf_index;
4084
4085     s->flags= avctx->flags;
4086     s->flags2= avctx->flags2;
4087
4088    /* end of stream, output what is still in the buffers */
4089  out:
4090     if (buf_size == 0) {
4091         Picture *out;
4092         int i, out_idx;
4093
4094         s->current_picture_ptr = NULL;
4095
4096 //FIXME factorize this with the output code below
4097         out = h->delayed_pic[0];
4098         out_idx = 0;
4099         for (i = 1; h->delayed_pic[i] && !h->delayed_pic[i]->f.key_frame && !h->delayed_pic[i]->mmco_reset; i++)
4100             if(h->delayed_pic[i]->poc < out->poc){
4101                 out = h->delayed_pic[i];
4102                 out_idx = i;
4103             }
4104
4105         for(i=out_idx; h->delayed_pic[i]; i++)
4106             h->delayed_pic[i] = h->delayed_pic[i+1];
4107
4108         if(out){
4109             *data_size = sizeof(AVFrame);
4110             *pict= *(AVFrame*)out;
4111         }
4112
4113         return 0;
4114     }
4115     if(h->is_avc && buf_size >= 9 && AV_RB32(buf)==0x0164001F && buf[5] && buf[8]==0x67)
4116         return ff_h264_decode_extradata(h, buf, buf_size);
4117
4118     buf_index=decode_nal_units(h, buf, buf_size);
4119     if(buf_index < 0)
4120         return -1;
4121
4122     if (!s->current_picture_ptr && h->nal_unit_type == NAL_END_SEQUENCE) {
4123         buf_size = 0;
4124         goto out;
4125     }
4126
4127     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
4128         if (avctx->skip_frame >= AVDISCARD_NONREF ||
4129             buf_size >= 4 && !memcmp("Q264", buf, 4))
4130             return 0;
4131         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
4132         return -1;
4133     }
4134
4135     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
4136
4137         if(s->flags2 & CODEC_FLAG2_CHUNKS) decode_postinit(h, 1);
4138
4139         field_end(h, 0);
4140
4141         *data_size = 0; /* Wait for second field. */
4142         if (h->next_output_pic && h->next_output_pic->sync) {
4143             if(h->sync>1 || h->next_output_pic->f.pict_type != AV_PICTURE_TYPE_B){
4144                 *data_size = sizeof(AVFrame);
4145                 *pict = *(AVFrame*)h->next_output_pic;
4146             }
4147         }
4148     }
4149
4150     assert(pict->data[0] || !*data_size);
4151     ff_print_debug_info(s, pict);
4152 //printf("out %d\n", (int)pict->data[0]);
4153
4154     return get_consumed_bytes(s, buf_index, buf_size);
4155 }
4156 #if 0
4157 static inline void fill_mb_avail(H264Context *h){
4158     MpegEncContext * const s = &h->s;
4159     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
4160
4161     if(s->mb_y){
4162         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
4163         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
4164         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
4165     }else{
4166         h->mb_avail[0]=
4167         h->mb_avail[1]=
4168         h->mb_avail[2]= 0;
4169     }
4170     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
4171     h->mb_avail[4]= 1; //FIXME move out
4172     h->mb_avail[5]= 0; //FIXME move out
4173 }
4174 #endif
4175
4176 #ifdef TEST
4177 #undef printf
4178 #undef random
4179 #define COUNT 8000
4180 #define SIZE (COUNT*40)
4181 extern AVCodec ff_h264_decoder;
4182 int main(void){
4183     int i;
4184     uint8_t temp[SIZE];
4185     PutBitContext pb;
4186     GetBitContext gb;
4187 //    int int_temp[10000];
4188     DSPContext dsp;
4189     AVCodecContext avctx;
4190
4191     avcodec_get_context_defaults3(&avctx, &ff_h264_decoder);
4192
4193     dsputil_init(&dsp, &avctx);
4194
4195     init_put_bits(&pb, temp, SIZE);
4196     printf("testing unsigned exp golomb\n");
4197     for(i=0; i<COUNT; i++){
4198         START_TIMER
4199         set_ue_golomb(&pb, i);
4200         STOP_TIMER("set_ue_golomb");
4201     }
4202     flush_put_bits(&pb);
4203
4204     init_get_bits(&gb, temp, 8*SIZE);
4205     for(i=0; i<COUNT; i++){
4206         int j, s;
4207
4208         s= show_bits(&gb, 24);
4209
4210         START_TIMER
4211         j= get_ue_golomb(&gb);
4212         if(j != i){
4213             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4214 //            return -1;
4215         }
4216         STOP_TIMER("get_ue_golomb");
4217     }
4218
4219
4220     init_put_bits(&pb, temp, SIZE);
4221     printf("testing signed exp golomb\n");
4222     for(i=0; i<COUNT; i++){
4223         START_TIMER
4224         set_se_golomb(&pb, i - COUNT/2);
4225         STOP_TIMER("set_se_golomb");
4226     }
4227     flush_put_bits(&pb);
4228
4229     init_get_bits(&gb, temp, 8*SIZE);
4230     for(i=0; i<COUNT; i++){
4231         int j, s;
4232
4233         s= show_bits(&gb, 24);
4234
4235         START_TIMER
4236         j= get_se_golomb(&gb);
4237         if(j != i - COUNT/2){
4238             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
4239 //            return -1;
4240         }
4241         STOP_TIMER("get_se_golomb");
4242     }
4243
4244     printf("Testing RBSP\n");
4245
4246
4247     return 0;
4248 }
4249 #endif /* TEST */
4250
4251
4252 av_cold void ff_h264_free_context(H264Context *h)
4253 {
4254     int i;
4255
4256     free_tables(h, 1); //FIXME cleanup init stuff perhaps
4257
4258     for(i = 0; i < MAX_SPS_COUNT; i++)
4259         av_freep(h->sps_buffers + i);
4260
4261     for(i = 0; i < MAX_PPS_COUNT; i++)
4262         av_freep(h->pps_buffers + i);
4263 }
4264
4265 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
4266 {
4267     H264Context *h = avctx->priv_data;
4268     MpegEncContext *s = &h->s;
4269
4270     ff_h264_remove_all_refs(h);
4271     ff_h264_free_context(h);
4272
4273     MPV_common_end(s);
4274
4275 //    memset(h, 0, sizeof(H264Context));
4276
4277     return 0;
4278 }
4279
4280 static const AVProfile profiles[] = {
4281     { FF_PROFILE_H264_BASELINE,             "Baseline"              },
4282     { FF_PROFILE_H264_CONSTRAINED_BASELINE, "Constrained Baseline"  },
4283     { FF_PROFILE_H264_MAIN,                 "Main"                  },
4284     { FF_PROFILE_H264_EXTENDED,             "Extended"              },
4285     { FF_PROFILE_H264_HIGH,                 "High"                  },
4286     { FF_PROFILE_H264_HIGH_10,              "High 10"               },
4287     { FF_PROFILE_H264_HIGH_10_INTRA,        "High 10 Intra"         },
4288     { FF_PROFILE_H264_HIGH_422,             "High 4:2:2"            },
4289     { FF_PROFILE_H264_HIGH_422_INTRA,       "High 4:2:2 Intra"      },
4290     { FF_PROFILE_H264_HIGH_444,             "High 4:4:4"            },
4291     { FF_PROFILE_H264_HIGH_444_PREDICTIVE,  "High 4:4:4 Predictive" },
4292     { FF_PROFILE_H264_HIGH_444_INTRA,       "High 4:4:4 Intra"      },
4293     { FF_PROFILE_H264_CAVLC_444,            "CAVLC 4:4:4"           },
4294     { FF_PROFILE_UNKNOWN },
4295 };
4296
4297 static const AVOption h264_options[] = {
4298     {"is_avc", "is avc", offsetof(H264Context, is_avc), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 1, 0},
4299     {"nal_length_size", "nal_length_size", offsetof(H264Context, nal_length_size), FF_OPT_TYPE_INT, {.dbl = 0}, 0, 4, 0},
4300     {NULL}
4301 };
4302
4303 static const AVClass h264_class = {
4304     "H264 Decoder",
4305     av_default_item_name,
4306     h264_options,
4307     LIBAVUTIL_VERSION_INT,
4308 };
4309
4310 static const AVClass h264_vdpau_class = {
4311     "H264 VDPAU Decoder",
4312     av_default_item_name,
4313     h264_options,
4314     LIBAVUTIL_VERSION_INT,
4315 };
4316
4317 AVCodec ff_h264_decoder = {
4318     .name           = "h264",
4319     .type           = AVMEDIA_TYPE_VIDEO,
4320     .id             = CODEC_ID_H264,
4321     .priv_data_size = sizeof(H264Context),
4322     .init           = ff_h264_decode_init,
4323     .close          = ff_h264_decode_end,
4324     .decode         = decode_frame,
4325     .capabilities   = /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY |
4326                       CODEC_CAP_SLICE_THREADS | CODEC_CAP_FRAME_THREADS,
4327     .flush= flush_dpb,
4328     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
4329     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(decode_init_thread_copy),
4330     .update_thread_context = ONLY_IF_THREADS_ENABLED(decode_update_thread_context),
4331     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4332     .priv_class     = &h264_class,
4333 };
4334
4335 #if CONFIG_H264_VDPAU_DECODER
4336 AVCodec ff_h264_vdpau_decoder = {
4337     .name           = "h264_vdpau",
4338     .type           = AVMEDIA_TYPE_VIDEO,
4339     .id             = CODEC_ID_H264,
4340     .priv_data_size = sizeof(H264Context),
4341     .init           = ff_h264_decode_init,
4342     .close          = ff_h264_decode_end,
4343     .decode         = decode_frame,
4344     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
4345     .flush= flush_dpb,
4346     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
4347     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
4348     .profiles = NULL_IF_CONFIG_SMALL(profiles),
4349     .priv_class     = &h264_vdpau_class,
4350 };
4351 #endif