FFmpeg: libavcodec/h264_direct.c Source File

00001 /*
00002  * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding
00003  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
00004  *
00005  * This file is part of FFmpeg.
00006  *
00007  * FFmpeg is free software; you can redistribute it and/or
00008  * modify it under the terms of the GNU Lesser General Public
00009  * License as published by the Free Software Foundation; either
00010  * version 2.1 of the License, or (at your option) any later version.
00011  *
00012  * FFmpeg is distributed in the hope that it will be useful,
00013  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00014  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00015  * Lesser General Public License for more details.
00016  *
00017  * You should have received a copy of the GNU Lesser General Public
00018  * License along with FFmpeg; if not, write to the Free Software
00019  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00020  */
00021 
00028 #include "internal.h"
00029 #include "dsputil.h"
00030 #include "avcodec.h"
00031 #include "mpegvideo.h"
00032 #include "h264.h"
00033 #include "rectangle.h"
00034 #include "thread.h"
00035 
00036 //#undef NDEBUG
00037 #include <assert.h>
00038 
00039 
00040 static int get_scale_factor(H264Context * const h, int poc, int poc1, int i){
00041     int poc0 = h->ref_list[0][i].poc;
00042     int td = av_clip(poc1 - poc0, -128, 127);
00043     if(td == 0 || h->ref_list[0][i].long_ref){
00044         return 256;
00045     }else{
00046         int tb = av_clip(poc - poc0, -128, 127);
00047         int tx = (16384 + (FFABS(td) >> 1)) / td;
00048         return av_clip((tb*tx + 32) >> 6, -1024, 1023);
00049     }
00050 }
00051 
00052 void ff_h264_direct_dist_scale_factor(H264Context * const h){
00053     MpegEncContext * const s = &h->s;
00054     const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
00055     const int poc1 = h->ref_list[1][0].poc;
00056     int i, field;
00057     for(field=0; field<2; field++){
00058         const int poc  = h->s.current_picture_ptr->field_poc[field];
00059         const int poc1 = h->ref_list[1][0].field_poc[field];
00060         for(i=0; i < 2*h->ref_count[0]; i++)
00061             h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
00062     }
00063 
00064     for(i=0; i<h->ref_count[0]; i++){
00065         h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
00066     }
00067 }
00068 
00069 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
00070     MpegEncContext * const s = &h->s;
00071     Picture * const ref1 = &h->ref_list[1][0];
00072     int j, old_ref, rfield;
00073     int start= mbafi ? 16                      : 0;
00074     int end  = mbafi ? 16+2*h->ref_count[0]    : h->ref_count[0];
00075     int interl= mbafi || s->picture_structure != PICT_FRAME;
00076 
00077     /* bogus; fills in for missing frames */
00078     memset(map[list], 0, sizeof(map[list]));
00079 
00080     for(rfield=0; rfield<2; rfield++){
00081         for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
00082             int poc = ref1->ref_poc[colfield][list][old_ref];
00083 
00084             if     (!interl)
00085                 poc |= 3;
00086             else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
00087                 poc= (poc&~3) + rfield + 1;
00088 
00089             for(j=start; j<end; j++){
00090                 if (4 * h->ref_list[0][j].frame_num + (h->ref_list[0][j].f.reference & 3) == poc) {
00091                     int cur_ref= mbafi ? (j-16)^field : j;
00092                     if(ref1->mbaff)
00093                         map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
00094                     if(rfield == field || !interl)
00095                         map[list][old_ref] = cur_ref;
00096                     break;
00097                 }
00098             }
00099         }
00100     }
00101 }
00102 
00103 void ff_h264_direct_ref_list_init(H264Context * const h){
00104     MpegEncContext * const s = &h->s;
00105     Picture * const ref1 = &h->ref_list[1][0];
00106     Picture * const cur = s->current_picture_ptr;
00107     int list, j, field;
00108     int sidx= (s->picture_structure&1)^1;
00109     int ref1sidx = (ref1->f.reference&1)^1;
00110 
00111     for(list=0; list<2; list++){
00112         cur->ref_count[sidx][list] = h->ref_count[list];
00113         for(j=0; j<h->ref_count[list]; j++)
00114             cur->ref_poc[sidx][list][j] = 4 * h->ref_list[list][j].frame_num + (h->ref_list[list][j].f.reference & 3);
00115     }
00116 
00117     if(s->picture_structure == PICT_FRAME){
00118         memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
00119         memcpy(cur->ref_poc  [1], cur->ref_poc  [0], sizeof(cur->ref_poc  [0]));
00120     }
00121 
00122     cur->mbaff= FRAME_MBAFF;
00123 
00124     h->col_fieldoff= 0;
00125     if(s->picture_structure == PICT_FRAME){
00126         int cur_poc = s->current_picture_ptr->poc;
00127         int *col_poc = h->ref_list[1]->field_poc;
00128         h->col_parity= (FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc));
00129         ref1sidx=sidx= h->col_parity;
00130     } else if (!(s->picture_structure & h->ref_list[1][0].f.reference) && !h->ref_list[1][0].mbaff) { // FL -> FL & differ parity
00131         h->col_fieldoff = 2 * h->ref_list[1][0].f.reference - 3;
00132     }
00133 
00134     if (cur->f.pict_type != AV_PICTURE_TYPE_B || h->direct_spatial_mv_pred)
00135         return;
00136 
00137     for(list=0; list<2; list++){
00138         fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
00139         if(FRAME_MBAFF)
00140         for(field=0; field<2; field++)
00141             fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
00142     }
00143 }
00144 
00145 static void await_reference_mb_row(H264Context * const h, Picture *ref, int mb_y)
00146 {
00147     int ref_field = ref->f.reference - 1;
00148     int ref_field_picture = ref->field_picture;
00149     int ref_height = 16*h->s.mb_height >> ref_field_picture;
00150 
00151     if(!HAVE_THREADS || !(h->s.avctx->active_thread_type&FF_THREAD_FRAME))
00152         return;
00153 
00154     //FIXME it can be safe to access mb stuff
00155     //even if pixels aren't deblocked yet
00156 
00157     ff_thread_await_progress(&ref->f,
00158                              FFMIN(16 * mb_y >> ref_field_picture, ref_height - 1),
00159                              ref_field_picture && ref_field);
00160 }
00161 
00162 static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
00163     MpegEncContext * const s = &h->s;
00164     int b8_stride = 2;
00165     int b4_stride = h->b_stride;
00166     int mb_xy = h->mb_xy, mb_y = s->mb_y;
00167     int mb_type_col[2];
00168     const int16_t (*l1mv0)[2], (*l1mv1)[2];
00169     const int8_t *l1ref0, *l1ref1;
00170     const int is_b8x8 = IS_8X8(*mb_type);
00171     unsigned int sub_mb_type= MB_TYPE_L0L1;
00172     int i8, i4;
00173     int ref[2];
00174     int mv[2];
00175     int list;
00176 
00177     assert(h->ref_list[1][0].f.reference & 3);
00178 
00179     await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type));
00180 
00181 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
00182 
00183 
00184     /* ref = min(neighbors) */
00185     for(list=0; list<2; list++){
00186         int left_ref = h->ref_cache[list][scan8[0] - 1];
00187         int top_ref  = h->ref_cache[list][scan8[0] - 8];
00188         int refc = h->ref_cache[list][scan8[0] - 8 + 4];
00189         const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
00190         if(refc == PART_NOT_AVAILABLE){
00191             refc = h->ref_cache[list][scan8[0] - 8 - 1];
00192             C    = h-> mv_cache[list][scan8[0] - 8 - 1];
00193         }
00194         ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
00195         if(ref[list] >= 0){
00196             //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
00197             const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
00198             const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
00199 
00200             int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
00201             if(match_count > 1){ //most common
00202                 mv[list]= pack16to32(mid_pred(A[0], B[0], C[0]),
00203                                      mid_pred(A[1], B[1], C[1]) );
00204             }else {
00205                 assert(match_count==1);
00206                 if(left_ref==ref[list]){
00207                     mv[list]= AV_RN32A(A);
00208                 }else if(top_ref==ref[list]){
00209                     mv[list]= AV_RN32A(B);
00210                 }else{
00211                     mv[list]= AV_RN32A(C);
00212                 }
00213             }
00214         }else{
00215             int mask= ~(MB_TYPE_L0 << (2*list));
00216             mv[list] = 0;
00217             ref[list] = -1;
00218             if(!is_b8x8)
00219                 *mb_type &= mask;
00220             sub_mb_type &= mask;
00221         }
00222     }
00223     if(ref[0] < 0 && ref[1] < 0){
00224         ref[0] = ref[1] = 0;
00225         if(!is_b8x8)
00226             *mb_type |= MB_TYPE_L0L1;
00227         sub_mb_type |= MB_TYPE_L0L1;
00228     }
00229 
00230     if(!(is_b8x8|mv[0]|mv[1])){
00231         fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
00232         fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
00233         fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
00234         fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
00235         *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;
00236         return;
00237     }
00238 
00239     if (IS_INTERLACED(h->ref_list[1][0].f.mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
00240         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
00241             mb_y = (s->mb_y&~1) + h->col_parity;
00242             mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
00243             b8_stride = 0;
00244         }else{
00245             mb_y  += h->col_fieldoff;
00246             mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
00247         }
00248         goto single_col;
00249     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
00250         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
00251             mb_y = s->mb_y&~1;
00252             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
00253             mb_type_col[0] = h->ref_list[1][0].f.mb_type[mb_xy];
00254             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy + s->mb_stride];
00255             b8_stride = 2+4*s->mb_stride;
00256             b4_stride *= 6;
00257             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
00258                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
00259                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
00260             }
00261 
00262             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00263             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
00264                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
00265                 && !is_b8x8){
00266                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
00267             }else{
00268                 *mb_type   |= MB_TYPE_8x8;
00269             }
00270         }else{                                           //     AFR/FR    -> AFR/FR
00271 single_col:
00272             mb_type_col[0] =
00273             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy];
00274 
00275             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00276             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
00277                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */
00278             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
00279                 *mb_type   |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
00280             }else{
00281                 if(!h->sps.direct_8x8_inference_flag){
00282                     /* FIXME save sub mb types from previous frames (or derive from MVs)
00283                     * so we know exactly what block size to use */
00284                     sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
00285                 }
00286                 *mb_type   |= MB_TYPE_8x8;
00287             }
00288         }
00289     }
00290 
00291     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
00292 
00293     l1mv0  = &h->ref_list[1][0].f.motion_val[0][h->mb2b_xy [mb_xy]];
00294     l1mv1  = &h->ref_list[1][0].f.motion_val[1][h->mb2b_xy [mb_xy]];
00295     l1ref0 = &h->ref_list[1][0].f.ref_index [0][4 * mb_xy];
00296     l1ref1 = &h->ref_list[1][0].f.ref_index [1][4 * mb_xy];
00297     if(!b8_stride){
00298         if(s->mb_y&1){
00299             l1ref0 += 2;
00300             l1ref1 += 2;
00301             l1mv0  +=  2*b4_stride;
00302             l1mv1  +=  2*b4_stride;
00303         }
00304     }
00305 
00306 
00307         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
00308             int n=0;
00309             for(i8=0; i8<4; i8++){
00310                 int x8 = i8&1;
00311                 int y8 = i8>>1;
00312                 int xy8 = x8+y8*b8_stride;
00313                 int xy4 = 3*x8+y8*b4_stride;
00314                 int a,b;
00315 
00316                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00317                     continue;
00318                 h->sub_mb_type[i8] = sub_mb_type;
00319 
00320                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
00321                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
00322                 if(!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref
00323                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
00324                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
00325                     a=b=0;
00326                     if(ref[0] > 0)
00327                         a= mv[0];
00328                     if(ref[1] > 0)
00329                         b= mv[1];
00330                     n++;
00331                 }else{
00332                     a= mv[0];
00333                     b= mv[1];
00334                 }
00335                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
00336                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
00337             }
00338             if(!is_b8x8 && !(n&3))
00339                 *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;
00340         }else if(IS_16X16(*mb_type)){
00341             int a,b;
00342 
00343             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
00344             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
00345             if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref
00346                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
00347                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
00348                        && h->x264_build>33U))){
00349                 a=b=0;
00350                 if(ref[0] > 0)
00351                     a= mv[0];
00352                 if(ref[1] > 0)
00353                     b= mv[1];
00354             }else{
00355                 a= mv[0];
00356                 b= mv[1];
00357             }
00358             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
00359             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
00360         }else{
00361             int n=0;
00362             for(i8=0; i8<4; i8++){
00363                 const int x8 = i8&1;
00364                 const int y8 = i8>>1;
00365 
00366                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00367                     continue;
00368                 h->sub_mb_type[i8] = sub_mb_type;
00369 
00370                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, mv[0], 4);
00371                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, mv[1], 4);
00372                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
00373                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
00374 
00375                 assert(b8_stride==2);
00376                 /* col_zero_flag */
00377                 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && (   l1ref0[i8] == 0
00378                                               || (l1ref0[i8] < 0 && l1ref1[i8] == 0
00379                                                   && h->x264_build>33U))){
00380                     const int16_t (*l1mv)[2]= l1ref0[i8] == 0 ? l1mv0 : l1mv1;
00381                     if(IS_SUB_8X8(sub_mb_type)){
00382                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
00383                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
00384                             if(ref[0] == 0)
00385                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00386                             if(ref[1] == 0)
00387                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00388                             n+=4;
00389                         }
00390                     }else{
00391                         int m=0;
00392                     for(i4=0; i4<4; i4++){
00393                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
00394                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
00395                             if(ref[0] == 0)
00396                                 AV_ZERO32(h->mv_cache[0][scan8[i8*4+i4]]);
00397                             if(ref[1] == 0)
00398                                 AV_ZERO32(h->mv_cache[1][scan8[i8*4+i4]]);
00399                             m++;
00400                         }
00401                     }
00402                     if(!(m&3))
00403                         h->sub_mb_type[i8]+= MB_TYPE_16x16 - MB_TYPE_8x8;
00404                     n+=m;
00405                     }
00406                 }
00407             }
00408             if(!is_b8x8 && !(n&15))
00409                 *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;
00410         }
00411 }
00412 
00413 static void pred_temp_direct_motion(H264Context * const h, int *mb_type){
00414     MpegEncContext * const s = &h->s;
00415     int b8_stride = 2;
00416     int b4_stride = h->b_stride;
00417     int mb_xy = h->mb_xy, mb_y = s->mb_y;
00418     int mb_type_col[2];
00419     const int16_t (*l1mv0)[2], (*l1mv1)[2];
00420     const int8_t *l1ref0, *l1ref1;
00421     const int is_b8x8 = IS_8X8(*mb_type);
00422     unsigned int sub_mb_type;
00423     int i8, i4;
00424 
00425     assert(h->ref_list[1][0].f.reference & 3);
00426 
00427     await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type));
00428 
00429     if (IS_INTERLACED(h->ref_list[1][0].f.mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
00430         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
00431             mb_y = (s->mb_y&~1) + h->col_parity;
00432             mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
00433             b8_stride = 0;
00434         }else{
00435             mb_y  += h->col_fieldoff;
00436             mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
00437         }
00438         goto single_col;
00439     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
00440         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
00441             mb_y = s->mb_y&~1;
00442             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
00443             mb_type_col[0] = h->ref_list[1][0].f.mb_type[mb_xy];
00444             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy + s->mb_stride];
00445             b8_stride = 2+4*s->mb_stride;
00446             b4_stride *= 6;
00447             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
00448                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
00449                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
00450             }
00451 
00452             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00453 
00454             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
00455                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
00456                 && !is_b8x8){
00457                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
00458             }else{
00459                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
00460             }
00461         }else{                                           //     AFR/FR    -> AFR/FR
00462 single_col:
00463             mb_type_col[0] =
00464             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy];
00465 
00466             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00467             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
00468                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
00469             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
00470                 *mb_type   |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
00471             }else{
00472                 if(!h->sps.direct_8x8_inference_flag){
00473                     /* FIXME save sub mb types from previous frames (or derive from MVs)
00474                     * so we know exactly what block size to use */
00475                     sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
00476                 }
00477                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
00478             }
00479         }
00480     }
00481 
00482     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
00483 
00484     l1mv0  = &h->ref_list[1][0].f.motion_val[0][h->mb2b_xy [mb_xy]];
00485     l1mv1  = &h->ref_list[1][0].f.motion_val[1][h->mb2b_xy [mb_xy]];
00486     l1ref0 = &h->ref_list[1][0].f.ref_index [0][4 * mb_xy];
00487     l1ref1 = &h->ref_list[1][0].f.ref_index [1][4 * mb_xy];
00488     if(!b8_stride){
00489         if(s->mb_y&1){
00490             l1ref0 += 2;
00491             l1ref1 += 2;
00492             l1mv0  +=  2*b4_stride;
00493             l1mv1  +=  2*b4_stride;
00494         }
00495     }
00496 
00497     {
00498         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
00499         const int *dist_scale_factor = h->dist_scale_factor;
00500         int ref_offset;
00501 
00502         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
00503             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
00504             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
00505             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
00506         }
00507         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
00508 
00509         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
00510             int y_shift  = 2*!IS_INTERLACED(*mb_type);
00511             assert(h->sps.direct_8x8_inference_flag);
00512 
00513             for(i8=0; i8<4; i8++){
00514                 const int x8 = i8&1;
00515                 const int y8 = i8>>1;
00516                 int ref0, scale;
00517                 const int16_t (*l1mv)[2]= l1mv0;
00518 
00519                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00520                     continue;
00521                 h->sub_mb_type[i8] = sub_mb_type;
00522 
00523                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
00524                 if(IS_INTRA(mb_type_col[y8])){
00525                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
00526                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00527                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00528                     continue;
00529                 }
00530 
00531                 ref0 = l1ref0[x8 + y8*b8_stride];
00532                 if(ref0 >= 0)
00533                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
00534                 else{
00535                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
00536                     l1mv= l1mv1;
00537                 }
00538                 scale = dist_scale_factor[ref0];
00539                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
00540 
00541                 {
00542                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
00543                     int my_col = (mv_col[1]<<y_shift)/2;
00544                     int mx = (scale * mv_col[0] + 128) >> 8;
00545                     int my = (scale * my_col + 128) >> 8;
00546                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
00547                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
00548                 }
00549             }
00550             return;
00551         }
00552 
00553         /* one-to-one mv scaling */
00554 
00555         if(IS_16X16(*mb_type)){
00556             int ref, mv0, mv1;
00557 
00558             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
00559             if(IS_INTRA(mb_type_col[0])){
00560                 ref=mv0=mv1=0;
00561             }else{
00562                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
00563                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
00564                 const int scale = dist_scale_factor[ref0];
00565                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
00566                 int mv_l0[2];
00567                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
00568                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
00569                 ref= ref0;
00570                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
00571                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
00572             }
00573             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
00574             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
00575             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
00576         }else{
00577             for(i8=0; i8<4; i8++){
00578                 const int x8 = i8&1;
00579                 const int y8 = i8>>1;
00580                 int ref0, scale;
00581                 const int16_t (*l1mv)[2]= l1mv0;
00582 
00583                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00584                     continue;
00585                 h->sub_mb_type[i8] = sub_mb_type;
00586                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
00587                 if(IS_INTRA(mb_type_col[0])){
00588                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
00589                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00590                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00591                     continue;
00592                 }
00593 
00594                 assert(b8_stride == 2);
00595                 ref0 = l1ref0[i8];
00596                 if(ref0 >= 0)
00597                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
00598                 else{
00599                     ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];
00600                     l1mv= l1mv1;
00601                 }
00602                 scale = dist_scale_factor[ref0];
00603 
00604                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
00605                 if(IS_SUB_8X8(sub_mb_type)){
00606                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
00607                     int mx = (scale * mv_col[0] + 128) >> 8;
00608                     int my = (scale * mv_col[1] + 128) >> 8;
00609                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
00610                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
00611                 }else
00612                 for(i4=0; i4<4; i4++){
00613                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
00614                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
00615                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
00616                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
00617                     AV_WN32A(h->mv_cache[1][scan8[i8*4+i4]],
00618                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]));
00619                 }
00620             }
00621         }
00622     }
00623 }
00624 
00625 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type){
00626     if(h->direct_spatial_mv_pred){
00627         pred_spatial_direct_motion(h, mb_type);
00628     }else{
00629         pred_temp_direct_motion(h, mb_type);
00630     }
00631 }