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

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/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 "libavcore/imgutils.h"
  29 #include "internal.h"
  30 #include "dsputil.h"
  31 #include "avcodec.h"
  32 #include "mpegvideo.h"
  33 #include "h264.h"
  34 #include "h264data.h"
  35 #include "h264_mvpred.h"
  36 #include "h264_parser.h"
  37 #include "golomb.h"
  38 #include "mathops.h"
  39 #include "rectangle.h"
  40 #include "vdpau_internal.h"
  41 #include "libavutil/avassert.h"
  42
  43 #include "cabac.h"
  44
  45 //#undef NDEBUG
  46 #include <assert.h>
  47
  48 static const uint8_t rem6[52]={
  49 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
  50 };
  51
  52 static const uint8_t div6[52]={
  53 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
  54 };
  55
  56 static const enum PixelFormat hwaccel_pixfmt_list_h264_jpeg_420[] = {
  57     PIX_FMT_DXVA2_VLD,
  58     PIX_FMT_VAAPI_VLD,
  59     PIX_FMT_YUVJ420P,
  60     PIX_FMT_NONE
  61 };
  62
  63 void ff_h264_write_back_intra_pred_mode(H264Context *h){
  64     int8_t *mode= h->intra4x4_pred_mode + h->mb2br_xy[h->mb_xy];
  65
  66     AV_COPY32(mode, h->intra4x4_pred_mode_cache + 4 + 8*4);
  67     mode[4]= h->intra4x4_pred_mode_cache[7+8*3];
  68     mode[5]= h->intra4x4_pred_mode_cache[7+8*2];
  69     mode[6]= h->intra4x4_pred_mode_cache[7+8*1];
  70 }
  71
  72 /**
  73  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
  74  */
  75 int ff_h264_check_intra4x4_pred_mode(H264Context *h){
  76     MpegEncContext * const s = &h->s;
  77     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
  78     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
  79     int i;
  80
  81     if(!(h->top_samples_available&0x8000)){
  82         for(i=0; i<4; i++){
  83             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
  84             if(status<0){
  85                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
  86                 return -1;
  87             } else if(status){
  88                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
  89             }
  90         }
  91     }
  92
  93     if((h->left_samples_available&0x8888)!=0x8888){
  94         static const int mask[4]={0x8000,0x2000,0x80,0x20};
  95         for(i=0; i<4; i++){
  96             if(!(h->left_samples_available&mask[i])){
  97                 int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
  98                 if(status<0){
  99                     av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
 100                     return -1;
 101                 } else if(status){
 102                     h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
 103                 }
 104             }
 105         }
 106     }
 107
 108     return 0;
 109 } //FIXME cleanup like ff_h264_check_intra_pred_mode
 110
 111 /**
 112  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
 113  */
 114 int ff_h264_check_intra_pred_mode(H264Context *h, int mode){
 115     MpegEncContext * const s = &h->s;
 116     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
 117     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
 118
 119     if(mode > 6U) {
 120         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
 121         return -1;
 122     }
 123
 124     if(!(h->top_samples_available&0x8000)){
 125         mode= top[ mode ];
 126         if(mode<0){
 127             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 128             return -1;
 129         }
 130     }
 131
 132     if((h->left_samples_available&0x8080) != 0x8080){
 133         mode= left[ mode ];
 134         if(h->left_samples_available&0x8080){ //mad cow disease mode, aka MBAFF + constrained_intra_pred
 135             mode= ALZHEIMER_DC_L0T_PRED8x8 + (!(h->left_samples_available&0x8000)) + 2*(mode == DC_128_PRED8x8);
 136         }
 137         if(mode<0){
 138             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
 139             return -1;
 140         }
 141     }
 142
 143     return mode;
 144 }
 145
 146 const uint8_t *ff_h264_decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
 147     int i, si, di;
 148     uint8_t *dst;
 149     int bufidx;
 150
 151 //    src[0]&0x80;                //forbidden bit
 152     h->nal_ref_idc= src[0]>>5;
 153     h->nal_unit_type= src[0]&0x1F;
 154
 155     src++; length--;
 156 #if 0
 157     for(i=0; i<length; i++)
 158         printf("%2X ", src[i]);
 159 #endif
 160
 161 #if HAVE_FAST_UNALIGNED
 162 # if HAVE_FAST_64BIT
 163 #   define RS 7
 164     for(i=0; i+1<length; i+=9){
 165         if(!((~AV_RN64A(src+i) & (AV_RN64A(src+i) - 0x0100010001000101ULL)) & 0x8000800080008080ULL))
 166 # else
 167 #   define RS 3
 168     for(i=0; i+1<length; i+=5){
 169         if(!((~AV_RN32A(src+i) & (AV_RN32A(src+i) - 0x01000101U)) & 0x80008080U))
 170 # endif
 171             continue;
 172         if(i>0 && !src[i]) i--;
 173         while(src[i]) i++;
 174 #else
 175 #   define RS 0
 176     for(i=0; i+1<length; i+=2){
 177         if(src[i]) continue;
 178         if(i>0 && src[i-1]==0) i--;
 179 #endif
 180         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
 181             if(src[i+2]!=3){
 182                 /* startcode, so we must be past the end */
 183                 length=i;
 184             }
 185             break;
 186         }
 187         i-= RS;
 188     }
 189
 190     if(i>=length-1){ //no escaped 0
 191         *dst_length= length;
 192         *consumed= length+1; //+1 for the header
 193         return src;
 194     }
 195
 196     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
 197     av_fast_malloc(&h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length+FF_INPUT_BUFFER_PADDING_SIZE);
 198     dst= h->rbsp_buffer[bufidx];
 199
 200     if (dst == NULL){
 201         return NULL;
 202     }
 203
 204 //printf("decoding esc\n");
 205     memcpy(dst, src, i);
 206     si=di=i;
 207     while(si+2<length){
 208         //remove escapes (very rare 1:2^22)
 209         if(src[si+2]>3){
 210             dst[di++]= src[si++];
 211             dst[di++]= src[si++];
 212         }else if(src[si]==0 && src[si+1]==0){
 213             if(src[si+2]==3){ //escape
 214                 dst[di++]= 0;
 215                 dst[di++]= 0;
 216                 si+=3;
 217                 continue;
 218             }else //next start code
 219                 goto nsc;
 220         }
 221
 222         dst[di++]= src[si++];
 223     }
 224     while(si<length)
 225         dst[di++]= src[si++];
 226 nsc:
 227
 228     memset(dst+di, 0, FF_INPUT_BUFFER_PADDING_SIZE);
 229
 230     *dst_length= di;
 231     *consumed= si + 1;//+1 for the header
 232 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
 233     return dst;
 234 }
 235
 236 int ff_h264_decode_rbsp_trailing(H264Context *h, const uint8_t *src){
 237     int v= *src;
 238     int r;
 239
 240     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
 241
 242     for(r=1; r<9; r++){
 243         if(v&1) return r;
 244         v>>=1;
 245     }
 246     return 0;
 247 }
 248
 249 /**
 250  * IDCT transforms the 16 dc values and dequantizes them.
 251  * @param qp quantization parameter
 252  */
 253 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
 254 #define stride 16
 255     int i;
 256     int temp[16]; //FIXME check if this is a good idea
 257     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 258     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 259
 260 //memset(block, 64, 2*256);
 261 //return;
 262     for(i=0; i<4; i++){
 263         const int offset= y_offset[i];
 264         const int z0= block[offset+stride*0] + block[offset+stride*4];
 265         const int z1= block[offset+stride*0] - block[offset+stride*4];
 266         const int z2= block[offset+stride*1] - block[offset+stride*5];
 267         const int z3= block[offset+stride*1] + block[offset+stride*5];
 268
 269         temp[4*i+0]= z0+z3;
 270         temp[4*i+1]= z1+z2;
 271         temp[4*i+2]= z1-z2;
 272         temp[4*i+3]= z0-z3;
 273     }
 274
 275     for(i=0; i<4; i++){
 276         const int offset= x_offset[i];
 277         const int z0= temp[4*0+i] + temp[4*2+i];
 278         const int z1= temp[4*0+i] - temp[4*2+i];
 279         const int z2= temp[4*1+i] - temp[4*3+i];
 280         const int z3= temp[4*1+i] + temp[4*3+i];
 281
 282         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
 283         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
 284         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
 285         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
 286     }
 287 }
 288
 289 #if 0
 290 /**
 291  * DCT transforms the 16 dc values.
 292  * @param qp quantization parameter ??? FIXME
 293  */
 294 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
 295 //    const int qmul= dequant_coeff[qp][0];
 296     int i;
 297     int temp[16]; //FIXME check if this is a good idea
 298     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
 299     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
 300
 301     for(i=0; i<4; i++){
 302         const int offset= y_offset[i];
 303         const int z0= block[offset+stride*0] + block[offset+stride*4];
 304         const int z1= block[offset+stride*0] - block[offset+stride*4];
 305         const int z2= block[offset+stride*1] - block[offset+stride*5];
 306         const int z3= block[offset+stride*1] + block[offset+stride*5];
 307
 308         temp[4*i+0]= z0+z3;
 309         temp[4*i+1]= z1+z2;
 310         temp[4*i+2]= z1-z2;
 311         temp[4*i+3]= z0-z3;
 312     }
 313
 314     for(i=0; i<4; i++){
 315         const int offset= x_offset[i];
 316         const int z0= temp[4*0+i] + temp[4*2+i];
 317         const int z1= temp[4*0+i] - temp[4*2+i];
 318         const int z2= temp[4*1+i] - temp[4*3+i];
 319         const int z3= temp[4*1+i] + temp[4*3+i];
 320
 321         block[stride*0 +offset]= (z0 + z3)>>1;
 322         block[stride*2 +offset]= (z1 + z2)>>1;
 323         block[stride*8 +offset]= (z1 - z2)>>1;
 324         block[stride*10+offset]= (z0 - z3)>>1;
 325     }
 326 }
 327 #endif
 328
 329 #undef xStride
 330 #undef stride
 331
 332 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
 333     const int stride= 16*2;
 334     const int xStride= 16;
 335     int a,b,c,d,e;
 336
 337     a= block[stride*0 + xStride*0];
 338     b= block[stride*0 + xStride*1];
 339     c= block[stride*1 + xStride*0];
 340     d= block[stride*1 + xStride*1];
 341
 342     e= a-b;
 343     a= a+b;
 344     b= c-d;
 345     c= c+d;
 346
 347     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
 348     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
 349     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
 350     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
 351 }
 352
 353 #if 0
 354 static void chroma_dc_dct_c(DCTELEM *block){
 355     const int stride= 16*2;
 356     const int xStride= 16;
 357     int a,b,c,d,e;
 358
 359     a= block[stride*0 + xStride*0];
 360     b= block[stride*0 + xStride*1];
 361     c= block[stride*1 + xStride*0];
 362     d= block[stride*1 + xStride*1];
 363
 364     e= a-b;
 365     a= a+b;
 366     b= c-d;
 367     c= c+d;
 368
 369     block[stride*0 + xStride*0]= (a+c);
 370     block[stride*0 + xStride*1]= (e+b);
 371     block[stride*1 + xStride*0]= (a-c);
 372     block[stride*1 + xStride*1]= (e-b);
 373 }
 374 #endif
 375
 376 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
 377                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 378                            int src_x_offset, int src_y_offset,
 379                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
 380     MpegEncContext * const s = &h->s;
 381     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
 382     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
 383     const int luma_xy= (mx&3) + ((my&3)<<2);
 384     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
 385     uint8_t * src_cb, * src_cr;
 386     int extra_width= h->emu_edge_width;
 387     int extra_height= h->emu_edge_height;
 388     int emu=0;
 389     const int full_mx= mx>>2;
 390     const int full_my= my>>2;
 391     const int pic_width  = 16*s->mb_width;
 392     const int pic_height = 16*s->mb_height >> MB_FIELD;
 393
 394     if(mx&7) extra_width -= 3;
 395     if(my&7) extra_height -= 3;
 396
 397     if(   full_mx < 0-extra_width
 398        || full_my < 0-extra_height
 399        || full_mx + 16/*FIXME*/ > pic_width + extra_width
 400        || full_my + 16/*FIXME*/ > pic_height + extra_height){
 401         ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
 402             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
 403         emu=1;
 404     }
 405
 406     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
 407     if(!square){
 408         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
 409     }
 410
 411     if(CONFIG_GRAY && s->flags&CODEC_FLAG_GRAY) return;
 412
 413     if(MB_FIELD){
 414         // chroma offset when predicting from a field of opposite parity
 415         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
 416         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
 417     }
 418     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
 419     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
 420
 421     if(emu){
 422         ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 423             src_cb= s->edge_emu_buffer;
 424     }
 425     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 426
 427     if(emu){
 428         ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
 429             src_cr= s->edge_emu_buffer;
 430     }
 431     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
 432 }
 433
 434 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
 435                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 436                            int x_offset, int y_offset,
 437                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 438                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 439                            int list0, int list1){
 440     MpegEncContext * const s = &h->s;
 441     qpel_mc_func *qpix_op=  qpix_put;
 442     h264_chroma_mc_func chroma_op= chroma_put;
 443
 444     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
 445     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
 446     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
 447     x_offset += 8*s->mb_x;
 448     y_offset += 8*(s->mb_y >> MB_FIELD);
 449
 450     if(list0){
 451         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
 452         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
 453                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 454                            qpix_op, chroma_op);
 455
 456         qpix_op=  qpix_avg;
 457         chroma_op= chroma_avg;
 458     }
 459
 460     if(list1){
 461         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
 462         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
 463                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
 464                            qpix_op, chroma_op);
 465     }
 466 }
 467
 468 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
 469                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 470                            int x_offset, int y_offset,
 471                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 472                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
 473                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
 474                            int list0, int list1){
 475     MpegEncContext * const s = &h->s;
 476
 477     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
 478     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
 479     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
 480     x_offset += 8*s->mb_x;
 481     y_offset += 8*(s->mb_y >> MB_FIELD);
 482
 483     if(list0 && list1){
 484         /* don't optimize for luma-only case, since B-frames usually
 485          * use implicit weights => chroma too. */
 486         uint8_t *tmp_cb = s->obmc_scratchpad;
 487         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
 488         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
 489         int refn0 = h->ref_cache[0][ scan8[n] ];
 490         int refn1 = h->ref_cache[1][ scan8[n] ];
 491
 492         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
 493                     dest_y, dest_cb, dest_cr,
 494                     x_offset, y_offset, qpix_put, chroma_put);
 495         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
 496                     tmp_y, tmp_cb, tmp_cr,
 497                     x_offset, y_offset, qpix_put, chroma_put);
 498
 499         if(h->use_weight == 2){
 500             int weight0 = h->implicit_weight[refn0][refn1][s->mb_y&1];
 501             int weight1 = 64 - weight0;
 502             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
 503             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
 504             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
 505         }else{
 506             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
 507                             h->luma_weight[refn0][0][0] , h->luma_weight[refn1][1][0],
 508                             h->luma_weight[refn0][0][1] + h->luma_weight[refn1][1][1]);
 509             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 510                             h->chroma_weight[refn0][0][0][0] , h->chroma_weight[refn1][1][0][0],
 511                             h->chroma_weight[refn0][0][0][1] + h->chroma_weight[refn1][1][0][1]);
 512             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 513                             h->chroma_weight[refn0][0][1][0] , h->chroma_weight[refn1][1][1][0],
 514                             h->chroma_weight[refn0][0][1][1] + h->chroma_weight[refn1][1][1][1]);
 515         }
 516     }else{
 517         int list = list1 ? 1 : 0;
 518         int refn = h->ref_cache[list][ scan8[n] ];
 519         Picture *ref= &h->ref_list[list][refn];
 520         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
 521                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
 522                     qpix_put, chroma_put);
 523
 524         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
 525                        h->luma_weight[refn][list][0], h->luma_weight[refn][list][1]);
 526         if(h->use_weight_chroma){
 527             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 528                              h->chroma_weight[refn][list][0][0], h->chroma_weight[refn][list][0][1]);
 529             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
 530                              h->chroma_weight[refn][list][1][0], h->chroma_weight[refn][list][1][1]);
 531         }
 532     }
 533 }
 534
 535 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
 536                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 537                            int x_offset, int y_offset,
 538                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
 539                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
 540                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
 541                            int list0, int list1){
 542     if((h->use_weight==2 && list0 && list1
 543         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ][h->s.mb_y&1] != 32))
 544        || h->use_weight==1)
 545         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 546                          x_offset, y_offset, qpix_put, chroma_put,
 547                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
 548     else
 549         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
 550                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
 551 }
 552
 553 static inline void prefetch_motion(H264Context *h, int list){
 554     /* fetch pixels for estimated mv 4 macroblocks ahead
 555      * optimized for 64byte cache lines */
 556     MpegEncContext * const s = &h->s;
 557     const int refn = h->ref_cache[list][scan8[0]];
 558     if(refn >= 0){
 559         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
 560         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
 561         uint8_t **src= h->ref_list[list][refn].data;
 562         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
 563         s->dsp.prefetch(src[0]+off, s->linesize, 4);
 564         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
 565         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
 566     }
 567 }
 568
 569 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
 570                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
 571                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
 572                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
 573     MpegEncContext * const s = &h->s;
 574     const int mb_xy= h->mb_xy;
 575     const int mb_type= s->current_picture.mb_type[mb_xy];
 576
 577     assert(IS_INTER(mb_type));
 578
 579     prefetch_motion(h, 0);
 580
 581     if(IS_16X16(mb_type)){
 582         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
 583                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
 584                 weight_op, weight_avg,
 585                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 586     }else if(IS_16X8(mb_type)){
 587         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
 588                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 589                 &weight_op[1], &weight_avg[1],
 590                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 591         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
 592                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
 593                 &weight_op[1], &weight_avg[1],
 594                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
 595     }else if(IS_8X16(mb_type)){
 596         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
 597                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 598                 &weight_op[2], &weight_avg[2],
 599                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
 600         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
 601                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 602                 &weight_op[2], &weight_avg[2],
 603                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
 604     }else{
 605         int i;
 606
 607         assert(IS_8X8(mb_type));
 608
 609         for(i=0; i<4; i++){
 610             const int sub_mb_type= h->sub_mb_type[i];
 611             const int n= 4*i;
 612             int x_offset= (i&1)<<2;
 613             int y_offset= (i&2)<<1;
 614
 615             if(IS_SUB_8X8(sub_mb_type)){
 616                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 617                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
 618                     &weight_op[3], &weight_avg[3],
 619                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 620             }else if(IS_SUB_8X4(sub_mb_type)){
 621                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 622                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 623                     &weight_op[4], &weight_avg[4],
 624                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 625                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
 626                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
 627                     &weight_op[4], &weight_avg[4],
 628                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 629             }else if(IS_SUB_4X8(sub_mb_type)){
 630                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
 631                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 632                     &weight_op[5], &weight_avg[5],
 633                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 634                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
 635                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 636                     &weight_op[5], &weight_avg[5],
 637                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 638             }else{
 639                 int j;
 640                 assert(IS_SUB_4X4(sub_mb_type));
 641                 for(j=0; j<4; j++){
 642                     int sub_x_offset= x_offset + 2*(j&1);
 643                     int sub_y_offset= y_offset +   (j&2);
 644                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
 645                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
 646                         &weight_op[6], &weight_avg[6],
 647                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
 648                 }
 649             }
 650         }
 651     }
 652
 653     prefetch_motion(h, 1);
 654 }
 655
 656
 657 static void free_tables(H264Context *h){
 658     int i;
 659     H264Context *hx;
 660     av_freep(&h->intra4x4_pred_mode);
 661     av_freep(&h->chroma_pred_mode_table);
 662     av_freep(&h->cbp_table);
 663     av_freep(&h->mvd_table[0]);
 664     av_freep(&h->mvd_table[1]);
 665     av_freep(&h->direct_table);
 666     av_freep(&h->non_zero_count);
 667     av_freep(&h->slice_table_base);
 668     h->slice_table= NULL;
 669     av_freep(&h->list_counts);
 670
 671     av_freep(&h->mb2b_xy);
 672     av_freep(&h->mb2br_xy);
 673
 674     for(i = 0; i < MAX_THREADS; i++) {
 675         hx = h->thread_context[i];
 676         if(!hx) continue;
 677         av_freep(&hx->top_borders[1]);
 678         av_freep(&hx->top_borders[0]);
 679         av_freep(&hx->s.obmc_scratchpad);
 680         av_freep(&hx->rbsp_buffer[1]);
 681         av_freep(&hx->rbsp_buffer[0]);
 682         hx->rbsp_buffer_size[0] = 0;
 683         hx->rbsp_buffer_size[1] = 0;
 684         if (i) av_freep(&h->thread_context[i]);
 685     }
 686 }
 687
 688 static void init_dequant8_coeff_table(H264Context *h){
 689     int i,q,x;
 690     const int transpose = (h->h264dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
 691     h->dequant8_coeff[0] = h->dequant8_buffer[0];
 692     h->dequant8_coeff[1] = h->dequant8_buffer[1];
 693
 694     for(i=0; i<2; i++ ){
 695         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
 696             h->dequant8_coeff[1] = h->dequant8_buffer[0];
 697             break;
 698         }
 699
 700         for(q=0; q<52; q++){
 701             int shift = div6[q];
 702             int idx = rem6[q];
 703             for(x=0; x<64; x++)
 704                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
 705                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
 706                     h->pps.scaling_matrix8[i][x]) << shift;
 707         }
 708     }
 709 }
 710
 711 static void init_dequant4_coeff_table(H264Context *h){
 712     int i,j,q,x;
 713     const int transpose = (h->h264dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
 714     for(i=0; i<6; i++ ){
 715         h->dequant4_coeff[i] = h->dequant4_buffer[i];
 716         for(j=0; j<i; j++){
 717             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
 718                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
 719                 break;
 720             }
 721         }
 722         if(j<i)
 723             continue;
 724
 725         for(q=0; q<52; q++){
 726             int shift = div6[q] + 2;
 727             int idx = rem6[q];
 728             for(x=0; x<16; x++)
 729                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
 730                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
 731                     h->pps.scaling_matrix4[i][x]) << shift;
 732         }
 733     }
 734 }
 735
 736 static void init_dequant_tables(H264Context *h){
 737     int i,x;
 738     init_dequant4_coeff_table(h);
 739     if(h->pps.transform_8x8_mode)
 740         init_dequant8_coeff_table(h);
 741     if(h->sps.transform_bypass){
 742         for(i=0; i<6; i++)
 743             for(x=0; x<16; x++)
 744                 h->dequant4_coeff[i][0][x] = 1<<6;
 745         if(h->pps.transform_8x8_mode)
 746             for(i=0; i<2; i++)
 747                 for(x=0; x<64; x++)
 748                     h->dequant8_coeff[i][0][x] = 1<<6;
 749     }
 750 }
 751
 752
 753 int ff_h264_alloc_tables(H264Context *h){
 754     MpegEncContext * const s = &h->s;
 755     const int big_mb_num= s->mb_stride * (s->mb_height+1);
 756     const int row_mb_num= 2*s->mb_stride*s->avctx->thread_count;
 757     int x,y;
 758
 759     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->intra4x4_pred_mode, row_mb_num * 8  * sizeof(uint8_t), fail)
 760
 761     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->non_zero_count    , big_mb_num * 32 * sizeof(uint8_t), fail)
 762     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(*h->slice_table_base), fail)
 763     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->cbp_table, big_mb_num * sizeof(uint16_t), fail)
 764
 765     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t), fail)
 766     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[0], 16*row_mb_num * sizeof(uint8_t), fail);
 767     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mvd_table[1], 16*row_mb_num * sizeof(uint8_t), fail);
 768     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->direct_table, 4*big_mb_num * sizeof(uint8_t) , fail);
 769     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->list_counts, big_mb_num * sizeof(uint8_t), fail)
 770
 771     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(*h->slice_table_base));
 772     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
 773
 774     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2b_xy  , big_mb_num * sizeof(uint32_t), fail);
 775     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->mb2br_xy , big_mb_num * sizeof(uint32_t), fail);
 776     for(y=0; y<s->mb_height; y++){
 777         for(x=0; x<s->mb_width; x++){
 778             const int mb_xy= x + y*s->mb_stride;
 779             const int b_xy = 4*x + 4*y*h->b_stride;
 780
 781             h->mb2b_xy [mb_xy]= b_xy;
 782             h->mb2br_xy[mb_xy]= 8*(FMO ? mb_xy : (mb_xy % (2*s->mb_stride)));
 783         }
 784     }
 785
 786     s->obmc_scratchpad = NULL;
 787
 788     if(!h->dequant4_coeff[0])
 789         init_dequant_tables(h);
 790
 791     return 0;
 792 fail:
 793     free_tables(h);
 794     return -1;
 795 }
 796
 797 /**
 798  * Mimic alloc_tables(), but for every context thread.
 799  */
 800 static void clone_tables(H264Context *dst, H264Context *src, int i){
 801     MpegEncContext * const s = &src->s;
 802     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode + i*8*2*s->mb_stride;
 803     dst->non_zero_count           = src->non_zero_count;
 804     dst->slice_table              = src->slice_table;
 805     dst->cbp_table                = src->cbp_table;
 806     dst->mb2b_xy                  = src->mb2b_xy;
 807     dst->mb2br_xy                 = src->mb2br_xy;
 808     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
 809     dst->mvd_table[0]             = src->mvd_table[0] + i*8*2*s->mb_stride;
 810     dst->mvd_table[1]             = src->mvd_table[1] + i*8*2*s->mb_stride;
 811     dst->direct_table             = src->direct_table;
 812     dst->list_counts              = src->list_counts;
 813
 814     dst->s.obmc_scratchpad = NULL;
 815     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
 816 }
 817
 818 /**
 819  * Init context
 820  * Allocate buffers which are not shared amongst multiple threads.
 821  */
 822 static int context_init(H264Context *h){
 823     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)
 824     FF_ALLOCZ_OR_GOTO(h->s.avctx, h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t), fail)
 825
 826     h->ref_cache[0][scan8[5 ]+1] = h->ref_cache[0][scan8[7 ]+1] = h->ref_cache[0][scan8[13]+1] =
 827     h->ref_cache[1][scan8[5 ]+1] = h->ref_cache[1][scan8[7 ]+1] = h->ref_cache[1][scan8[13]+1] = PART_NOT_AVAILABLE;
 828
 829     return 0;
 830 fail:
 831     return -1; // free_tables will clean up for us
 832 }
 833
 834 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size);
 835
 836 static av_cold void common_init(H264Context *h){
 837     MpegEncContext * const s = &h->s;
 838
 839     s->width = s->avctx->width;
 840     s->height = s->avctx->height;
 841     s->codec_id= s->avctx->codec->id;
 842
 843     ff_h264dsp_init(&h->h264dsp);
 844     ff_h264_pred_init(&h->hpc, s->codec_id);
 845
 846     h->dequant_coeff_pps= -1;
 847     s->unrestricted_mv=1;
 848     s->decode=1; //FIXME
 849
 850     dsputil_init(&s->dsp, s->avctx); // needed so that idct permutation is known early
 851
 852     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
 853     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
 854 }
 855
 856 int ff_h264_decode_extradata(H264Context *h)
 857 {
 858     AVCodecContext *avctx = h->s.avctx;
 859
 860     if(*(char *)avctx->extradata == 1){
 861         int i, cnt, nalsize;
 862         unsigned char *p = avctx->extradata;
 863
 864         h->is_avc = 1;
 865
 866         if(avctx->extradata_size < 7) {
 867             av_log(avctx, AV_LOG_ERROR, "avcC too short\n");
 868             return -1;
 869         }
 870         /* sps and pps in the avcC always have length coded with 2 bytes,
 871            so put a fake nal_length_size = 2 while parsing them */
 872         h->nal_length_size = 2;
 873         // Decode sps from avcC
 874         cnt = *(p+5) & 0x1f; // Number of sps
 875         p += 6;
 876         for (i = 0; i < cnt; i++) {
 877             nalsize = AV_RB16(p) + 2;
 878             if(decode_nal_units(h, p, nalsize) < 0) {
 879                 av_log(avctx, AV_LOG_ERROR, "Decoding sps %d from avcC failed\n", i);
 880                 return -1;
 881             }
 882             p += nalsize;
 883         }
 884         // Decode pps from avcC
 885         cnt = *(p++); // Number of pps
 886         for (i = 0; i < cnt; i++) {
 887             nalsize = AV_RB16(p) + 2;
 888             if(decode_nal_units(h, p, nalsize)  != nalsize) {
 889                 av_log(avctx, AV_LOG_ERROR, "Decoding pps %d from avcC failed\n", i);
 890                 return -1;
 891             }
 892             p += nalsize;
 893         }
 894         // Now store right nal length size, that will be use to parse all other nals
 895         h->nal_length_size = ((*(((char*)(avctx->extradata))+4))&0x03)+1;
 896     } else {
 897         h->is_avc = 0;
 898         if(decode_nal_units(h, avctx->extradata, avctx->extradata_size) < 0)
 899             return -1;
 900     }
 901     return 0;
 902 }
 903
 904 av_cold int ff_h264_decode_init(AVCodecContext *avctx){
 905     H264Context *h= avctx->priv_data;
 906     MpegEncContext * const s = &h->s;
 907
 908     MPV_decode_defaults(s);
 909
 910     s->avctx = avctx;
 911     common_init(h);
 912
 913     s->out_format = FMT_H264;
 914     s->workaround_bugs= avctx->workaround_bugs;
 915
 916     // set defaults
 917 //    s->decode_mb= ff_h263_decode_mb;
 918     s->quarter_sample = 1;
 919     if(!avctx->has_b_frames)
 920     s->low_delay= 1;
 921
 922     avctx->chroma_sample_location = AVCHROMA_LOC_LEFT;
 923
 924     ff_h264_decode_init_vlc();
 925
 926     h->thread_context[0] = h;
 927     h->outputed_poc = INT_MIN;
 928     h->prev_poc_msb= 1<<16;
 929     h->x264_build = -1;
 930     ff_h264_reset_sei(h);
 931     if(avctx->codec_id == CODEC_ID_H264){
 932         if(avctx->ticks_per_frame == 1){
 933             s->avctx->time_base.den *=2;
 934         }
 935         avctx->ticks_per_frame = 2;
 936     }
 937
 938     if(avctx->extradata_size > 0 && avctx->extradata &&
 939         ff_h264_decode_extradata(h))
 940         return -1;
 941
 942     if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames < h->sps.num_reorder_frames){
 943         s->avctx->has_b_frames = h->sps.num_reorder_frames;
 944         s->low_delay = 0;
 945     }
 946
 947     return 0;
 948 }
 949
 950 int ff_h264_frame_start(H264Context *h){
 951     MpegEncContext * const s = &h->s;
 952     int i;
 953
 954     if(MPV_frame_start(s, s->avctx) < 0)
 955         return -1;
 956     ff_er_frame_start(s);
 957     /*
 958      * MPV_frame_start uses pict_type to derive key_frame.
 959      * This is incorrect for H.264; IDR markings must be used.
 960      * Zero here; IDR markings per slice in frame or fields are ORed in later.
 961      * See decode_nal_units().
 962      */
 963     s->current_picture_ptr->key_frame= 0;
 964     s->current_picture_ptr->mmco_reset= 0;
 965
 966     assert(s->linesize && s->uvlinesize);
 967
 968     for(i=0; i<16; i++){
 969         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
 970         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
 971     }
 972     for(i=0; i<4; i++){
 973         h->block_offset[16+i]=
 974         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
 975         h->block_offset[24+16+i]=
 976         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
 977     }
 978
 979     /* can't be in alloc_tables because linesize isn't known there.
 980      * FIXME: redo bipred weight to not require extra buffer? */
 981     for(i = 0; i < s->avctx->thread_count; i++)
 982         if(!h->thread_context[i]->s.obmc_scratchpad)
 983             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
 984
 985     /* some macroblocks can be accessed before they're available in case of lost slices, mbaff or threading*/
 986     memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(*h->slice_table));
 987
 988 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
 989
 990     // We mark the current picture as non-reference after allocating it, so
 991     // that if we break out due to an error it can be released automatically
 992     // in the next MPV_frame_start().
 993     // SVQ3 as well as most other codecs have only last/next/current and thus
 994     // get released even with set reference, besides SVQ3 and others do not
 995     // mark frames as reference later "naturally".
 996     if(s->codec_id != CODEC_ID_SVQ3)
 997         s->current_picture_ptr->reference= 0;
 998
 999     s->current_picture_ptr->field_poc[0]=
1000     s->current_picture_ptr->field_poc[1]= INT_MAX;
1001     assert(s->current_picture_ptr->long_ref==0);
1002
1003     return 0;
1004 }
1005
1006 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
1007     MpegEncContext * const s = &h->s;
1008     uint8_t *top_border;
1009     int top_idx = 1;
1010
1011     src_y  -=   linesize;
1012     src_cb -= uvlinesize;
1013     src_cr -= uvlinesize;
1014
1015     if(!simple && FRAME_MBAFF){
1016         if(s->mb_y&1){
1017             if(!MB_MBAFF){
1018                 top_border = h->top_borders[0][s->mb_x];
1019                 AV_COPY128(top_border, src_y + 15*linesize);
1020                 if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1021                     AV_COPY64(top_border+16, src_cb+7*uvlinesize);
1022                     AV_COPY64(top_border+24, src_cr+7*uvlinesize);
1023                 }
1024             }
1025         }else if(MB_MBAFF){
1026             top_idx = 0;
1027         }else
1028             return;
1029     }
1030
1031     top_border = h->top_borders[top_idx][s->mb_x];
1032     // There are two lines saved, the line above the the top macroblock of a pair,
1033     // and the line above the bottom macroblock
1034     AV_COPY128(top_border, src_y + 16*linesize);
1035
1036     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1037         AV_COPY64(top_border+16, src_cb+8*uvlinesize);
1038         AV_COPY64(top_border+24, src_cr+8*uvlinesize);
1039     }
1040 }
1041
1042 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg, int simple){
1043     MpegEncContext * const s = &h->s;
1044     int deblock_left;
1045     int deblock_top;
1046     int top_idx = 1;
1047     uint8_t *top_border_m1;
1048     uint8_t *top_border;
1049
1050     if(!simple && FRAME_MBAFF){
1051         if(s->mb_y&1){
1052             if(!MB_MBAFF)
1053                 return;
1054         }else{
1055             top_idx = MB_MBAFF ? 0 : 1;
1056         }
1057     }
1058
1059     if(h->deblocking_filter == 2) {
1060         deblock_left = h->left_type[0];
1061         deblock_top  = h->top_type;
1062     } else {
1063         deblock_left = (s->mb_x > 0);
1064         deblock_top =  (s->mb_y > !!MB_FIELD);
1065     }
1066
1067     src_y  -=   linesize + 1;
1068     src_cb -= uvlinesize + 1;
1069     src_cr -= uvlinesize + 1;
1070
1071     top_border_m1 = h->top_borders[top_idx][s->mb_x-1];
1072     top_border    = h->top_borders[top_idx][s->mb_x];
1073
1074 #define XCHG(a,b,xchg)\
1075 if (xchg) AV_SWAP64(b,a);\
1076 else      AV_COPY64(b,a);
1077
1078     if(deblock_top){
1079         if(deblock_left){
1080             XCHG(top_border_m1+8, src_y -7, 1);
1081         }
1082         XCHG(top_border+0, src_y +1, xchg);
1083         XCHG(top_border+8, src_y +9, 1);
1084         if(s->mb_x+1 < s->mb_width){
1085             XCHG(h->top_borders[top_idx][s->mb_x+1], src_y +17, 1);
1086         }
1087     }
1088
1089     if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1090         if(deblock_top){
1091             if(deblock_left){
1092                 XCHG(top_border_m1+16, src_cb -7, 1);
1093                 XCHG(top_border_m1+24, src_cr -7, 1);
1094             }
1095             XCHG(top_border+16, src_cb+1, 1);
1096             XCHG(top_border+24, src_cr+1, 1);
1097         }
1098     }
1099 }
1100
1101 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
1102     MpegEncContext * const s = &h->s;
1103     const int mb_x= s->mb_x;
1104     const int mb_y= s->mb_y;
1105     const int mb_xy= h->mb_xy;
1106     const int mb_type= s->current_picture.mb_type[mb_xy];
1107     uint8_t  *dest_y, *dest_cb, *dest_cr;
1108     int linesize, uvlinesize /*dct_offset*/;
1109     int i;
1110     int *block_offset = &h->block_offset[0];
1111     const int transform_bypass = !simple && (s->qscale == 0 && h->sps.transform_bypass);
1112     /* is_h264 should always be true if SVQ3 is disabled. */
1113     const int is_h264 = !CONFIG_SVQ3_DECODER || simple || s->codec_id == CODEC_ID_H264;
1114     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
1115     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
1116
1117     dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
1118     dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
1119     dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
1120
1121     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
1122     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
1123
1124     h->list_counts[mb_xy]= h->list_count;
1125
1126     if (!simple && MB_FIELD) {
1127         linesize   = h->mb_linesize   = s->linesize * 2;
1128         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
1129         block_offset = &h->block_offset[24];
1130         if(mb_y&1){ //FIXME move out of this function?
1131             dest_y -= s->linesize*15;
1132             dest_cb-= s->uvlinesize*7;
1133             dest_cr-= s->uvlinesize*7;
1134         }
1135         if(FRAME_MBAFF) {
1136             int list;
1137             for(list=0; list<h->list_count; list++){
1138                 if(!USES_LIST(mb_type, list))
1139                     continue;
1140                 if(IS_16X16(mb_type)){
1141                     int8_t *ref = &h->ref_cache[list][scan8[0]];
1142                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
1143                 }else{
1144                     for(i=0; i<16; i+=4){
1145                         int ref = h->ref_cache[list][scan8[i]];
1146                         if(ref >= 0)
1147                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
1148                     }
1149                 }
1150             }
1151         }
1152     } else {
1153         linesize   = h->mb_linesize   = s->linesize;
1154         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
1155 //        dct_offset = s->linesize * 16;
1156     }
1157
1158     if (!simple && IS_INTRA_PCM(mb_type)) {
1159         for (i=0; i<16; i++) {
1160             memcpy(dest_y + i*  linesize, h->mb       + i*8, 16);
1161         }
1162         for (i=0; i<8; i++) {
1163             memcpy(dest_cb+ i*uvlinesize, h->mb + 128 + i*4,  8);
1164             memcpy(dest_cr+ i*uvlinesize, h->mb + 160 + i*4,  8);
1165         }
1166     } else {
1167         if(IS_INTRA(mb_type)){
1168             if(h->deblocking_filter)
1169                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
1170
1171             if(simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
1172                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
1173                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
1174             }
1175
1176             if(IS_INTRA4x4(mb_type)){
1177                 if(simple || !s->encoding){
1178                     if(IS_8x8DCT(mb_type)){
1179                         if(transform_bypass){
1180                             idct_dc_add =
1181                             idct_add    = s->dsp.add_pixels8;
1182                         }else{
1183                             idct_dc_add = h->h264dsp.h264_idct8_dc_add;
1184                             idct_add    = h->h264dsp.h264_idct8_add;
1185                         }
1186                         for(i=0; i<16; i+=4){
1187                             uint8_t * const ptr= dest_y + block_offset[i];
1188                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1189                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1190                                 h->hpc.pred8x8l_add[dir](ptr, h->mb + i*16, linesize);
1191                             }else{
1192                                 const int nnz = h->non_zero_count_cache[ scan8[i] ];
1193                                 h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
1194                                                             (h->topright_samples_available<<i)&0x4000, linesize);
1195                                 if(nnz){
1196                                     if(nnz == 1 && h->mb[i*16])
1197                                         idct_dc_add(ptr, h->mb + i*16, linesize);
1198                                     else
1199                                         idct_add   (ptr, h->mb + i*16, linesize);
1200                                 }
1201                             }
1202                         }
1203                     }else{
1204                         if(transform_bypass){
1205                             idct_dc_add =
1206                             idct_add    = s->dsp.add_pixels4;
1207                         }else{
1208                             idct_dc_add = h->h264dsp.h264_idct_dc_add;
1209                             idct_add    = h->h264dsp.h264_idct_add;
1210                         }
1211                         for(i=0; i<16; i++){
1212                             uint8_t * const ptr= dest_y + block_offset[i];
1213                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
1214
1215                             if(transform_bypass && h->sps.profile_idc==244 && dir<=1){
1216                                 h->hpc.pred4x4_add[dir](ptr, h->mb + i*16, linesize);
1217                             }else{
1218                                 uint8_t *topright;
1219                                 int nnz, tr;
1220                                 if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
1221                                     const int topright_avail= (h->topright_samples_available<<i)&0x8000;
1222                                     assert(mb_y || linesize <= block_offset[i]);
1223                                     if(!topright_avail){
1224                                         tr= ptr[3 - linesize]*0x01010101;
1225                                         topright= (uint8_t*) &tr;
1226                                     }else
1227                                         topright= ptr + 4 - linesize;
1228                                 }else
1229                                     topright= NULL;
1230
1231                                 h->hpc.pred4x4[ dir ](ptr, topright, linesize);
1232                                 nnz = h->non_zero_count_cache[ scan8[i] ];
1233                                 if(nnz){
1234                                     if(is_h264){
1235                                         if(nnz == 1 && h->mb[i*16])
1236                                             idct_dc_add(ptr, h->mb + i*16, linesize);
1237                                         else
1238                                             idct_add   (ptr, h->mb + i*16, linesize);
1239                                     }else
1240                                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
1241                                 }
1242                             }
1243                         }
1244                     }
1245                 }
1246             }else{
1247                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
1248                 if(is_h264){
1249                     if(!transform_bypass)
1250                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
1251                 }else
1252                     ff_svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
1253             }
1254             if(h->deblocking_filter)
1255                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
1256         }else if(is_h264){
1257             hl_motion(h, dest_y, dest_cb, dest_cr,
1258                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
1259                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
1260                       h->h264dsp.weight_h264_pixels_tab, h->h264dsp.biweight_h264_pixels_tab);
1261         }
1262
1263
1264         if(!IS_INTRA4x4(mb_type)){
1265             if(is_h264){
1266                 if(IS_INTRA16x16(mb_type)){
1267                     if(transform_bypass){
1268                         if(h->sps.profile_idc==244 && (h->intra16x16_pred_mode==VERT_PRED8x8 || h->intra16x16_pred_mode==HOR_PRED8x8)){
1269                             h->hpc.pred16x16_add[h->intra16x16_pred_mode](dest_y, block_offset, h->mb, linesize);
1270                         }else{
1271                             for(i=0; i<16; i++){
1272                                 if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
1273                                     s->dsp.add_pixels4(dest_y + block_offset[i], h->mb + i*16, linesize);
1274                             }
1275                         }
1276                     }else{
1277                          h->h264dsp.h264_idct_add16intra(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1278                     }
1279                 }else if(h->cbp&15){
1280                     if(transform_bypass){
1281                         const int di = IS_8x8DCT(mb_type) ? 4 : 1;
1282                         idct_add= IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
1283                         for(i=0; i<16; i+=di){
1284                             if(h->non_zero_count_cache[ scan8[i] ]){
1285                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
1286                             }
1287                         }
1288                     }else{
1289                         if(IS_8x8DCT(mb_type)){
1290                             h->h264dsp.h264_idct8_add4(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1291                         }else{
1292                             h->h264dsp.h264_idct_add16(dest_y, block_offset, h->mb, linesize, h->non_zero_count_cache);
1293                         }
1294                     }
1295                 }
1296             }else{
1297                 for(i=0; i<16; i++){
1298                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
1299                         uint8_t * const ptr= dest_y + block_offset[i];
1300                         ff_svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
1301                     }
1302                 }
1303             }
1304         }
1305
1306         if((simple || !CONFIG_GRAY || !(s->flags&CODEC_FLAG_GRAY)) && (h->cbp&0x30)){
1307             uint8_t *dest[2] = {dest_cb, dest_cr};
1308             if(transform_bypass){
1309                 if(IS_INTRA(mb_type) && h->sps.profile_idc==244 && (h->chroma_pred_mode==VERT_PRED8x8 || h->chroma_pred_mode==HOR_PRED8x8)){
1310                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[0], block_offset + 16, h->mb + 16*16, uvlinesize);
1311                     h->hpc.pred8x8_add[h->chroma_pred_mode](dest[1], block_offset + 20, h->mb + 20*16, uvlinesize);
1312                 }else{
1313                     idct_add = s->dsp.add_pixels4;
1314                     for(i=16; i<16+8; i++){
1315                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16])
1316                             idct_add   (dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
1317                     }
1318                 }
1319             }else{
1320                 chroma_dc_dequant_idct_c(h->mb + 16*16, h->chroma_qp[0], h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
1321                 chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->chroma_qp[1], h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
1322                 if(is_h264){
1323                     h->h264dsp.h264_idct_add8(dest, block_offset,
1324                                               h->mb, uvlinesize,
1325                                               h->non_zero_count_cache);
1326                 }else{
1327                     for(i=16; i<16+8; i++){
1328                         if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
1329                             uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
1330                             ff_svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, ff_h264_chroma_qp[s->qscale + 12] - 12, 2);
1331                         }
1332                     }
1333                 }
1334             }
1335         }
1336     }
1337     if(h->cbp || IS_INTRA(mb_type))
1338         s->dsp.clear_blocks(h->mb);
1339 }
1340
1341 /**
1342  * Process a macroblock; this case avoids checks for expensive uncommon cases.
1343  */
1344 static void hl_decode_mb_simple(H264Context *h){
1345     hl_decode_mb_internal(h, 1);
1346 }
1347
1348 /**
1349  * Process a macroblock; this handles edge cases, such as interlacing.
1350  */
1351 static void av_noinline hl_decode_mb_complex(H264Context *h){
1352     hl_decode_mb_internal(h, 0);
1353 }
1354
1355 void ff_h264_hl_decode_mb(H264Context *h){
1356     MpegEncContext * const s = &h->s;
1357     const int mb_xy= h->mb_xy;
1358     const int mb_type= s->current_picture.mb_type[mb_xy];
1359     int is_complex = CONFIG_SMALL || h->is_complex || IS_INTRA_PCM(mb_type) || s->qscale == 0;
1360
1361     if (is_complex)
1362         hl_decode_mb_complex(h);
1363     else hl_decode_mb_simple(h);
1364 }
1365
1366 static int pred_weight_table(H264Context *h){
1367     MpegEncContext * const s = &h->s;
1368     int list, i;
1369     int luma_def, chroma_def;
1370
1371     h->use_weight= 0;
1372     h->use_weight_chroma= 0;
1373     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
1374     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
1375     luma_def = 1<<h->luma_log2_weight_denom;
1376     chroma_def = 1<<h->chroma_log2_weight_denom;
1377
1378     for(list=0; list<2; list++){
1379         h->luma_weight_flag[list]   = 0;
1380         h->chroma_weight_flag[list] = 0;
1381         for(i=0; i<h->ref_count[list]; i++){
1382             int luma_weight_flag, chroma_weight_flag;
1383
1384             luma_weight_flag= get_bits1(&s->gb);
1385             if(luma_weight_flag){
1386                 h->luma_weight[i][list][0]= get_se_golomb(&s->gb);
1387                 h->luma_weight[i][list][1]= get_se_golomb(&s->gb);
1388                 if(   h->luma_weight[i][list][0] != luma_def
1389                    || h->luma_weight[i][list][1] != 0) {
1390                     h->use_weight= 1;
1391                     h->luma_weight_flag[list]= 1;
1392                 }
1393             }else{
1394                 h->luma_weight[i][list][0]= luma_def;
1395                 h->luma_weight[i][list][1]= 0;
1396             }
1397
1398             if(CHROMA){
1399                 chroma_weight_flag= get_bits1(&s->gb);
1400                 if(chroma_weight_flag){
1401                     int j;
1402                     for(j=0; j<2; j++){
1403                         h->chroma_weight[i][list][j][0]= get_se_golomb(&s->gb);
1404                         h->chroma_weight[i][list][j][1]= get_se_golomb(&s->gb);
1405                         if(   h->chroma_weight[i][list][j][0] != chroma_def
1406                            || h->chroma_weight[i][list][j][1] != 0) {
1407                             h->use_weight_chroma= 1;
1408                             h->chroma_weight_flag[list]= 1;
1409                         }
1410                     }
1411                 }else{
1412                     int j;
1413                     for(j=0; j<2; j++){
1414                         h->chroma_weight[i][list][j][0]= chroma_def;
1415                         h->chroma_weight[i][list][j][1]= 0;
1416                     }
1417                 }
1418             }
1419         }
1420         if(h->slice_type_nos != FF_B_TYPE) break;
1421     }
1422     h->use_weight= h->use_weight || h->use_weight_chroma;
1423     return 0;
1424 }
1425
1426 /**
1427  * Initialize implicit_weight table.
1428  * @param field  0/1 initialize the weight for interlaced MBAFF
1429  *                -1 initializes the rest
1430  */
1431 static void implicit_weight_table(H264Context *h, int field){
1432     MpegEncContext * const s = &h->s;
1433     int ref0, ref1, i, cur_poc, ref_start, ref_count0, ref_count1;
1434
1435     for (i = 0; i < 2; i++) {
1436         h->luma_weight_flag[i]   = 0;
1437         h->chroma_weight_flag[i] = 0;
1438     }
1439
1440     if(field < 0){
1441         cur_poc = s->current_picture_ptr->poc;
1442     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1 && !FRAME_MBAFF
1443        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
1444         h->use_weight= 0;
1445         h->use_weight_chroma= 0;
1446         return;
1447     }
1448         ref_start= 0;
1449         ref_count0= h->ref_count[0];
1450         ref_count1= h->ref_count[1];
1451     }else{
1452         cur_poc = s->current_picture_ptr->field_poc[field];
1453         ref_start= 16;
1454         ref_count0= 16+2*h->ref_count[0];
1455         ref_count1= 16+2*h->ref_count[1];
1456     }
1457
1458     h->use_weight= 2;
1459     h->use_weight_chroma= 2;
1460     h->luma_log2_weight_denom= 5;
1461     h->chroma_log2_weight_denom= 5;
1462
1463     for(ref0=ref_start; ref0 < ref_count0; ref0++){
1464         int poc0 = h->ref_list[0][ref0].poc;
1465         for(ref1=ref_start; ref1 < ref_count1; ref1++){
1466             int poc1 = h->ref_list[1][ref1].poc;
1467             int td = av_clip(poc1 - poc0, -128, 127);
1468             int w= 32;
1469             if(td){
1470                 int tb = av_clip(cur_poc - poc0, -128, 127);
1471                 int tx = (16384 + (FFABS(td) >> 1)) / td;
1472                 int dist_scale_factor = (tb*tx + 32) >> 8;
1473                 if(dist_scale_factor >= -64 && dist_scale_factor <= 128)
1474                     w = 64 - dist_scale_factor;
1475             }
1476             if(field<0){
1477                 h->implicit_weight[ref0][ref1][0]=
1478                 h->implicit_weight[ref0][ref1][1]= w;
1479             }else{
1480                 h->implicit_weight[ref0][ref1][field]=w;
1481             }
1482         }
1483     }
1484 }
1485
1486 /**
1487  * instantaneous decoder refresh.
1488  */
1489 static void idr(H264Context *h){
1490     ff_h264_remove_all_refs(h);
1491     h->prev_frame_num= 0;
1492     h->prev_frame_num_offset= 0;
1493     h->prev_poc_msb=
1494     h->prev_poc_lsb= 0;
1495 }
1496
1497 /* forget old pics after a seek */
1498 static void flush_dpb(AVCodecContext *avctx){
1499     H264Context *h= avctx->priv_data;
1500     int i;
1501     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
1502         if(h->delayed_pic[i])
1503             h->delayed_pic[i]->reference= 0;
1504         h->delayed_pic[i]= NULL;
1505     }
1506     h->outputed_poc= INT_MIN;
1507     h->prev_interlaced_frame = 1;
1508     idr(h);
1509     if(h->s.current_picture_ptr)
1510         h->s.current_picture_ptr->reference= 0;
1511     h->s.first_field= 0;
1512     ff_h264_reset_sei(h);
1513     ff_mpeg_flush(avctx);
1514 }
1515
1516 static int init_poc(H264Context *h){
1517     MpegEncContext * const s = &h->s;
1518     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
1519     int field_poc[2];
1520     Picture *cur = s->current_picture_ptr;
1521
1522     h->frame_num_offset= h->prev_frame_num_offset;
1523     if(h->frame_num < h->prev_frame_num)
1524         h->frame_num_offset += max_frame_num;
1525
1526     if(h->sps.poc_type==0){
1527         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
1528
1529         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
1530             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
1531         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
1532             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
1533         else
1534             h->poc_msb = h->prev_poc_msb;
1535 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
1536         field_poc[0] =
1537         field_poc[1] = h->poc_msb + h->poc_lsb;
1538         if(s->picture_structure == PICT_FRAME)
1539             field_poc[1] += h->delta_poc_bottom;
1540     }else if(h->sps.poc_type==1){
1541         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
1542         int i;
1543
1544         if(h->sps.poc_cycle_length != 0)
1545             abs_frame_num = h->frame_num_offset + h->frame_num;
1546         else
1547             abs_frame_num = 0;
1548
1549         if(h->nal_ref_idc==0 && abs_frame_num > 0)
1550             abs_frame_num--;
1551
1552         expected_delta_per_poc_cycle = 0;
1553         for(i=0; i < h->sps.poc_cycle_length; i++)
1554             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
1555
1556         if(abs_frame_num > 0){
1557             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
1558             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
1559
1560             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
1561             for(i = 0; i <= frame_num_in_poc_cycle; i++)
1562                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
1563         } else
1564             expectedpoc = 0;
1565
1566         if(h->nal_ref_idc == 0)
1567             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
1568
1569         field_poc[0] = expectedpoc + h->delta_poc[0];
1570         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
1571
1572         if(s->picture_structure == PICT_FRAME)
1573             field_poc[1] += h->delta_poc[1];
1574     }else{
1575         int poc= 2*(h->frame_num_offset + h->frame_num);
1576
1577         if(!h->nal_ref_idc)
1578             poc--;
1579
1580         field_poc[0]= poc;
1581         field_poc[1]= poc;
1582     }
1583
1584     if(s->picture_structure != PICT_BOTTOM_FIELD)
1585         s->current_picture_ptr->field_poc[0]= field_poc[0];
1586     if(s->picture_structure != PICT_TOP_FIELD)
1587         s->current_picture_ptr->field_poc[1]= field_poc[1];
1588     cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
1589
1590     return 0;
1591 }
1592
1593
1594 /**
1595  * initialize scan tables
1596  */
1597 static void init_scan_tables(H264Context *h){
1598     int i;
1599     if(h->h264dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
1600         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
1601         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
1602     }else{
1603         for(i=0; i<16; i++){
1604 #define T(x) (x>>2) | ((x<<2) & 0xF)
1605             h->zigzag_scan[i] = T(zigzag_scan[i]);
1606             h-> field_scan[i] = T( field_scan[i]);
1607 #undef T
1608         }
1609     }
1610     if(h->h264dsp.h264_idct8_add == ff_h264_idct8_add_c){
1611         memcpy(h->zigzag_scan8x8,       ff_zigzag_direct,     64*sizeof(uint8_t));
1612         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
1613         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
1614         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
1615     }else{
1616         for(i=0; i<64; i++){
1617 #define T(x) (x>>3) | ((x&7)<<3)
1618             h->zigzag_scan8x8[i]       = T(ff_zigzag_direct[i]);
1619             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
1620             h->field_scan8x8[i]        = T(field_scan8x8[i]);
1621             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
1622 #undef T
1623         }
1624     }
1625     if(h->sps.transform_bypass){ //FIXME same ugly
1626         h->zigzag_scan_q0          = zigzag_scan;
1627         h->zigzag_scan8x8_q0       = ff_zigzag_direct;
1628         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
1629         h->field_scan_q0           = field_scan;
1630         h->field_scan8x8_q0        = field_scan8x8;
1631         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
1632     }else{
1633         h->zigzag_scan_q0          = h->zigzag_scan;
1634         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
1635         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
1636         h->field_scan_q0           = h->field_scan;
1637         h->field_scan8x8_q0        = h->field_scan8x8;
1638         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
1639     }
1640 }
1641
1642 static void field_end(H264Context *h){
1643     MpegEncContext * const s = &h->s;
1644     AVCodecContext * const avctx= s->avctx;
1645     s->mb_y= 0;
1646
1647     s->current_picture_ptr->qscale_type= FF_QSCALE_TYPE_H264;
1648     s->current_picture_ptr->pict_type= s->pict_type;
1649
1650     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1651         ff_vdpau_h264_set_reference_frames(s);
1652
1653     if(!s->dropable) {
1654         ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1655         h->prev_poc_msb= h->poc_msb;
1656         h->prev_poc_lsb= h->poc_lsb;
1657     }
1658     h->prev_frame_num_offset= h->frame_num_offset;
1659     h->prev_frame_num= h->frame_num;
1660
1661     if (avctx->hwaccel) {
1662         if (avctx->hwaccel->end_frame(avctx) < 0)
1663             av_log(avctx, AV_LOG_ERROR, "hardware accelerator failed to decode picture\n");
1664     }
1665
1666     if (CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
1667         ff_vdpau_h264_picture_complete(s);
1668
1669     /*
1670      * FIXME: Error handling code does not seem to support interlaced
1671      * when slices span multiple rows
1672      * The ff_er_add_slice calls don't work right for bottom
1673      * fields; they cause massive erroneous error concealing
1674      * Error marking covers both fields (top and bottom).
1675      * This causes a mismatched s->error_count
1676      * and a bad error table. Further, the error count goes to
1677      * INT_MAX when called for bottom field, because mb_y is
1678      * past end by one (callers fault) and resync_mb_y != 0
1679      * causes problems for the first MB line, too.
1680      */
1681     if (!FIELD_PICTURE)
1682         ff_er_frame_end(s);
1683
1684     MPV_frame_end(s);
1685
1686     h->current_slice=0;
1687 }
1688
1689 /**
1690  * Replicate H264 "master" context to thread contexts.
1691  */
1692 static void clone_slice(H264Context *dst, H264Context *src)
1693 {
1694     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
1695     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
1696     dst->s.current_picture      = src->s.current_picture;
1697     dst->s.linesize             = src->s.linesize;
1698     dst->s.uvlinesize           = src->s.uvlinesize;
1699     dst->s.first_field          = src->s.first_field;
1700
1701     dst->prev_poc_msb           = src->prev_poc_msb;
1702     dst->prev_poc_lsb           = src->prev_poc_lsb;
1703     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
1704     dst->prev_frame_num         = src->prev_frame_num;
1705     dst->short_ref_count        = src->short_ref_count;
1706
1707     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
1708     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
1709     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
1710     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
1711
1712     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
1713     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
1714 }
1715
1716 /**
1717  * decodes a slice header.
1718  * This will also call MPV_common_init() and frame_start() as needed.
1719  *
1720  * @param h h264context
1721  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
1722  *
1723  * @return 0 if okay, <0 if an error occurred, 1 if decoding must not be multithreaded
1724  */
1725 static int decode_slice_header(H264Context *h, H264Context *h0){
1726     MpegEncContext * const s = &h->s;
1727     MpegEncContext * const s0 = &h0->s;
1728     unsigned int first_mb_in_slice;
1729     unsigned int pps_id;
1730     int num_ref_idx_active_override_flag;
1731     unsigned int slice_type, tmp, i, j;
1732     int default_ref_list_done = 0;
1733     int last_pic_structure;
1734
1735     s->dropable= h->nal_ref_idc == 0;
1736
1737     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
1738         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
1739         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
1740     }else{
1741         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
1742         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
1743     }
1744
1745     first_mb_in_slice= get_ue_golomb(&s->gb);
1746
1747     if(first_mb_in_slice == 0){ //FIXME better field boundary detection
1748         if(h0->current_slice && FIELD_PICTURE){
1749             field_end(h);
1750         }
1751
1752         h0->current_slice = 0;
1753         if (!s0->first_field)
1754             s->current_picture_ptr= NULL;
1755     }
1756
1757     slice_type= get_ue_golomb_31(&s->gb);
1758     if(slice_type > 9){
1759         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);
1760         return -1;
1761     }
1762     if(slice_type > 4){
1763         slice_type -= 5;
1764         h->slice_type_fixed=1;
1765     }else
1766         h->slice_type_fixed=0;
1767
1768     slice_type= golomb_to_pict_type[ slice_type ];
1769     if (slice_type == FF_I_TYPE
1770         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
1771         default_ref_list_done = 1;
1772     }
1773     h->slice_type= slice_type;
1774     h->slice_type_nos= slice_type & 3;
1775
1776     s->pict_type= h->slice_type; // to make a few old functions happy, it's wrong though
1777
1778     pps_id= get_ue_golomb(&s->gb);
1779     if(pps_id>=MAX_PPS_COUNT){
1780         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
1781         return -1;
1782     }
1783     if(!h0->pps_buffers[pps_id]) {
1784         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing PPS %u referenced\n", pps_id);
1785         return -1;
1786     }
1787     h->pps= *h0->pps_buffers[pps_id];
1788
1789     if(!h0->sps_buffers[h->pps.sps_id]) {
1790         av_log(h->s.avctx, AV_LOG_ERROR, "non-existing SPS %u referenced\n", h->pps.sps_id);
1791         return -1;
1792     }
1793     h->sps = *h0->sps_buffers[h->pps.sps_id];
1794
1795     s->avctx->profile = h->sps.profile_idc;
1796     s->avctx->level   = h->sps.level_idc;
1797     s->avctx->refs    = h->sps.ref_frame_count;
1798
1799     if(h == h0 && h->dequant_coeff_pps != pps_id){
1800         h->dequant_coeff_pps = pps_id;
1801         init_dequant_tables(h);
1802     }
1803
1804     s->mb_width= h->sps.mb_width;
1805     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
1806
1807     h->b_stride=  s->mb_width*4;
1808
1809     s->width = 16*s->mb_width - 2*FFMIN(h->sps.crop_right, 7);
1810     if(h->sps.frame_mbs_only_flag)
1811         s->height= 16*s->mb_height - 2*FFMIN(h->sps.crop_bottom, 7);
1812     else
1813         s->height= 16*s->mb_height - 4*FFMIN(h->sps.crop_bottom, 3);
1814
1815     if (s->context_initialized
1816         && (   s->width != s->avctx->width || s->height != s->avctx->height
1817             || av_cmp_q(h->sps.sar, s->avctx->sample_aspect_ratio))) {
1818         if(h != h0)
1819             return -1;   // width / height changed during parallelized decoding
1820         free_tables(h);
1821         flush_dpb(s->avctx);
1822         MPV_common_end(s);
1823     }
1824     if (!s->context_initialized) {
1825         if(h != h0)
1826             return -1;  // we cant (re-)initialize context during parallel decoding
1827
1828         avcodec_set_dimensions(s->avctx, s->width, s->height);
1829         s->avctx->sample_aspect_ratio= h->sps.sar;
1830         av_assert0(s->avctx->sample_aspect_ratio.den);
1831
1832         if(h->sps.video_signal_type_present_flag){
1833             s->avctx->color_range = h->sps.full_range ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;
1834             if(h->sps.colour_description_present_flag){
1835                 s->avctx->color_primaries = h->sps.color_primaries;
1836                 s->avctx->color_trc       = h->sps.color_trc;
1837                 s->avctx->colorspace      = h->sps.colorspace;
1838             }
1839         }
1840
1841         if(h->sps.timing_info_present_flag){
1842             int64_t den= h->sps.time_scale;
1843             if(h->x264_build < 44U)
1844                 den *= 2;
1845             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
1846                       h->sps.num_units_in_tick, den, 1<<30);
1847         }
1848         s->avctx->pix_fmt = s->avctx->get_format(s->avctx,
1849                                                  s->avctx->codec->pix_fmts ?
1850                                                  s->avctx->codec->pix_fmts :
1851                                                  s->avctx->color_range == AVCOL_RANGE_JPEG ?
1852                                                  hwaccel_pixfmt_list_h264_jpeg_420 :
1853                                                  ff_hwaccel_pixfmt_list_420);
1854         s->avctx->hwaccel = ff_find_hwaccel(s->avctx->codec->id, s->avctx->pix_fmt);
1855
1856         if (MPV_common_init(s) < 0)
1857             return -1;
1858         s->first_field = 0;
1859         h->prev_interlaced_frame = 1;
1860
1861         init_scan_tables(h);
1862         ff_h264_alloc_tables(h);
1863
1864         for(i = 1; i < s->avctx->thread_count; i++) {
1865             H264Context *c;
1866             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
1867             memcpy(c, h->s.thread_context[i], sizeof(MpegEncContext));
1868             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
1869             c->h264dsp = h->h264dsp;
1870             c->sps = h->sps;
1871             c->pps = h->pps;
1872             init_scan_tables(c);
1873             clone_tables(c, h, i);
1874         }
1875
1876         for(i = 0; i < s->avctx->thread_count; i++)
1877             if(context_init(h->thread_context[i]) < 0)
1878                 return -1;
1879     }
1880
1881     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
1882
1883     h->mb_mbaff = 0;
1884     h->mb_aff_frame = 0;
1885     last_pic_structure = s0->picture_structure;
1886     if(h->sps.frame_mbs_only_flag){
1887         s->picture_structure= PICT_FRAME;
1888     }else{
1889         if(get_bits1(&s->gb)) { //field_pic_flag
1890             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
1891         } else {
1892             s->picture_structure= PICT_FRAME;
1893             h->mb_aff_frame = h->sps.mb_aff;
1894         }
1895     }
1896     h->mb_field_decoding_flag= s->picture_structure != PICT_FRAME;
1897
1898     if(h0->current_slice == 0){
1899         while(h->frame_num !=  h->prev_frame_num &&
1900               h->frame_num != (h->prev_frame_num+1)%(1<<h->sps.log2_max_frame_num)){
1901             Picture *prev = h->short_ref_count ? h->short_ref[0] : NULL;
1902             av_log(h->s.avctx, AV_LOG_DEBUG, "Frame num gap %d %d\n", h->frame_num, h->prev_frame_num);
1903             if (ff_h264_frame_start(h) < 0)
1904                 return -1;
1905             h->prev_frame_num++;
1906             h->prev_frame_num %= 1<<h->sps.log2_max_frame_num;
1907             s->current_picture_ptr->frame_num= h->prev_frame_num;
1908             ff_generate_sliding_window_mmcos(h);
1909             ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
1910             /* Error concealment: if a ref is missing, copy the previous ref in its place.
1911              * FIXME: avoiding a memcpy would be nice, but ref handling makes many assumptions
1912              * about there being no actual duplicates.
1913              * FIXME: this doesn't copy padding for out-of-frame motion vectors.  Given we're
1914              * concealing a lost frame, this probably isn't noticable by comparison, but it should
1915              * be fixed. */
1916             if (h->short_ref_count) {
1917                 if (prev) {
1918                     av_image_copy(h->short_ref[0]->data, h->short_ref[0]->linesize,
1919                                   (const uint8_t**)prev->data, prev->linesize,
1920                                   s->avctx->pix_fmt, s->mb_width*16, s->mb_height*16);
1921                     h->short_ref[0]->poc = prev->poc+2;
1922                 }
1923                 h->short_ref[0]->frame_num = h->prev_frame_num;
1924             }
1925         }
1926
1927         /* See if we have a decoded first field looking for a pair... */
1928         if (s0->first_field) {
1929             assert(s0->current_picture_ptr);
1930             assert(s0->current_picture_ptr->data[0]);
1931             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
1932
1933             /* figure out if we have a complementary field pair */
1934             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
1935                 /*
1936                  * Previous field is unmatched. Don't display it, but let it
1937                  * remain for reference if marked as such.
1938                  */
1939                 s0->current_picture_ptr = NULL;
1940                 s0->first_field = FIELD_PICTURE;
1941
1942             } else {
1943                 if (h->nal_ref_idc &&
1944                         s0->current_picture_ptr->reference &&
1945                         s0->current_picture_ptr->frame_num != h->frame_num) {
1946                     /*
1947                      * This and previous field were reference, but had
1948                      * different frame_nums. Consider this field first in
1949                      * pair. Throw away previous field except for reference
1950                      * purposes.
1951                      */
1952                     s0->first_field = 1;
1953                     s0->current_picture_ptr = NULL;
1954
1955                 } else {
1956                     /* Second field in complementary pair */
1957                     s0->first_field = 0;
1958                 }
1959             }
1960
1961         } else {
1962             /* Frame or first field in a potentially complementary pair */
1963             assert(!s0->current_picture_ptr);
1964             s0->first_field = FIELD_PICTURE;
1965         }
1966
1967         if((!FIELD_PICTURE || s0->first_field) && ff_h264_frame_start(h) < 0) {
1968             s0->first_field = 0;
1969             return -1;
1970         }
1971     }
1972     if(h != h0)
1973         clone_slice(h, h0);
1974
1975     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
1976
1977     assert(s->mb_num == s->mb_width * s->mb_height);
1978     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
1979        first_mb_in_slice                    >= s->mb_num){
1980         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
1981         return -1;
1982     }
1983     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
1984     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
1985     if (s->picture_structure == PICT_BOTTOM_FIELD)
1986         s->resync_mb_y = s->mb_y = s->mb_y + 1;
1987     assert(s->mb_y < s->mb_height);
1988
1989     if(s->picture_structure==PICT_FRAME){
1990         h->curr_pic_num=   h->frame_num;
1991         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
1992     }else{
1993         h->curr_pic_num= 2*h->frame_num + 1;
1994         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
1995     }
1996
1997     if(h->nal_unit_type == NAL_IDR_SLICE){
1998         get_ue_golomb(&s->gb); /* idr_pic_id */
1999     }
2000
2001     if(h->sps.poc_type==0){
2002         h->poc_lsb= get_bits(&s->gb, h->sps.log2_max_poc_lsb);
2003
2004         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME){
2005             h->delta_poc_bottom= get_se_golomb(&s->gb);
2006         }
2007     }
2008
2009     if(h->sps.poc_type==1 && !h->sps.delta_pic_order_always_zero_flag){
2010         h->delta_poc[0]= get_se_golomb(&s->gb);
2011
2012         if(h->pps.pic_order_present==1 && s->picture_structure==PICT_FRAME)
2013             h->delta_poc[1]= get_se_golomb(&s->gb);
2014     }
2015
2016     init_poc(h);
2017
2018     if(h->pps.redundant_pic_cnt_present){
2019         h->redundant_pic_count= get_ue_golomb(&s->gb);
2020     }
2021
2022     //set defaults, might be overridden a few lines later
2023     h->ref_count[0]= h->pps.ref_count[0];
2024     h->ref_count[1]= h->pps.ref_count[1];
2025
2026     if(h->slice_type_nos != FF_I_TYPE){
2027         if(h->slice_type_nos == FF_B_TYPE){
2028             h->direct_spatial_mv_pred= get_bits1(&s->gb);
2029         }
2030         num_ref_idx_active_override_flag= get_bits1(&s->gb);
2031
2032         if(num_ref_idx_active_override_flag){
2033             h->ref_count[0]= get_ue_golomb(&s->gb) + 1;
2034             if(h->slice_type_nos==FF_B_TYPE)
2035                 h->ref_count[1]= get_ue_golomb(&s->gb) + 1;
2036
2037             if(h->ref_count[0]-1 > 32-1 || h->ref_count[1]-1 > 32-1){
2038                 av_log(h->s.avctx, AV_LOG_ERROR, "reference overflow\n");
2039                 h->ref_count[0]= h->ref_count[1]= 1;
2040                 return -1;
2041             }
2042         }
2043         if(h->slice_type_nos == FF_B_TYPE)
2044             h->list_count= 2;
2045         else
2046             h->list_count= 1;
2047     }else
2048         h->list_count= 0;
2049
2050     if(!default_ref_list_done){
2051         ff_h264_fill_default_ref_list(h);
2052     }
2053
2054     if(h->slice_type_nos!=FF_I_TYPE && ff_h264_decode_ref_pic_list_reordering(h) < 0)
2055         return -1;
2056
2057     if(h->slice_type_nos!=FF_I_TYPE){
2058         s->last_picture_ptr= &h->ref_list[0][0];
2059         ff_copy_picture(&s->last_picture, s->last_picture_ptr);
2060     }
2061     if(h->slice_type_nos==FF_B_TYPE){
2062         s->next_picture_ptr= &h->ref_list[1][0];
2063         ff_copy_picture(&s->next_picture, s->next_picture_ptr);
2064     }
2065
2066     if(   (h->pps.weighted_pred          && h->slice_type_nos == FF_P_TYPE )
2067        ||  (h->pps.weighted_bipred_idc==1 && h->slice_type_nos== FF_B_TYPE ) )
2068         pred_weight_table(h);
2069     else if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE){
2070         implicit_weight_table(h, -1);
2071     }else {
2072         h->use_weight = 0;
2073         for (i = 0; i < 2; i++) {
2074             h->luma_weight_flag[i]   = 0;
2075             h->chroma_weight_flag[i] = 0;
2076         }
2077     }
2078
2079     if(h->nal_ref_idc)
2080         ff_h264_decode_ref_pic_marking(h0, &s->gb);
2081
2082     if(FRAME_MBAFF){
2083         ff_h264_fill_mbaff_ref_list(h);
2084
2085         if(h->pps.weighted_bipred_idc==2 && h->slice_type_nos== FF_B_TYPE){
2086             implicit_weight_table(h, 0);
2087             implicit_weight_table(h, 1);
2088         }
2089     }
2090
2091     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
2092         ff_h264_direct_dist_scale_factor(h);
2093     ff_h264_direct_ref_list_init(h);
2094
2095     if( h->slice_type_nos != FF_I_TYPE && h->pps.cabac ){
2096         tmp = get_ue_golomb_31(&s->gb);
2097         if(tmp > 2){
2098             av_log(s->avctx, AV_LOG_ERROR, "cabac_init_idc overflow\n");
2099             return -1;
2100         }
2101         h->cabac_init_idc= tmp;
2102     }
2103
2104     h->last_qscale_diff = 0;
2105     tmp = h->pps.init_qp + get_se_golomb(&s->gb);
2106     if(tmp>51){
2107         av_log(s->avctx, AV_LOG_ERROR, "QP %u out of range\n", tmp);
2108         return -1;
2109     }
2110     s->qscale= tmp;
2111     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2112     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2113     //FIXME qscale / qp ... stuff
2114     if(h->slice_type == FF_SP_TYPE){
2115         get_bits1(&s->gb); /* sp_for_switch_flag */
2116     }
2117     if(h->slice_type==FF_SP_TYPE || h->slice_type == FF_SI_TYPE){
2118         get_se_golomb(&s->gb); /* slice_qs_delta */
2119     }
2120
2121     h->deblocking_filter = 1;
2122     h->slice_alpha_c0_offset = 52;
2123     h->slice_beta_offset = 52;
2124     if( h->pps.deblocking_filter_parameters_present ) {
2125         tmp= get_ue_golomb_31(&s->gb);
2126         if(tmp > 2){
2127             av_log(s->avctx, AV_LOG_ERROR, "deblocking_filter_idc %u out of range\n", tmp);
2128             return -1;
2129         }
2130         h->deblocking_filter= tmp;
2131         if(h->deblocking_filter < 2)
2132             h->deblocking_filter^= 1; // 1<->0
2133
2134         if( h->deblocking_filter ) {
2135             h->slice_alpha_c0_offset += get_se_golomb(&s->gb) << 1;
2136             h->slice_beta_offset     += get_se_golomb(&s->gb) << 1;
2137             if(   h->slice_alpha_c0_offset > 104U
2138                || h->slice_beta_offset     > 104U){
2139                 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);
2140                 return -1;
2141             }
2142         }
2143     }
2144
2145     if(   s->avctx->skip_loop_filter >= AVDISCARD_ALL
2146        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONKEY && h->slice_type_nos != FF_I_TYPE)
2147        ||(s->avctx->skip_loop_filter >= AVDISCARD_BIDIR  && h->slice_type_nos == FF_B_TYPE)
2148        ||(s->avctx->skip_loop_filter >= AVDISCARD_NONREF && h->nal_ref_idc == 0))
2149         h->deblocking_filter= 0;
2150
2151     if(h->deblocking_filter == 1 && h0->max_contexts > 1) {
2152         if(s->avctx->flags2 & CODEC_FLAG2_FAST) {
2153             /* Cheat slightly for speed:
2154                Do not bother to deblock across slices. */
2155             h->deblocking_filter = 2;
2156         } else {
2157             h0->max_contexts = 1;
2158             if(!h0->single_decode_warning) {
2159                 av_log(s->avctx, AV_LOG_INFO, "Cannot parallelize deblocking type 1, decoding such frames in sequential order\n");
2160                 h0->single_decode_warning = 1;
2161             }
2162             if(h != h0)
2163                 return 1; // deblocking switched inside frame
2164         }
2165     }
2166     h->qp_thresh= 15 + 52 - FFMIN(h->slice_alpha_c0_offset, h->slice_beta_offset) - FFMAX3(0, h->pps.chroma_qp_index_offset[0], h->pps.chroma_qp_index_offset[1]);
2167
2168 #if 0 //FMO
2169     if( h->pps.num_slice_groups > 1  && h->pps.mb_slice_group_map_type >= 3 && h->pps.mb_slice_group_map_type <= 5)
2170         slice_group_change_cycle= get_bits(&s->gb, ?);
2171 #endif
2172
2173     h0->last_slice_type = slice_type;
2174     h->slice_num = ++h0->current_slice;
2175     if(h->slice_num >= MAX_SLICES){
2176         av_log(s->avctx, AV_LOG_ERROR, "Too many slices, increase MAX_SLICES and recompile\n");
2177     }
2178
2179     for(j=0; j<2; j++){
2180         int id_list[16];
2181         int *ref2frm= h->ref2frm[h->slice_num&(MAX_SLICES-1)][j];
2182         for(i=0; i<16; i++){
2183             id_list[i]= 60;
2184             if(h->ref_list[j][i].data[0]){
2185                 int k;
2186                 uint8_t *base= h->ref_list[j][i].base[0];
2187                 for(k=0; k<h->short_ref_count; k++)
2188                     if(h->short_ref[k]->base[0] == base){
2189                         id_list[i]= k;
2190                         break;
2191                     }
2192                 for(k=0; k<h->long_ref_count; k++)
2193                     if(h->long_ref[k] && h->long_ref[k]->base[0] == base){
2194                         id_list[i]= h->short_ref_count + k;
2195                         break;
2196                     }
2197             }
2198         }
2199
2200         ref2frm[0]=
2201         ref2frm[1]= -1;
2202         for(i=0; i<16; i++)
2203             ref2frm[i+2]= 4*id_list[i]
2204                           +(h->ref_list[j][i].reference&3);
2205         ref2frm[18+0]=
2206         ref2frm[18+1]= -1;
2207         for(i=16; i<48; i++)
2208             ref2frm[i+4]= 4*id_list[(i-16)>>1]
2209                           +(h->ref_list[j][i].reference&3);
2210     }
2211
2212     h->emu_edge_width= (s->flags&CODEC_FLAG_EMU_EDGE) ? 0 : 16;
2213     h->emu_edge_height= (FRAME_MBAFF || FIELD_PICTURE) ? 0 : h->emu_edge_width;
2214
2215     if(s->avctx->debug&FF_DEBUG_PICT_INFO){
2216         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",
2217                h->slice_num,
2218                (s->picture_structure==PICT_FRAME ? "F" : s->picture_structure==PICT_TOP_FIELD ? "T" : "B"),
2219                first_mb_in_slice,
2220                av_get_pict_type_char(h->slice_type), h->slice_type_fixed ? " fix" : "", h->nal_unit_type == NAL_IDR_SLICE ? " IDR" : "",
2221                pps_id, h->frame_num,
2222                s->current_picture_ptr->field_poc[0], s->current_picture_ptr->field_poc[1],
2223                h->ref_count[0], h->ref_count[1],
2224                s->qscale,
2225                h->deblocking_filter, h->slice_alpha_c0_offset/2-26, h->slice_beta_offset/2-26,
2226                h->use_weight,
2227                h->use_weight==1 && h->use_weight_chroma ? "c" : "",
2228                h->slice_type == FF_B_TYPE ? (h->direct_spatial_mv_pred ? "SPAT" : "TEMP") : ""
2229                );
2230     }
2231
2232     return 0;
2233 }
2234
2235 int ff_h264_get_slice_type(const H264Context *h)
2236 {
2237     switch (h->slice_type) {
2238     case FF_P_TYPE:  return 0;
2239     case FF_B_TYPE:  return 1;
2240     case FF_I_TYPE:  return 2;
2241     case FF_SP_TYPE: return 3;
2242     case FF_SI_TYPE: return 4;
2243     default:         return -1;
2244     }
2245 }
2246
2247 /**
2248  *
2249  * @return non zero if the loop filter can be skiped
2250  */
2251 static int fill_filter_caches(H264Context *h, int mb_type){
2252     MpegEncContext * const s = &h->s;
2253     const int mb_xy= h->mb_xy;
2254     int top_xy, left_xy[2];
2255     int top_type, left_type[2];
2256
2257     top_xy     = mb_xy  - (s->mb_stride << MB_FIELD);
2258
2259     //FIXME deblocking could skip the intra and nnz parts.
2260
2261     /* Wow, what a mess, why didn't they simplify the interlacing & intra
2262      * stuff, I can't imagine that these complex rules are worth it. */
2263
2264     left_xy[1] = left_xy[0] = mb_xy-1;
2265     if(FRAME_MBAFF){
2266         const int left_mb_field_flag     = IS_INTERLACED(s->current_picture.mb_type[mb_xy-1]);
2267         const int curr_mb_field_flag     = IS_INTERLACED(mb_type);
2268         if(s->mb_y&1){
2269             if (left_mb_field_flag != curr_mb_field_flag) {
2270                 left_xy[0] -= s->mb_stride;
2271             }
2272         }else{
2273             if(curr_mb_field_flag){
2274                 top_xy      += s->mb_stride & (((s->current_picture.mb_type[top_xy    ]>>7)&1)-1);
2275             }
2276             if (left_mb_field_flag != curr_mb_field_flag) {
2277                 left_xy[1] += s->mb_stride;
2278             }
2279         }
2280     }
2281
2282     h->top_mb_xy = top_xy;
2283     h->left_mb_xy[0] = left_xy[0];
2284     h->left_mb_xy[1] = left_xy[1];
2285     {
2286         //for sufficiently low qp, filtering wouldn't do anything
2287         //this is a conservative estimate: could also check beta_offset and more accurate chroma_qp
2288         int qp_thresh = h->qp_thresh; //FIXME strictly we should store qp_thresh for each mb of a slice
2289         int qp = s->current_picture.qscale_table[mb_xy];
2290         if(qp <= qp_thresh
2291            && (left_xy[0]<0 || ((qp + s->current_picture.qscale_table[left_xy[0]] + 1)>>1) <= qp_thresh)
2292            && (top_xy   < 0 || ((qp + s->current_picture.qscale_table[top_xy    ] + 1)>>1) <= qp_thresh)){
2293             if(!FRAME_MBAFF)
2294                 return 1;
2295             if(   (left_xy[0]< 0            || ((qp + s->current_picture.qscale_table[left_xy[1]             ] + 1)>>1) <= qp_thresh)
2296                && (top_xy    < s->mb_stride || ((qp + s->current_picture.qscale_table[top_xy    -s->mb_stride] + 1)>>1) <= qp_thresh))
2297                 return 1;
2298         }
2299     }
2300
2301     top_type     = s->current_picture.mb_type[top_xy]    ;
2302     left_type[0] = s->current_picture.mb_type[left_xy[0]];
2303     left_type[1] = s->current_picture.mb_type[left_xy[1]];
2304     if(h->deblocking_filter == 2){
2305         if(h->slice_table[top_xy     ] != h->slice_num) top_type= 0;
2306         if(h->slice_table[left_xy[0] ] != h->slice_num) left_type[0]= left_type[1]= 0;
2307     }else{
2308         if(h->slice_table[top_xy     ] == 0xFFFF) top_type= 0;
2309         if(h->slice_table[left_xy[0] ] == 0xFFFF) left_type[0]= left_type[1] =0;
2310     }
2311     h->top_type    = top_type    ;
2312     h->left_type[0]= left_type[0];
2313     h->left_type[1]= left_type[1];
2314
2315     if(IS_INTRA(mb_type))
2316         return 0;
2317
2318     AV_COPY64(&h->non_zero_count_cache[0+8*1], &h->non_zero_count[mb_xy][ 0]);
2319     AV_COPY64(&h->non_zero_count_cache[0+8*2], &h->non_zero_count[mb_xy][ 8]);
2320     AV_COPY32(&h->non_zero_count_cache[0+8*5], &h->non_zero_count[mb_xy][16]);
2321     AV_COPY32(&h->non_zero_count_cache[4+8*3], &h->non_zero_count[mb_xy][20]);
2322     AV_COPY64(&h->non_zero_count_cache[0+8*4], &h->non_zero_count[mb_xy][24]);
2323
2324     h->cbp= h->cbp_table[mb_xy];
2325
2326     {
2327         int list;
2328         for(list=0; list<h->list_count; list++){
2329             int8_t *ref;
2330             int y, b_stride;
2331             int16_t (*mv_dst)[2];
2332             int16_t (*mv_src)[2];
2333
2334             if(!USES_LIST(mb_type, list)){
2335                 fill_rectangle(  h->mv_cache[list][scan8[0]], 4, 4, 8, pack16to32(0,0), 4);
2336                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2337                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2338                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2339                 AV_WN32A(&h->ref_cache[list][scan8[10]], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2340                 continue;
2341             }
2342
2343             ref = &s->current_picture.ref_index[list][4*mb_xy];
2344             {
2345                 int (*ref2frm)[64] = h->ref2frm[ h->slice_num&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2346                 AV_WN32A(&h->ref_cache[list][scan8[ 0]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2347                 AV_WN32A(&h->ref_cache[list][scan8[ 2]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2348                 ref += 2;
2349                 AV_WN32A(&h->ref_cache[list][scan8[ 8]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2350                 AV_WN32A(&h->ref_cache[list][scan8[10]], (pack16to32(ref2frm[list][ref[0]],ref2frm[list][ref[1]])&0x00FF00FF)*0x0101);
2351             }
2352
2353             b_stride = h->b_stride;
2354             mv_dst   = &h->mv_cache[list][scan8[0]];
2355             mv_src   = &s->current_picture.motion_val[list][4*s->mb_x + 4*s->mb_y*b_stride];
2356             for(y=0; y<4; y++){
2357                 AV_COPY128(mv_dst + 8*y, mv_src + y*b_stride);
2358             }
2359
2360         }
2361     }
2362
2363
2364 /*
2365 0 . T T. T T T T
2366 1 L . .L . . . .
2367 2 L . .L . . . .
2368 3 . T TL . . . .
2369 4 L . .L . . . .
2370 5 L . .. . . . .
2371 */
2372 //FIXME constraint_intra_pred & partitioning & nnz (let us hope this is just a typo in the spec)
2373     if(top_type){
2374         AV_COPY32(&h->non_zero_count_cache[4+8*0], &h->non_zero_count[top_xy][4+3*8]);
2375     }
2376
2377     if(left_type[0]){
2378         h->non_zero_count_cache[3+8*1]= h->non_zero_count[left_xy[0]][7+0*8];
2379         h->non_zero_count_cache[3+8*2]= h->non_zero_count[left_xy[0]][7+1*8];
2380         h->non_zero_count_cache[3+8*3]= h->non_zero_count[left_xy[0]][7+2*8];
2381         h->non_zero_count_cache[3+8*4]= h->non_zero_count[left_xy[0]][7+3*8];
2382     }
2383
2384     // CAVLC 8x8dct requires NNZ values for residual decoding that differ from what the loop filter needs
2385     if(!CABAC && h->pps.transform_8x8_mode){
2386         if(IS_8x8DCT(top_type)){
2387             h->non_zero_count_cache[4+8*0]=
2388             h->non_zero_count_cache[5+8*0]= h->cbp_table[top_xy] & 4;
2389             h->non_zero_count_cache[6+8*0]=
2390             h->non_zero_count_cache[7+8*0]= h->cbp_table[top_xy] & 8;
2391         }
2392         if(IS_8x8DCT(left_type[0])){
2393             h->non_zero_count_cache[3+8*1]=
2394             h->non_zero_count_cache[3+8*2]= h->cbp_table[left_xy[0]]&2; //FIXME check MBAFF
2395         }
2396         if(IS_8x8DCT(left_type[1])){
2397             h->non_zero_count_cache[3+8*3]=
2398             h->non_zero_count_cache[3+8*4]= h->cbp_table[left_xy[1]]&8; //FIXME check MBAFF
2399         }
2400
2401         if(IS_8x8DCT(mb_type)){
2402             h->non_zero_count_cache[scan8[0   ]]= h->non_zero_count_cache[scan8[1   ]]=
2403             h->non_zero_count_cache[scan8[2   ]]= h->non_zero_count_cache[scan8[3   ]]= h->cbp & 1;
2404
2405             h->non_zero_count_cache[scan8[0+ 4]]= h->non_zero_count_cache[scan8[1+ 4]]=
2406             h->non_zero_count_cache[scan8[2+ 4]]= h->non_zero_count_cache[scan8[3+ 4]]= h->cbp & 2;
2407
2408             h->non_zero_count_cache[scan8[0+ 8]]= h->non_zero_count_cache[scan8[1+ 8]]=
2409             h->non_zero_count_cache[scan8[2+ 8]]= h->non_zero_count_cache[scan8[3+ 8]]= h->cbp & 4;
2410
2411             h->non_zero_count_cache[scan8[0+12]]= h->non_zero_count_cache[scan8[1+12]]=
2412             h->non_zero_count_cache[scan8[2+12]]= h->non_zero_count_cache[scan8[3+12]]= h->cbp & 8;
2413         }
2414     }
2415
2416     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
2417         int list;
2418         for(list=0; list<h->list_count; list++){
2419             if(USES_LIST(top_type, list)){
2420                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
2421                 const int b8_xy= 4*top_xy + 2;
2422                 int (*ref2frm)[64] = h->ref2frm[ h->slice_table[top_xy]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2423                 AV_COPY128(h->mv_cache[list][scan8[0] + 0 - 1*8], s->current_picture.motion_val[list][b_xy + 0]);
2424                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
2425                 h->ref_cache[list][scan8[0] + 1 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 0]];
2426                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
2427                 h->ref_cache[list][scan8[0] + 3 - 1*8]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 1]];
2428             }else{
2429                 AV_ZERO128(h->mv_cache[list][scan8[0] + 0 - 1*8]);
2430                 AV_WN32A(&h->ref_cache[list][scan8[0] + 0 - 1*8], ((LIST_NOT_USED)&0xFF)*0x01010101u);
2431             }
2432
2433             if(!IS_INTERLACED(mb_type^left_type[0])){
2434                 if(USES_LIST(left_type[0], list)){
2435                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
2436                     const int b8_xy= 4*left_xy[0] + 1;
2437                     int (*ref2frm)[64] = h->ref2frm[ h->slice_table[left_xy[0]]&(MAX_SLICES-1) ][0] + (MB_MBAFF ? 20 : 2);
2438                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 0 ], s->current_picture.motion_val[list][b_xy + h->b_stride*0]);
2439                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 + 8 ], s->current_picture.motion_val[list][b_xy + h->b_stride*1]);
2440                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +16 ], s->current_picture.motion_val[list][b_xy + h->b_stride*2]);
2441                     AV_COPY32(h->mv_cache[list][scan8[0] - 1 +24 ], s->current_picture.motion_val[list][b_xy + h->b_stride*3]);
2442                     h->ref_cache[list][scan8[0] - 1 + 0 ]=
2443                     h->ref_cache[list][scan8[0] - 1 + 8 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*0]];
2444                     h->ref_cache[list][scan8[0] - 1 +16 ]=
2445                     h->ref_cache[list][scan8[0] - 1 +24 ]= ref2frm[list][s->current_picture.ref_index[list][b8_xy + 2*1]];
2446                 }else{
2447                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 0 ]);
2448                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 + 8 ]);
2449                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +16 ]);
2450                     AV_ZERO32(h->mv_cache [list][scan8[0] - 1 +24 ]);
2451                     h->ref_cache[list][scan8[0] - 1 + 0  ]=
2452                     h->ref_cache[list][scan8[0] - 1 + 8  ]=
2453                     h->ref_cache[list][scan8[0] - 1 + 16 ]=
2454                     h->ref_cache[list][scan8[0] - 1 + 24 ]= LIST_NOT_USED;
2455                 }
2456             }
2457         }
2458     }
2459
2460     return 0;
2461 }
2462
2463 static void loop_filter(H264Context *h){
2464     MpegEncContext * const s = &h->s;
2465     uint8_t  *dest_y, *dest_cb, *dest_cr;
2466     int linesize, uvlinesize, mb_x, mb_y;
2467     const int end_mb_y= s->mb_y + FRAME_MBAFF;
2468     const int old_slice_type= h->slice_type;
2469
2470     if(h->deblocking_filter) {
2471         for(mb_x= 0; mb_x<s->mb_width; mb_x++){
2472             for(mb_y=end_mb_y - FRAME_MBAFF; mb_y<= end_mb_y; mb_y++){
2473                 int mb_xy, mb_type;
2474                 mb_xy = h->mb_xy = mb_x + mb_y*s->mb_stride;
2475                 h->slice_num= h->slice_table[mb_xy];
2476                 mb_type= s->current_picture.mb_type[mb_xy];
2477                 h->list_count= h->list_counts[mb_xy];
2478
2479                 if(FRAME_MBAFF)
2480                     h->mb_mbaff = h->mb_field_decoding_flag = !!IS_INTERLACED(mb_type);
2481
2482                 s->mb_x= mb_x;
2483                 s->mb_y= mb_y;
2484                 dest_y  = s->current_picture.data[0] + (mb_x + mb_y * s->linesize  ) * 16;
2485                 dest_cb = s->current_picture.data[1] + (mb_x + mb_y * s->uvlinesize) * 8;
2486                 dest_cr = s->current_picture.data[2] + (mb_x + mb_y * s->uvlinesize) * 8;
2487                     //FIXME simplify above
2488
2489                 if (MB_FIELD) {
2490                     linesize   = h->mb_linesize   = s->linesize * 2;
2491                     uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2492                     if(mb_y&1){ //FIXME move out of this function?
2493                         dest_y -= s->linesize*15;
2494                         dest_cb-= s->uvlinesize*7;
2495                         dest_cr-= s->uvlinesize*7;
2496                     }
2497                 } else {
2498                     linesize   = h->mb_linesize   = s->linesize;
2499                     uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2500                 }
2501                 backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0);
2502                 if(fill_filter_caches(h, mb_type))
2503                     continue;
2504                 h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2505                 h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2506
2507                 if (FRAME_MBAFF) {
2508                     ff_h264_filter_mb     (h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2509                 } else {
2510                     ff_h264_filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2511                 }
2512             }
2513         }
2514     }
2515     h->slice_type= old_slice_type;
2516     s->mb_x= 0;
2517     s->mb_y= end_mb_y - FRAME_MBAFF;
2518     h->chroma_qp[0] = get_chroma_qp(h, 0, s->qscale);
2519     h->chroma_qp[1] = get_chroma_qp(h, 1, s->qscale);
2520 }
2521
2522 static void predict_field_decoding_flag(H264Context *h){
2523     MpegEncContext * const s = &h->s;
2524     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
2525     int mb_type = (h->slice_table[mb_xy-1] == h->slice_num)
2526                 ? s->current_picture.mb_type[mb_xy-1]
2527                 : (h->slice_table[mb_xy-s->mb_stride] == h->slice_num)
2528                 ? s->current_picture.mb_type[mb_xy-s->mb_stride]
2529                 : 0;
2530     h->mb_mbaff = h->mb_field_decoding_flag = IS_INTERLACED(mb_type) ? 1 : 0;
2531 }
2532
2533 static int decode_slice(struct AVCodecContext *avctx, void *arg){
2534     H264Context *h = *(void**)arg;
2535     MpegEncContext * const s = &h->s;
2536     const int part_mask= s->partitioned_frame ? (AC_END|AC_ERROR) : 0x7F;
2537
2538     s->mb_skip_run= -1;
2539
2540     h->is_complex = FRAME_MBAFF || s->picture_structure != PICT_FRAME || s->codec_id != CODEC_ID_H264 ||
2541                     (CONFIG_GRAY && (s->flags&CODEC_FLAG_GRAY));
2542
2543     if( h->pps.cabac ) {
2544         /* realign */
2545         align_get_bits( &s->gb );
2546
2547         /* init cabac */
2548         ff_init_cabac_states( &h->cabac);
2549         ff_init_cabac_decoder( &h->cabac,
2550                                s->gb.buffer + get_bits_count(&s->gb)/8,
2551                                (get_bits_left(&s->gb) + 7)/8);
2552
2553         ff_h264_init_cabac_states(h);
2554
2555         for(;;){
2556 //START_TIMER
2557             int ret = ff_h264_decode_mb_cabac(h);
2558             int eos;
2559 //STOP_TIMER("decode_mb_cabac")
2560
2561             if(ret>=0) ff_h264_hl_decode_mb(h);
2562
2563             if( ret >= 0 && FRAME_MBAFF ) { //FIXME optimal? or let mb_decode decode 16x32 ?
2564                 s->mb_y++;
2565
2566                 ret = ff_h264_decode_mb_cabac(h);
2567
2568                 if(ret>=0) ff_h264_hl_decode_mb(h);
2569                 s->mb_y--;
2570             }
2571             eos = get_cabac_terminate( &h->cabac );
2572
2573             if((s->workaround_bugs & FF_BUG_TRUNCATED) && h->cabac.bytestream > h->cabac.bytestream_end + 2){
2574                 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);
2575                 return 0;
2576             }
2577             if( ret < 0 || h->cabac.bytestream > h->cabac.bytestream_end + 2) {
2578                 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);
2579                 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);
2580                 return -1;
2581             }
2582
2583             if( ++s->mb_x >= s->mb_width ) {
2584                 s->mb_x = 0;
2585                 loop_filter(h);
2586                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2587                 ++s->mb_y;
2588                 if(FIELD_OR_MBAFF_PICTURE) {
2589                     ++s->mb_y;
2590                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2591                         predict_field_decoding_flag(h);
2592                 }
2593             }
2594
2595             if( eos || s->mb_y >= s->mb_height ) {
2596                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2597                 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);
2598                 return 0;
2599             }
2600         }
2601
2602     } else {
2603         for(;;){
2604             int ret = ff_h264_decode_mb_cavlc(h);
2605
2606             if(ret>=0) ff_h264_hl_decode_mb(h);
2607
2608             if(ret>=0 && FRAME_MBAFF){ //FIXME optimal? or let mb_decode decode 16x32 ?
2609                 s->mb_y++;
2610                 ret = ff_h264_decode_mb_cavlc(h);
2611
2612                 if(ret>=0) ff_h264_hl_decode_mb(h);
2613                 s->mb_y--;
2614             }
2615
2616             if(ret<0){
2617                 av_log(h->s.avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2618                 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);
2619
2620                 return -1;
2621             }
2622
2623             if(++s->mb_x >= s->mb_width){
2624                 s->mb_x=0;
2625                 loop_filter(h);
2626                 ff_draw_horiz_band(s, 16*s->mb_y, 16);
2627                 ++s->mb_y;
2628                 if(FIELD_OR_MBAFF_PICTURE) {
2629                     ++s->mb_y;
2630                     if(FRAME_MBAFF && s->mb_y < s->mb_height)
2631                         predict_field_decoding_flag(h);
2632                 }
2633                 if(s->mb_y >= s->mb_height){
2634                     tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2635
2636                     if(get_bits_count(&s->gb) == s->gb.size_in_bits ) {
2637                         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);
2638
2639                         return 0;
2640                     }else{
2641                         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);
2642
2643                         return -1;
2644                     }
2645                 }
2646             }
2647
2648             if(get_bits_count(&s->gb) >= s->gb.size_in_bits && s->mb_skip_run<=0){
2649                 tprintf(s->avctx, "slice end %d %d\n", get_bits_count(&s->gb), s->gb.size_in_bits);
2650                 if(get_bits_count(&s->gb) == s->gb.size_in_bits ){
2651                     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);
2652
2653                     return 0;
2654                 }else{
2655                     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);
2656
2657                     return -1;
2658                 }
2659             }
2660         }
2661     }
2662
2663 #if 0
2664     for(;s->mb_y < s->mb_height; s->mb_y++){
2665         for(;s->mb_x < s->mb_width; s->mb_x++){
2666             int ret= decode_mb(h);
2667
2668             ff_h264_hl_decode_mb(h);
2669
2670             if(ret<0){
2671                 av_log(s->avctx, AV_LOG_ERROR, "error while decoding MB %d %d\n", s->mb_x, s->mb_y);
2672                 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);
2673
2674                 return -1;
2675             }
2676
2677             if(++s->mb_x >= s->mb_width){
2678                 s->mb_x=0;
2679                 if(++s->mb_y >= s->mb_height){
2680                     if(get_bits_count(s->gb) == s->gb.size_in_bits){
2681                         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);
2682
2683                         return 0;
2684                     }else{
2685                         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);
2686
2687                         return -1;
2688                     }
2689                 }
2690             }
2691
2692             if(get_bits_count(s->?gb) >= s->gb?.size_in_bits){
2693                 if(get_bits_count(s->gb) == s->gb.size_in_bits){
2694                     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);
2695
2696                     return 0;
2697                 }else{
2698                     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);
2699
2700                     return -1;
2701                 }
2702             }
2703         }
2704         s->mb_x=0;
2705         ff_draw_horiz_band(s, 16*s->mb_y, 16);
2706     }
2707 #endif
2708     return -1; //not reached
2709 }
2710
2711 /**
2712  * Call decode_slice() for each context.
2713  *
2714  * @param h h264 master context
2715  * @param context_count number of contexts to execute
2716  */
2717 static void execute_decode_slices(H264Context *h, int context_count){
2718     MpegEncContext * const s = &h->s;
2719     AVCodecContext * const avctx= s->avctx;
2720     H264Context *hx;
2721     int i;
2722
2723     if (s->avctx->hwaccel)
2724         return;
2725     if(s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2726         return;
2727     if(context_count == 1) {
2728         decode_slice(avctx, &h);
2729     } else {
2730         for(i = 1; i < context_count; i++) {
2731             hx = h->thread_context[i];
2732             hx->s.error_recognition = avctx->error_recognition;
2733             hx->s.error_count = 0;
2734         }
2735
2736         avctx->execute(avctx, (void *)decode_slice,
2737                        h->thread_context, NULL, context_count, sizeof(void*));
2738
2739         /* pull back stuff from slices to master context */
2740         hx = h->thread_context[context_count - 1];
2741         s->mb_x = hx->s.mb_x;
2742         s->mb_y = hx->s.mb_y;
2743         s->dropable = hx->s.dropable;
2744         s->picture_structure = hx->s.picture_structure;
2745         for(i = 1; i < context_count; i++)
2746             h->s.error_count += h->thread_context[i]->s.error_count;
2747     }
2748 }
2749
2750
2751 static int decode_nal_units(H264Context *h, const uint8_t *buf, int buf_size){
2752     MpegEncContext * const s = &h->s;
2753     AVCodecContext * const avctx= s->avctx;
2754     int buf_index=0;
2755     H264Context *hx; ///< thread context
2756     int context_count = 0;
2757     int next_avc= h->is_avc ? 0 : buf_size;
2758
2759     h->max_contexts = avctx->thread_count;
2760 #if 0
2761     int i;
2762     for(i=0; i<50; i++){
2763         av_log(NULL, AV_LOG_ERROR,"%02X ", buf[i]);
2764     }
2765 #endif
2766     if(!(s->flags2 & CODEC_FLAG2_CHUNKS)){
2767         h->current_slice = 0;
2768         if (!s->first_field)
2769             s->current_picture_ptr= NULL;
2770         ff_h264_reset_sei(h);
2771     }
2772
2773     for(;;){
2774         int consumed;
2775         int dst_length;
2776         int bit_length;
2777         const uint8_t *ptr;
2778         int i, nalsize = 0;
2779         int err;
2780
2781         if(buf_index >= next_avc) {
2782             if(buf_index >= buf_size) break;
2783             nalsize = 0;
2784             for(i = 0; i < h->nal_length_size; i++)
2785                 nalsize = (nalsize << 8) | buf[buf_index++];
2786             if(nalsize <= 1 || nalsize > buf_size - buf_index){
2787                 if(nalsize == 1){
2788                     buf_index++;
2789                     continue;
2790                 }else{
2791                     av_log(h->s.avctx, AV_LOG_ERROR, "AVC: nal size %d\n", nalsize);
2792                     break;
2793                 }
2794             }
2795             next_avc= buf_index + nalsize;
2796         } else {
2797             // start code prefix search
2798             for(; buf_index + 3 < next_avc; buf_index++){
2799                 // This should always succeed in the first iteration.
2800                 if(buf[buf_index] == 0 && buf[buf_index+1] == 0 && buf[buf_index+2] == 1)
2801                     break;
2802             }
2803
2804             if(buf_index+3 >= buf_size) break;
2805
2806             buf_index+=3;
2807             if(buf_index >= next_avc) continue;
2808         }
2809
2810         hx = h->thread_context[context_count];
2811
2812         ptr= ff_h264_decode_nal(hx, buf + buf_index, &dst_length, &consumed, next_avc - buf_index);
2813         if (ptr==NULL || dst_length < 0){
2814             return -1;
2815         }
2816         i= buf_index + consumed;
2817         if((s->workaround_bugs & FF_BUG_AUTODETECT) && i+3<next_avc &&
2818            buf[i]==0x00 && buf[i+1]==0x00 && buf[i+2]==0x01 && buf[i+3]==0xE0)
2819             s->workaround_bugs |= FF_BUG_TRUNCATED;
2820
2821         if(!(s->workaround_bugs & FF_BUG_TRUNCATED)){
2822         while(ptr[dst_length - 1] == 0 && dst_length > 0)
2823             dst_length--;
2824         }
2825         bit_length= !dst_length ? 0 : (8*dst_length - ff_h264_decode_rbsp_trailing(h, ptr + dst_length - 1));
2826
2827         if(s->avctx->debug&FF_DEBUG_STARTCODE){
2828             av_log(h->s.avctx, AV_LOG_DEBUG, "NAL %d at %d/%d length %d\n", hx->nal_unit_type, buf_index, buf_size, dst_length);
2829         }
2830
2831         if (h->is_avc && (nalsize != consumed) && nalsize){
2832             av_log(h->s.avctx, AV_LOG_DEBUG, "AVC: Consumed only %d bytes instead of %d\n", consumed, nalsize);
2833         }
2834
2835         buf_index += consumed;
2836
2837         if(  (s->hurry_up == 1 && h->nal_ref_idc  == 0) //FIXME do not discard SEI id
2838            ||(avctx->skip_frame >= AVDISCARD_NONREF && h->nal_ref_idc  == 0))
2839             continue;
2840
2841       again:
2842         err = 0;
2843         switch(hx->nal_unit_type){
2844         case NAL_IDR_SLICE:
2845             if (h->nal_unit_type != NAL_IDR_SLICE) {
2846                 av_log(h->s.avctx, AV_LOG_ERROR, "Invalid mix of idr and non-idr slices");
2847                 return -1;
2848             }
2849             idr(h); //FIXME ensure we don't loose some frames if there is reordering
2850         case NAL_SLICE:
2851             init_get_bits(&hx->s.gb, ptr, bit_length);
2852             hx->intra_gb_ptr=
2853             hx->inter_gb_ptr= &hx->s.gb;
2854             hx->s.data_partitioning = 0;
2855
2856             if((err = decode_slice_header(hx, h)))
2857                break;
2858
2859             if (h->current_slice == 1) {
2860                 if (s->avctx->hwaccel && s->avctx->hwaccel->start_frame(s->avctx, NULL, 0) < 0)
2861                     return -1;
2862                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU)
2863                     ff_vdpau_h264_picture_start(s);
2864             }
2865
2866             s->current_picture_ptr->key_frame |=
2867                     (hx->nal_unit_type == NAL_IDR_SLICE) ||
2868                     (h->sei_recovery_frame_cnt >= 0);
2869             if(hx->redundant_pic_count==0 && hx->s.hurry_up < 5
2870                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2871                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=FF_B_TYPE)
2872                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==FF_I_TYPE)
2873                && avctx->skip_frame < AVDISCARD_ALL){
2874                 if(avctx->hwaccel) {
2875                     if (avctx->hwaccel->decode_slice(avctx, &buf[buf_index - consumed], consumed) < 0)
2876                         return -1;
2877                 }else
2878                 if(CONFIG_H264_VDPAU_DECODER && s->avctx->codec->capabilities&CODEC_CAP_HWACCEL_VDPAU){
2879                     static const uint8_t start_code[] = {0x00, 0x00, 0x01};
2880                     ff_vdpau_add_data_chunk(s, start_code, sizeof(start_code));
2881                     ff_vdpau_add_data_chunk(s, &buf[buf_index - consumed], consumed );
2882                 }else
2883                     context_count++;
2884             }
2885             break;
2886         case NAL_DPA:
2887             init_get_bits(&hx->s.gb, ptr, bit_length);
2888             hx->intra_gb_ptr=
2889             hx->inter_gb_ptr= NULL;
2890
2891             if ((err = decode_slice_header(hx, h)) < 0)
2892                 break;
2893
2894             hx->s.data_partitioning = 1;
2895
2896             break;
2897         case NAL_DPB:
2898             init_get_bits(&hx->intra_gb, ptr, bit_length);
2899             hx->intra_gb_ptr= &hx->intra_gb;
2900             break;
2901         case NAL_DPC:
2902             init_get_bits(&hx->inter_gb, ptr, bit_length);
2903             hx->inter_gb_ptr= &hx->inter_gb;
2904
2905             if(hx->redundant_pic_count==0 && hx->intra_gb_ptr && hx->s.data_partitioning
2906                && s->context_initialized
2907                && s->hurry_up < 5
2908                && (avctx->skip_frame < AVDISCARD_NONREF || hx->nal_ref_idc)
2909                && (avctx->skip_frame < AVDISCARD_BIDIR  || hx->slice_type_nos!=FF_B_TYPE)
2910                && (avctx->skip_frame < AVDISCARD_NONKEY || hx->slice_type_nos==FF_I_TYPE)
2911                && avctx->skip_frame < AVDISCARD_ALL)
2912                 context_count++;
2913             break;
2914         case NAL_SEI:
2915             init_get_bits(&s->gb, ptr, bit_length);
2916             ff_h264_decode_sei(h);
2917             break;
2918         case NAL_SPS:
2919             init_get_bits(&s->gb, ptr, bit_length);
2920             ff_h264_decode_seq_parameter_set(h);
2921
2922             if(s->flags& CODEC_FLAG_LOW_DELAY)
2923                 s->low_delay=1;
2924
2925             if(avctx->has_b_frames < 2)
2926                 avctx->has_b_frames= !s->low_delay;
2927             break;
2928         case NAL_PPS:
2929             init_get_bits(&s->gb, ptr, bit_length);
2930
2931             ff_h264_decode_picture_parameter_set(h, bit_length);
2932
2933             break;
2934         case NAL_AUD:
2935         case NAL_END_SEQUENCE:
2936         case NAL_END_STREAM:
2937         case NAL_FILLER_DATA:
2938         case NAL_SPS_EXT:
2939         case NAL_AUXILIARY_SLICE:
2940             break;
2941         default:
2942             av_log(avctx, AV_LOG_DEBUG, "Unknown NAL code: %d (%d bits)\n", hx->nal_unit_type, bit_length);
2943         }
2944
2945         if(context_count == h->max_contexts) {
2946             execute_decode_slices(h, context_count);
2947             context_count = 0;
2948         }
2949
2950         if (err < 0)
2951             av_log(h->s.avctx, AV_LOG_ERROR, "decode_slice_header error\n");
2952         else if(err == 1) {
2953             /* Slice could not be decoded in parallel mode, copy down
2954              * NAL unit stuff to context 0 and restart. Note that
2955              * rbsp_buffer is not transferred, but since we no longer
2956              * run in parallel mode this should not be an issue. */
2957             h->nal_unit_type = hx->nal_unit_type;
2958             h->nal_ref_idc   = hx->nal_ref_idc;
2959             hx = h;
2960             goto again;
2961         }
2962     }
2963     if(context_count)
2964         execute_decode_slices(h, context_count);
2965     return buf_index;
2966 }
2967
2968 /**
2969  * returns the number of bytes consumed for building the current frame
2970  */
2971 static int get_consumed_bytes(MpegEncContext *s, int pos, int buf_size){
2972         if(pos==0) pos=1; //avoid infinite loops (i doubt that is needed but ...)
2973         if(pos+10>buf_size) pos=buf_size; // oops ;)
2974
2975         return pos;
2976 }
2977
2978 static int decode_frame(AVCodecContext *avctx,
2979                              void *data, int *data_size,
2980                              AVPacket *avpkt)
2981 {
2982     const uint8_t *buf = avpkt->data;
2983     int buf_size = avpkt->size;
2984     H264Context *h = avctx->priv_data;
2985     MpegEncContext *s = &h->s;
2986     AVFrame *pict = data;
2987     int buf_index;
2988
2989     s->flags= avctx->flags;
2990     s->flags2= avctx->flags2;
2991
2992    /* end of stream, output what is still in the buffers */
2993     if (buf_size == 0) {
2994         Picture *out;
2995         int i, out_idx;
2996
2997 //FIXME factorize this with the output code below
2998         out = h->delayed_pic[0];
2999         out_idx = 0;
3000         for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3001             if(h->delayed_pic[i]->poc < out->poc){
3002                 out = h->delayed_pic[i];
3003                 out_idx = i;
3004             }
3005
3006         for(i=out_idx; h->delayed_pic[i]; i++)
3007             h->delayed_pic[i] = h->delayed_pic[i+1];
3008
3009         if(out){
3010             *data_size = sizeof(AVFrame);
3011             *pict= *(AVFrame*)out;
3012         }
3013
3014         return 0;
3015     }
3016
3017     buf_index=decode_nal_units(h, buf, buf_size);
3018     if(buf_index < 0)
3019         return -1;
3020
3021     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) && !s->current_picture_ptr){
3022         if (avctx->skip_frame >= AVDISCARD_NONREF || s->hurry_up) return 0;
3023         av_log(avctx, AV_LOG_ERROR, "no frame!\n");
3024         return -1;
3025     }
3026
3027     if(!(s->flags2 & CODEC_FLAG2_CHUNKS) || (s->mb_y >= s->mb_height && s->mb_height)){
3028         Picture *out = s->current_picture_ptr;
3029         Picture *cur = s->current_picture_ptr;
3030         int i, pics, out_of_order, out_idx;
3031
3032         field_end(h);
3033
3034         if (cur->field_poc[0]==INT_MAX || cur->field_poc[1]==INT_MAX) {
3035             /* Wait for second field. */
3036             *data_size = 0;
3037
3038         } else {
3039             cur->interlaced_frame = 0;
3040             cur->repeat_pict = 0;
3041
3042             /* Signal interlacing information externally. */
3043             /* Prioritize picture timing SEI information over used decoding process if it exists. */
3044
3045             if(h->sps.pic_struct_present_flag){
3046                 switch (h->sei_pic_struct)
3047                 {
3048                 case SEI_PIC_STRUCT_FRAME:
3049                     break;
3050                 case SEI_PIC_STRUCT_TOP_FIELD:
3051                 case SEI_PIC_STRUCT_BOTTOM_FIELD:
3052                     cur->interlaced_frame = 1;
3053                     break;
3054                 case SEI_PIC_STRUCT_TOP_BOTTOM:
3055                 case SEI_PIC_STRUCT_BOTTOM_TOP:
3056                     if (FIELD_OR_MBAFF_PICTURE)
3057                         cur->interlaced_frame = 1;
3058                     else
3059                         // try to flag soft telecine progressive
3060                         cur->interlaced_frame = h->prev_interlaced_frame;
3061                     break;
3062                 case SEI_PIC_STRUCT_TOP_BOTTOM_TOP:
3063                 case SEI_PIC_STRUCT_BOTTOM_TOP_BOTTOM:
3064                     // Signal the possibility of telecined film externally (pic_struct 5,6)
3065                     // From these hints, let the applications decide if they apply deinterlacing.
3066                     cur->repeat_pict = 1;
3067                     break;
3068                 case SEI_PIC_STRUCT_FRAME_DOUBLING:
3069                     // Force progressive here, as doubling interlaced frame is a bad idea.
3070                     cur->repeat_pict = 2;
3071                     break;
3072                 case SEI_PIC_STRUCT_FRAME_TRIPLING:
3073                     cur->repeat_pict = 4;
3074                     break;
3075                 }
3076
3077                 if ((h->sei_ct_type & 3) && h->sei_pic_struct <= SEI_PIC_STRUCT_BOTTOM_TOP)
3078                     cur->interlaced_frame = (h->sei_ct_type & (1<<1)) != 0;
3079             }else{
3080                 /* Derive interlacing flag from used decoding process. */
3081                 cur->interlaced_frame = FIELD_OR_MBAFF_PICTURE;
3082             }
3083             h->prev_interlaced_frame = cur->interlaced_frame;
3084
3085             if (cur->field_poc[0] != cur->field_poc[1]){
3086                 /* Derive top_field_first from field pocs. */
3087                 cur->top_field_first = cur->field_poc[0] < cur->field_poc[1];
3088             }else{
3089                 if(cur->interlaced_frame || h->sps.pic_struct_present_flag){
3090                     /* Use picture timing SEI information. Even if it is a information of a past frame, better than nothing. */
3091                     if(h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM
3092                       || h->sei_pic_struct == SEI_PIC_STRUCT_TOP_BOTTOM_TOP)
3093                         cur->top_field_first = 1;
3094                     else
3095                         cur->top_field_first = 0;
3096                 }else{
3097                     /* Most likely progressive */
3098                     cur->top_field_first = 0;
3099                 }
3100             }
3101
3102         //FIXME do something with unavailable reference frames
3103
3104             /* Sort B-frames into display order */
3105
3106             if(h->sps.bitstream_restriction_flag
3107                && s->avctx->has_b_frames < h->sps.num_reorder_frames){
3108                 s->avctx->has_b_frames = h->sps.num_reorder_frames;
3109                 s->low_delay = 0;
3110             }
3111
3112             if(   s->avctx->strict_std_compliance >= FF_COMPLIANCE_STRICT
3113                && !h->sps.bitstream_restriction_flag){
3114                 s->avctx->has_b_frames= MAX_DELAYED_PIC_COUNT;
3115                 s->low_delay= 0;
3116             }
3117
3118             pics = 0;
3119             while(h->delayed_pic[pics]) pics++;
3120
3121             assert(pics <= MAX_DELAYED_PIC_COUNT);
3122
3123             h->delayed_pic[pics++] = cur;
3124             if(cur->reference == 0)
3125                 cur->reference = DELAYED_PIC_REF;
3126
3127             out = h->delayed_pic[0];
3128             out_idx = 0;
3129             for(i=1; h->delayed_pic[i] && !h->delayed_pic[i]->key_frame && !h->delayed_pic[i]->mmco_reset; i++)
3130                 if(h->delayed_pic[i]->poc < out->poc){
3131                     out = h->delayed_pic[i];
3132                     out_idx = i;
3133                 }
3134             if(s->avctx->has_b_frames == 0 && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset))
3135                 h->outputed_poc= INT_MIN;
3136             out_of_order = out->poc < h->outputed_poc;
3137
3138             if(h->sps.bitstream_restriction_flag && s->avctx->has_b_frames >= h->sps.num_reorder_frames)
3139                 { }
3140             else if((out_of_order && pics-1 == s->avctx->has_b_frames && s->avctx->has_b_frames < MAX_DELAYED_PIC_COUNT)
3141                || (s->low_delay &&
3142                 ((h->outputed_poc != INT_MIN && out->poc > h->outputed_poc + 2)
3143                  || cur->pict_type == FF_B_TYPE)))
3144             {
3145                 s->low_delay = 0;
3146                 s->avctx->has_b_frames++;
3147             }
3148
3149             if(out_of_order || pics > s->avctx->has_b_frames){
3150                 out->reference &= ~DELAYED_PIC_REF;
3151                 for(i=out_idx; h->delayed_pic[i]; i++)
3152                     h->delayed_pic[i] = h->delayed_pic[i+1];
3153             }
3154             if(!out_of_order && pics > s->avctx->has_b_frames){
3155                 *data_size = sizeof(AVFrame);
3156
3157                 if(out_idx==0 && h->delayed_pic[0] && (h->delayed_pic[0]->key_frame || h->delayed_pic[0]->mmco_reset)) {
3158                     h->outputed_poc = INT_MIN;
3159                 } else
3160                     h->outputed_poc = out->poc;
3161                 *pict= *(AVFrame*)out;
3162             }else{
3163                 av_log(avctx, AV_LOG_DEBUG, "no picture\n");
3164             }
3165         }
3166     }
3167
3168     assert(pict->data[0] || !*data_size);
3169     ff_print_debug_info(s, pict);
3170 //printf("out %d\n", (int)pict->data[0]);
3171
3172     return get_consumed_bytes(s, buf_index, buf_size);
3173 }
3174 #if 0
3175 static inline void fill_mb_avail(H264Context *h){
3176     MpegEncContext * const s = &h->s;
3177     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
3178
3179     if(s->mb_y){
3180         h->mb_avail[0]= s->mb_x                 && h->slice_table[mb_xy - s->mb_stride - 1] == h->slice_num;
3181         h->mb_avail[1]=                            h->slice_table[mb_xy - s->mb_stride    ] == h->slice_num;
3182         h->mb_avail[2]= s->mb_x+1 < s->mb_width && h->slice_table[mb_xy - s->mb_stride + 1] == h->slice_num;
3183     }else{
3184         h->mb_avail[0]=
3185         h->mb_avail[1]=
3186         h->mb_avail[2]= 0;
3187     }
3188     h->mb_avail[3]= s->mb_x && h->slice_table[mb_xy - 1] == h->slice_num;
3189     h->mb_avail[4]= 1; //FIXME move out
3190     h->mb_avail[5]= 0; //FIXME move out
3191 }
3192 #endif
3193
3194 #ifdef TEST
3195 #undef printf
3196 #undef random
3197 #define COUNT 8000
3198 #define SIZE (COUNT*40)
3199 int main(void){
3200     int i;
3201     uint8_t temp[SIZE];
3202     PutBitContext pb;
3203     GetBitContext gb;
3204 //    int int_temp[10000];
3205     DSPContext dsp;
3206     AVCodecContext avctx;
3207
3208     dsputil_init(&dsp, &avctx);
3209
3210     init_put_bits(&pb, temp, SIZE);
3211     printf("testing unsigned exp golomb\n");
3212     for(i=0; i<COUNT; i++){
3213         START_TIMER
3214         set_ue_golomb(&pb, i);
3215         STOP_TIMER("set_ue_golomb");
3216     }
3217     flush_put_bits(&pb);
3218
3219     init_get_bits(&gb, temp, 8*SIZE);
3220     for(i=0; i<COUNT; i++){
3221         int j, s;
3222
3223         s= show_bits(&gb, 24);
3224
3225         START_TIMER
3226         j= get_ue_golomb(&gb);
3227         if(j != i){
3228             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3229 //            return -1;
3230         }
3231         STOP_TIMER("get_ue_golomb");
3232     }
3233
3234
3235     init_put_bits(&pb, temp, SIZE);
3236     printf("testing signed exp golomb\n");
3237     for(i=0; i<COUNT; i++){
3238         START_TIMER
3239         set_se_golomb(&pb, i - COUNT/2);
3240         STOP_TIMER("set_se_golomb");
3241     }
3242     flush_put_bits(&pb);
3243
3244     init_get_bits(&gb, temp, 8*SIZE);
3245     for(i=0; i<COUNT; i++){
3246         int j, s;
3247
3248         s= show_bits(&gb, 24);
3249
3250         START_TIMER
3251         j= get_se_golomb(&gb);
3252         if(j != i - COUNT/2){
3253             printf("mismatch! at %d (%d should be %d) bits:%6X\n", i, j, i, s);
3254 //            return -1;
3255         }
3256         STOP_TIMER("get_se_golomb");
3257     }
3258
3259 #if 0
3260     printf("testing 4x4 (I)DCT\n");
3261
3262     DCTELEM block[16];
3263     uint8_t src[16], ref[16];
3264     uint64_t error= 0, max_error=0;
3265
3266     for(i=0; i<COUNT; i++){
3267         int j;
3268 //        printf("%d %d %d\n", r1, r2, (r2-r1)*16);
3269         for(j=0; j<16; j++){
3270             ref[j]= random()%255;
3271             src[j]= random()%255;
3272         }
3273
3274         h264_diff_dct_c(block, src, ref, 4);
3275
3276         //normalize
3277         for(j=0; j<16; j++){
3278 //            printf("%d ", block[j]);
3279             block[j]= block[j]*4;
3280             if(j&1) block[j]= (block[j]*4 + 2)/5;
3281             if(j&4) block[j]= (block[j]*4 + 2)/5;
3282         }
3283 //        printf("\n");
3284
3285         h->h264dsp.h264_idct_add(ref, block, 4);
3286 /*        for(j=0; j<16; j++){
3287             printf("%d ", ref[j]);
3288         }
3289         printf("\n");*/
3290
3291         for(j=0; j<16; j++){
3292             int diff= FFABS(src[j] - ref[j]);
3293
3294             error+= diff*diff;
3295             max_error= FFMAX(max_error, diff);
3296         }
3297     }
3298     printf("error=%f max_error=%d\n", ((float)error)/COUNT/16, (int)max_error );
3299     printf("testing quantizer\n");
3300     for(qp=0; qp<52; qp++){
3301         for(i=0; i<16; i++)
3302             src1_block[i]= src2_block[i]= random()%255;
3303
3304     }
3305     printf("Testing NAL layer\n");
3306
3307     uint8_t bitstream[COUNT];
3308     uint8_t nal[COUNT*2];
3309     H264Context h;
3310     memset(&h, 0, sizeof(H264Context));
3311
3312     for(i=0; i<COUNT; i++){
3313         int zeros= i;
3314         int nal_length;
3315         int consumed;
3316         int out_length;
3317         uint8_t *out;
3318         int j;
3319
3320         for(j=0; j<COUNT; j++){
3321             bitstream[j]= (random() % 255) + 1;
3322         }
3323
3324         for(j=0; j<zeros; j++){
3325             int pos= random() % COUNT;
3326             while(bitstream[pos] == 0){
3327                 pos++;
3328                 pos %= COUNT;
3329             }
3330             bitstream[pos]=0;
3331         }
3332
3333         START_TIMER
3334
3335         nal_length= encode_nal(&h, nal, bitstream, COUNT, COUNT*2);
3336         if(nal_length<0){
3337             printf("encoding failed\n");
3338             return -1;
3339         }
3340
3341         out= ff_h264_decode_nal(&h, nal, &out_length, &consumed, nal_length);
3342
3343         STOP_TIMER("NAL")
3344
3345         if(out_length != COUNT){
3346             printf("incorrect length %d %d\n", out_length, COUNT);
3347             return -1;
3348         }
3349
3350         if(consumed != nal_length){
3351             printf("incorrect consumed length %d %d\n", nal_length, consumed);
3352             return -1;
3353         }
3354
3355         if(memcmp(bitstream, out, COUNT)){
3356             printf("mismatch\n");
3357             return -1;
3358         }
3359     }
3360 #endif
3361
3362     printf("Testing RBSP\n");
3363
3364
3365     return 0;
3366 }
3367 #endif /* TEST */
3368
3369
3370 av_cold void ff_h264_free_context(H264Context *h)
3371 {
3372     int i;
3373
3374     free_tables(h); //FIXME cleanup init stuff perhaps
3375
3376     for(i = 0; i < MAX_SPS_COUNT; i++)
3377         av_freep(h->sps_buffers + i);
3378
3379     for(i = 0; i < MAX_PPS_COUNT; i++)
3380         av_freep(h->pps_buffers + i);
3381 }
3382
3383 av_cold int ff_h264_decode_end(AVCodecContext *avctx)
3384 {
3385     H264Context *h = avctx->priv_data;
3386     MpegEncContext *s = &h->s;
3387
3388     ff_h264_free_context(h);
3389
3390     MPV_common_end(s);
3391
3392 //    memset(h, 0, sizeof(H264Context));
3393
3394     return 0;
3395 }
3396
3397
3398 AVCodec h264_decoder = {
3399     "h264",
3400     AVMEDIA_TYPE_VIDEO,
3401     CODEC_ID_H264,
3402     sizeof(H264Context),
3403     ff_h264_decode_init,
3404     NULL,
3405     ff_h264_decode_end,
3406     decode_frame,
3407     /*CODEC_CAP_DRAW_HORIZ_BAND |*/ CODEC_CAP_DR1 | CODEC_CAP_DELAY,
3408     .flush= flush_dpb,
3409     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10"),
3410 };
3411
3412 #if CONFIG_H264_VDPAU_DECODER
3413 AVCodec h264_vdpau_decoder = {
3414     "h264_vdpau",
3415     AVMEDIA_TYPE_VIDEO,
3416     CODEC_ID_H264,
3417     sizeof(H264Context),
3418     ff_h264_decode_init,
3419     NULL,
3420     ff_h264_decode_end,
3421     decode_frame,
3422     CODEC_CAP_DR1 | CODEC_CAP_DELAY | CODEC_CAP_HWACCEL_VDPAU,
3423     .flush= flush_dpb,
3424     .long_name = NULL_IF_CONFIG_SMALL("H.264 / AVC / MPEG-4 AVC / MPEG-4 part 10 (VDPAU acceleration)"),
3425     .pix_fmts = (const enum PixelFormat[]){PIX_FMT_VDPAU_H264, PIX_FMT_NONE},
3426 };
3427 #endif