2 * H.26L/H.264/AVC/JVT/14496-10/... direct mb/block decoding
3 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
5 * This file is part of FFmpeg.
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.
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.
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
23 * @file libavcodec/h264_direct.c
24 * H.264 / AVC / MPEG4 part10 direct mb/block decoding.
25 * @author Michael Niedermayer <michaelni@gmx.at>
31 #include "mpegvideo.h"
33 #include "rectangle.h"
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){
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);
51 void ff_h264_direct_dist_scale_factor(H264Context * const h){
52 MpegEncContext * const s = &h->s;
53 const int poc = h->s.current_picture_ptr->field_poc[ s->picture_structure == PICT_BOTTOM_FIELD ];
54 const int poc1 = h->ref_list[1][0].poc;
56 for(field=0; field<2; field++){
57 const int poc = h->s.current_picture_ptr->field_poc[field];
58 const int poc1 = h->ref_list[1][0].field_poc[field];
59 for(i=0; i < 2*h->ref_count[0]; i++)
60 h->dist_scale_factor_field[field][i^field] = get_scale_factor(h, poc, poc1, i+16);
63 for(i=0; i<h->ref_count[0]; i++){
64 h->dist_scale_factor[i] = get_scale_factor(h, poc, poc1, i);
68 static void fill_colmap(H264Context *h, int map[2][16+32], int list, int field, int colfield, int mbafi){
69 MpegEncContext * const s = &h->s;
70 Picture * const ref1 = &h->ref_list[1][0];
71 int j, old_ref, rfield;
72 int start= mbafi ? 16 : 0;
73 int end = mbafi ? 16+2*h->ref_count[0] : h->ref_count[0];
74 int interl= mbafi || s->picture_structure != PICT_FRAME;
76 /* bogus; fills in for missing frames */
77 memset(map[list], 0, sizeof(map[list]));
79 for(rfield=0; rfield<2; rfield++){
80 for(old_ref=0; old_ref<ref1->ref_count[colfield][list]; old_ref++){
81 int poc = ref1->ref_poc[colfield][list][old_ref];
85 else if( interl && (poc&3) == 3) //FIXME store all MBAFF references so this isnt needed
86 poc= (poc&~3) + rfield + 1;
88 for(j=start; j<end; j++){
89 if(4*h->ref_list[0][j].frame_num + (h->ref_list[0][j].reference&3) == poc){
90 int cur_ref= mbafi ? (j-16)^field : j;
91 map[list][2*old_ref + (rfield^field) + 16] = cur_ref;
92 if(rfield == field || !interl)
93 map[list][old_ref] = cur_ref;
101 void ff_h264_direct_ref_list_init(H264Context * const h){
102 MpegEncContext * const s = &h->s;
103 Picture * const ref1 = &h->ref_list[1][0];
104 Picture * const cur = s->current_picture_ptr;
106 int sidx= (s->picture_structure&1)^1;
107 int ref1sidx= (ref1->reference&1)^1;
109 for(list=0; list<2; list++){
110 cur->ref_count[sidx][list] = h->ref_count[list];
111 for(j=0; j<h->ref_count[list]; j++)
112 cur->ref_poc[sidx][list][j] = 4*h->ref_list[list][j].frame_num + (h->ref_list[list][j].reference&3);
115 if(s->picture_structure == PICT_FRAME){
116 memcpy(cur->ref_count[1], cur->ref_count[0], sizeof(cur->ref_count[0]));
117 memcpy(cur->ref_poc [1], cur->ref_poc [0], sizeof(cur->ref_poc [0]));
120 cur->mbaff= FRAME_MBAFF;
123 if(s->picture_structure == PICT_FRAME){
124 int cur_poc = s->current_picture_ptr->poc;
125 int *col_poc = h->ref_list[1]->field_poc;
126 h->col_parity= (FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc));
127 ref1sidx=sidx= h->col_parity;
128 }else if(!(s->picture_structure & h->ref_list[1][0].reference) && !h->ref_list[1][0].mbaff){ // FL -> FL & differ parity
129 h->col_fieldoff= s->mb_stride*(2*(h->ref_list[1][0].reference) - 3);
132 if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
135 for(list=0; list<2; list++){
136 fill_colmap(h, h->map_col_to_list0, list, sidx, ref1sidx, 0);
138 for(field=0; field<2; field++)
139 fill_colmap(h, h->map_col_to_list0_field[field], list, field, field, 1);
143 void ff_h264_pred_direct_motion(H264Context * const h, int *mb_type){
144 MpegEncContext * const s = &h->s;
145 int b8_stride = h->b8_stride;
146 int b4_stride = h->b_stride;
147 int mb_xy = h->mb_xy;
149 const int16_t (*l1mv0)[2], (*l1mv1)[2];
150 const int8_t *l1ref0, *l1ref1;
151 const int is_b8x8 = IS_8X8(*mb_type);
152 unsigned int sub_mb_type;
155 assert(h->ref_list[1][0].reference&3);
157 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
159 if(IS_INTERLACED(h->ref_list[1][0].mb_type[mb_xy])){ // AFL/AFR/FR/FL -> AFL/FL
160 if(!IS_INTERLACED(*mb_type)){ // AFR/FR -> AFL/FL
161 mb_xy= s->mb_x + ((s->mb_y&~1) + h->col_parity)*s->mb_stride;
164 mb_xy += h->col_fieldoff; // non zero for FL -> FL & differ parity
167 }else{ // AFL/AFR/FR/FL -> AFR/FR
168 if(IS_INTERLACED(*mb_type)){ // AFL /FL -> AFR/FR
169 mb_xy= s->mb_x + (s->mb_y&~1)*s->mb_stride;
170 mb_type_col[0] = h->ref_list[1][0].mb_type[mb_xy];
171 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy + s->mb_stride];
175 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
176 if( (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)
177 && (mb_type_col[1] & MB_TYPE_16x16_OR_INTRA)
179 *mb_type |= MB_TYPE_16x8 |MB_TYPE_L0L1|MB_TYPE_DIRECT2; /* B_16x8 */
181 *mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
183 }else{ // AFR/FR -> AFR/FR
186 mb_type_col[1] = h->ref_list[1][0].mb_type[mb_xy];
188 sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
189 if(!is_b8x8 && (mb_type_col[0] & MB_TYPE_16x16_OR_INTRA)){
190 *mb_type |= MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
191 }else if(!is_b8x8 && (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16))){
192 *mb_type |= MB_TYPE_L0L1|MB_TYPE_DIRECT2 | (mb_type_col[0] & (MB_TYPE_16x8|MB_TYPE_8x16));
194 if(!h->sps.direct_8x8_inference_flag){
195 /* FIXME save sub mb types from previous frames (or derive from MVs)
196 * so we know exactly what block size to use */
197 sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
199 *mb_type |= MB_TYPE_8x8|MB_TYPE_L0L1;
204 l1mv0 = &h->ref_list[1][0].motion_val[0][h->mb2b_xy [mb_xy]];
205 l1mv1 = &h->ref_list[1][0].motion_val[1][h->mb2b_xy [mb_xy]];
206 l1ref0 = &h->ref_list[1][0].ref_index [0][h->mb2b8_xy[mb_xy]];
207 l1ref1 = &h->ref_list[1][0].ref_index [1][h->mb2b8_xy[mb_xy]];
210 l1ref0 += h->b8_stride;
211 l1ref1 += h->b8_stride;
212 l1mv0 += 2*b4_stride;
213 l1mv1 += 2*b4_stride;
217 if(h->direct_spatial_mv_pred){
222 /* ref = min(neighbors) */
223 for(list=0; list<2; list++){
224 int left_ref = h->ref_cache[list][scan8[0] - 1];
225 int top_ref = h->ref_cache[list][scan8[0] - 8];
226 int refc = h->ref_cache[list][scan8[0] - 8 + 4];
227 const int16_t *C= h->mv_cache[list][ scan8[0] - 8 + 4];
228 if(refc == PART_NOT_AVAILABLE){
229 refc = h->ref_cache[list][scan8[0] - 8 - 1];
230 C = h-> mv_cache[list][scan8[0] - 8 - 1];
232 ref[list] = FFMIN3((unsigned)left_ref, (unsigned)top_ref, (unsigned)refc);
234 //this is just pred_motion() but with the cases removed that cannot happen for direct blocks
235 const int16_t * const A= h->mv_cache[list][ scan8[0] - 1 ];
236 const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
238 int match_count= (left_ref==ref[list]) + (top_ref==ref[list]) + (refc==ref[list]);
239 if(match_count > 1){ //most common
240 mv[list]= (mid_pred(A[0], B[0], C[0])&0xFFFF)
241 +(mid_pred(A[1], B[1], C[1])<<16);
243 assert(match_count==1);
244 if(left_ref==ref[list]){
245 mv[list]= *(uint32_t*)A;
246 }else if(top_ref==ref[list]){
247 mv[list]= *(uint32_t*)B;
249 mv[list]= *(uint32_t*)C;
253 int mask= ~(MB_TYPE_L0 << (2*list));
261 if(ref[0] < 0 && ref[1] < 0){
264 *mb_type |= MB_TYPE_L0L1;
265 sub_mb_type |= MB_TYPE_L0L1;
268 if(!is_b8x8 && (mv[0]|mv[1]) == 0){
269 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
270 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
271 fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, 0, 4);
272 fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, 0, 4);
273 *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;
274 }else if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
276 for(i8=0; i8<4; i8++){
279 int xy8 = x8+y8*b8_stride;
280 int xy4 = 3*x8+y8*b4_stride;
283 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
285 h->sub_mb_type[i8] = sub_mb_type;
287 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
288 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
289 if(!IS_INTRA(mb_type_col[y8]) && !h->ref_list[1][0].long_ref
290 && ( (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
291 || (l1ref0[xy8] < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
302 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
303 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
305 if(!is_b8x8 && !(n&3))
306 *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;
307 }else if(IS_16X16(*mb_type)){
310 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
311 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
312 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref
313 && ( (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
314 || (l1ref0[0] < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
315 && h->x264_build>33U))){
325 fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
326 fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
329 for(i8=0; i8<4; i8++){
331 const int y8 = i8>>1;
333 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
335 h->sub_mb_type[i8] = sub_mb_type;
337 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, mv[0], 4);
338 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, mv[1], 4);
339 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
340 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
343 if(!IS_INTRA(mb_type_col[0]) && !h->ref_list[1][0].long_ref && ( l1ref0[x8 + y8*b8_stride] == 0
344 || (l1ref0[x8 + y8*b8_stride] < 0 && l1ref1[x8 + y8*b8_stride] == 0
345 && h->x264_build>33U))){
346 const int16_t (*l1mv)[2]= l1ref0[x8 + y8*b8_stride] == 0 ? l1mv0 : l1mv1;
347 if(IS_SUB_8X8(sub_mb_type)){
348 const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
349 if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
351 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
353 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
358 for(i4=0; i4<4; i4++){
359 const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
360 if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
362 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
364 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
369 h->sub_mb_type[i8]+= MB_TYPE_16x16 - MB_TYPE_8x8;
374 if(!is_b8x8 && !(n&15))
375 *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;
377 }else{ /* direct temporal mv pred */
378 const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
379 const int *dist_scale_factor = h->dist_scale_factor;
382 if(FRAME_MBAFF && IS_INTERLACED(*mb_type)){
383 map_col_to_list0[0] = h->map_col_to_list0_field[s->mb_y&1][0];
384 map_col_to_list0[1] = h->map_col_to_list0_field[s->mb_y&1][1];
385 dist_scale_factor =h->dist_scale_factor_field[s->mb_y&1];
387 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
389 if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col[0])){
390 int y_shift = 2*!IS_INTERLACED(*mb_type);
391 assert(h->sps.direct_8x8_inference_flag);
393 for(i8=0; i8<4; i8++){
395 const int y8 = i8>>1;
397 const int16_t (*l1mv)[2]= l1mv0;
399 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
401 h->sub_mb_type[i8] = sub_mb_type;
403 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
404 if(IS_INTRA(mb_type_col[y8])){
405 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
406 fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
407 fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
411 ref0 = l1ref0[x8 + y8*b8_stride];
413 ref0 = map_col_to_list0[0][ref0 + ref_offset];
415 ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
418 scale = dist_scale_factor[ref0];
419 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
422 const int16_t *mv_col = l1mv[x8*3 + y8*b4_stride];
423 int my_col = (mv_col[1]<<y_shift)/2;
424 int mx = (scale * mv_col[0] + 128) >> 8;
425 int my = (scale * my_col + 128) >> 8;
426 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
427 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
433 /* one-to-one mv scaling */
435 if(IS_16X16(*mb_type)){
438 fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
439 if(IS_INTRA(mb_type_col[0])){
442 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0] + ref_offset]
443 : map_col_to_list0[1][l1ref1[0] + ref_offset];
444 const int scale = dist_scale_factor[ref0];
445 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
447 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
448 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
450 mv0= pack16to32(mv_l0[0],mv_l0[1]);
451 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
453 fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
454 fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
455 fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
457 for(i8=0; i8<4; i8++){
459 const int y8 = i8>>1;
461 const int16_t (*l1mv)[2]= l1mv0;
463 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
465 h->sub_mb_type[i8] = sub_mb_type;
466 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
467 if(IS_INTRA(mb_type_col[0])){
468 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
469 fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
470 fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
474 ref0 = l1ref0[x8 + y8*b8_stride];
476 ref0 = map_col_to_list0[0][ref0 + ref_offset];
478 ref0 = map_col_to_list0[1][l1ref1[x8 + y8*b8_stride] + ref_offset];
481 scale = dist_scale_factor[ref0];
483 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
484 if(IS_SUB_8X8(sub_mb_type)){
485 const int16_t *mv_col = l1mv[x8*3 + y8*3*b4_stride];
486 int mx = (scale * mv_col[0] + 128) >> 8;
487 int my = (scale * mv_col[1] + 128) >> 8;
488 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
489 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
491 for(i4=0; i4<4; i4++){
492 const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*b4_stride];
493 int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
494 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
495 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
496 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
497 pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);