Don't test for H264 encoding unless an encoder exists.
[ffmpeg.git] / libavcodec / h264.c
1 /*
2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21
22 /**
23  * @file h264.c
24  * H.264 / AVC / MPEG4 part10 codec.
25  * @author Michael Niedermayer <michaelni@gmx.at>
26  */
27
28 #include "dsputil.h"
29 #include "avcodec.h"
30 #include "mpegvideo.h"
31 #include "h264.h"
32 #include "h264data.h"
33 #include "h264_parser.h"
34 #include "golomb.h"
35 #include "rectangle.h"
36
37 #include "cabac.h"
38 #ifdef ARCH_X86
39 #include "i386/h264_i386.h"
40 #endif
41
42 //#undef NDEBUG
43 #include <assert.h>
44
45 /**
46  * Value of Picture.reference when Picture is not a reference picture, but
47  * is held for delayed output.
48  */
49 #define DELAYED_PIC_REF 4
50
51 static VLC coeff_token_vlc[4];
52 static VLC chroma_dc_coeff_token_vlc;
53
54 static VLC total_zeros_vlc[15];
55 static VLC chroma_dc_total_zeros_vlc[3];
56
57 static VLC run_vlc[6];
58 static VLC run7_vlc;
59
60 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
61 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
62 static void filter_mb( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
63 static void filter_mb_fast( H264Context *h, int mb_x, int mb_y, uint8_t *img_y, uint8_t *img_cb, uint8_t *img_cr, unsigned int linesize, unsigned int uvlinesize);
64
65 static av_always_inline uint32_t pack16to32(int a, int b){
66 #ifdef WORDS_BIGENDIAN
67    return (b&0xFFFF) + (a<<16);
68 #else
69    return (a&0xFFFF) + (b<<16);
70 #endif
71 }
72
73 const uint8_t ff_rem6[52]={
74 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3,
75 };
76
77 const uint8_t ff_div6[52]={
78 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
79 };
80
81
82 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
83     MpegEncContext * const s = &h->s;
84     const int mb_xy= h->mb_xy;
85     int topleft_xy, top_xy, topright_xy, left_xy[2];
86     int topleft_type, top_type, topright_type, left_type[2];
87     int left_block[8];
88     int topleft_partition= -1;
89     int i;
90
91     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
92
93     //FIXME deblocking could skip the intra and nnz parts.
94     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
95         return;
96
97     /* Wow, what a mess, why didn't they simplify the interlacing & intra
98      * stuff, I can't imagine that these complex rules are worth it. */
99
100     topleft_xy = top_xy - 1;
101     topright_xy= top_xy + 1;
102     left_xy[1] = left_xy[0] = mb_xy-1;
103     left_block[0]= 0;
104     left_block[1]= 1;
105     left_block[2]= 2;
106     left_block[3]= 3;
107     left_block[4]= 7;
108     left_block[5]= 10;
109     left_block[6]= 8;
110     left_block[7]= 11;
111     if(FRAME_MBAFF){
112         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
113         const int top_pair_xy      = pair_xy     - s->mb_stride;
114         const int topleft_pair_xy  = top_pair_xy - 1;
115         const int topright_pair_xy = top_pair_xy + 1;
116         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
117         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
118         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
119         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
120         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
121         const int bottom = (s->mb_y & 1);
122         tprintf(s->avctx, "fill_caches: curr_mb_frame_flag:%d, left_mb_frame_flag:%d, topleft_mb_frame_flag:%d, top_mb_frame_flag:%d, topright_mb_frame_flag:%d\n", curr_mb_frame_flag, left_mb_frame_flag, topleft_mb_frame_flag, top_mb_frame_flag, topright_mb_frame_flag);
123         if (bottom
124                 ? !curr_mb_frame_flag // bottom macroblock
125                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
126                 ) {
127             top_xy -= s->mb_stride;
128         }
129         if (bottom
130                 ? !curr_mb_frame_flag // bottom macroblock
131                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
132                 ) {
133             topleft_xy -= s->mb_stride;
134         } else if(bottom && curr_mb_frame_flag && !left_mb_frame_flag) {
135             topleft_xy += s->mb_stride;
136             // take topleft mv from the middle of the mb, as opposed to all other modes which use the bottom-right partition
137             topleft_partition = 0;
138         }
139         if (bottom
140                 ? !curr_mb_frame_flag // bottom macroblock
141                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
142                 ) {
143             topright_xy -= s->mb_stride;
144         }
145         if (left_mb_frame_flag != curr_mb_frame_flag) {
146             left_xy[1] = left_xy[0] = pair_xy - 1;
147             if (curr_mb_frame_flag) {
148                 if (bottom) {
149                     left_block[0]= 2;
150                     left_block[1]= 2;
151                     left_block[2]= 3;
152                     left_block[3]= 3;
153                     left_block[4]= 8;
154                     left_block[5]= 11;
155                     left_block[6]= 8;
156                     left_block[7]= 11;
157                 } else {
158                     left_block[0]= 0;
159                     left_block[1]= 0;
160                     left_block[2]= 1;
161                     left_block[3]= 1;
162                     left_block[4]= 7;
163                     left_block[5]= 10;
164                     left_block[6]= 7;
165                     left_block[7]= 10;
166                 }
167             } else {
168                 left_xy[1] += s->mb_stride;
169                 //left_block[0]= 0;
170                 left_block[1]= 2;
171                 left_block[2]= 0;
172                 left_block[3]= 2;
173                 //left_block[4]= 7;
174                 left_block[5]= 10;
175                 left_block[6]= 7;
176                 left_block[7]= 10;
177             }
178         }
179     }
180
181     h->top_mb_xy = top_xy;
182     h->left_mb_xy[0] = left_xy[0];
183     h->left_mb_xy[1] = left_xy[1];
184     if(for_deblock){
185         topleft_type = 0;
186         topright_type = 0;
187         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
188         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
189         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
190
191         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
192             int list;
193             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
194             for(i=0; i<16; i++)
195                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
196             for(list=0; list<h->list_count; list++){
197                 if(USES_LIST(mb_type,list)){
198                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
199                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
200                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
201                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
202                         dst[0] = src[0];
203                         dst[1] = src[1];
204                         dst[2] = src[2];
205                         dst[3] = src[3];
206                     }
207                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
208                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
209                     ref += h->b8_stride;
210                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
211                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
212                 }else{
213                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
214                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
215                 }
216             }
217         }
218     }else{
219         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
220         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
221         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
222         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
223         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
224     }
225
226     if(IS_INTRA(mb_type)){
227         h->topleft_samples_available=
228         h->top_samples_available=
229         h->left_samples_available= 0xFFFF;
230         h->topright_samples_available= 0xEEEA;
231
232         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
233             h->topleft_samples_available= 0xB3FF;
234             h->top_samples_available= 0x33FF;
235             h->topright_samples_available= 0x26EA;
236         }
237         for(i=0; i<2; i++){
238             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
239                 h->topleft_samples_available&= 0xDF5F;
240                 h->left_samples_available&= 0x5F5F;
241             }
242         }
243
244         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
245             h->topleft_samples_available&= 0x7FFF;
246
247         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
248             h->topright_samples_available&= 0xFBFF;
249
250         if(IS_INTRA4x4(mb_type)){
251             if(IS_INTRA4x4(top_type)){
252                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
253                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
254                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
255                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
256             }else{
257                 int pred;
258                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
259                     pred= -1;
260                 else{
261                     pred= 2;
262                 }
263                 h->intra4x4_pred_mode_cache[4+8*0]=
264                 h->intra4x4_pred_mode_cache[5+8*0]=
265                 h->intra4x4_pred_mode_cache[6+8*0]=
266                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
267             }
268             for(i=0; i<2; i++){
269                 if(IS_INTRA4x4(left_type[i])){
270                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
271                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
272                 }else{
273                     int pred;
274                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
275                         pred= -1;
276                     else{
277                         pred= 2;
278                     }
279                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
280                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
281                 }
282             }
283         }
284     }
285
286
287 /*
288 0 . T T. T T T T
289 1 L . .L . . . .
290 2 L . .L . . . .
291 3 . T TL . . . .
292 4 L . .L . . . .
293 5 L . .. . . . .
294 */
295 //FIXME constraint_intra_pred & partitioning & nnz (lets hope this is just a typo in the spec)
296     if(top_type){
297         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
298         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
299         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
300         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
301
302         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
303         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
304
305         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
306         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
307
308     }else{
309         h->non_zero_count_cache[4+8*0]=
310         h->non_zero_count_cache[5+8*0]=
311         h->non_zero_count_cache[6+8*0]=
312         h->non_zero_count_cache[7+8*0]=
313
314         h->non_zero_count_cache[1+8*0]=
315         h->non_zero_count_cache[2+8*0]=
316
317         h->non_zero_count_cache[1+8*3]=
318         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
319
320     }
321
322     for (i=0; i<2; i++) {
323         if(left_type[i]){
324             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
325             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
326             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
327             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
328         }else{
329             h->non_zero_count_cache[3+8*1 + 2*8*i]=
330             h->non_zero_count_cache[3+8*2 + 2*8*i]=
331             h->non_zero_count_cache[0+8*1 +   8*i]=
332             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
333         }
334     }
335
336     if( h->pps.cabac ) {
337         // top_cbp
338         if(top_type) {
339             h->top_cbp = h->cbp_table[top_xy];
340         } else if(IS_INTRA(mb_type)) {
341             h->top_cbp = 0x1C0;
342         } else {
343             h->top_cbp = 0;
344         }
345         // left_cbp
346         if (left_type[0]) {
347             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
348         } else if(IS_INTRA(mb_type)) {
349             h->left_cbp = 0x1C0;
350         } else {
351             h->left_cbp = 0;
352         }
353         if (left_type[0]) {
354             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
355         }
356         if (left_type[1]) {
357             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
358         }
359     }
360
361 #if 1
362     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
363         int list;
364         for(list=0; list<h->list_count; list++){
365             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
366                 /*if(!h->mv_cache_clean[list]){
367                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
368                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
369                     h->mv_cache_clean[list]= 1;
370                 }*/
371                 continue;
372             }
373             h->mv_cache_clean[list]= 0;
374
375             if(USES_LIST(top_type, list)){
376                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
377                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
378                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
379                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
380                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
381                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
382                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
383                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
384                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
385                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
386             }else{
387                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
388                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
389                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
390                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
391                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
392             }
393
394             for(i=0; i<2; i++){
395                 int cache_idx = scan8[0] - 1 + i*2*8;
396                 if(USES_LIST(left_type[i], list)){
397                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
398                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
399                     *(uint32_t*)h->mv_cache[list][cache_idx  ]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[0+i*2]];
400                     *(uint32_t*)h->mv_cache[list][cache_idx+8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + h->b_stride*left_block[1+i*2]];
401                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
402                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
403                 }else{
404                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
405                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
406                     h->ref_cache[list][cache_idx  ]=
407                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
408                 }
409             }
410
411             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
412                 continue;
413
414             if(USES_LIST(topleft_type, list)){
415                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + h->b_stride + (topleft_partition & 2*h->b_stride);
416                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + (topleft_partition & h->b8_stride);
417                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
418                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
419             }else{
420                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
421                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
422             }
423
424             if(USES_LIST(topright_type, list)){
425                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
426                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
427                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
428                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
429             }else{
430                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
431                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
432             }
433
434             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
435                 continue;
436
437             h->ref_cache[list][scan8[5 ]+1] =
438             h->ref_cache[list][scan8[7 ]+1] =
439             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
440             h->ref_cache[list][scan8[4 ]] =
441             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
442             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
443             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
444             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
445             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
446             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
447
448             if( h->pps.cabac ) {
449                 /* XXX beurk, Load mvd */
450                 if(USES_LIST(top_type, list)){
451                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
452                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
453                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
454                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
455                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
456                 }else{
457                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
458                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
459                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
460                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
461                 }
462                 if(USES_LIST(left_type[0], list)){
463                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
464                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 0*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[0]];
465                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[1]];
466                 }else{
467                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
468                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
469                 }
470                 if(USES_LIST(left_type[1], list)){
471                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
472                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 2*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[2]];
473                     *(uint32_t*)h->mvd_cache[list][scan8[0] - 1 + 3*8]= *(uint32_t*)h->mvd_table[list][b_xy + h->b_stride*left_block[3]];
474                 }else{
475                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
476                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
477                 }
478                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
479                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
480                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
481                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
482                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
483
484                 if(h->slice_type == FF_B_TYPE){
485                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
486
487                     if(IS_DIRECT(top_type)){
488                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
489                     }else if(IS_8X8(top_type)){
490                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
491                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
492                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
493                     }else{
494                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
495                     }
496
497                     if(IS_DIRECT(left_type[0]))
498                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
499                     else if(IS_8X8(left_type[0]))
500                         h->direct_cache[scan8[0] - 1 + 0*8]= h->direct_table[h->mb2b8_xy[left_xy[0]] + 1 + h->b8_stride*(left_block[0]>>1)];
501                     else
502                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
503
504                     if(IS_DIRECT(left_type[1]))
505                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
506                     else if(IS_8X8(left_type[1]))
507                         h->direct_cache[scan8[0] - 1 + 2*8]= h->direct_table[h->mb2b8_xy[left_xy[1]] + 1 + h->b8_stride*(left_block[2]>>1)];
508                     else
509                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
510                 }
511             }
512
513             if(FRAME_MBAFF){
514 #define MAP_MVS\
515                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
516                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
517                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
518                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
519                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
520                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
521                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
522                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
523                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
524                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
525                 if(MB_FIELD){
526 #define MAP_F2F(idx, mb_type)\
527                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
528                         h->ref_cache[list][idx] <<= 1;\
529                         h->mv_cache[list][idx][1] /= 2;\
530                         h->mvd_cache[list][idx][1] /= 2;\
531                     }
532                     MAP_MVS
533 #undef MAP_F2F
534                 }else{
535 #define MAP_F2F(idx, mb_type)\
536                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
537                         h->ref_cache[list][idx] >>= 1;\
538                         h->mv_cache[list][idx][1] <<= 1;\
539                         h->mvd_cache[list][idx][1] <<= 1;\
540                     }
541                     MAP_MVS
542 #undef MAP_F2F
543                 }
544             }
545         }
546     }
547 #endif
548
549     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
550 }
551
552 static inline void write_back_intra_pred_mode(H264Context *h){
553     const int mb_xy= h->mb_xy;
554
555     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
556     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
557     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
558     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
559     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
560     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
561     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
562 }
563
564 /**
565  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
566  */
567 static inline int check_intra4x4_pred_mode(H264Context *h){
568     MpegEncContext * const s = &h->s;
569     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
570     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
571     int i;
572
573     if(!(h->top_samples_available&0x8000)){
574         for(i=0; i<4; i++){
575             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
576             if(status<0){
577                 av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
578                 return -1;
579             } else if(status){
580                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
581             }
582         }
583     }
584
585     if(!(h->left_samples_available&0x8000)){
586         for(i=0; i<4; i++){
587             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
588             if(status<0){
589                 av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra4x4 mode %d at %d %d\n", status, s->mb_x, s->mb_y);
590                 return -1;
591             } else if(status){
592                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
593             }
594         }
595     }
596
597     return 0;
598 } //FIXME cleanup like next
599
600 /**
601  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
602  */
603 static inline int check_intra_pred_mode(H264Context *h, int mode){
604     MpegEncContext * const s = &h->s;
605     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
606     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
607
608     if(mode > 6U) {
609         av_log(h->s.avctx, AV_LOG_ERROR, "out of range intra chroma pred mode at %d %d\n", s->mb_x, s->mb_y);
610         return -1;
611     }
612
613     if(!(h->top_samples_available&0x8000)){
614         mode= top[ mode ];
615         if(mode<0){
616             av_log(h->s.avctx, AV_LOG_ERROR, "top block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
617             return -1;
618         }
619     }
620
621     if(!(h->left_samples_available&0x8000)){
622         mode= left[ mode ];
623         if(mode<0){
624             av_log(h->s.avctx, AV_LOG_ERROR, "left block unavailable for requested intra mode at %d %d\n", s->mb_x, s->mb_y);
625             return -1;
626         }
627     }
628
629     return mode;
630 }
631
632 /**
633  * gets the predicted intra4x4 prediction mode.
634  */
635 static inline int pred_intra_mode(H264Context *h, int n){
636     const int index8= scan8[n];
637     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
638     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
639     const int min= FFMIN(left, top);
640
641     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
642
643     if(min<0) return DC_PRED;
644     else      return min;
645 }
646
647 static inline void write_back_non_zero_count(H264Context *h){
648     const int mb_xy= h->mb_xy;
649
650     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
651     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
652     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
653     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
654     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
655     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
656     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
657
658     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
659     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
660     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
661
662     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
663     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
664     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
665
666     if(FRAME_MBAFF){
667         // store all luma nnzs, for deblocking
668         int v = 0, i;
669         for(i=0; i<16; i++)
670             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
671         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
672     }
673 }
674
675 /**
676  * gets the predicted number of non zero coefficients.
677  * @param n block index
678  */
679 static inline int pred_non_zero_count(H264Context *h, int n){
680     const int index8= scan8[n];
681     const int left= h->non_zero_count_cache[index8 - 1];
682     const int top = h->non_zero_count_cache[index8 - 8];
683     int i= left + top;
684
685     if(i<64) i= (i+1)>>1;
686
687     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
688
689     return i&31;
690 }
691
692 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
693     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
694     MpegEncContext *s = &h->s;
695
696     /* there is no consistent mapping of mvs to neighboring locations that will
697      * make mbaff happy, so we can't move all this logic to fill_caches */
698     if(FRAME_MBAFF){
699         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
700         const int16_t *mv;
701         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
702         *C = h->mv_cache[list][scan8[0]-2];
703
704         if(!MB_FIELD
705            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
706             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
707             if(IS_INTERLACED(mb_types[topright_xy])){
708 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
709                 const int x4 = X4, y4 = Y4;\
710                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
711                 if(!USES_LIST(mb_type,list))\
712                     return LIST_NOT_USED;\
713                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
714                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
715                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
716                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
717
718                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
719             }
720         }
721         if(topright_ref == PART_NOT_AVAILABLE
722            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
723            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
724             if(!MB_FIELD
725                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
726                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
727             }
728             if(MB_FIELD
729                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
730                && i >= scan8[0]+8){
731                 // leftshift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's ok.
732                 SET_DIAG_MV(/2, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
733             }
734         }
735 #undef SET_DIAG_MV
736     }
737
738     if(topright_ref != PART_NOT_AVAILABLE){
739         *C= h->mv_cache[list][ i - 8 + part_width ];
740         return topright_ref;
741     }else{
742         tprintf(s->avctx, "topright MV not available\n");
743
744         *C= h->mv_cache[list][ i - 8 - 1 ];
745         return h->ref_cache[list][ i - 8 - 1 ];
746     }
747 }
748
749 /**
750  * gets the predicted MV.
751  * @param n the block index
752  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
753  * @param mx the x component of the predicted motion vector
754  * @param my the y component of the predicted motion vector
755  */
756 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
757     const int index8= scan8[n];
758     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
759     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
760     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
761     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
762     const int16_t * C;
763     int diagonal_ref, match_count;
764
765     assert(part_width==1 || part_width==2 || part_width==4);
766
767 /* mv_cache
768   B . . A T T T T
769   U . . L . . , .
770   U . . L . . . .
771   U . . L . . , .
772   . . . L . . . .
773 */
774
775     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
776     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
777     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
778     if(match_count > 1){ //most common
779         *mx= mid_pred(A[0], B[0], C[0]);
780         *my= mid_pred(A[1], B[1], C[1]);
781     }else if(match_count==1){
782         if(left_ref==ref){
783             *mx= A[0];
784             *my= A[1];
785         }else if(top_ref==ref){
786             *mx= B[0];
787             *my= B[1];
788         }else{
789             *mx= C[0];
790             *my= C[1];
791         }
792     }else{
793         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
794             *mx= A[0];
795             *my= A[1];
796         }else{
797             *mx= mid_pred(A[0], B[0], C[0]);
798             *my= mid_pred(A[1], B[1], C[1]);
799         }
800     }
801
802     tprintf(h->s.avctx, "pred_motion (%2d %2d %2d) (%2d %2d %2d) (%2d %2d %2d) -> (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1],                    diagonal_ref, C[0], C[1], left_ref, A[0], A[1], ref, *mx, *my, h->s.mb_x, h->s.mb_y, n, list);
803 }
804
805 /**
806  * gets the directionally predicted 16x8 MV.
807  * @param n the block index
808  * @param mx the x component of the predicted motion vector
809  * @param my the y component of the predicted motion vector
810  */
811 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
812     if(n==0){
813         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
814         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
815
816         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", top_ref, B[0], B[1], h->s.mb_x, h->s.mb_y, n, list);
817
818         if(top_ref == ref){
819             *mx= B[0];
820             *my= B[1];
821             return;
822         }
823     }else{
824         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
825         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
826
827         tprintf(h->s.avctx, "pred_16x8: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
828
829         if(left_ref == ref){
830             *mx= A[0];
831             *my= A[1];
832             return;
833         }
834     }
835
836     //RARE
837     pred_motion(h, n, 4, list, ref, mx, my);
838 }
839
840 /**
841  * gets the directionally predicted 8x16 MV.
842  * @param n the block index
843  * @param mx the x component of the predicted motion vector
844  * @param my the y component of the predicted motion vector
845  */
846 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
847     if(n==0){
848         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
849         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
850
851         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", left_ref, A[0], A[1], h->s.mb_x, h->s.mb_y, n, list);
852
853         if(left_ref == ref){
854             *mx= A[0];
855             *my= A[1];
856             return;
857         }
858     }else{
859         const int16_t * C;
860         int diagonal_ref;
861
862         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
863
864         tprintf(h->s.avctx, "pred_8x16: (%2d %2d %2d) at %2d %2d %d list %d\n", diagonal_ref, C[0], C[1], h->s.mb_x, h->s.mb_y, n, list);
865
866         if(diagonal_ref == ref){
867             *mx= C[0];
868             *my= C[1];
869             return;
870         }
871     }
872
873     //RARE
874     pred_motion(h, n, 2, list, ref, mx, my);
875 }
876
877 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
878     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
879     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
880
881     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
882
883     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
884        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
885        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
886
887         *mx = *my = 0;
888         return;
889     }
890
891     pred_motion(h, 0, 4, 0, 0, mx, my);
892
893     return;
894 }
895
896 static inline void direct_dist_scale_factor(H264Context * const h){
897     const int poc = h->s.current_picture_ptr->poc;
898     const int poc1 = h->ref_list[1][0].poc;
899     int i;
900     for(i=0; i<h->ref_count[0]; i++){
901         int poc0 = h->ref_list[0][i].poc;
902         int td = av_clip(poc1 - poc0, -128, 127);
903         if(td == 0 /* FIXME || pic0 is a long-term ref */){
904             h->dist_scale_factor[i] = 256;
905         }else{
906             int tb = av_clip(poc - poc0, -128, 127);
907             int tx = (16384 + (FFABS(td) >> 1)) / td;
908             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
909         }
910     }
911     if(FRAME_MBAFF){
912         for(i=0; i<h->ref_count[0]; i++){
913             h->dist_scale_factor_field[2*i] =
914             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
915         }
916     }
917 }
918 static inline void direct_ref_list_init(H264Context * const h){
919     MpegEncContext * const s = &h->s;
920     Picture * const ref1 = &h->ref_list[1][0];
921     Picture * const cur = s->current_picture_ptr;
922     int list, i, j;
923     if(cur->pict_type == FF_I_TYPE)
924         cur->ref_count[0] = 0;
925     if(cur->pict_type != FF_B_TYPE)
926         cur->ref_count[1] = 0;
927     for(list=0; list<2; list++){
928         cur->ref_count[list] = h->ref_count[list];
929         for(j=0; j<h->ref_count[list]; j++)
930             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
931     }
932     if(cur->pict_type != FF_B_TYPE || h->direct_spatial_mv_pred)
933         return;
934     for(list=0; list<2; list++){
935         for(i=0; i<ref1->ref_count[list]; i++){
936             const int poc = ref1->ref_poc[list][i];
937             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
938             for(j=0; j<h->ref_count[list]; j++)
939                 if(h->ref_list[list][j].poc == poc){
940                     h->map_col_to_list0[list][i] = j;
941                     break;
942                 }
943         }
944     }
945     if(FRAME_MBAFF){
946         for(list=0; list<2; list++){
947             for(i=0; i<ref1->ref_count[list]; i++){
948                 j = h->map_col_to_list0[list][i];
949                 h->map_col_to_list0_field[list][2*i] = 2*j;
950                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
951             }
952         }
953     }
954 }
955
956 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
957     MpegEncContext * const s = &h->s;
958     const int mb_xy =   h->mb_xy;
959     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
960     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
961     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
962     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
963     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
964     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
965     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
966     const int is_b8x8 = IS_8X8(*mb_type);
967     unsigned int sub_mb_type;
968     int i8, i4;
969
970 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
971     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
972         /* FIXME save sub mb types from previous frames (or derive from MVs)
973          * so we know exactly what block size to use */
974         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
975         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
976     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
977         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
978         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
979     }else{
980         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
981         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
982     }
983     if(!is_b8x8)
984         *mb_type |= MB_TYPE_DIRECT2;
985     if(MB_FIELD)
986         *mb_type |= MB_TYPE_INTERLACED;
987
988     tprintf(s->avctx, "mb_type = %08x, sub_mb_type = %08x, is_b8x8 = %d, mb_type_col = %08x\n", *mb_type, sub_mb_type, is_b8x8, mb_type_col);
989
990     if(h->direct_spatial_mv_pred){
991         int ref[2];
992         int mv[2][2];
993         int list;
994
995         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
996
997         /* ref = min(neighbors) */
998         for(list=0; list<2; list++){
999             int refa = h->ref_cache[list][scan8[0] - 1];
1000             int refb = h->ref_cache[list][scan8[0] - 8];
1001             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1002             if(refc == -2)
1003                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1004             ref[list] = refa;
1005             if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
1006                 ref[list] = refb;
1007             if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
1008                 ref[list] = refc;
1009             if(ref[list] < 0)
1010                 ref[list] = -1;
1011         }
1012
1013         if(ref[0] < 0 && ref[1] < 0){
1014             ref[0] = ref[1] = 0;
1015             mv[0][0] = mv[0][1] =
1016             mv[1][0] = mv[1][1] = 0;
1017         }else{
1018             for(list=0; list<2; list++){
1019                 if(ref[list] >= 0)
1020                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1021                 else
1022                     mv[list][0] = mv[list][1] = 0;
1023             }
1024         }
1025
1026         if(ref[1] < 0){
1027             if(!is_b8x8)
1028                 *mb_type &= ~MB_TYPE_L1;
1029             sub_mb_type &= ~MB_TYPE_L1;
1030         }else if(ref[0] < 0){
1031             if(!is_b8x8)
1032                 *mb_type &= ~MB_TYPE_L0;
1033             sub_mb_type &= ~MB_TYPE_L0;
1034         }
1035
1036         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1037             int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1038             int mb_types_col[2];
1039             int b8_stride = h->b8_stride;
1040             int b4_stride = h->b_stride;
1041
1042             *mb_type = (*mb_type & ~MB_TYPE_16x16) | MB_TYPE_8x8;
1043
1044             if(IS_INTERLACED(*mb_type)){
1045                 mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1046                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1047                 if(s->mb_y&1){
1048                     l1ref0 -= 2*b8_stride;
1049                     l1ref1 -= 2*b8_stride;
1050                     l1mv0 -= 4*b4_stride;
1051                     l1mv1 -= 4*b4_stride;
1052                 }
1053                 b8_stride *= 3;
1054                 b4_stride *= 6;
1055             }else{
1056                 int cur_poc = s->current_picture_ptr->poc;
1057                 int *col_poc = h->ref_list[1]->field_poc;
1058                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1059                 int dy = 2*col_parity - (s->mb_y&1);
1060                 mb_types_col[0] =
1061                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy + col_parity*s->mb_stride];
1062                 l1ref0 += dy*b8_stride;
1063                 l1ref1 += dy*b8_stride;
1064                 l1mv0 += 2*dy*b4_stride;
1065                 l1mv1 += 2*dy*b4_stride;
1066                 b8_stride = 0;
1067             }
1068
1069             for(i8=0; i8<4; i8++){
1070                 int x8 = i8&1;
1071                 int y8 = i8>>1;
1072                 int xy8 = x8+y8*b8_stride;
1073                 int xy4 = 3*x8+y8*b4_stride;
1074                 int a=0, b=0;
1075
1076                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1077                     continue;
1078                 h->sub_mb_type[i8] = sub_mb_type;
1079
1080                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1081                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1082                 if(!IS_INTRA(mb_types_col[y8])
1083                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1084                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1085                     if(ref[0] > 0)
1086                         a= pack16to32(mv[0][0],mv[0][1]);
1087                     if(ref[1] > 0)
1088                         b= pack16to32(mv[1][0],mv[1][1]);
1089                 }else{
1090                     a= pack16to32(mv[0][0],mv[0][1]);
1091                     b= pack16to32(mv[1][0],mv[1][1]);
1092                 }
1093                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1094                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1095             }
1096         }else if(IS_16X16(*mb_type)){
1097             int a=0, b=0;
1098
1099             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1100             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1101             if(!IS_INTRA(mb_type_col)
1102                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1103                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1104                        && (h->x264_build>33 || !h->x264_build)))){
1105                 if(ref[0] > 0)
1106                     a= pack16to32(mv[0][0],mv[0][1]);
1107                 if(ref[1] > 0)
1108                     b= pack16to32(mv[1][0],mv[1][1]);
1109             }else{
1110                 a= pack16to32(mv[0][0],mv[0][1]);
1111                 b= pack16to32(mv[1][0],mv[1][1]);
1112             }
1113             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1114             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1115         }else{
1116             for(i8=0; i8<4; i8++){
1117                 const int x8 = i8&1;
1118                 const int y8 = i8>>1;
1119
1120                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1121                     continue;
1122                 h->sub_mb_type[i8] = sub_mb_type;
1123
1124                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1125                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1126                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1127                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1128
1129                 /* col_zero_flag */
1130                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1131                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1132                                                   && (h->x264_build>33 || !h->x264_build)))){
1133                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1134                     if(IS_SUB_8X8(sub_mb_type)){
1135                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1136                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1137                             if(ref[0] == 0)
1138                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1139                             if(ref[1] == 0)
1140                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1141                         }
1142                     }else
1143                     for(i4=0; i4<4; i4++){
1144                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1145                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1146                             if(ref[0] == 0)
1147                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1148                             if(ref[1] == 0)
1149                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1150                         }
1151                     }
1152                 }
1153             }
1154         }
1155     }else{ /* direct temporal mv pred */
1156         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1157         const int *dist_scale_factor = h->dist_scale_factor;
1158
1159         if(FRAME_MBAFF){
1160             if(IS_INTERLACED(*mb_type)){
1161                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1162                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1163                 dist_scale_factor = h->dist_scale_factor_field;
1164             }
1165             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1166                 /* FIXME assumes direct_8x8_inference == 1 */
1167                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1168                 int mb_types_col[2];
1169                 int y_shift;
1170
1171                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1172                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1173                          | (*mb_type & MB_TYPE_INTERLACED);
1174                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1175
1176                 if(IS_INTERLACED(*mb_type)){
1177                     /* frame to field scaling */
1178                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1179                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1180                     if(s->mb_y&1){
1181                         l1ref0 -= 2*h->b8_stride;
1182                         l1ref1 -= 2*h->b8_stride;
1183                         l1mv0 -= 4*h->b_stride;
1184                         l1mv1 -= 4*h->b_stride;
1185                     }
1186                     y_shift = 0;
1187
1188                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1189                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1190                        && !is_b8x8)
1191                         *mb_type |= MB_TYPE_16x8;
1192                     else
1193                         *mb_type |= MB_TYPE_8x8;
1194                 }else{
1195                     /* field to frame scaling */
1196                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1197                      * but in MBAFF, top and bottom POC are equal */
1198                     int dy = (s->mb_y&1) ? 1 : 2;
1199                     mb_types_col[0] =
1200                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1201                     l1ref0 += dy*h->b8_stride;
1202                     l1ref1 += dy*h->b8_stride;
1203                     l1mv0 += 2*dy*h->b_stride;
1204                     l1mv1 += 2*dy*h->b_stride;
1205                     y_shift = 2;
1206
1207                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1208                        && !is_b8x8)
1209                         *mb_type |= MB_TYPE_16x16;
1210                     else
1211                         *mb_type |= MB_TYPE_8x8;
1212                 }
1213
1214                 for(i8=0; i8<4; i8++){
1215                     const int x8 = i8&1;
1216                     const int y8 = i8>>1;
1217                     int ref0, scale;
1218                     const int16_t (*l1mv)[2]= l1mv0;
1219
1220                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1221                         continue;
1222                     h->sub_mb_type[i8] = sub_mb_type;
1223
1224                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1225                     if(IS_INTRA(mb_types_col[y8])){
1226                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1227                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1228                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1229                         continue;
1230                     }
1231
1232                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1233                     if(ref0 >= 0)
1234                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1235                     else{
1236                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1237                         l1mv= l1mv1;
1238                     }
1239                     scale = dist_scale_factor[ref0];
1240                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1241
1242                     {
1243                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1244                         int my_col = (mv_col[1]<<y_shift)/2;
1245                         int mx = (scale * mv_col[0] + 128) >> 8;
1246                         int my = (scale * my_col + 128) >> 8;
1247                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1248                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1249                     }
1250                 }
1251                 return;
1252             }
1253         }
1254
1255         /* one-to-one mv scaling */
1256
1257         if(IS_16X16(*mb_type)){
1258             int ref, mv0, mv1;
1259
1260             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1261             if(IS_INTRA(mb_type_col)){
1262                 ref=mv0=mv1=0;
1263             }else{
1264                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1265                                                 : map_col_to_list0[1][l1ref1[0]];
1266                 const int scale = dist_scale_factor[ref0];
1267                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1268                 int mv_l0[2];
1269                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1270                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1271                 ref= ref0;
1272                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1273                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1274             }
1275             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1276             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1277             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1278         }else{
1279             for(i8=0; i8<4; i8++){
1280                 const int x8 = i8&1;
1281                 const int y8 = i8>>1;
1282                 int ref0, scale;
1283                 const int16_t (*l1mv)[2]= l1mv0;
1284
1285                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1286                     continue;
1287                 h->sub_mb_type[i8] = sub_mb_type;
1288                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1289                 if(IS_INTRA(mb_type_col)){
1290                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1291                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1292                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1293                     continue;
1294                 }
1295
1296                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1297                 if(ref0 >= 0)
1298                     ref0 = map_col_to_list0[0][ref0];
1299                 else{
1300                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1301                     l1mv= l1mv1;
1302                 }
1303                 scale = dist_scale_factor[ref0];
1304
1305                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1306                 if(IS_SUB_8X8(sub_mb_type)){
1307                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1308                     int mx = (scale * mv_col[0] + 128) >> 8;
1309                     int my = (scale * mv_col[1] + 128) >> 8;
1310                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1311                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1312                 }else
1313                 for(i4=0; i4<4; i4++){
1314                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1315                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1316                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1317                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1318                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1319                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1320                 }
1321             }
1322         }
1323     }
1324 }
1325
1326 static inline void write_back_motion(H264Context *h, int mb_type){
1327     MpegEncContext * const s = &h->s;
1328     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1329     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1330     int list;
1331
1332     if(!USES_LIST(mb_type, 0))
1333         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1334
1335     for(list=0; list<h->list_count; list++){
1336         int y;
1337         if(!USES_LIST(mb_type, list))
1338             continue;
1339
1340         for(y=0; y<4; y++){
1341             *(uint64_t*)s->current_picture.motion_val[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+0 + 8*y];
1342             *(uint64_t*)s->current_picture.motion_val[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mv_cache[list][scan8[0]+2 + 8*y];
1343         }
1344         if( h->pps.cabac ) {
1345             if(IS_SKIP(mb_type))
1346                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1347             else
1348             for(y=0; y<4; y++){
1349                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1350                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1351             }
1352         }
1353
1354         {
1355             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1356             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1357             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1358             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1359             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1360         }
1361     }
1362
1363     if(h->slice_type == FF_B_TYPE && h->pps.cabac){
1364         if(IS_8X8(mb_type)){
1365             uint8_t *direct_table = &h->direct_table[b8_xy];
1366             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1367             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1368             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1369         }
1370     }
1371 }
1372
1373 /**
1374  * Decodes a network abstraction layer unit.
1375  * @param consumed is the number of bytes used as input
1376  * @param length is the length of the array
1377  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1378  * @returns decoded bytes, might be src+1 if no escapes
1379  */
1380 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1381     int i, si, di;
1382     uint8_t *dst;
1383     int bufidx;
1384
1385 //    src[0]&0x80;                //forbidden bit
1386     h->nal_ref_idc= src[0]>>5;
1387     h->nal_unit_type= src[0]&0x1F;
1388
1389     src++; length--;
1390 #if 0
1391     for(i=0; i<length; i++)
1392         printf("%2X ", src[i]);
1393 #endif
1394     for(i=0; i+1<length; i+=2){
1395         if(src[i]) continue;
1396         if(i>0 && src[i-1]==0) i--;
1397         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1398             if(src[i+2]!=3){
1399                 /* startcode, so we must be past the end */
1400                 length=i;
1401             }
1402             break;
1403         }
1404     }
1405
1406     if(i>=length-1){ //no escaped 0
1407         *dst_length= length;
1408         *consumed= length+1; //+1 for the header
1409         return src;
1410     }
1411
1412     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1413     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1414     dst= h->rbsp_buffer[bufidx];
1415
1416     if (dst == NULL){
1417         return NULL;
1418     }
1419
1420 //printf("decoding esc\n");
1421     si=di=0;
1422     while(si<length){
1423         //remove escapes (very rare 1:2^22)
1424         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1425             if(src[si+2]==3){ //escape
1426                 dst[di++]= 0;
1427                 dst[di++]= 0;
1428                 si+=3;
1429                 continue;
1430             }else //next start code
1431                 break;
1432         }
1433
1434         dst[di++]= src[si++];
1435     }
1436
1437     *dst_length= di;
1438     *consumed= si + 1;//+1 for the header
1439 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1440     return dst;
1441 }
1442
1443 /**
1444  * identifies the exact end of the bitstream
1445  * @return the length of the trailing, or 0 if damaged
1446  */
1447 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1448     int v= *src;
1449     int r;
1450
1451     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1452
1453     for(r=1; r<9; r++){
1454         if(v&1) return r;
1455         v>>=1;
1456     }
1457     return 0;
1458 }
1459
1460 /**
1461  * idct tranforms the 16 dc values and dequantize them.
1462  * @param qp quantization parameter
1463  */
1464 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1465 #define stride 16
1466     int i;
1467     int temp[16]; //FIXME check if this is a good idea
1468     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1469     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1470
1471 //memset(block, 64, 2*256);
1472 //return;
1473     for(i=0; i<4; i++){
1474         const int offset= y_offset[i];
1475         const int z0= block[offset+stride*0] + block[offset+stride*4];
1476         const int z1= block[offset+stride*0] - block[offset+stride*4];
1477         const int z2= block[offset+stride*1] - block[offset+stride*5];
1478         const int z3= block[offset+stride*1] + block[offset+stride*5];
1479
1480         temp[4*i+0]= z0+z3;
1481         temp[4*i+1]= z1+z2;
1482         temp[4*i+2]= z1-z2;
1483         temp[4*i+3]= z0-z3;
1484     }
1485
1486     for(i=0; i<4; i++){
1487         const int offset= x_offset[i];
1488         const int z0= temp[4*0+i] + temp[4*2+i];
1489         const int z1= temp[4*0+i] - temp[4*2+i];
1490         const int z2= temp[4*1+i] - temp[4*3+i];
1491         const int z3= temp[4*1+i] + temp[4*3+i];
1492
1493         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_resdual
1494         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1495         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1496         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1497     }
1498 }
1499
1500 #if 0
1501 /**
1502  * dct tranforms the 16 dc values.
1503  * @param qp quantization parameter ??? FIXME
1504  */
1505 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1506 //    const int qmul= dequant_coeff[qp][0];
1507     int i;
1508     int temp[16]; //FIXME check if this is a good idea
1509     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1510     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1511
1512     for(i=0; i<4; i++){
1513         const int offset= y_offset[i];
1514         const int z0= block[offset+stride*0] + block[offset+stride*4];
1515         const int z1= block[offset+stride*0] - block[offset+stride*4];
1516         const int z2= block[offset+stride*1] - block[offset+stride*5];
1517         const int z3= block[offset+stride*1] + block[offset+stride*5];
1518
1519         temp[4*i+0]= z0+z3;
1520         temp[4*i+1]= z1+z2;
1521         temp[4*i+2]= z1-z2;
1522         temp[4*i+3]= z0-z3;
1523     }
1524
1525     for(i=0; i<4; i++){
1526         const int offset= x_offset[i];
1527         const int z0= temp[4*0+i] + temp[4*2+i];
1528         const int z1= temp[4*0+i] - temp[4*2+i];
1529         const int z2= temp[4*1+i] - temp[4*3+i];
1530         const int z3= temp[4*1+i] + temp[4*3+i];
1531
1532         block[stride*0 +offset]= (z0 + z3)>>1;
1533         block[stride*2 +offset]= (z1 + z2)>>1;
1534         block[stride*8 +offset]= (z1 - z2)>>1;
1535         block[stride*10+offset]= (z0 - z3)>>1;
1536     }
1537 }
1538 #endif
1539
1540 #undef xStride
1541 #undef stride
1542
1543 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1544     const int stride= 16*2;
1545     const int xStride= 16;
1546     int a,b,c,d,e;
1547
1548     a= block[stride*0 + xStride*0];
1549     b= block[stride*0 + xStride*1];
1550     c= block[stride*1 + xStride*0];
1551     d= block[stride*1 + xStride*1];
1552
1553     e= a-b;
1554     a= a+b;
1555     b= c-d;
1556     c= c+d;
1557
1558     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1559     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1560     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1561     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1562 }
1563
1564 #if 0
1565 static void chroma_dc_dct_c(DCTELEM *block){
1566     const int stride= 16*2;
1567     const int xStride= 16;
1568     int a,b,c,d,e;
1569
1570     a= block[stride*0 + xStride*0];
1571     b= block[stride*0 + xStride*1];
1572     c= block[stride*1 + xStride*0];
1573     d= block[stride*1 + xStride*1];
1574
1575     e= a-b;
1576     a= a+b;
1577     b= c-d;
1578     c= c+d;
1579
1580     block[stride*0 + xStride*0]= (a+c);
1581     block[stride*0 + xStride*1]= (e+b);
1582     block[stride*1 + xStride*0]= (a-c);
1583     block[stride*1 + xStride*1]= (e-b);
1584 }
1585 #endif
1586
1587 /**
1588  * gets the chroma qp.
1589  */
1590 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1591     return h->pps.chroma_qp_table[t][qscale & 0xff];
1592 }
1593
1594 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1595 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1596 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1597     int i;
1598     const int * const quant_table= quant_coeff[qscale];
1599     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1600     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1601     const unsigned int threshold2= (threshold1<<1);
1602     int last_non_zero;
1603
1604     if(separate_dc){
1605         if(qscale<=18){
1606             //avoid overflows
1607             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1608             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1609             const unsigned int dc_threshold2= (dc_threshold1<<1);
1610
1611             int level= block[0]*quant_coeff[qscale+18][0];
1612             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1613                 if(level>0){
1614                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1615                     block[0]= level;
1616                 }else{
1617                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1618                     block[0]= -level;
1619                 }
1620 //                last_non_zero = i;
1621             }else{
1622                 block[0]=0;
1623             }
1624         }else{
1625             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1626             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1627             const unsigned int dc_threshold2= (dc_threshold1<<1);
1628
1629             int level= block[0]*quant_table[0];
1630             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1631                 if(level>0){
1632                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1633                     block[0]= level;
1634                 }else{
1635                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1636                     block[0]= -level;
1637                 }
1638 //                last_non_zero = i;
1639             }else{
1640                 block[0]=0;
1641             }
1642         }
1643         last_non_zero= 0;
1644         i=1;
1645     }else{
1646         last_non_zero= -1;
1647         i=0;
1648     }
1649
1650     for(; i<16; i++){
1651         const int j= scantable[i];
1652         int level= block[j]*quant_table[j];
1653
1654 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1655 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1656         if(((unsigned)(level+threshold1))>threshold2){
1657             if(level>0){
1658                 level= (bias + level)>>QUANT_SHIFT;
1659                 block[j]= level;
1660             }else{
1661                 level= (bias - level)>>QUANT_SHIFT;
1662                 block[j]= -level;
1663             }
1664             last_non_zero = i;
1665         }else{
1666             block[j]=0;
1667         }
1668     }
1669
1670     return last_non_zero;
1671 }
1672
1673 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1674                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1675                            int src_x_offset, int src_y_offset,
1676                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1677     MpegEncContext * const s = &h->s;
1678     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1679     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1680     const int luma_xy= (mx&3) + ((my&3)<<2);
1681     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1682     uint8_t * src_cb, * src_cr;
1683     int extra_width= h->emu_edge_width;
1684     int extra_height= h->emu_edge_height;
1685     int emu=0;
1686     const int full_mx= mx>>2;
1687     const int full_my= my>>2;
1688     const int pic_width  = 16*s->mb_width;
1689     const int pic_height = 16*s->mb_height >> MB_FIELD;
1690
1691     if(!pic->data[0]) //FIXME this is unacceptable, some senseable error concealment must be done for missing reference frames
1692         return;
1693
1694     if(mx&7) extra_width -= 3;
1695     if(my&7) extra_height -= 3;
1696
1697     if(   full_mx < 0-extra_width
1698        || full_my < 0-extra_height
1699        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1700        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1701         ff_emulated_edge_mc(s->edge_emu_buffer, src_y - 2 - 2*h->mb_linesize, h->mb_linesize, 16+5, 16+5/*FIXME*/, full_mx-2, full_my-2, pic_width, pic_height);
1702             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1703         emu=1;
1704     }
1705
1706     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1707     if(!square){
1708         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1709     }
1710
1711     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1712
1713     if(MB_FIELD){
1714         // chroma offset when predicting from a field of opposite parity
1715         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1716         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1717     }
1718     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1719     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1720
1721     if(emu){
1722         ff_emulated_edge_mc(s->edge_emu_buffer, src_cb, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
1723             src_cb= s->edge_emu_buffer;
1724     }
1725     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1726
1727     if(emu){
1728         ff_emulated_edge_mc(s->edge_emu_buffer, src_cr, h->mb_uvlinesize, 9, 9/*FIXME*/, (mx>>3), (my>>3), pic_width>>1, pic_height>>1);
1729             src_cr= s->edge_emu_buffer;
1730     }
1731     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1732 }
1733
1734 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1735                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1736                            int x_offset, int y_offset,
1737                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1738                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1739                            int list0, int list1){
1740     MpegEncContext * const s = &h->s;
1741     qpel_mc_func *qpix_op=  qpix_put;
1742     h264_chroma_mc_func chroma_op= chroma_put;
1743
1744     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1745     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1746     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1747     x_offset += 8*s->mb_x;
1748     y_offset += 8*(s->mb_y >> MB_FIELD);
1749
1750     if(list0){
1751         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1752         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1753                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1754                            qpix_op, chroma_op);
1755
1756         qpix_op=  qpix_avg;
1757         chroma_op= chroma_avg;
1758     }
1759
1760     if(list1){
1761         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1762         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1763                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1764                            qpix_op, chroma_op);
1765     }
1766 }
1767
1768 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1769                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1770                            int x_offset, int y_offset,
1771                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1772                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1773                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1774                            int list0, int list1){
1775     MpegEncContext * const s = &h->s;
1776
1777     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1778     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1779     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1780     x_offset += 8*s->mb_x;
1781     y_offset += 8*(s->mb_y >> MB_FIELD);
1782
1783     if(list0 && list1){
1784         /* don't optimize for luma-only case, since B-frames usually
1785          * use implicit weights => chroma too. */
1786         uint8_t *tmp_cb = s->obmc_scratchpad;
1787         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1788         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1789         int refn0 = h->ref_cache[0][ scan8[n] ];
1790         int refn1 = h->ref_cache[1][ scan8[n] ];
1791
1792         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1793                     dest_y, dest_cb, dest_cr,
1794                     x_offset, y_offset, qpix_put, chroma_put);
1795         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1796                     tmp_y, tmp_cb, tmp_cr,
1797                     x_offset, y_offset, qpix_put, chroma_put);
1798
1799         if(h->use_weight == 2){
1800             int weight0 = h->implicit_weight[refn0][refn1];
1801             int weight1 = 64 - weight0;
1802             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1803             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1804             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1805         }else{
1806             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1807                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1808                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1809             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1810                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1811                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1812             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1813                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1814                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1815         }
1816     }else{
1817         int list = list1 ? 1 : 0;
1818         int refn = h->ref_cache[list][ scan8[n] ];
1819         Picture *ref= &h->ref_list[list][refn];
1820         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1821                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1822                     qpix_put, chroma_put);
1823
1824         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1825                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1826         if(h->use_weight_chroma){
1827             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1828                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1829             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1830                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1831         }
1832     }
1833 }
1834
1835 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1836                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1837                            int x_offset, int y_offset,
1838                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1839                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1840                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1841                            int list0, int list1){
1842     if((h->use_weight==2 && list0 && list1
1843         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1844        || h->use_weight==1)
1845         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1846                          x_offset, y_offset, qpix_put, chroma_put,
1847                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1848     else
1849         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1850                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1851 }
1852
1853 static inline void prefetch_motion(H264Context *h, int list){
1854     /* fetch pixels for estimated mv 4 macroblocks ahead
1855      * optimized for 64byte cache lines */
1856     MpegEncContext * const s = &h->s;
1857     const int refn = h->ref_cache[list][scan8[0]];
1858     if(refn >= 0){
1859         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1860         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1861         uint8_t **src= h->ref_list[list][refn].data;
1862         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1863         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1864         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1865         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1866     }
1867 }
1868
1869 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1870                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1871                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1872                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1873     MpegEncContext * const s = &h->s;
1874     const int mb_xy= h->mb_xy;
1875     const int mb_type= s->current_picture.mb_type[mb_xy];
1876
1877     assert(IS_INTER(mb_type));
1878
1879     prefetch_motion(h, 0);
1880
1881     if(IS_16X16(mb_type)){
1882         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1883                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1884                 &weight_op[0], &weight_avg[0],
1885                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1886     }else if(IS_16X8(mb_type)){
1887         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1888                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1889                 &weight_op[1], &weight_avg[1],
1890                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1891         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1892                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1893                 &weight_op[1], &weight_avg[1],
1894                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1895     }else if(IS_8X16(mb_type)){
1896         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1897                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1898                 &weight_op[2], &weight_avg[2],
1899                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1900         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1901                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1902                 &weight_op[2], &weight_avg[2],
1903                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1904     }else{
1905         int i;
1906
1907         assert(IS_8X8(mb_type));
1908
1909         for(i=0; i<4; i++){
1910             const int sub_mb_type= h->sub_mb_type[i];
1911             const int n= 4*i;
1912             int x_offset= (i&1)<<2;
1913             int y_offset= (i&2)<<1;
1914
1915             if(IS_SUB_8X8(sub_mb_type)){
1916                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1917                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1918                     &weight_op[3], &weight_avg[3],
1919                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1920             }else if(IS_SUB_8X4(sub_mb_type)){
1921                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1922                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1923                     &weight_op[4], &weight_avg[4],
1924                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1925                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1926                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1927                     &weight_op[4], &weight_avg[4],
1928                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1929             }else if(IS_SUB_4X8(sub_mb_type)){
1930                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1931                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1932                     &weight_op[5], &weight_avg[5],
1933                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1934                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1935                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1936                     &weight_op[5], &weight_avg[5],
1937                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1938             }else{
1939                 int j;
1940                 assert(IS_SUB_4X4(sub_mb_type));
1941                 for(j=0; j<4; j++){
1942                     int sub_x_offset= x_offset + 2*(j&1);
1943                     int sub_y_offset= y_offset +   (j&2);
1944                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1945                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1946                         &weight_op[6], &weight_avg[6],
1947                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1948                 }
1949             }
1950         }
1951     }
1952
1953     prefetch_motion(h, 1);
1954 }
1955
1956 static av_cold void decode_init_vlc(void){
1957     static int done = 0;
1958
1959     if (!done) {
1960         int i;
1961         done = 1;
1962
1963         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1964                  &chroma_dc_coeff_token_len [0], 1, 1,
1965                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1966
1967         for(i=0; i<4; i++){
1968             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1969                      &coeff_token_len [i][0], 1, 1,
1970                      &coeff_token_bits[i][0], 1, 1, 1);
1971         }
1972
1973         for(i=0; i<3; i++){
1974             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1975                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1976                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1977         }
1978         for(i=0; i<15; i++){
1979             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1980                      &total_zeros_len [i][0], 1, 1,
1981                      &total_zeros_bits[i][0], 1, 1, 1);
1982         }
1983
1984         for(i=0; i<6; i++){
1985             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
1986                      &run_len [i][0], 1, 1,
1987                      &run_bits[i][0], 1, 1, 1);
1988         }
1989         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1990                  &run_len [6][0], 1, 1,
1991                  &run_bits[6][0], 1, 1, 1);
1992     }
1993 }
1994
1995 static void free_tables(H264Context *h){
1996     int i;
1997     H264Context *hx;
1998     av_freep(&h->intra4x4_pred_mode);
1999     av_freep(&h->chroma_pred_mode_table);
2000     av_freep(&h->cbp_table);
2001     av_freep(&h->mvd_table[0]);
2002     av_freep(&h->mvd_table[1]);
2003     av_freep(&h->direct_table);
2004     av_freep(&h->non_zero_count);
2005     av_freep(&h->slice_table_base);
2006     h->slice_table= NULL;
2007
2008     av_freep(&h->mb2b_xy);
2009     av_freep(&h->mb2b8_xy);
2010
2011     for(i = 0; i < MAX_SPS_COUNT; i++)
2012         av_freep(h->sps_buffers + i);
2013
2014     for(i = 0; i < MAX_PPS_COUNT; i++)
2015         av_freep(h->pps_buffers + i);
2016
2017     for(i = 0; i < h->s.avctx->thread_count; i++) {
2018         hx = h->thread_context[i];
2019         if(!hx) continue;
2020         av_freep(&hx->top_borders[1]);
2021         av_freep(&hx->top_borders[0]);
2022         av_freep(&hx->s.obmc_scratchpad);
2023     }
2024 }
2025
2026 static void init_dequant8_coeff_table(H264Context *h){
2027     int i,q,x;
2028     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2029     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2030     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2031
2032     for(i=0; i<2; i++ ){
2033         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2034             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2035             break;
2036         }
2037
2038         for(q=0; q<52; q++){
2039             int shift = ff_div6[q];
2040             int idx = ff_rem6[q];
2041             for(x=0; x<64; x++)
2042                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2043                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2044                     h->pps.scaling_matrix8[i][x]) << shift;
2045         }
2046     }
2047 }
2048
2049 static void init_dequant4_coeff_table(H264Context *h){
2050     int i,j,q,x;
2051     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2052     for(i=0; i<6; i++ ){
2053         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2054         for(j=0; j<i; j++){
2055             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2056                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2057                 break;
2058             }
2059         }
2060         if(j<i)
2061             continue;
2062
2063         for(q=0; q<52; q++){
2064             int shift = ff_div6[q] + 2;
2065             int idx = ff_rem6[q];
2066             for(x=0; x<16; x++)
2067                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2068                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2069                     h->pps.scaling_matrix4[i][x]) << shift;
2070         }
2071     }
2072 }
2073
2074 static void init_dequant_tables(H264Context *h){
2075     int i,x;
2076     init_dequant4_coeff_table(h);
2077     if(h->pps.transform_8x8_mode)
2078         init_dequant8_coeff_table(h);
2079     if(h->sps.transform_bypass){
2080         for(i=0; i<6; i++)
2081             for(x=0; x<16; x++)
2082                 h->dequant4_coeff[i][0][x] = 1<<6;
2083         if(h->pps.transform_8x8_mode)
2084             for(i=0; i<2; i++)
2085                 for(x=0; x<64; x++)
2086                     h->dequant8_coeff[i][0][x] = 1<<6;
2087     }
2088 }
2089
2090
2091 /**
2092  * allocates tables.
2093  * needs width/height
2094  */
2095 static int alloc_tables(H264Context *h){
2096     MpegEncContext * const s = &h->s;
2097     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2098     int x,y;
2099
2100     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2101
2102     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2103     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2104     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2105
2106     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2107     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2108     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2109     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2110
2111     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2112     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2113
2114     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2115     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2116     for(y=0; y<s->mb_height; y++){
2117         for(x=0; x<s->mb_width; x++){
2118             const int mb_xy= x + y*s->mb_stride;
2119             const int b_xy = 4*x + 4*y*h->b_stride;
2120             const int b8_xy= 2*x + 2*y*h->b8_stride;
2121
2122             h->mb2b_xy [mb_xy]= b_xy;
2123             h->mb2b8_xy[mb_xy]= b8_xy;
2124         }
2125     }
2126
2127     s->obmc_scratchpad = NULL;
2128
2129     if(!h->dequant4_coeff[0])
2130         init_dequant_tables(h);
2131
2132     return 0;
2133 fail:
2134     free_tables(h);
2135     return -1;
2136 }
2137
2138 /**
2139  * Mimic alloc_tables(), but for every context thread.
2140  */
2141 static void clone_tables(H264Context *dst, H264Context *src){
2142     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2143     dst->non_zero_count           = src->non_zero_count;
2144     dst->slice_table              = src->slice_table;
2145     dst->cbp_table                = src->cbp_table;
2146     dst->mb2b_xy                  = src->mb2b_xy;
2147     dst->mb2b8_xy                 = src->mb2b8_xy;
2148     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2149     dst->mvd_table[0]             = src->mvd_table[0];
2150     dst->mvd_table[1]             = src->mvd_table[1];
2151     dst->direct_table             = src->direct_table;
2152
2153     dst->s.obmc_scratchpad = NULL;
2154     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2155 }
2156
2157 /**
2158  * Init context
2159  * Allocate buffers which are not shared amongst multiple threads.
2160  */
2161 static int context_init(H264Context *h){
2162     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2163     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2164
2165     return 0;
2166 fail:
2167     return -1; // free_tables will clean up for us
2168 }
2169
2170 static av_cold void common_init(H264Context *h){
2171     MpegEncContext * const s = &h->s;
2172
2173     s->width = s->avctx->width;
2174     s->height = s->avctx->height;
2175     s->codec_id= s->avctx->codec->id;
2176
2177     ff_h264_pred_init(&h->hpc, s->codec_id);
2178
2179     h->dequant_coeff_pps= -1;
2180     s->unrestricted_mv=1;
2181     s->decode=1; //FIXME
2182
2183     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2184     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2185 }
2186
2187 static av_cold int decode_init(AVCodecContext *avctx){
2188     H264Context *h= avctx->priv_data;
2189     MpegEncContext * const s = &h->s;
2190
2191     MPV_decode_defaults(s);
2192
2193     s->avctx = avctx;
2194     common_init(h);
2195
2196     s->out_format = FMT_H264;
2197     s->workaround_bugs= avctx->workaround_bugs;
2198
2199     // set defaults
2200 //    s->decode_mb= ff_h263_decode_mb;
2201     s->quarter_sample = 1;
2202     s->low_delay= 1;
2203     avctx->pix_fmt= PIX_FMT_YUV420P;
2204
2205     decode_init_vlc();
2206
2207     if(avctx->extradata_size > 0 && avctx->extradata &&
2208        *(char *)avctx->extradata == 1){
2209         h->is_avc = 1;
2210         h->got_avcC = 0;
2211     } else {
2212         h->is_avc = 0;
2213     }
2214
2215     h->thread_context[0] = h;
2216     return 0;
2217 }
2218
2219 static int frame_start(H264Context *h){
2220     MpegEncContext * const s = &h->s;
2221     int i;
2222
2223     if(MPV_frame_start(s, s->avctx) < 0)
2224         return -1;
2225     ff_er_frame_start(s);
2226     /*
2227      * MPV_frame_start uses pict_type to derive key_frame.
2228      * This is incorrect for H.264; IDR markings must be used.
2229      * Zero here; IDR markings per slice in frame or fields are OR'd in later.
2230      * See decode_nal_units().
2231      */
2232     s->current_picture_ptr->key_frame= 0;
2233
2234     assert(s->linesize && s->uvlinesize);
2235
2236     for(i=0; i<16; i++){
2237         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2238         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2239     }
2240     for(i=0; i<4; i++){
2241         h->block_offset[16+i]=
2242         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2243         h->block_offset[24+16+i]=
2244         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2245     }
2246
2247     /* can't be in alloc_tables because linesize isn't known there.
2248      * FIXME: redo bipred weight to not require extra buffer? */
2249     for(i = 0; i < s->avctx->thread_count; i++)
2250         if(!h->thread_context[i]->s.obmc_scratchpad)
2251             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2252
2253     /* some macroblocks will be accessed before they're available */
2254     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2255         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2256
2257 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2258     return 0;
2259 }
2260
2261 static inline void backup_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int simple){
2262     MpegEncContext * const s = &h->s;
2263     int i;
2264
2265     src_y  -=   linesize;
2266     src_cb -= uvlinesize;
2267     src_cr -= uvlinesize;
2268
2269     // There are two lines saved, the line above the the top macroblock of a pair,
2270     // and the line above the bottom macroblock
2271     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2272     for(i=1; i<17; i++){
2273         h->left_border[i]= src_y[15+i*  linesize];
2274     }
2275
2276     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2277     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2278
2279     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2280         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2281         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2282         for(i=1; i<9; i++){
2283             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2284             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2285         }
2286         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2287         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2288     }
2289 }
2290
2291 static inline void xchg_mb_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg, int simple){
2292     MpegEncContext * const s = &h->s;
2293     int temp8, i;
2294     uint64_t temp64;
2295     int deblock_left;
2296     int deblock_top;
2297     int mb_xy;
2298
2299     if(h->deblocking_filter == 2) {
2300         mb_xy = h->mb_xy;
2301         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2302         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2303     } else {
2304         deblock_left = (s->mb_x > 0);
2305         deblock_top =  (s->mb_y > 0);
2306     }
2307
2308     src_y  -=   linesize + 1;
2309     src_cb -= uvlinesize + 1;
2310     src_cr -= uvlinesize + 1;
2311
2312 #define XCHG(a,b,t,xchg)\
2313 t= a;\
2314 if(xchg)\
2315     a= b;\
2316 b= t;
2317
2318     if(deblock_left){
2319         for(i = !deblock_top; i<17; i++){
2320             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2321         }
2322     }
2323
2324     if(deblock_top){
2325         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2326         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2327         if(s->mb_x+1 < s->mb_width){
2328             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2329         }
2330     }
2331
2332     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2333         if(deblock_left){
2334             for(i = !deblock_top; i<9; i++){
2335                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2336                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2337             }
2338         }
2339         if(deblock_top){
2340             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2341             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2342         }
2343     }
2344 }
2345
2346 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2347     MpegEncContext * const s = &h->s;
2348     int i;
2349
2350     src_y  -= 2 *   linesize;
2351     src_cb -= 2 * uvlinesize;
2352     src_cr -= 2 * uvlinesize;
2353
2354     // There are two lines saved, the line above the the top macroblock of a pair,
2355     // and the line above the bottom macroblock
2356     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2357     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2358     for(i=2; i<34; i++){
2359         h->left_border[i]= src_y[15+i*  linesize];
2360     }
2361
2362     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2363     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2364     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2365     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2366
2367     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2368         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2369         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2370         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2371         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2372         for(i=2; i<18; i++){
2373             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2374             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2375         }
2376         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2377         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2378         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2379         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2380     }
2381 }
2382
2383 static inline void xchg_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize, int xchg){
2384     MpegEncContext * const s = &h->s;
2385     int temp8, i;
2386     uint64_t temp64;
2387     int deblock_left = (s->mb_x > 0);
2388     int deblock_top  = (s->mb_y > 1);
2389
2390     tprintf(s->avctx, "xchg_pair_border: src_y:%p src_cb:%p src_cr:%p ls:%d uvls:%d\n", src_y, src_cb, src_cr, linesize, uvlinesize);
2391
2392     src_y  -= 2 *   linesize + 1;
2393     src_cb -= 2 * uvlinesize + 1;
2394     src_cr -= 2 * uvlinesize + 1;
2395
2396 #define XCHG(a,b,t,xchg)\
2397 t= a;\
2398 if(xchg)\
2399     a= b;\
2400 b= t;
2401
2402     if(deblock_left){
2403         for(i = (!deblock_top)<<1; i<34; i++){
2404             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2405         }
2406     }
2407
2408     if(deblock_top){
2409         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2410         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2411         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2412         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2413         if(s->mb_x+1 < s->mb_width){
2414             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2415             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2416         }
2417     }
2418
2419     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2420         if(deblock_left){
2421             for(i = (!deblock_top) << 1; i<18; i++){
2422                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2423                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2424             }
2425         }
2426         if(deblock_top){
2427             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2428             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2429             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2430             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2431         }
2432     }
2433 }
2434
2435 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2436     MpegEncContext * const s = &h->s;
2437     const int mb_x= s->mb_x;
2438     const int mb_y= s->mb_y;
2439     const int mb_xy= h->mb_xy;
2440     const int mb_type= s->current_picture.mb_type[mb_xy];
2441     uint8_t  *dest_y, *dest_cb, *dest_cr;
2442     int linesize, uvlinesize /*dct_offset*/;
2443     int i;
2444     int *block_offset = &h->block_offset[0];
2445     const unsigned int bottom = mb_y & 1;
2446     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2447     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2448     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2449
2450     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2451     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2452     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2453
2454     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2455     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2456
2457     if (!simple && MB_FIELD) {
2458         linesize   = h->mb_linesize   = s->linesize * 2;
2459         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2460         block_offset = &h->block_offset[24];
2461         if(mb_y&1){ //FIXME move out of this func?
2462             dest_y -= s->linesize*15;
2463             dest_cb-= s->uvlinesize*7;
2464             dest_cr-= s->uvlinesize*7;
2465         }
2466         if(FRAME_MBAFF) {
2467             int list;
2468             for(list=0; list<h->list_count; list++){
2469                 if(!USES_LIST(mb_type, list))
2470                     continue;
2471                 if(IS_16X16(mb_type)){
2472                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2473                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2474                 }else{
2475                     for(i=0; i<16; i+=4){
2476                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2477                         int ref = h->ref_cache[list][scan8[i]];
2478                         if(ref >= 0)
2479                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2480                     }
2481                 }
2482             }
2483         }
2484     } else {
2485         linesize   = h->mb_linesize   = s->linesize;
2486         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2487 //        dct_offset = s->linesize * 16;
2488     }
2489
2490     if(transform_bypass){
2491         idct_dc_add =
2492         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2493     }else if(IS_8x8DCT(mb_type)){
2494         idct_dc_add = s->dsp.h264_idct8_dc_add;
2495         idct_add = s->dsp.h264_idct8_add;
2496     }else{
2497         idct_dc_add = s->dsp.h264_idct_dc_add;
2498         idct_add = s->dsp.h264_idct_add;
2499     }
2500
2501     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2502        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2503         int mbt_y = mb_y&~1;
2504         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2505         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2506         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2507         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2508     }
2509
2510     if (!simple && IS_INTRA_PCM(mb_type)) {
2511         unsigned int x, y;
2512
2513         // The pixels are stored in h->mb array in the same order as levels,
2514         // copy them in output in the correct order.
2515         for(i=0; i<16; i++) {
2516             for (y=0; y<4; y++) {
2517                 for (x=0; x<4; x++) {
2518                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2519                 }
2520             }
2521         }
2522         for(i=16; i<16+4; i++) {
2523             for (y=0; y<4; y++) {
2524                 for (x=0; x<4; x++) {
2525                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2526                 }
2527             }
2528         }
2529         for(i=20; i<20+4; i++) {
2530             for (y=0; y<4; y++) {
2531                 for (x=0; x<4; x++) {
2532                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2533                 }
2534             }
2535         }
2536     } else {
2537         if(IS_INTRA(mb_type)){
2538             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2539                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2540
2541             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2542                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2543                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2544             }
2545
2546             if(IS_INTRA4x4(mb_type)){
2547                 if(simple || !s->encoding){
2548                     if(IS_8x8DCT(mb_type)){
2549                         for(i=0; i<16; i+=4){
2550                             uint8_t * const ptr= dest_y + block_offset[i];
2551                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2552                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2553                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2554                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2555                             if(nnz){
2556                                 if(nnz == 1 && h->mb[i*16])
2557                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2558                                 else
2559                                     idct_add(ptr, h->mb + i*16, linesize);
2560                             }
2561                         }
2562                     }else
2563                     for(i=0; i<16; i++){
2564                         uint8_t * const ptr= dest_y + block_offset[i];
2565                         uint8_t *topright;
2566                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2567                         int nnz, tr;
2568
2569                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2570                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2571                             assert(mb_y || linesize <= block_offset[i]);
2572                             if(!topright_avail){
2573                                 tr= ptr[3 - linesize]*0x01010101;
2574                                 topright= (uint8_t*) &tr;
2575                             }else
2576                                 topright= ptr + 4 - linesize;
2577                         }else
2578                             topright= NULL;
2579
2580                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2581                         nnz = h->non_zero_count_cache[ scan8[i] ];
2582                         if(nnz){
2583                             if(is_h264){
2584                                 if(nnz == 1 && h->mb[i*16])
2585                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2586                                 else
2587                                     idct_add(ptr, h->mb + i*16, linesize);
2588                             }else
2589                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2590                         }
2591                     }
2592                 }
2593             }else{
2594                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2595                 if(is_h264){
2596                     if(!transform_bypass)
2597                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2598                 }else
2599                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2600             }
2601             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2602                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2603         }else if(is_h264){
2604             hl_motion(h, dest_y, dest_cb, dest_cr,
2605                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2606                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2607                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2608         }
2609
2610
2611         if(!IS_INTRA4x4(mb_type)){
2612             if(is_h264){
2613                 if(IS_INTRA16x16(mb_type)){
2614                     for(i=0; i<16; i++){
2615                         if(h->non_zero_count_cache[ scan8[i] ])
2616                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2617                         else if(h->mb[i*16])
2618                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2619                     }
2620                 }else{
2621                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2622                     for(i=0; i<16; i+=di){
2623                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2624                         if(nnz){
2625                             if(nnz==1 && h->mb[i*16])
2626                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2627                             else
2628                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2629                         }
2630                     }
2631                 }
2632             }else{
2633                 for(i=0; i<16; i++){
2634                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2635                         uint8_t * const ptr= dest_y + block_offset[i];
2636                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2637                     }
2638                 }
2639             }
2640         }
2641
2642         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2643             uint8_t *dest[2] = {dest_cb, dest_cr};
2644             if(transform_bypass){
2645                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2646             }else{
2647                 idct_add = s->dsp.h264_idct_add;
2648                 idct_dc_add = s->dsp.h264_idct_dc_add;
2649                 chroma_dc_dequant_idct_c(h->mb + 16*16, h->chroma_qp[0], h->dequant4_coeff[IS_INTRA(mb_type) ? 1:4][h->chroma_qp[0]][0]);
2650                 chroma_dc_dequant_idct_c(h->mb + 16*16+4*16, h->chroma_qp[1], h->dequant4_coeff[IS_INTRA(mb_type) ? 2:5][h->chroma_qp[1]][0]);
2651             }
2652             if(is_h264){
2653                 for(i=16; i<16+8; i++){
2654                     if(h->non_zero_count_cache[ scan8[i] ])
2655                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2656                     else if(h->mb[i*16])
2657                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2658                 }
2659             }else{
2660                 for(i=16; i<16+8; i++){
2661                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2662                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2663                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2664                     }
2665                 }
2666             }
2667         }
2668     }
2669     if(h->deblocking_filter) {
2670         if (!simple && FRAME_MBAFF) {
2671             //FIXME try deblocking one mb at a time?
2672             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2673             const int mb_y = s->mb_y - 1;
2674             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2675             const int mb_xy= mb_x + mb_y*s->mb_stride;
2676             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2677             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2678             if (!bottom) return;
2679             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2680             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2681             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2682
2683             if(IS_INTRA(mb_type_top | mb_type_bottom))
2684                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2685
2686             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2687             // deblock a pair
2688             // top
2689             s->mb_y--; h->mb_xy -= s->mb_stride;
2690             tprintf(h->s.avctx, "call mbaff filter_mb mb_x:%d mb_y:%d pair_dest_y = %p, dest_y = %p\n", mb_x, mb_y, pair_dest_y, dest_y);
2691             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2692             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2693             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2694             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2695             // bottom
2696             s->mb_y++; h->mb_xy += s->mb_stride;
2697             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2698             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2699             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2700             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2701             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2702         } else {
2703             tprintf(h->s.avctx, "call filter_mb\n");
2704             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2705             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2706             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2707         }
2708     }
2709 }
2710
2711 /**
2712  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2713  */
2714 static void hl_decode_mb_simple(H264Context *h){
2715     hl_decode_mb_internal(h, 1);
2716 }
2717
2718 /**
2719  * Process a macroblock; this handles edge cases, such as interlacing.
2720  */
2721 static void av_noinline hl_decode_mb_complex(H264Context *h){
2722     hl_decode_mb_internal(h, 0);
2723 }
2724
2725 static void hl_decode_mb(H264Context *h){
2726     MpegEncContext * const s = &h->s;
2727     const int mb_xy= h->mb_xy;
2728     const int mb_type= s->current_picture.mb_type[mb_xy];
2729     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 || (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding);
2730
2731     if(ENABLE_H264_ENCODER && !s->decode)
2732         return;
2733
2734     if (is_complex)
2735         hl_decode_mb_complex(h);
2736     else hl_decode_mb_simple(h);
2737 }
2738
2739 static void pic_as_field(Picture *pic, const int parity){
2740     int i;
2741     for (i = 0; i < 4; ++i) {
2742         if (parity == PICT_BOTTOM_FIELD)
2743             pic->data[i] += pic->linesize[i];
2744         pic->reference = parity;
2745         pic->linesize[i] *= 2;
2746     }
2747 }
2748
2749 static int split_field_copy(Picture *dest, Picture *src,
2750                             int parity, int id_add){
2751     int match = !!(src->reference & parity);
2752
2753     if (match) {
2754         *dest = *src;
2755         pic_as_field(dest, parity);
2756         dest->pic_id *= 2;
2757         dest->pic_id += id_add;
2758     }
2759
2760     return match;
2761 }
2762
2763 /**
2764  * Split one reference list into field parts, interleaving by parity
2765  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2766  * set to look at the actual start of data for that field.
2767  *
2768  * @param dest output list
2769  * @param dest_len maximum number of fields to put in dest
2770  * @param src the source reference list containing fields and/or field pairs
2771  *            (aka short_ref/long_ref, or
2772  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2773  * @param src_len number of Picture's in source (pairs and unmatched fields)
2774  * @param parity the parity of the picture being decoded/needing
2775  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2776  * @return number of fields placed in dest
2777  */
2778 static int split_field_half_ref_list(Picture *dest, int dest_len,
2779                                      Picture *src,  int src_len,  int parity){
2780     int same_parity   = 1;
2781     int same_i        = 0;
2782     int opp_i         = 0;
2783     int out_i;
2784     int field_output;
2785
2786     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2787         if (same_parity && same_i < src_len) {
2788             field_output = split_field_copy(dest + out_i, src + same_i,
2789                                             parity, 1);
2790             same_parity = !field_output;
2791             same_i++;
2792
2793         } else if (opp_i < src_len) {
2794             field_output = split_field_copy(dest + out_i, src + opp_i,
2795                                             PICT_FRAME - parity, 0);
2796             same_parity = field_output;
2797             opp_i++;
2798
2799         } else {
2800             break;
2801         }
2802     }
2803
2804     return out_i;
2805 }
2806
2807 /**
2808  * Split the reference frame list into a reference field list.
2809  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2810  * The input list contains both reference field pairs and
2811  * unmatched reference fields; it is ordered as spec describes
2812  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2813  * unmatched field pairs are also present. Conceptually this is equivalent
2814  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2815  *
2816  * @param dest output reference list where ordered fields are to be placed
2817  * @param dest_len max number of fields to place at dest
2818  * @param src source reference list, as described above
2819  * @param src_len number of pictures (pairs and unmatched fields) in src
2820  * @param parity parity of field being currently decoded
2821  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2822  * @param long_i index into src array that holds first long reference picture,
2823  *        or src_len if no long refs present.
2824  */
2825 static int split_field_ref_list(Picture *dest, int dest_len,
2826                                 Picture *src,  int src_len,
2827                                 int parity,    int long_i){
2828
2829     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2830     dest += i;
2831     dest_len -= i;
2832
2833     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2834                                    src_len - long_i, parity);
2835     return i;
2836 }
2837
2838 /**
2839  * fills the default_ref_list.
2840  */
2841 static int fill_default_ref_list(H264Context *h){
2842     MpegEncContext * const s = &h->s;
2843     int i;
2844     int smallest_poc_greater_than_current = -1;
2845     int structure_sel;
2846     Picture sorted_short_ref[32];
2847     Picture field_entry_list[2][32];
2848     Picture *frame_list[2];
2849
2850     if (FIELD_PICTURE) {
2851         structure_sel = PICT_FRAME;
2852         frame_list[0] = field_entry_list[0];
2853         frame_list[1] = field_entry_list[1];
2854     } else {
2855         structure_sel = 0;
2856         frame_list[0] = h->default_ref_list[0];
2857         frame_list[1] = h->default_ref_list[1];
2858     }
2859
2860     if(h->slice_type==FF_B_TYPE){
2861         int list;
2862         int len[2];
2863         int short_len[2];
2864         int out_i;
2865         int limit= INT_MIN;
2866
2867         /* sort frame according to poc in B slice */
2868         for(out_i=0; out_i<h->short_ref_count; out_i++){
2869             int best_i=INT_MIN;
2870             int best_poc=INT_MAX;
2871
2872             for(i=0; i<h->short_ref_count; i++){
2873                 const int poc= h->short_ref[i]->poc;
2874                 if(poc > limit && poc < best_poc){
2875                     best_poc= poc;
2876                     best_i= i;
2877                 }
2878             }
2879
2880             assert(best_i != INT_MIN);
2881
2882             limit= best_poc;
2883             sorted_short_ref[out_i]= *h->short_ref[best_i];
2884             tprintf(h->s.avctx, "sorted poc: %d->%d poc:%d fn:%d\n", best_i, out_i, sorted_short_ref[out_i].poc, sorted_short_ref[out_i].frame_num);
2885             if (-1 == smallest_poc_greater_than_current) {
2886                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2887                     smallest_poc_greater_than_current = out_i;
2888                 }
2889             }
2890         }
2891
2892         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2893
2894         // find the largest poc
2895         for(list=0; list<2; list++){
2896             int index = 0;
2897             int j= -99;
2898             int step= list ? -1 : 1;
2899
2900             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2901                 int sel;
2902                 while(j<0 || j>= h->short_ref_count){
2903                     if(j != -99 && step == (list ? -1 : 1))
2904                         return -1;
2905                     step = -step;
2906                     j= smallest_poc_greater_than_current + (step>>1);
2907                 }
2908                 sel = sorted_short_ref[j].reference | structure_sel;
2909                 if(sel != PICT_FRAME) continue;
2910                 frame_list[list][index  ]= sorted_short_ref[j];
2911                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2912             }
2913             short_len[list] = index;
2914
2915             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2916                 int sel;
2917                 if(h->long_ref[i] == NULL) continue;
2918                 sel = h->long_ref[i]->reference | structure_sel;
2919                 if(sel != PICT_FRAME) continue;
2920
2921                 frame_list[ list ][index  ]= *h->long_ref[i];
2922                 frame_list[ list ][index++].pic_id= i;
2923             }
2924             len[list] = index;
2925         }
2926
2927         for(list=0; list<2; list++){
2928             if (FIELD_PICTURE)
2929                 len[list] = split_field_ref_list(h->default_ref_list[list],
2930                                                  h->ref_count[list],
2931                                                  frame_list[list],
2932                                                  len[list],
2933                                                  s->picture_structure,
2934                                                  short_len[list]);
2935
2936             // swap the two first elements of L1 when L0 and L1 are identical
2937             if(list && len[0] > 1 && len[0] == len[1])
2938                 for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0]; i++)
2939                     if(i == len[0]){
2940                         FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2941                         break;
2942                     }
2943
2944             if(len[list] < h->ref_count[ list ])
2945                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2946         }
2947
2948
2949     }else{
2950         int index=0;
2951         int short_len;
2952         for(i=0; i<h->short_ref_count; i++){
2953             int sel;
2954             sel = h->short_ref[i]->reference | structure_sel;
2955             if(sel != PICT_FRAME) continue;
2956             frame_list[0][index  ]= *h->short_ref[i];
2957             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2958         }
2959         short_len = index;
2960         for(i = 0; i < 16; i++){
2961             int sel;
2962             if(h->long_ref[i] == NULL) continue;
2963             sel = h->long_ref[i]->reference | structure_sel;
2964             if(sel != PICT_FRAME) continue;
2965             frame_list[0][index  ]= *h->long_ref[i];
2966             frame_list[0][index++].pic_id= i;
2967         }
2968
2969         if (FIELD_PICTURE)
2970             index = split_field_ref_list(h->default_ref_list[0],
2971                                          h->ref_count[0], frame_list[0],
2972                                          index, s->picture_structure,
2973                                          short_len);
2974
2975         if(index < h->ref_count[0])
2976             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
2977     }
2978 #ifdef TRACE
2979     for (i=0; i<h->ref_count[0]; i++) {
2980         tprintf(h->s.avctx, "List0: %s fn:%d 0x%p\n", (h->default_ref_list[0][i].long_ref ? "LT" : "ST"), h->default_ref_list[0][i].pic_id, h->default_ref_list[0][i].data[0]);
2981     }
2982     if(h->slice_type==FF_B_TYPE){
2983         for (i=0; i<h->ref_count[1]; i++) {
2984             tprintf(h->s.avctx, "List1: %s fn:%d 0x%p\n", (h->default_ref_list[1][i].long_ref ? "LT" : "ST"), h->default_ref_list[1][i].pic_id, h->default_ref_list[1][i].data[0]);
2985         }
2986     }
2987 #endif
2988     return 0;
2989 }
2990
2991 static void print_short_term(H264Context *h);
2992 static void print_long_term(H264Context *h);
2993
2994 /**
2995  * Extract structure information about the picture described by pic_num in
2996  * the current decoding context (frame or field). Note that pic_num is
2997  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2998  * @param pic_num picture number for which to extract structure information
2999  * @param structure one of PICT_XXX describing structure of picture
3000  *                      with pic_num
3001  * @return frame number (short term) or long term index of picture
3002  *         described by pic_num
3003  */
3004 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3005     MpegEncContext * const s = &h->s;
3006
3007     *structure = s->picture_structure;
3008     if(FIELD_PICTURE){
3009         if (!(pic_num & 1))
3010             /* opposite field */
3011             *structure ^= PICT_FRAME;
3012         pic_num >>= 1;
3013     }
3014
3015     return pic_num;
3016 }
3017
3018 static int decode_ref_pic_list_reordering(H264Context *h){
3019     MpegEncContext * const s = &h->s;
3020     int list, index, pic_structure;
3021
3022     print_short_term(h);
3023     print_long_term(h);
3024     if(h->slice_type==FF_I_TYPE || h->slice_type==FF_SI_TYPE) return 0; //FIXME move before func
3025
3026     for(list=0; list<h->list_count; list++){
3027         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3028
3029         if(get_bits1(&s->gb)){
3030             int pred= h->curr_pic_num;
3031
3032             for(index=0; ; index++){
3033                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3034                 unsigned int pic_id;
3035                 int i;
3036                 Picture *ref = NULL;
3037
3038                 if(reordering_of_pic_nums_idc==3)
3039                     break;
3040
3041                 if(index >= h->ref_count[list]){
3042                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3043                     return -1;
3044                 }
3045
3046                 if(reordering_of_pic_nums_idc<3){
3047                     if(reordering_of_pic_nums_idc<2){
3048                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3049                         int frame_num;
3050
3051                         if(abs_diff_pic_num > h->max_pic_num){
3052                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3053                             return -1;
3054                         }
3055
3056                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3057                         else                                pred+= abs_diff_pic_num;
3058                         pred &= h->max_pic_num - 1;
3059
3060                         frame_num = pic_num_extract(h, pred, &pic_structure);
3061
3062                         for(i= h->short_ref_count-1; i>=0; i--){
3063                             ref = h->short_ref[i];
3064                             assert(ref->reference);
3065                             assert(!ref->long_ref);
3066                             if(ref->data[0] != NULL &&
3067                                    ref->frame_num == frame_num &&
3068                                    (ref->reference & pic_structure) &&
3069                                    ref->long_ref == 0) // ignore non existing pictures by testing data[0] pointer
3070                                 break;
3071                         }
3072                         if(i>=0)
3073                             ref->pic_id= pred;
3074                     }else{
3075                         int long_idx;
3076                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3077
3078                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3079
3080                         if(long_idx>31){
3081                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3082                             return -1;
3083                         }
3084                         ref = h->long_ref[long_idx];
3085                         assert(!(ref && !ref->reference));
3086                         if(ref && (ref->reference & pic_structure)){
3087                             ref->pic_id= pic_id;
3088                             assert(ref->long_ref);
3089                             i=0;
3090                         }else{
3091                             i=-1;
3092                         }
3093                     }
3094
3095                     if (i < 0) {
3096                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3097                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3098                     } else {
3099                         for(i=index; i+1<h->ref_count[list]; i++){
3100                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3101                                 break;
3102                         }
3103                         for(; i > index; i--){
3104                             h->ref_list[list][i]= h->ref_list[list][i-1];
3105                         }
3106                         h->ref_list[list][index]= *ref;
3107                         if (FIELD_PICTURE){
3108                             pic_as_field(&h->ref_list[list][index], pic_structure);
3109                         }
3110                     }
3111                 }else{
3112                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3113                     return -1;
3114                 }
3115             }
3116         }
3117     }
3118     for(list=0; list<h->list_count; list++){
3119         for(index= 0; index < h->ref_count[list]; index++){
3120             if(!h->ref_list[list][index].data[0])
3121                 h->ref_list[list][index]= s->current_picture;
3122         }
3123     }
3124
3125     if(h->slice_type==FF_B_TYPE && !h->direct_spatial_mv_pred)
3126         direct_dist_scale_factor(h);
3127     direct_ref_list_init(h);
3128     return 0;
3129 }
3130
3131 static void fill_mbaff_ref_list(H264Context *h){
3132     int list, i, j;
3133     for(list=0; list<2; list++){ //FIXME try list_count
3134         for(i=0; i<h->ref_count[list]; i++){
3135             Picture *frame = &h->ref_list[list][i];
3136             Picture *field = &h->ref_list[list][16+2*i];
3137             field[0] = *frame;
3138             for(j=0; j<3; j++)
3139                 field[0].linesize[j] <<= 1;
3140             field[0].reference = PICT_TOP_FIELD;
3141             field[1] = field[0];
3142             for(j=0; j<3; j++)
3143                 field[1].data[j] += frame->linesize[j];
3144             field[1].reference = PICT_BOTTOM_FIELD;
3145
3146             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3147             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3148             for(j=0; j<2; j++){
3149                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3150                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3151             }
3152         }
3153     }
3154     for(j=0; j<h->ref_count[1]; j++){
3155         for(i=0; i<h->ref_count[0]; i++)
3156             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3157         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3158         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3159     }
3160 }
3161
3162 static int pred_weight_table(H264Context *h){
3163     MpegEncContext * const s = &h->s;
3164     int list, i;
3165     int luma_def, chroma_def;
3166
3167     h->use_weight= 0;
3168     h->use_weight_chroma= 0;
3169     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3170     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3171     luma_def = 1<<h->luma_log2_weight_denom;
3172     chroma_def = 1<<h->chroma_log2_weight_denom;
3173
3174     for(list=0; list<2; list++){
3175         for(i=0; i<h->ref_count[list]; i++){
3176             int luma_weight_flag, chroma_weight_flag;
3177
3178             luma_weight_flag= get_bits1(&s->gb);
3179             if(luma_weight_flag){
3180                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3181                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3182                 if(   h->luma_weight[list][i] != luma_def
3183                    || h->luma_offset[list][i] != 0)
3184                     h->use_weight= 1;
3185             }else{
3186                 h->luma_weight[list][i]= luma_def;
3187                 h->luma_offset[list][i]= 0;
3188             }
3189
3190             chroma_weight_flag= get_bits1(&s->gb);
3191             if(chroma_weight_flag){
3192                 int j;
3193                 for(j=0; j<2; j++){
3194                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3195                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3196                     if(   h->chroma_weight[list][i][j] != chroma_def
3197                        || h->chroma_offset[list][i][j] != 0)
3198                         h->use_weight_chroma= 1;
3199                 }
3200             }else{
3201                 int j;
3202                 for(j=0; j<2; j++){
3203                     h->chroma_weight[list][i][j]= chroma_def;
3204                     h->chroma_offset[list][i][j]= 0;
3205                 }
3206             }
3207         }
3208         if(h->slice_type != FF_B_TYPE) break;
3209     }
3210     h->use_weight= h->use_weight || h->use_weight_chroma;
3211     return 0;
3212 }
3213
3214 static void implicit_weight_table(H264Context *h){
3215     MpegEncContext * const s = &h->s;
3216     int ref0, ref1;
3217     int cur_poc = s->current_picture_ptr->poc;
3218
3219     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3220        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3221         h->use_weight= 0;
3222         h->use_weight_chroma= 0;
3223         return;
3224     }
3225
3226     h->use_weight= 2;
3227     h->use_weight_chroma= 2;
3228     h->luma_log2_weight_denom= 5;
3229     h->chroma_log2_weight_denom= 5;
3230
3231     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3232         int poc0 = h->ref_list[0][ref0].poc;
3233         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3234             int poc1 = h->ref_list[1][ref1].poc;
3235             int td = av_clip(poc1 - poc0, -128, 127);
3236             if(td){
3237                 int tb = av_clip(cur_poc - poc0, -128, 127);
3238                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3239                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3240                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3241                     h->implicit_weight[ref0][ref1] = 32;
3242                 else
3243                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3244             }else
3245                 h->implicit_weight[ref0][ref1] = 32;
3246         }
3247     }
3248 }
3249
3250 /**
3251  * Mark a picture as no longer needed for reference. The refmask
3252  * argument allows unreferencing of individual fields or the whole frame.
3253  * If the picture becomes entirely unreferenced, but is being held for
3254  * display purposes, it is marked as such.
3255  * @param refmask mask of fields to unreference; the mask is bitwise
3256  *                anded with the reference marking of pic
3257  * @return non-zero if pic becomes entirely unreferenced (except possibly
3258  *         for display purposes) zero if one of the fields remains in
3259  *         reference
3260  */
3261 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3262     int i;
3263     if (pic->reference &= refmask) {
3264         return 0;
3265     } else {
3266         if(pic == h->delayed_output_pic)
3267             pic->reference=DELAYED_PIC_REF;
3268         else{
3269             for(i = 0; h->delayed_pic[i]; i++)
3270                 if(pic == h->delayed_pic[i]){
3271                     pic->reference=DELAYED_PIC_REF;
3272                     break;
3273                 }
3274         }
3275         return 1;
3276     }
3277 }
3278
3279 /**
3280  * instantaneous decoder refresh.
3281  */
3282 static void idr(H264Context *h){
3283     int i;
3284
3285     for(i=0; i<16; i++){
3286         if (h->long_ref[i] != NULL) {
3287             unreference_pic(h, h->long_ref[i], 0);
3288             h->long_ref[i]= NULL;
3289         }
3290     }
3291     h->long_ref_count=0;
3292
3293     for(i=0; i<h->short_ref_count; i++){
3294         unreference_pic(h, h->short_ref[i], 0);
3295         h->short_ref[i]= NULL;
3296     }
3297     h->short_ref_count=0;
3298 }
3299
3300 /* forget old pics after a seek */
3301 static void flush_dpb(AVCodecContext *avctx){
3302     H264Context *h= avctx->priv_data;
3303     int i;
3304     for(i=0; i<16; i++) {
3305         if(h->delayed_pic[i])
3306             h->delayed_pic[i]->reference= 0;
3307         h->delayed_pic[i]= NULL;
3308     }
3309     if(h->delayed_output_pic)
3310         h->delayed_output_pic->reference= 0;
3311     h->delayed_output_pic= NULL;
3312     idr(h);
3313     if(h->s.current_picture_ptr)
3314         h->s.current_picture_ptr->reference= 0;
3315     h->s.first_field= 0;
3316     ff_mpeg_flush(avctx);
3317 }
3318
3319 /**
3320  * Find a Picture in the short term reference list by frame number.
3321  * @param frame_num frame number to search for
3322  * @param idx the index into h->short_ref where returned picture is found
3323  *            undefined if no picture found.
3324  * @return pointer to the found picture, or NULL if no pic with the provided
3325  *                 frame number is found
3326  */
3327 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3328     MpegEncContext * const s = &h->s;
3329     int i;
3330
3331     for(i=0; i<h->short_ref_count; i++){
3332         Picture *pic= h->short_ref[i];
3333         if(s->avctx->debug&FF_DEBUG_MMCO)
3334             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3335         if(pic->frame_num == frame_num) {
3336             *idx = i;
3337             return pic;
3338         }
3339     }
3340     return NULL;
3341 }
3342
3343 /**
3344  * Remove a picture from the short term reference list by its index in
3345  * that list.  This does no checking on the provided index; it is assumed
3346  * to be valid. Other list entries are shifted down.
3347  * @param i index into h->short_ref of picture to remove.
3348  */
3349 static void remove_short_at_index(H264Context *h, int i){
3350     assert(i > 0 && i < h->short_ref_count);
3351     h->short_ref[i]= NULL;
3352     if (--h->short_ref_count)
3353         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3354 }
3355
3356 /**
3357  *
3358  * @return the removed picture or NULL if an error occurs
3359  */
3360 static Picture * remove_short(H264Context *h, int frame_num){
3361     MpegEncContext * const s = &h->s;
3362     Picture *pic;
3363     int i;
3364
3365     if(s->avctx->debug&FF_DEBUG_MMCO)
3366         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3367
3368     pic = find_short(h, frame_num, &i);
3369     if (pic)
3370         remove_short_at_index(h, i);
3371
3372     return pic;
3373 }
3374
3375 /**
3376  * Remove a picture from the long term reference list by its index in
3377  * that list.  This does no checking on the provided index; it is assumed
3378  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3379  * @param i index into h->long_ref of picture to remove.
3380  */
3381 static void remove_long_at_index(H264Context *h, int i){
3382     h->long_ref[i]= NULL;
3383     h->long_ref_count--;
3384 }
3385
3386 /**
3387  *
3388  * @return the removed picture or NULL if an error occurs
3389  */
3390 static Picture * remove_long(H264Context *h, int i){
3391     Picture *pic;
3392
3393     pic= h->long_ref[i];
3394     if (pic)
3395         remove_long_at_index(h, i);
3396
3397     return pic;
3398 }
3399
3400 /**
3401  * print short term list
3402  */
3403 static void print_short_term(H264Context *h) {
3404     uint32_t i;
3405     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3406         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3407         for(i=0; i<h->short_ref_count; i++){
3408             Picture *pic= h->short_ref[i];
3409             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3410         }
3411     }
3412 }
3413
3414 /**
3415  * print long term list
3416  */
3417 static void print_long_term(H264Context *h) {
3418     uint32_t i;
3419     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3420         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3421         for(i = 0; i < 16; i++){
3422             Picture *pic= h->long_ref[i];
3423             if (pic) {
3424                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);