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((AVFrame*)ref, FFMIN(16*mb_y >> ref_field_picture, ref_height-1),
00158                              ref_field_picture && ref_field);
00159 }
00160 
00161 static void pred_spatial_direct_motion(H264Context * const h, int *mb_type){
00162     MpegEncContext * const s = &h->s;
00163     int b8_stride = 2;
00164     int b4_stride = h->b_stride;
00165     int mb_xy = h->mb_xy, mb_y = s->mb_y;
00166     int mb_type_col[2];
00167     const int16_t (*l1mv0)[2], (*l1mv1)[2];
00168     const int8_t *l1ref0, *l1ref1;
00169     const int is_b8x8 = IS_8X8(*mb_type);
00170     unsigned int sub_mb_type= MB_TYPE_L0L1;
00171     int i8, i4;
00172     int ref[2];
00173     int mv[2];
00174     int list;
00175 
00176     assert(h->ref_list[1][0].f.reference & 3);
00177 
00178     await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type));
00179 
00180 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
00181 
00182 
00183     /* ref = min(neighbors) */
00184     for(list=0; list<2; list++){
00185         int left_ref = h->ref_cache[list][scan8[0] - 1];
00186         int top_ref  = h->ref_cache[list][scan8[0] - 8];
00187         int refc = h->ref_cache[list][scan8[0] - 8 + 4];
00188         const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
00189         if(refc == PART_NOT_AVAILABLE){
00190             refc = h->ref_cache[list][scan8[0] - 8 - 1];
00191             C    = h-> mv_cache[list][scan8[0] - 8 - 1];
00192         }
00193         ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
00194         if(ref[list] >= 0){
00195             //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
00196             const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
00197             const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
00198 
00199             int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
00200             if(match_count > 1){ //most common
00201                 mv[list]= pack16to32(mid_pred(A[0], B[0], C[0]),
00202                                      mid_pred(A[1], B[1], C[1]) );
00203             }else {
00204                 assert(match_count==1);
00205                 if(left_ref==ref[list]){
00206                     mv[list]= AV_RN32A(A);
00207                 }else if(top_ref==ref[list]){
00208                     mv[list]= AV_RN32A(B);
00209                 }else{
00210                     mv[list]= AV_RN32A(C);
00211                 }
00212             }
00213         }else{
00214             int mask= ~(MB_TYPE_L0 << (2*list));
00215             mv[list] = 0;
00216             ref[list] = -1;
00217             if(!is_b8x8)
00218                 *mb_type &= mask;
00219             sub_mb_type &= mask;
00220         }
00221     }
00222     if(ref[0] < 0 && ref[1] < 0){
00223         ref[0] = ref[1] = 0;
00224         if(!is_b8x8)
00225             *mb_type |= MB_TYPE_L0L1;
00226         sub_mb_type |= MB_TYPE_L0L1;
00227     }
00228 
00229     if(!(is_b8x8|mv[0]|mv[1])){
00230         fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
00231         fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
00232         fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
00233         fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
00234         *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;
00235         return;
00236     }
00237 
00238     if (IS_INTERLACED(h->ref_list[1][0].f.mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
00239         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
00240             mb_y = (s->mb_y&~1) + h->col_parity;
00241             mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
00242             b8_stride = 0;
00243         }else{
00244             mb_y  += h->col_fieldoff;
00245             mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
00246         }
00247         goto single_col;
00248     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
00249         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
00250             mb_y = s->mb_y&~1;
00251             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
00252             mb_type_col[0] = h->ref_list[1][0].f.mb_type[mb_xy];
00253             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy + s->mb_stride];
00254             b8_stride = 2+4*s->mb_stride;
00255             b4_stride *= 6;
00256             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
00257                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
00258                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
00259             }
00260 
00261             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00262             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
00263                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
00264                 && !is_b8x8){
00265                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_DIRECT2; /* B_16x8 */
00266             }else{
00267                 *mb_type   |= MB_TYPE_8x8;
00268             }
00269         }else{                                           //     AFR/FR    -> AFR/FR
00270 single_col:
00271             mb_type_col[0] =
00272             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy];
00273 
00274             sub_mb_type |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00275             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
00276                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_DIRECT2; /* B_16x16 */
00277             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
00278                 *mb_type   |= MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
00279             }else{
00280                 if(!h->sps.direct_8x8_inference_flag){
00281                     /* FIXME save sub mb types from previous frames (or derive from MVs)
00282                     * so we know exactly what block size to use */
00283                     sub_mb_type += (MB_TYPE_8x8-MB_TYPE_16x16); /* B_SUB_4x4 */
00284                 }
00285                 *mb_type   |= MB_TYPE_8x8;
00286             }
00287         }
00288     }
00289 
00290     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
00291 
00292     l1mv0  = &h->ref_list[1][0].f.motion_val[0][h->mb2b_xy [mb_xy]];
00293     l1mv1  = &h->ref_list[1][0].f.motion_val[1][h->mb2b_xy [mb_xy]];
00294     l1ref0 = &h->ref_list[1][0].f.ref_index [0][4 * mb_xy];
00295     l1ref1 = &h->ref_list[1][0].f.ref_index [1][4 * mb_xy];
00296     if(!b8_stride){
00297         if(s->mb_y&1){
00298             l1ref0 += 2;
00299             l1ref1 += 2;
00300             l1mv0  +=  2*b4_stride;
00301             l1mv1  +=  2*b4_stride;
00302         }
00303     }
00304 
00305 
00306         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
00307             int n=0;
00308             for(i8=0; i8<4; i8++){
00309                 int x8 = i8&1;
00310                 int y8 = i8>>1;
00311                 int xy8 = x8+y8*b8_stride;
00312                 int xy4 = 3*x8+y8*b4_stride;
00313                 int a,b;
00314 
00315                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00316                     continue;
00317                 h->sub_mb_type[i8] = sub_mb_type;
00318 
00319                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
00320                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
00321                 if(!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref
00322                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
00323                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
00324                     a=b=0;
00325                     if(ref[0] > 0)
00326                         a= mv[0];
00327                     if(ref[1] > 0)
00328                         b= mv[1];
00329                     n++;
00330                 }else{
00331                     a= mv[0];
00332                     b= mv[1];
00333                 }
00334                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
00335                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
00336             }
00337             if(!is_b8x8 && !(n&3))
00338                 *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;
00339         }else if(IS_16X16(*mb_type)){
00340             int a,b;
00341 
00342             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
00343             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
00344             if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref
00345                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
00346                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
00347                        && h->x264_build>33U))){
00348                 a=b=0;
00349                 if(ref[0] > 0)
00350                     a= mv[0];
00351                 if(ref[1] > 0)
00352                     b= mv[1];
00353             }else{
00354                 a= mv[0];
00355                 b= mv[1];
00356             }
00357             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
00358             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
00359         }else{
00360             int n=0;
00361             for(i8=0; i8<4; i8++){
00362                 const int x8 = i8&1;
00363                 const int y8 = i8>>1;
00364 
00365                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00366                     continue;
00367                 h->sub_mb_type[i8] = sub_mb_type;
00368 
00369                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, mv[0], 4);
00370                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, mv[1], 4);
00371                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
00372                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
00373 
00374                 assert(b8_stride==2);
00375                 /* col_zero_flag */
00376                 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && (   l1ref0[i8] == 0
00377                                               || (l1ref0[i8] < 0 && l1ref1[i8] == 0
00378                                                   && h->x264_build>33U))){
00379                     const int16_t (*l1mv)[2]= l1ref0[i8] == 0 ? l1mv0 : l1mv1;
00380                     if(IS_SUB_8X8(sub_mb_type)){
00381                         const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
00382                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
00383                             if(ref[0] == 0)
00384                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00385                             if(ref[1] == 0)
00386                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00387                             n+=4;
00388                         }
00389                     }else{
00390                         int m=0;
00391                     for(i4=0; i4<4; i4++){
00392                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
00393                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
00394                             if(ref[0] == 0)
00395                                 AV_ZERO32(h->mv_cache[0][scan8[i8*4+i4]]);
00396                             if(ref[1] == 0)
00397                                 AV_ZERO32(h->mv_cache[1][scan8[i8*4+i4]]);
00398                             m++;
00399                         }
00400                     }
00401                     if(!(m&3))
00402                         h->sub_mb_type[i8]+= MB_TYPE_16x16 - MB_TYPE_8x8;
00403                     n+=m;
00404                     }
00405                 }
00406             }
00407             if(!is_b8x8 && !(n&15))
00408                 *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;
00409         }
00410 }
00411 
00412 static void pred_temp_direct_motion(H264Context * const h, int *mb_type){
00413     MpegEncContext * const s = &h->s;
00414     int b8_stride = 2;
00415     int b4_stride = h->b_stride;
00416     int mb_xy = h->mb_xy, mb_y = s->mb_y;
00417     int mb_type_col[2];
00418     const int16_t (*l1mv0)[2], (*l1mv1)[2];
00419     const int8_t *l1ref0, *l1ref1;
00420     const int is_b8x8 = IS_8X8(*mb_type);
00421     unsigned int sub_mb_type;
00422     int i8, i4;
00423 
00424     assert(h->ref_list[1][0].f.reference & 3);
00425 
00426     await_reference_mb_row(h, &h->ref_list[1][0], s->mb_y + !!IS_INTERLACED(*mb_type));
00427 
00428     if (IS_INTERLACED(h->ref_list[1][0].f.mb_type[mb_xy])) { // AFL/AFR/FR/FL -> AFL/FL
00429         if (!IS_INTERLACED(*mb_type)) {                          //     AFR/FR    -> AFL/FL
00430             mb_y = (s->mb_y&~1) + h->col_parity;
00431             mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
00432             b8_stride = 0;
00433         }else{
00434             mb_y  += h->col_fieldoff;
00435             mb_xy += s->mb_stride*h->col_fieldoff; // non zero for FL -> FL & differ parity
00436         }
00437         goto single_col;
00438     }else{                                               // AFL/AFR/FR/FL -> AFR/FR
00439         if(IS_INTERLACED(*mb_type)){                     // AFL       /FL -> AFR/FR
00440             mb_y = s->mb_y&~1;
00441             mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
00442             mb_type_col[0] = h->ref_list[1][0].f.mb_type[mb_xy];
00443             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy + s->mb_stride];
00444             b8_stride = 2+4*s->mb_stride;
00445             b4_stride *= 6;
00446             if (IS_INTERLACED(mb_type_col[0]) != IS_INTERLACED(mb_type_col[1])) {
00447                 mb_type_col[0] &= ~MB_TYPE_INTERLACED;
00448                 mb_type_col[1] &= ~MB_TYPE_INTERLACED;
00449             }
00450 
00451             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00452 
00453             if(    (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
00454                 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
00455                 && !is_b8x8){
00456                 *mb_type   |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
00457             }else{
00458                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
00459             }
00460         }else{                                           //     AFR/FR    -> AFR/FR
00461 single_col:
00462             mb_type_col[0] =
00463             mb_type_col[1] = h->ref_list[1][0].f.mb_type[mb_xy];
00464 
00465             sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
00466             if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
00467                 *mb_type   |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
00468             }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
00469                 *mb_type   |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
00470             }else{
00471                 if(!h->sps.direct_8x8_inference_flag){
00472                     /* FIXME save sub mb types from previous frames (or derive from MVs)
00473                     * so we know exactly what block size to use */
00474                     sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
00475                 }
00476                 *mb_type   |= MB_TYPE_8x8|MB_TYPE_L0L1;
00477             }
00478         }
00479     }
00480 
00481     await_reference_mb_row(h, &h->ref_list[1][0], mb_y);
00482 
00483     l1mv0  = &h->ref_list[1][0].f.motion_val[0][h->mb2b_xy [mb_xy]];
00484     l1mv1  = &h->ref_list[1][0].f.motion_val[1][h->mb2b_xy [mb_xy]];
00485     l1ref0 = &h->ref_list[1][0].f.ref_index [0][4 * mb_xy];
00486     l1ref1 = &h->ref_list[1][0].f.ref_index [1][4 * mb_xy];
00487     if(!b8_stride){
00488         if(s->mb_y&1){
00489             l1ref0 += 2;
00490             l1ref1 += 2;
00491             l1mv0  +=  2*b4_stride;
00492             l1mv1  +=  2*b4_stride;
00493         }
00494     }
00495 
00496     {
00497         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
00498         const int *dist_scale_factor = h->dist_scale_factor;
00499         int ref_offset;
00500 
00501         if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
00502             map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
00503             map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
00504             dist_scale_factor   =h->dist_scale_factor_field[s->mb_y&1];
00505         }
00506         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
00507 
00508         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
00509             int y_shift  = 2*!IS_INTERLACED(*mb_type);
00510             assert(h->sps.direct_8x8_inference_flag);
00511 
00512             for(i8=0; i8<4; i8++){
00513                 const int x8 = i8&1;
00514                 const int y8 = i8>>1;
00515                 int ref0, scale;
00516                 const int16_t (*l1mv)[2]= l1mv0;
00517 
00518                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00519                     continue;
00520                 h->sub_mb_type[i8] = sub_mb_type;
00521 
00522                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
00523                 if(IS_INTRA(mb_type_col[y8])){
00524                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
00525                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00526                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00527                     continue;
00528                 }
00529 
00530                 ref0 = l1ref0[x8 + y8*b8_stride];
00531                 if(ref0 >= 0)
00532                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
00533                 else{
00534                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
00535                     l1mv= l1mv1;
00536                 }
00537                 scale = dist_scale_factor[ref0];
00538                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
00539 
00540                 {
00541                     const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
00542                     int my_col = (mv_col[1]<<y_shift)/2;
00543                     int mx = (scale * mv_col[0] + 128) >> 8;
00544                     int my = (scale * my_col + 128) >> 8;
00545                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
00546                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
00547                 }
00548             }
00549             return;
00550         }
00551 
00552         /* one-to-one mv scaling */
00553 
00554         if(IS_16X16(*mb_type)){
00555             int ref, mv0, mv1;
00556 
00557             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
00558             if(IS_INTRA(mb_type_col[0])){
00559                 ref=mv0=mv1=0;
00560             }else{
00561                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
00562                                                 : map_col_to_list0[1][l1ref1[0] + ref_offset];
00563                 const int scale = dist_scale_factor[ref0];
00564                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
00565                 int mv_l0[2];
00566                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
00567                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
00568                 ref= ref0;
00569                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
00570                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
00571             }
00572             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
00573             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
00574             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
00575         }else{
00576             for(i8=0; i8<4; i8++){
00577                 const int x8 = i8&1;
00578                 const int y8 = i8>>1;
00579                 int ref0, scale;
00580                 const int16_t (*l1mv)[2]= l1mv0;
00581 
00582                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
00583                     continue;
00584                 h->sub_mb_type[i8] = sub_mb_type;
00585                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
00586                 if(IS_INTRA(mb_type_col[0])){
00587                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
00588                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
00589                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
00590                     continue;
00591                 }
00592 
00593                 assert(b8_stride == 2);
00594                 ref0 = l1ref0[i8];
00595                 if(ref0 >= 0)
00596                     ref0 = map_col_to_list0[0][ref0 + ref_offset];
00597                 else{
00598                     ref0 = map_col_to_list0[1][l1ref1[i8] + ref_offset];
00599                     l1mv= l1mv1;
00600                 }
00601                 scale = dist_scale_factor[ref0];
00602 
00603                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
00604                 if(IS_SUB_8X8(sub_mb_type)){
00605                     const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
00606                     int mx = (scale * mv_col[0] + 128) >> 8;
00607                     int my = (scale * mv_col[1] + 128) >> 8;
00608                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
00609                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
00610                 }else
00611                 for(i4=0; i4<4; i4++){
00612                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
00613                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
00614                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
00615                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
00616                     AV_WN32A(h->mv_cache[1][scan8[i8*4+i4]],
00617                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]));
00618                 }
00619             }
00620         }
00621     }
00622 }
00623 
00624 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type){
00625     if(h->direct_spatial_mv_pred){
00626         pred_spatial_direct_motion(h, mb_type);
00627     }else{
00628         pred_temp_direct_motion(h, mb_type);
00629     }
00630 }