libavcodec/h264_direct.c

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * Libav is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with Libav; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 direct mb/block decoding.
  25  * @author Michael Niedermayer <michaelni@gmx.at>
  26  */
  27
  28 #include "internal.h"
  29 #include "avcodec.h"
  30 #include "mpegvideo.h"
  31 #include "h264.h"
  32 #include "rectangle.h"
  33 #include "thread.h"
  34
  35 //#undef NDEBUG
  36 #include <assert.h>
  37
  38
  39 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
  40     int poc0 = h->ref_list[0][i].poc;
  41     int td = av_clip(poc1 - poc0, -128, 127);
  42     if(td == 0 || h->ref_list[0][i].long_ref){
  43         return 256;
  44     }else{
  45         int tb = av_clip(poc - poc0, -128, 127);
  46         int tx = (16384 + (FFABS(td) >> 1)) / td;
  47         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
  48     }
  49 }
  50
  51 void ff_h264_direct_dist_scale_factor(H264Context * const h){
  52     const int poc = h->cur_pic_ptr->field_poc[h->picture_structure == PICT_BOTTOM_FIELD];
  53     const int poc1 = h->ref_list[1][0].poc;
  54     int i, field;
  55
  56     if (FRAME_MBAFF)
  57         for (field = 0; field < 2; field++){
  58             const int poc  = h->cur_pic_ptr->field_poc[field];
  59             const int poc1 = h->ref_list[1][0].field_poc[field];
  60             for (i = 0; i < 2 * h->ref_count[0]; i++)
  61                 h->dist_scale_factor_field[field][i^field] =
  62                     get_scale_factor(h, poc, poc1, i+16);
  63         }
  64
  65     for (i = 0; i < h->ref_count[0]; i++){
  66         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
  67     }
  68 }
  69
  70 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
  71     Picture * const ref1 = &h->ref_list[1][0];
  72     int j, old_ref, rfield;
  73     int start= mbafi ? 16                      : 0;
  74     int end  = mbafi ? 16+2*h->ref_count[0]    : h->ref_count[0];
  75     int interl= mbafi || h->picture_structure != PICT_FRAME;
  76
  77     /* bogus; fills in for missing frames */
  78     memset(map[list], 0, sizeof(map[list]));
  79
  80     for(rfield=0; rfield<2; rfield++){
  81         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
  82             int poc = ref1->ref_poc[colfield][list][old_ref];
  83
  84             if     (!interl)
  85                 poc |= 3;
  86             else if( interl && (poc&3) == 3) // FIXME: store all MBAFF references so this is not needed
  87                 poc= (poc&~3) + rfield + 1;
  88
  89             for(j=start; j<end; j++){
  90                 if (4 * h->ref_list[0][j].frame_num + (h->ref_list[0][j].reference & 3) == poc) {
  91                     int cur_ref= mbafi ? (j-16)^field : j;
  92                     if (ref1->mbaff)
  93                         map[list][2 * old_ref + (rfield^field) + 16] = cur_ref;
  94                     if(rfield == field || !interl)
  95                         map[list][old_ref] = cur_ref;
  96                     break;
  97                 }
  98             }
  99         }
 100     }
 101 }
 102
 103 void ff_h264_direct_ref_list_init(H264Context * const h){
 104     Picture * const ref1 = &h->ref_list[1][0];
 105     Picture * const cur = h->cur_pic_ptr;
 106     int list, j, field;
 107     int sidx= (h->picture_structure&1)^1;
 108     int ref1sidx = (ref1->reference&1)^1;
 109
 110     for(list=0; list<2; list++){
 111         cur->ref_count[sidx][list] = h->ref_count[list];
 112         for(j=0; j<h->ref_count[list]; j++)
 113             cur->ref_poc[sidx][list][j] = 4 * h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference & 3);
 114     }
 115
 116     if(h->picture_structure == PICT_FRAME){
 117         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
 118         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
 119     }
 120
 121     cur->mbaff= FRAME_MBAFF;
 122
 123     h->col_fieldoff= 0;
 124     if(h->picture_structure == PICT_FRAME){
 125         int cur_poc = h->cur_pic_ptr->poc;
 126         int *col_poc = h->ref_list[1]->field_poc;
 127         h->col_parity= (FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc));
 128         ref1sidx=sidx= h->col_parity;
 129     } else if (!(h->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff) { // FL -> FL & differ parity
 130         h->col_fieldoff = 2 * h->ref_list[1][0].reference - 3;
 131     }
 132
 133     if (h->slice_type_nos != AV_PICTURE_TYPE_B || h->direct_spatial_mv_pred)
 134         return;
 135
 136     for(list=0; list<2; list++){
 137         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
 138         if(FRAME_MBAFF)
 139         for(field=0; field<2; field++)
 140             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
 141     }
 142 }
 143
 144 static void await_reference_mb_row(H264Context * const h, Picture *ref, int mb_y)
 145 {
 146     int ref_field = ref->reference - 1;
 147     int ref_field_picture = ref->field_picture;
 148     int ref_height = 16*h->mb_height >> ref_field_picture;
 149
 150     if(!HAVE_THREADS || !(h->avctx->active_thread_type&FF_THREAD_FRAME))
 151         return;
 152
 153     //FIXME it can be safe to access mb stuff
 154     //even if pixels aren't deblocked yet
 155
 156     ff_thread_await_progress(&ref->tf,
 157                              FFMIN(16 * mb_y >> ref_field_picture, ref_height - 1),
 158                              ref_field_picture && ref_field);
 159 }
 160
 161 static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
 162     int b8_stride = 2;
 163     int b4_stride = h->b_stride;
 164     int mb_xy = h->mb_xy, mb_y = h->mb_y;
 165     int mb_type_col[2];
 166     const int16_t (*l1mv0)[2], (*l1mv1)[2];
 167     const int8_t *l1ref0, *l1ref1;
 168     const int is_b8x8 = IS_8X8(*mb_type);
 169     unsigned int sub_mb_type= MB_TYPE_L0L1;
 170     int i8, i4;
 171     int ref[2];
 172     int mv[2];
 173     int list;
 174
 175     assert(h->ref_list[1][0].reference & 3);
 176
 177     await_reference_mb_row(h, &h->ref_list[1][0], h->mb_y + !!IS_INTERLACED(*mb_type));
 178
 179 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
 180
 181
 182     /* ref = min(neighbors) */
 183     for(list=0; list<2; list++){
 184         int left_ref = h->ref_cache[list][scan8[0] - 1];
 185         int top_ref  = h->ref_cache[list][scan8[0] - 8];
 186         int refc = h->ref_cache[list][scan8[0] - 8 + 4];
 187         const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
 188         if(refc == PART_NOT_AVAILABLE){
 189             refc = h->ref_cache[list][scan8[0] - 8 - 1];
 190             C    = h-> mv_cache[list][scan8[0] - 8 - 1];
 191         }
 192         ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
 193         if(ref[list] >= 0){
 194             //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
 195             const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
 196             const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
 197
 198             int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
 199             if(match_count > 1){ //most common
 200                 mv[list]= pack16to32(mid_pred(A[0], B[0], C[0]),
 201                                      mid_pred(A[1], B[1], C[1]) );
 202             }else {
 203                 assert(match_count==1);
 204                 if(left_ref==ref[list]){
 205                     mv[list]= AV_RN32A(A);
 206                 }else if(top_ref==ref[list]){
 207                     mv[list]= AV_RN32A(B);
 208                 }else{
 209                     mv[list]= AV_RN32A(C);
 210                 }
 211             }
 212         }else{
 213             int mask= ~(MB_TYPE_L0 << (2*list));
 214             mv[list] = 0;
 215             ref[list] = -1;
 216             if(!is_b8x8)
 217                 *mb_type &= mask;
 218             sub_mb_type &= mask;
 219         }
 220     }
 221     if(ref[0] < 0 && ref[1] < 0){
 222         ref[0] = ref[1] = 0;
 223         if(!is_b8x8)
 224             *mb_type |= MB_TYPE_L0L1;
 225         sub_mb_type |= MB_TYPE_L0L1;
 226     }
 227
 228     if(!(is_b8x8|mv[0]|mv[1])){
 229         fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
 230         fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
 231         fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
 232         fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
 233         *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
 234         return;
 235     }
 236
 237     if (IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
 238         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
 239             mb_y = (h->mb_y&~1) + h->col_parity;
 240             mb_xy= h->mb_x + ((h->mb_y&~1) + h->col_parity)*h->mb_stride;
 241             b8_stride = 0;
 242         }else{
 243             mb_y  += h->col_fieldoff;
 244             mb_xy += h->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
 245         }
 246         goto single_col;
 247     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
 248         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
 249             mb_y = h->mb_y&~1;
 250             mb_xy= h->mb_x + (h->mb_y&~1)*h->mb_stride;
 251             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
 252             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + h->mb_stride];
 253             b8_stride = 2+4*h->mb_stride;
 254             b4_stride *= 6;
 255             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
 256                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
 257                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
 258             }
 259
 260             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
 261             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
 262                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
 263                 && !is_b8x8){
 264                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
 265             }else{
 266                 *mb_type   |= MB_TYPE_8x8;
 267             }
 268         }else{                                           //     AFR/FR    -> AFR/FR
 269 single_col:
 270             mb_type_col[0] =
 271             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
 272
 273             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
 274             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
 275                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */
 276             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
 277                 *mb_type   |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
 278             }else{
 279                 if(!h->sps.direct_8x8_inference_flag){
 280                     /* FIXME save sub mb types from previous frames (or derive from MVs)
 281                     * so we know exactly what block size to use */
 282                     sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
 283                 }
 284                 *mb_type   |= MB_TYPE_8x8;
 285             }
 286         }
 287     }
 288
 289     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
 290
 291     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
 292     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
 293     l1ref0 = &h->ref_list[1][0].ref_index [0][4 * mb_xy];
 294     l1ref1 = &h->ref_list[1][0].ref_index [1][4 * mb_xy];
 295     if(!b8_stride){
 296         if(h->mb_y&1){
 297             l1ref0 += 2;
 298             l1ref1 += 2;
 299             l1mv0  +=  2*b4_stride;
 300             l1mv1  +=  2*b4_stride;
 301         }
 302     }
 303
 304
 305         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
 306             int n=0;
 307             for(i8=0; i8<4; i8++){
 308                 int x8 = i8&1;
 309                 int y8 = i8>>1;
 310                 int xy8 = x8+y8*b8_stride;
 311                 int xy4 = 3*x8+y8*b4_stride;
 312                 int a,b;
 313
 314                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
 315                     continue;
 316                 h->sub_mb_type[i8] = sub_mb_type;
 317
 318                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
 319                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
 320                 if(!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref
 321                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
 322                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
 323                     a=b=0;
 324                     if(ref[0] > 0)
 325                         a= mv[0];
 326                     if(ref[1] > 0)
 327                         b= mv[1];
 328                     n++;
 329                 }else{
 330                     a= mv[0];
 331                     b= mv[1];
 332                 }
 333                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
 334                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
 335             }
 336             if(!is_b8x8 && !(n&3))
 337                 *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
 338         }else if(IS_16X16(*mb_type)){
 339             int a,b;
 340
 341             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
 342             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
 343             if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref
 344                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
 345                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
 346                        && h->x264_build>33U))){
 347                 a=b=0;
 348                 if(ref[0] > 0)
 349                     a= mv[0];
 350                 if(ref[1] > 0)
 351                     b= mv[1];
 352             }else{
 353                 a= mv[0];
 354                 b= mv[1];
 355             }
 356             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
 357             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
 358         }else{
 359             int n=0;
 360             for(i8=0; i8<4; i8++){
 361                 const int x8 = i8&1;
 362                 const int y8 = i8>>1;
 363
 364                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
 365                     continue;
 366                 h->sub_mb_type[i8] = sub_mb_type;
 367
 368                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, mv[0], 4);
 369                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, mv[1], 4);
 370                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
 371                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
 372
 373                 assert(b8_stride==2);
 374                 /* col_zero_flag */
 375                 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && (   l1ref0[i8] == 0
 376                                               || (l1ref0[i8] < 0 && l1ref1[i8] == 0
 377                                                   && h->x264_build>33U))){
 378                     const int16_t (*l1mv)[2]= l1ref0[i8] == 0 ? l1mv0 : l1mv1;
 379                     if(IS_SUB_8X8(sub_mb_type)){
 380                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
 381                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
 382                             if(ref[0] == 0)
 383                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
 384                             if(ref[1] == 0)
 385                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
 386                             n+=4;
 387                         }
 388                     }else{
 389                         int m=0;
 390                     for(i4=0; i4<4; i4++){
 391                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
 392                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
 393                             if(ref[0] == 0)
 394                                 AV_ZERO32(h->mv_cache[0][scan8[i8*4+i4]]);
 395                             if(ref[1] == 0)
 396                                 AV_ZERO32(h->mv_cache[1][scan8[i8*4+i4]]);
 397                             m++;
 398                         }
 399                     }
 400                     if(!(m&3))
 401                         h->sub_mb_type[i8]+= MB_TYPE_16x16 - MB_TYPE_8x8;
 402                     n+=m;
 403                     }
 404                 }
 405             }
 406             if(!is_b8x8 && !(n&15))
 407                 *mb_type= (*mb_type & ~(MB_TYPE_8x8|MB_TYPE_16x8|MB_TYPE_8x16|MB_TYPE_P1L0|MB_TYPE_P1L1))|MB_TYPE_16x16|MB_TYPE_DIRECT2;
 408         }
 409 }
 410
 411 static void pred_temp_direct_motion(H264Context * const h, int *mb_type){
 412     int b8_stride = 2;
 413     int b4_stride = h->b_stride;
 414     int mb_xy = h->mb_xy, mb_y = h->mb_y;
 415     int mb_type_col[2];
 416     const int16_t (*l1mv0)[2], (*l1mv1)[2];
 417     const int8_t *l1ref0, *l1ref1;
 418     const int is_b8x8 = IS_8X8(*mb_type);
 419     unsigned int sub_mb_type;
 420     int i8, i4;
 421
 422     assert(h->ref_list[1][0].reference & 3);
 423
 424     await_reference_mb_row(h, &h->ref_list[1][0], h->mb_y + !!IS_INTERLACED(*mb_type));
 425
 426     if (IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
 427         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
 428             mb_y = (h->mb_y&~1) + h->col_parity;
 429             mb_xy= h->mb_x + ((h->mb_y&~1) + h->col_parity)*h->mb_stride;
 430             b8_stride = 0;
 431         }else{
 432             mb_y  += h->col_fieldoff;
 433             mb_xy += h->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
 434         }
 435         goto single_col;
 436     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
 437         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
 438             mb_y = h->mb_y&~1;
 439             mb_xy= h->mb_x + (h->mb_y&~1)*h->mb_stride;
 440             mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
 441             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + h->mb_stride];
 442             b8_stride = 2+4*h->mb_stride;
 443             b4_stride *= 6;
 444             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
 445                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
 446                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
 447             }
 448
 449             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
 450
 451             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
 452                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
 453                 && !is_b8x8){
 454                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
 455             }else{
 456                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
 457             }
 458         }else{                                           //     AFR/FR    -> AFR/FR
 459 single_col:
 460             mb_type_col[0] =
 461             mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
 462
 463             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
 464             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
 465                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
 466             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
 467                 *mb_type   |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
 468             }else{
 469                 if(!h->sps.direct_8x8_inference_flag){
 470                     /* FIXME save sub mb types from previous frames (or derive from MVs)
 471                     * so we know exactly what block size to use */
 472                     sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
 473                 }
 474                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
 475             }
 476         }
 477     }
 478
 479     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
 480
 481     l1mv0  = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
 482     l1mv1  = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
 483     l1ref0 = &h->ref_list[1][0].ref_index [0][4 * mb_xy];
 484     l1ref1 = &h->ref_list[1][0].ref_index [1][4 * mb_xy];
 485     if(!b8_stride){
 486         if(h->mb_y&1){
 487             l1ref0 += 2;
 488             l1ref1 += 2;
 489             l1mv0  +=  2*b4_stride;
 490             l1mv1  +=  2*b4_stride;
 491         }
 492     }
 493
 494     {
 495         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
 496         const int *dist_scale_factor = h->dist_scale_factor;
 497         int ref_offset;
 498
 499         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
 500             map_col_to_list0[0] = h->map_col_to_list0_field[h->mb_y&1][0];
 501             map_col_to_list0[1] = h->map_col_to_list0_field[h->mb_y&1][1];
 502             dist_scale_factor   =h->dist_scale_factor_field[h->mb_y&1];
 503         }
 504         ref_offset = (h->ref_list[1][0].mbaff<<4) & (mb_type_col[0]>>3); //if(h->ref_list[1][0].mbaff && IS_INTERLACED(mb_type_col[0])) ref_offset=16 else 0
 505
 506         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
 507             int y_shift  = 2*!IS_INTERLACED(*mb_type);
 508             assert(h->sps.direct_8x8_inference_flag);
 509
 510             for(i8=0; i8<4; i8++){
 511                 const int x8 = i8&1;
 512                 const int y8 = i8>>1;
 513                 int ref0, scale;
 514                 const int16_t (*l1mv)[2]= l1mv0;
 515
 516                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
 517                     continue;
 518                 h->sub_mb_type[i8] = sub_mb_type;
 519
 520                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
 521                 if(IS_INTRA(mb_type_col[y8])){
 522                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
 523                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
 524                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
 525                     continue;
 526                 }
 527
 528                 ref0 = l1ref0[x8 + y8*b8_stride];
 529                 if(ref0 >= 0)
 530                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
 531                 else{
 532                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
 533                     l1mv= l1mv1;
 534                 }
 535                 scale = dist_scale_factor[ref0];
 536                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
 537
 538                 {
 539                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
 540                     int my_col = (mv_col[1]<<y_shift)/2;
 541                     int mx = (scale * mv_col[0] + 128) >> 8;
 542                     int my = (scale * my_col + 128) >> 8;
 543                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
 544                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
 545                 }
 546             }
 547             return;
 548         }
 549
 550         /* one-to-one mv scaling */
 551
 552         if(IS_16X16(*mb_type)){
 553             int ref, mv0, mv1;
 554
 555             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
 556             if(IS_INTRA(mb_type_col[0])){
 557                 ref=mv0=mv1=0;
 558             }else{
 559                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
 560                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
 561                 const int scale = dist_scale_factor[ref0];
 562                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
 563                 int mv_l0[2];
 564                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
 565                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
 566                 ref= ref0;
 567                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
 568                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
 569             }
 570             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
 571             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
 572             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
 573         }else{
 574             for(i8=0; i8<4; i8++){
 575                 const int x8 = i8&1;
 576                 const int y8 = i8>>1;
 577                 int ref0, scale;
 578                 const int16_t (*l1mv)[2]= l1mv0;
 579
 580                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
 581                     continue;
 582                 h->sub_mb_type[i8] = sub_mb_type;
 583                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
 584                 if(IS_INTRA(mb_type_col[0])){
 585                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
 586                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
 587                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
 588                     continue;
 589                 }
 590
 591                 assert(b8_stride == 2);
 592                 ref0 = l1ref0[i8];
 593                 if(ref0 >= 0)
 594                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
 595                 else{
 596                     ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];
 597                     l1mv= l1mv1;
 598                 }
 599                 scale = dist_scale_factor[ref0];
 600
 601                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
 602                 if(IS_SUB_8X8(sub_mb_type)){
 603                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
 604                     int mx = (scale * mv_col[0] + 128) >> 8;
 605                     int my = (scale * mv_col[1] + 128) >> 8;
 606                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
 607                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
 608                 }else
 609                 for(i4=0; i4<4; i4++){
 610                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
 611                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
 612                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
 613                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
 614                     AV_WN32A(h->mv_cache[1][scan8[i8*4+i4]],
 615                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]));
 616                 }
 617             }
 618         }
 619     }
 620 }
 621
 622 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type){
 623     if(h->direct_spatial_mv_pred){
 624         pred_spatial_direct_motion(h, mb_type);
 625     }else{
 626         pred_temp_direct_motion(h, mb_type);
 627     }
 628 }