Fix loop filter with CAVLC 8x8dct.
[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 top left 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 (let us 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_nos == 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                 // left shift 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] = FFMIN3((unsigned)refa, (unsigned)refb, (unsigned)refc);
1005             if(ref[list] < 0)
1006                 ref[list] = -1;
1007         }
1008
1009         if(ref[0] < 0 && ref[1] < 0){
1010             ref[0] = ref[1] = 0;
1011             mv[0][0] = mv[0][1] =
1012             mv[1][0] = mv[1][1] = 0;
1013         }else{
1014             for(list=0; list<2; list++){
1015                 if(ref[list] >= 0)
1016                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1017                 else
1018                     mv[list][0] = mv[list][1] = 0;
1019             }
1020         }
1021
1022         if(ref[1] < 0){
1023             if(!is_b8x8)
1024                 *mb_type &= ~MB_TYPE_L1;
1025             sub_mb_type &= ~MB_TYPE_L1;
1026         }else if(ref[0] < 0){
1027             if(!is_b8x8)
1028                 *mb_type &= ~MB_TYPE_L0;
1029             sub_mb_type &= ~MB_TYPE_L0;
1030         }
1031
1032         if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1033             int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1034             int mb_types_col[2];
1035             int b8_stride = h->b8_stride;
1036             int b4_stride = h->b_stride;
1037
1038             *mb_type = (*mb_type & ~MB_TYPE_16x16) | MB_TYPE_8x8;
1039
1040             if(IS_INTERLACED(*mb_type)){
1041                 mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1042                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1043                 if(s->mb_y&1){
1044                     l1ref0 -= 2*b8_stride;
1045                     l1ref1 -= 2*b8_stride;
1046                     l1mv0 -= 4*b4_stride;
1047                     l1mv1 -= 4*b4_stride;
1048                 }
1049                 b8_stride *= 3;
1050                 b4_stride *= 6;
1051             }else{
1052                 int cur_poc = s->current_picture_ptr->poc;
1053                 int *col_poc = h->ref_list[1]->field_poc;
1054                 int col_parity = FFABS(col_poc[0] - cur_poc) >= FFABS(col_poc[1] - cur_poc);
1055                 int dy = 2*col_parity - (s->mb_y&1);
1056                 mb_types_col[0] =
1057                 mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy + col_parity*s->mb_stride];
1058                 l1ref0 += dy*b8_stride;
1059                 l1ref1 += dy*b8_stride;
1060                 l1mv0 += 2*dy*b4_stride;
1061                 l1mv1 += 2*dy*b4_stride;
1062                 b8_stride = 0;
1063             }
1064
1065             for(i8=0; i8<4; i8++){
1066                 int x8 = i8&1;
1067                 int y8 = i8>>1;
1068                 int xy8 = x8+y8*b8_stride;
1069                 int xy4 = 3*x8+y8*b4_stride;
1070                 int a=0, b=0;
1071
1072                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1073                     continue;
1074                 h->sub_mb_type[i8] = sub_mb_type;
1075
1076                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1077                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1078                 if(!IS_INTRA(mb_types_col[y8])
1079                    && (   (l1ref0[xy8] == 0 && FFABS(l1mv0[xy4][0]) <= 1 && FFABS(l1mv0[xy4][1]) <= 1)
1080                        || (l1ref0[xy8]  < 0 && l1ref1[xy8] == 0 && FFABS(l1mv1[xy4][0]) <= 1 && FFABS(l1mv1[xy4][1]) <= 1))){
1081                     if(ref[0] > 0)
1082                         a= pack16to32(mv[0][0],mv[0][1]);
1083                     if(ref[1] > 0)
1084                         b= pack16to32(mv[1][0],mv[1][1]);
1085                 }else{
1086                     a= pack16to32(mv[0][0],mv[0][1]);
1087                     b= pack16to32(mv[1][0],mv[1][1]);
1088                 }
1089                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, a, 4);
1090                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, b, 4);
1091             }
1092         }else if(IS_16X16(*mb_type)){
1093             int a=0, b=0;
1094
1095             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1096             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1097             if(!IS_INTRA(mb_type_col)
1098                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1099                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1100                        && (h->x264_build>33 || !h->x264_build)))){
1101                 if(ref[0] > 0)
1102                     a= pack16to32(mv[0][0],mv[0][1]);
1103                 if(ref[1] > 0)
1104                     b= pack16to32(mv[1][0],mv[1][1]);
1105             }else{
1106                 a= pack16to32(mv[0][0],mv[0][1]);
1107                 b= pack16to32(mv[1][0],mv[1][1]);
1108             }
1109             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1110             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1111         }else{
1112             for(i8=0; i8<4; i8++){
1113                 const int x8 = i8&1;
1114                 const int y8 = i8>>1;
1115
1116                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1117                     continue;
1118                 h->sub_mb_type[i8] = sub_mb_type;
1119
1120                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1121                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1122                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1123                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1124
1125                 /* col_zero_flag */
1126                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1127                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1128                                                   && (h->x264_build>33 || !h->x264_build)))){
1129                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1130                     if(IS_SUB_8X8(sub_mb_type)){
1131                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1132                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1133                             if(ref[0] == 0)
1134                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1135                             if(ref[1] == 0)
1136                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1137                         }
1138                     }else
1139                     for(i4=0; i4<4; i4++){
1140                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1141                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1142                             if(ref[0] == 0)
1143                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1144                             if(ref[1] == 0)
1145                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1146                         }
1147                     }
1148                 }
1149             }
1150         }
1151     }else{ /* direct temporal mv pred */
1152         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1153         const int *dist_scale_factor = h->dist_scale_factor;
1154
1155         if(FRAME_MBAFF){
1156             if(IS_INTERLACED(*mb_type)){
1157                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1158                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1159                 dist_scale_factor = h->dist_scale_factor_field;
1160             }
1161             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1162                 /* FIXME assumes direct_8x8_inference == 1 */
1163                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1164                 int mb_types_col[2];
1165                 int y_shift;
1166
1167                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1168                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1169                          | (*mb_type & MB_TYPE_INTERLACED);
1170                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1171
1172                 if(IS_INTERLACED(*mb_type)){
1173                     /* frame to field scaling */
1174                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1175                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1176                     if(s->mb_y&1){
1177                         l1ref0 -= 2*h->b8_stride;
1178                         l1ref1 -= 2*h->b8_stride;
1179                         l1mv0 -= 4*h->b_stride;
1180                         l1mv1 -= 4*h->b_stride;
1181                     }
1182                     y_shift = 0;
1183
1184                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1185                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1186                        && !is_b8x8)
1187                         *mb_type |= MB_TYPE_16x8;
1188                     else
1189                         *mb_type |= MB_TYPE_8x8;
1190                 }else{
1191                     /* field to frame scaling */
1192                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1193                      * but in MBAFF, top and bottom POC are equal */
1194                     int dy = (s->mb_y&1) ? 1 : 2;
1195                     mb_types_col[0] =
1196                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1197                     l1ref0 += dy*h->b8_stride;
1198                     l1ref1 += dy*h->b8_stride;
1199                     l1mv0 += 2*dy*h->b_stride;
1200                     l1mv1 += 2*dy*h->b_stride;
1201                     y_shift = 2;
1202
1203                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1204                        && !is_b8x8)
1205                         *mb_type |= MB_TYPE_16x16;
1206                     else
1207                         *mb_type |= MB_TYPE_8x8;
1208                 }
1209
1210                 for(i8=0; i8<4; i8++){
1211                     const int x8 = i8&1;
1212                     const int y8 = i8>>1;
1213                     int ref0, scale;
1214                     const int16_t (*l1mv)[2]= l1mv0;
1215
1216                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1217                         continue;
1218                     h->sub_mb_type[i8] = sub_mb_type;
1219
1220                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1221                     if(IS_INTRA(mb_types_col[y8])){
1222                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1223                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1224                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1225                         continue;
1226                     }
1227
1228                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1229                     if(ref0 >= 0)
1230                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1231                     else{
1232                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1233                         l1mv= l1mv1;
1234                     }
1235                     scale = dist_scale_factor[ref0];
1236                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1237
1238                     {
1239                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1240                         int my_col = (mv_col[1]<<y_shift)/2;
1241                         int mx = (scale * mv_col[0] + 128) >> 8;
1242                         int my = (scale * my_col + 128) >> 8;
1243                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1244                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1245                     }
1246                 }
1247                 return;
1248             }
1249         }
1250
1251         /* one-to-one mv scaling */
1252
1253         if(IS_16X16(*mb_type)){
1254             int ref, mv0, mv1;
1255
1256             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1257             if(IS_INTRA(mb_type_col)){
1258                 ref=mv0=mv1=0;
1259             }else{
1260                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1261                                                 : map_col_to_list0[1][l1ref1[0]];
1262                 const int scale = dist_scale_factor[ref0];
1263                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1264                 int mv_l0[2];
1265                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1266                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1267                 ref= ref0;
1268                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1269                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1270             }
1271             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1272             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1273             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1274         }else{
1275             for(i8=0; i8<4; i8++){
1276                 const int x8 = i8&1;
1277                 const int y8 = i8>>1;
1278                 int ref0, scale;
1279                 const int16_t (*l1mv)[2]= l1mv0;
1280
1281                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1282                     continue;
1283                 h->sub_mb_type[i8] = sub_mb_type;
1284                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1285                 if(IS_INTRA(mb_type_col)){
1286                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1287                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1288                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1289                     continue;
1290                 }
1291
1292                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1293                 if(ref0 >= 0)
1294                     ref0 = map_col_to_list0[0][ref0];
1295                 else{
1296                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1297                     l1mv= l1mv1;
1298                 }
1299                 scale = dist_scale_factor[ref0];
1300
1301                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1302                 if(IS_SUB_8X8(sub_mb_type)){
1303                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1304                     int mx = (scale * mv_col[0] + 128) >> 8;
1305                     int my = (scale * mv_col[1] + 128) >> 8;
1306                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1307                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1308                 }else
1309                 for(i4=0; i4<4; i4++){
1310                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1311                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1312                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1313                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1314                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1315                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1316                 }
1317             }
1318         }
1319     }
1320 }
1321
1322 static inline void write_back_motion(H264Context *h, int mb_type){
1323     MpegEncContext * const s = &h->s;
1324     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1325     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1326     int list;
1327
1328     if(!USES_LIST(mb_type, 0))
1329         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1330
1331     for(list=0; list<h->list_count; list++){
1332         int y;
1333         if(!USES_LIST(mb_type, list))
1334             continue;
1335
1336         for(y=0; y<4; y++){
1337             *(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];
1338             *(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];
1339         }
1340         if( h->pps.cabac ) {
1341             if(IS_SKIP(mb_type))
1342                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1343             else
1344             for(y=0; y<4; y++){
1345                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1346                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1347             }
1348         }
1349
1350         {
1351             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1352             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1353             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1354             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1355             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1356         }
1357     }
1358
1359     if(h->slice_type_nos == FF_B_TYPE && h->pps.cabac){
1360         if(IS_8X8(mb_type)){
1361             uint8_t *direct_table = &h->direct_table[b8_xy];
1362             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1363             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1364             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1365         }
1366     }
1367 }
1368
1369 /**
1370  * Decodes a network abstraction layer unit.
1371  * @param consumed is the number of bytes used as input
1372  * @param length is the length of the array
1373  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1374  * @returns decoded bytes, might be src+1 if no escapes
1375  */
1376 static const uint8_t *decode_nal(H264Context *h, const uint8_t *src, int *dst_length, int *consumed, int length){
1377     int i, si, di;
1378     uint8_t *dst;
1379     int bufidx;
1380
1381 //    src[0]&0x80;                //forbidden bit
1382     h->nal_ref_idc= src[0]>>5;
1383     h->nal_unit_type= src[0]&0x1F;
1384
1385     src++; length--;
1386 #if 0
1387     for(i=0; i<length; i++)
1388         printf("%2X ", src[i]);
1389 #endif
1390     for(i=0; i+1<length; i+=2){
1391         if(src[i]) continue;
1392         if(i>0 && src[i-1]==0) i--;
1393         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1394             if(src[i+2]!=3){
1395                 /* startcode, so we must be past the end */
1396                 length=i;
1397             }
1398             break;
1399         }
1400     }
1401
1402     if(i>=length-1){ //no escaped 0
1403         *dst_length= length;
1404         *consumed= length+1; //+1 for the header
1405         return src;
1406     }
1407
1408     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1409     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1410     dst= h->rbsp_buffer[bufidx];
1411
1412     if (dst == NULL){
1413         return NULL;
1414     }
1415
1416 //printf("decoding esc\n");
1417     si=di=0;
1418     while(si<length){
1419         //remove escapes (very rare 1:2^22)
1420         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1421             if(src[si+2]==3){ //escape
1422                 dst[di++]= 0;
1423                 dst[di++]= 0;
1424                 si+=3;
1425                 continue;
1426             }else //next start code
1427                 break;
1428         }
1429
1430         dst[di++]= src[si++];
1431     }
1432
1433     *dst_length= di;
1434     *consumed= si + 1;//+1 for the header
1435 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1436     return dst;
1437 }
1438
1439 /**
1440  * identifies the exact end of the bitstream
1441  * @return the length of the trailing, or 0 if damaged
1442  */
1443 static int decode_rbsp_trailing(H264Context *h, const uint8_t *src){
1444     int v= *src;
1445     int r;
1446
1447     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1448
1449     for(r=1; r<9; r++){
1450         if(v&1) return r;
1451         v>>=1;
1452     }
1453     return 0;
1454 }
1455
1456 /**
1457  * IDCT transforms the 16 dc values and dequantizes them.
1458  * @param qp quantization parameter
1459  */
1460 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1461 #define stride 16
1462     int i;
1463     int temp[16]; //FIXME check if this is a good idea
1464     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1465     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1466
1467 //memset(block, 64, 2*256);
1468 //return;
1469     for(i=0; i<4; i++){
1470         const int offset= y_offset[i];
1471         const int z0= block[offset+stride*0] + block[offset+stride*4];
1472         const int z1= block[offset+stride*0] - block[offset+stride*4];
1473         const int z2= block[offset+stride*1] - block[offset+stride*5];
1474         const int z3= block[offset+stride*1] + block[offset+stride*5];
1475
1476         temp[4*i+0]= z0+z3;
1477         temp[4*i+1]= z1+z2;
1478         temp[4*i+2]= z1-z2;
1479         temp[4*i+3]= z0-z3;
1480     }
1481
1482     for(i=0; i<4; i++){
1483         const int offset= x_offset[i];
1484         const int z0= temp[4*0+i] + temp[4*2+i];
1485         const int z1= temp[4*0+i] - temp[4*2+i];
1486         const int z2= temp[4*1+i] - temp[4*3+i];
1487         const int z3= temp[4*1+i] + temp[4*3+i];
1488
1489         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_residual
1490         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1491         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1492         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1493     }
1494 }
1495
1496 #if 0
1497 /**
1498  * DCT transforms the 16 dc values.
1499  * @param qp quantization parameter ??? FIXME
1500  */
1501 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1502 //    const int qmul= dequant_coeff[qp][0];
1503     int i;
1504     int temp[16]; //FIXME check if this is a good idea
1505     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1506     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1507
1508     for(i=0; i<4; i++){
1509         const int offset= y_offset[i];
1510         const int z0= block[offset+stride*0] + block[offset+stride*4];
1511         const int z1= block[offset+stride*0] - block[offset+stride*4];
1512         const int z2= block[offset+stride*1] - block[offset+stride*5];
1513         const int z3= block[offset+stride*1] + block[offset+stride*5];
1514
1515         temp[4*i+0]= z0+z3;
1516         temp[4*i+1]= z1+z2;
1517         temp[4*i+2]= z1-z2;
1518         temp[4*i+3]= z0-z3;
1519     }
1520
1521     for(i=0; i<4; i++){
1522         const int offset= x_offset[i];
1523         const int z0= temp[4*0+i] + temp[4*2+i];
1524         const int z1= temp[4*0+i] - temp[4*2+i];
1525         const int z2= temp[4*1+i] - temp[4*3+i];
1526         const int z3= temp[4*1+i] + temp[4*3+i];
1527
1528         block[stride*0 +offset]= (z0 + z3)>>1;
1529         block[stride*2 +offset]= (z1 + z2)>>1;
1530         block[stride*8 +offset]= (z1 - z2)>>1;
1531         block[stride*10+offset]= (z0 - z3)>>1;
1532     }
1533 }
1534 #endif
1535
1536 #undef xStride
1537 #undef stride
1538
1539 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1540     const int stride= 16*2;
1541     const int xStride= 16;
1542     int a,b,c,d,e;
1543
1544     a= block[stride*0 + xStride*0];
1545     b= block[stride*0 + xStride*1];
1546     c= block[stride*1 + xStride*0];
1547     d= block[stride*1 + xStride*1];
1548
1549     e= a-b;
1550     a= a+b;
1551     b= c-d;
1552     c= c+d;
1553
1554     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1555     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1556     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1557     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1558 }
1559
1560 #if 0
1561 static void chroma_dc_dct_c(DCTELEM *block){
1562     const int stride= 16*2;
1563     const int xStride= 16;
1564     int a,b,c,d,e;
1565
1566     a= block[stride*0 + xStride*0];
1567     b= block[stride*0 + xStride*1];
1568     c= block[stride*1 + xStride*0];
1569     d= block[stride*1 + xStride*1];
1570
1571     e= a-b;
1572     a= a+b;
1573     b= c-d;
1574     c= c+d;
1575
1576     block[stride*0 + xStride*0]= (a+c);
1577     block[stride*0 + xStride*1]= (e+b);
1578     block[stride*1 + xStride*0]= (a-c);
1579     block[stride*1 + xStride*1]= (e-b);
1580 }
1581 #endif
1582
1583 /**
1584  * gets the chroma qp.
1585  */
1586 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1587     return h->pps.chroma_qp_table[t][qscale];
1588 }
1589
1590 //FIXME need to check that this does not overflow signed 32 bit for low qp, I am not sure, it's very close
1591 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1592 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1593     int i;
1594     const int * const quant_table= quant_coeff[qscale];
1595     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1596     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1597     const unsigned int threshold2= (threshold1<<1);
1598     int last_non_zero;
1599
1600     if(separate_dc){
1601         if(qscale<=18){
1602             //avoid overflows
1603             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1604             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1605             const unsigned int dc_threshold2= (dc_threshold1<<1);
1606
1607             int level= block[0]*quant_coeff[qscale+18][0];
1608             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1609                 if(level>0){
1610                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1611                     block[0]= level;
1612                 }else{
1613                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1614                     block[0]= -level;
1615                 }
1616 //                last_non_zero = i;
1617             }else{
1618                 block[0]=0;
1619             }
1620         }else{
1621             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1622             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1623             const unsigned int dc_threshold2= (dc_threshold1<<1);
1624
1625             int level= block[0]*quant_table[0];
1626             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1627                 if(level>0){
1628                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1629                     block[0]= level;
1630                 }else{
1631                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1632                     block[0]= -level;
1633                 }
1634 //                last_non_zero = i;
1635             }else{
1636                 block[0]=0;
1637             }
1638         }
1639         last_non_zero= 0;
1640         i=1;
1641     }else{
1642         last_non_zero= -1;
1643         i=0;
1644     }
1645
1646     for(; i<16; i++){
1647         const int j= scantable[i];
1648         int level= block[j]*quant_table[j];
1649
1650 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1651 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1652         if(((unsigned)(level+threshold1))>threshold2){
1653             if(level>0){
1654                 level= (bias + level)>>QUANT_SHIFT;
1655                 block[j]= level;
1656             }else{
1657                 level= (bias - level)>>QUANT_SHIFT;
1658                 block[j]= -level;
1659             }
1660             last_non_zero = i;
1661         }else{
1662             block[j]=0;
1663         }
1664     }
1665
1666     return last_non_zero;
1667 }
1668
1669 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1670                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1671                            int src_x_offset, int src_y_offset,
1672                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1673     MpegEncContext * const s = &h->s;
1674     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1675     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1676     const int luma_xy= (mx&3) + ((my&3)<<2);
1677     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1678     uint8_t * src_cb, * src_cr;
1679     int extra_width= h->emu_edge_width;
1680     int extra_height= h->emu_edge_height;
1681     int emu=0;
1682     const int full_mx= mx>>2;
1683     const int full_my= my>>2;
1684     const int pic_width  = 16*s->mb_width;
1685     const int pic_height = 16*s->mb_height >> MB_FIELD;
1686
1687     if(!pic->data[0]) //FIXME this is unacceptable, some sensible error concealment must be done for missing reference frames
1688         return;
1689
1690     if(mx&7) extra_width -= 3;
1691     if(my&7) extra_height -= 3;
1692
1693     if(   full_mx < 0-extra_width
1694        || full_my < 0-extra_height
1695        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1696        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1697         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);
1698             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1699         emu=1;
1700     }
1701
1702     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1703     if(!square){
1704         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1705     }
1706
1707     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1708
1709     if(MB_FIELD){
1710         // chroma offset when predicting from a field of opposite parity
1711         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1712         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1713     }
1714     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1715     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1716
1717     if(emu){
1718         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);
1719             src_cb= s->edge_emu_buffer;
1720     }
1721     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1722
1723     if(emu){
1724         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);
1725             src_cr= s->edge_emu_buffer;
1726     }
1727     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1728 }
1729
1730 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1731                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1732                            int x_offset, int y_offset,
1733                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1734                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1735                            int list0, int list1){
1736     MpegEncContext * const s = &h->s;
1737     qpel_mc_func *qpix_op=  qpix_put;
1738     h264_chroma_mc_func chroma_op= chroma_put;
1739
1740     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1741     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1742     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1743     x_offset += 8*s->mb_x;
1744     y_offset += 8*(s->mb_y >> MB_FIELD);
1745
1746     if(list0){
1747         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1748         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1749                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1750                            qpix_op, chroma_op);
1751
1752         qpix_op=  qpix_avg;
1753         chroma_op= chroma_avg;
1754     }
1755
1756     if(list1){
1757         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1758         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1759                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1760                            qpix_op, chroma_op);
1761     }
1762 }
1763
1764 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1765                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1766                            int x_offset, int y_offset,
1767                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1768                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1769                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1770                            int list0, int list1){
1771     MpegEncContext * const s = &h->s;
1772
1773     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1774     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1775     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1776     x_offset += 8*s->mb_x;
1777     y_offset += 8*(s->mb_y >> MB_FIELD);
1778
1779     if(list0 && list1){
1780         /* don't optimize for luma-only case, since B-frames usually
1781          * use implicit weights => chroma too. */
1782         uint8_t *tmp_cb = s->obmc_scratchpad;
1783         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1784         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1785         int refn0 = h->ref_cache[0][ scan8[n] ];
1786         int refn1 = h->ref_cache[1][ scan8[n] ];
1787
1788         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1789                     dest_y, dest_cb, dest_cr,
1790                     x_offset, y_offset, qpix_put, chroma_put);
1791         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1792                     tmp_y, tmp_cb, tmp_cr,
1793                     x_offset, y_offset, qpix_put, chroma_put);
1794
1795         if(h->use_weight == 2){
1796             int weight0 = h->implicit_weight[refn0][refn1];
1797             int weight1 = 64 - weight0;
1798             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1799             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1800             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1801         }else{
1802             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1803                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1804                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1805             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1806                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1807                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1808             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1809                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1810                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1811         }
1812     }else{
1813         int list = list1 ? 1 : 0;
1814         int refn = h->ref_cache[list][ scan8[n] ];
1815         Picture *ref= &h->ref_list[list][refn];
1816         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1817                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1818                     qpix_put, chroma_put);
1819
1820         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1821                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1822         if(h->use_weight_chroma){
1823             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1824                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1825             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1826                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1827         }
1828     }
1829 }
1830
1831 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1832                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1833                            int x_offset, int y_offset,
1834                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1835                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1836                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1837                            int list0, int list1){
1838     if((h->use_weight==2 && list0 && list1
1839         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1840        || h->use_weight==1)
1841         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1842                          x_offset, y_offset, qpix_put, chroma_put,
1843                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1844     else
1845         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1846                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1847 }
1848
1849 static inline void prefetch_motion(H264Context *h, int list){
1850     /* fetch pixels for estimated mv 4 macroblocks ahead
1851      * optimized for 64byte cache lines */
1852     MpegEncContext * const s = &h->s;
1853     const int refn = h->ref_cache[list][scan8[0]];
1854     if(refn >= 0){
1855         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1856         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1857         uint8_t **src= h->ref_list[list][refn].data;
1858         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1859         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1860         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1861         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1862     }
1863 }
1864
1865 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1866                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1867                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1868                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1869     MpegEncContext * const s = &h->s;
1870     const int mb_xy= h->mb_xy;
1871     const int mb_type= s->current_picture.mb_type[mb_xy];
1872
1873     assert(IS_INTER(mb_type));
1874
1875     prefetch_motion(h, 0);
1876
1877     if(IS_16X16(mb_type)){
1878         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1879                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1880                 &weight_op[0], &weight_avg[0],
1881                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1882     }else if(IS_16X8(mb_type)){
1883         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1884                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1885                 &weight_op[1], &weight_avg[1],
1886                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1887         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
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, 1, 0), IS_DIR(mb_type, 1, 1));
1891     }else if(IS_8X16(mb_type)){
1892         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1893                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1894                 &weight_op[2], &weight_avg[2],
1895                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1896         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 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, 1, 0), IS_DIR(mb_type, 1, 1));
1900     }else{
1901         int i;
1902
1903         assert(IS_8X8(mb_type));
1904
1905         for(i=0; i<4; i++){
1906             const int sub_mb_type= h->sub_mb_type[i];
1907             const int n= 4*i;
1908             int x_offset= (i&1)<<2;
1909             int y_offset= (i&2)<<1;
1910
1911             if(IS_SUB_8X8(sub_mb_type)){
1912                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1913                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1914                     &weight_op[3], &weight_avg[3],
1915                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1916             }else if(IS_SUB_8X4(sub_mb_type)){
1917                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1918                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1919                     &weight_op[4], &weight_avg[4],
1920                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1921                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
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             }else if(IS_SUB_4X8(sub_mb_type)){
1926                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1927                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1928                     &weight_op[5], &weight_avg[5],
1929                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1930                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, 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             }else{
1935                 int j;
1936                 assert(IS_SUB_4X4(sub_mb_type));
1937                 for(j=0; j<4; j++){
1938                     int sub_x_offset= x_offset + 2*(j&1);
1939                     int sub_y_offset= y_offset +   (j&2);
1940                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1941                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1942                         &weight_op[6], &weight_avg[6],
1943                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1944                 }
1945             }
1946         }
1947     }
1948
1949     prefetch_motion(h, 1);
1950 }
1951
1952 static av_cold void decode_init_vlc(void){
1953     static int done = 0;
1954
1955     if (!done) {
1956         int i;
1957         done = 1;
1958
1959         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1960                  &chroma_dc_coeff_token_len [0], 1, 1,
1961                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1962
1963         for(i=0; i<4; i++){
1964             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1965                      &coeff_token_len [i][0], 1, 1,
1966                      &coeff_token_bits[i][0], 1, 1, 1);
1967         }
1968
1969         for(i=0; i<3; i++){
1970             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1971                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1972                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1973         }
1974         for(i=0; i<15; i++){
1975             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1976                      &total_zeros_len [i][0], 1, 1,
1977                      &total_zeros_bits[i][0], 1, 1, 1);
1978         }
1979
1980         for(i=0; i<6; i++){
1981             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
1982                      &run_len [i][0], 1, 1,
1983                      &run_bits[i][0], 1, 1, 1);
1984         }
1985         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1986                  &run_len [6][0], 1, 1,
1987                  &run_bits[6][0], 1, 1, 1);
1988     }
1989 }
1990
1991 static void free_tables(H264Context *h){
1992     int i;
1993     H264Context *hx;
1994     av_freep(&h->intra4x4_pred_mode);
1995     av_freep(&h->chroma_pred_mode_table);
1996     av_freep(&h->cbp_table);
1997     av_freep(&h->mvd_table[0]);
1998     av_freep(&h->mvd_table[1]);
1999     av_freep(&h->direct_table);
2000     av_freep(&h->non_zero_count);
2001     av_freep(&h->slice_table_base);
2002     h->slice_table= NULL;
2003
2004     av_freep(&h->mb2b_xy);
2005     av_freep(&h->mb2b8_xy);
2006
2007     for(i = 0; i < MAX_SPS_COUNT; i++)
2008         av_freep(h->sps_buffers + i);
2009
2010     for(i = 0; i < MAX_PPS_COUNT; i++)
2011         av_freep(h->pps_buffers + i);
2012
2013     for(i = 0; i < h->s.avctx->thread_count; i++) {
2014         hx = h->thread_context[i];
2015         if(!hx) continue;
2016         av_freep(&hx->top_borders[1]);
2017         av_freep(&hx->top_borders[0]);
2018         av_freep(&hx->s.obmc_scratchpad);
2019     }
2020 }
2021
2022 static void init_dequant8_coeff_table(H264Context *h){
2023     int i,q,x;
2024     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2025     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2026     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2027
2028     for(i=0; i<2; i++ ){
2029         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2030             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2031             break;
2032         }
2033
2034         for(q=0; q<52; q++){
2035             int shift = ff_div6[q];
2036             int idx = ff_rem6[q];
2037             for(x=0; x<64; x++)
2038                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2039                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2040                     h->pps.scaling_matrix8[i][x]) << shift;
2041         }
2042     }
2043 }
2044
2045 static void init_dequant4_coeff_table(H264Context *h){
2046     int i,j,q,x;
2047     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2048     for(i=0; i<6; i++ ){
2049         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2050         for(j=0; j<i; j++){
2051             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2052                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2053                 break;
2054             }
2055         }
2056         if(j<i)
2057             continue;
2058
2059         for(q=0; q<52; q++){
2060             int shift = ff_div6[q] + 2;
2061             int idx = ff_rem6[q];
2062             for(x=0; x<16; x++)
2063                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2064                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2065                     h->pps.scaling_matrix4[i][x]) << shift;
2066         }
2067     }
2068 }
2069
2070 static void init_dequant_tables(H264Context *h){
2071     int i,x;
2072     init_dequant4_coeff_table(h);
2073     if(h->pps.transform_8x8_mode)
2074         init_dequant8_coeff_table(h);
2075     if(h->sps.transform_bypass){
2076         for(i=0; i<6; i++)
2077             for(x=0; x<16; x++)
2078                 h->dequant4_coeff[i][0][x] = 1<<6;
2079         if(h->pps.transform_8x8_mode)
2080             for(i=0; i<2; i++)
2081                 for(x=0; x<64; x++)
2082                     h->dequant8_coeff[i][0][x] = 1<<6;
2083     }
2084 }
2085
2086
2087 /**
2088  * allocates tables.
2089  * needs width/height
2090  */
2091 static int alloc_tables(H264Context *h){
2092     MpegEncContext * const s = &h->s;
2093     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2094     int x,y;
2095
2096     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2097
2098     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2099     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2100     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2101
2102     CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2103     CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2104     CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2105     CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2106
2107     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2108     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2109
2110     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2111     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2112     for(y=0; y<s->mb_height; y++){
2113         for(x=0; x<s->mb_width; x++){
2114             const int mb_xy= x + y*s->mb_stride;
2115             const int b_xy = 4*x + 4*y*h->b_stride;
2116             const int b8_xy= 2*x + 2*y*h->b8_stride;
2117
2118             h->mb2b_xy [mb_xy]= b_xy;
2119             h->mb2b8_xy[mb_xy]= b8_xy;
2120         }
2121     }
2122
2123     s->obmc_scratchpad = NULL;
2124
2125     if(!h->dequant4_coeff[0])
2126         init_dequant_tables(h);
2127
2128     return 0;
2129 fail:
2130     free_tables(h);
2131     return -1;
2132 }
2133
2134 /**
2135  * Mimic alloc_tables(), but for every context thread.
2136  */
2137 static void clone_tables(H264Context *dst, H264Context *src){
2138     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2139     dst->non_zero_count           = src->non_zero_count;
2140     dst->slice_table              = src->slice_table;
2141     dst->cbp_table                = src->cbp_table;
2142     dst->mb2b_xy                  = src->mb2b_xy;
2143     dst->mb2b8_xy                 = src->mb2b8_xy;
2144     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2145     dst->mvd_table[0]             = src->mvd_table[0];
2146     dst->mvd_table[1]             = src->mvd_table[1];
2147     dst->direct_table             = src->direct_table;
2148
2149     dst->s.obmc_scratchpad = NULL;
2150     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2151 }
2152
2153 /**
2154  * Init context
2155  * Allocate buffers which are not shared amongst multiple threads.
2156  */
2157 static int context_init(H264Context *h){
2158     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2159     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2160
2161     return 0;
2162 fail:
2163     return -1; // free_tables will clean up for us
2164 }
2165
2166 static av_cold void common_init(H264Context *h){
2167     MpegEncContext * const s = &h->s;
2168
2169     s->width = s->avctx->width;
2170     s->height = s->avctx->height;
2171     s->codec_id= s->avctx->codec->id;
2172
2173     ff_h264_pred_init(&h->hpc, s->codec_id);
2174
2175     h->dequant_coeff_pps= -1;
2176     s->unrestricted_mv=1;
2177     s->decode=1; //FIXME
2178
2179     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2180     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2181 }
2182
2183 static av_cold int decode_init(AVCodecContext *avctx){
2184     H264Context *h= avctx->priv_data;
2185     MpegEncContext * const s = &h->s;
2186
2187     MPV_decode_defaults(s);
2188
2189     s->avctx = avctx;
2190     common_init(h);
2191
2192     s->out_format = FMT_H264;
2193     s->workaround_bugs= avctx->workaround_bugs;
2194
2195     // set defaults
2196 //    s->decode_mb= ff_h263_decode_mb;
2197     s->quarter_sample = 1;
2198     s->low_delay= 1;
2199
2200     if(avctx->codec_id == CODEC_ID_SVQ3)
2201         avctx->pix_fmt= PIX_FMT_YUVJ420P;
2202     else
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 ORed 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
2259     // We mark the current picture as non-reference after allocating it, so
2260     // that if we break out due to an error it can be released automatically
2261     // in the next MPV_frame_start().
2262     // SVQ3 as well as most other codecs have only last/next/current and thus
2263     // get released even with set reference, besides SVQ3 and others do not
2264     // mark frames as reference later "naturally".
2265     if(s->codec_id != CODEC_ID_SVQ3)
2266         s->current_picture_ptr->reference= 0;
2267     return 0;
2268 }
2269
2270 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){
2271     MpegEncContext * const s = &h->s;
2272     int i;
2273
2274     src_y  -=   linesize;
2275     src_cb -= uvlinesize;
2276     src_cr -= uvlinesize;
2277
2278     // There are two lines saved, the line above the the top macroblock of a pair,
2279     // and the line above the bottom macroblock
2280     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2281     for(i=1; i<17; i++){
2282         h->left_border[i]= src_y[15+i*  linesize];
2283     }
2284
2285     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2286     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2287
2288     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2289         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2290         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2291         for(i=1; i<9; i++){
2292             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2293             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2294         }
2295         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2296         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2297     }
2298 }
2299
2300 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){
2301     MpegEncContext * const s = &h->s;
2302     int temp8, i;
2303     uint64_t temp64;
2304     int deblock_left;
2305     int deblock_top;
2306     int mb_xy;
2307
2308     if(h->deblocking_filter == 2) {
2309         mb_xy = h->mb_xy;
2310         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2311         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2312     } else {
2313         deblock_left = (s->mb_x > 0);
2314         deblock_top =  (s->mb_y > 0);
2315     }
2316
2317     src_y  -=   linesize + 1;
2318     src_cb -= uvlinesize + 1;
2319     src_cr -= uvlinesize + 1;
2320
2321 #define XCHG(a,b,t,xchg)\
2322 t= a;\
2323 if(xchg)\
2324     a= b;\
2325 b= t;
2326
2327     if(deblock_left){
2328         for(i = !deblock_top; i<17; i++){
2329             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2330         }
2331     }
2332
2333     if(deblock_top){
2334         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2335         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2336         if(s->mb_x+1 < s->mb_width){
2337             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2338         }
2339     }
2340
2341     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2342         if(deblock_left){
2343             for(i = !deblock_top; i<9; i++){
2344                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2345                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2346             }
2347         }
2348         if(deblock_top){
2349             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2350             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2351         }
2352     }
2353 }
2354
2355 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2356     MpegEncContext * const s = &h->s;
2357     int i;
2358
2359     src_y  -= 2 *   linesize;
2360     src_cb -= 2 * uvlinesize;
2361     src_cr -= 2 * uvlinesize;
2362
2363     // There are two lines saved, the line above the the top macroblock of a pair,
2364     // and the line above the bottom macroblock
2365     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2366     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2367     for(i=2; i<34; i++){
2368         h->left_border[i]= src_y[15+i*  linesize];
2369     }
2370
2371     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2372     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2373     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2374     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2375
2376     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2377         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2378         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2379         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2380         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2381         for(i=2; i<18; i++){
2382             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2383             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2384         }
2385         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2386         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2387         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2388         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2389     }
2390 }
2391
2392 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){
2393     MpegEncContext * const s = &h->s;
2394     int temp8, i;
2395     uint64_t temp64;
2396     int deblock_left = (s->mb_x > 0);
2397     int deblock_top  = (s->mb_y > 1);
2398
2399     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);
2400
2401     src_y  -= 2 *   linesize + 1;
2402     src_cb -= 2 * uvlinesize + 1;
2403     src_cr -= 2 * uvlinesize + 1;
2404
2405 #define XCHG(a,b,t,xchg)\
2406 t= a;\
2407 if(xchg)\
2408     a= b;\
2409 b= t;
2410
2411     if(deblock_left){
2412         for(i = (!deblock_top)<<1; i<34; i++){
2413             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2414         }
2415     }
2416
2417     if(deblock_top){
2418         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2419         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2420         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2421         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2422         if(s->mb_x+1 < s->mb_width){
2423             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2424             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2425         }
2426     }
2427
2428     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2429         if(deblock_left){
2430             for(i = (!deblock_top) << 1; i<18; i++){
2431                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2432                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2433             }
2434         }
2435         if(deblock_top){
2436             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2437             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2438             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2439             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2440         }
2441     }
2442 }
2443
2444 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2445     MpegEncContext * const s = &h->s;
2446     const int mb_x= s->mb_x;
2447     const int mb_y= s->mb_y;
2448     const int mb_xy= h->mb_xy;
2449     const int mb_type= s->current_picture.mb_type[mb_xy];
2450     uint8_t  *dest_y, *dest_cb, *dest_cr;
2451     int linesize, uvlinesize /*dct_offset*/;
2452     int i;
2453     int *block_offset = &h->block_offset[0];
2454     const unsigned int bottom = mb_y & 1;
2455     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2456     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2457     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2458
2459     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2460     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2461     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2462
2463     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2464     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2465
2466     if (!simple && MB_FIELD) {
2467         linesize   = h->mb_linesize   = s->linesize * 2;
2468         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2469         block_offset = &h->block_offset[24];
2470         if(mb_y&1){ //FIXME move out of this function?
2471             dest_y -= s->linesize*15;
2472             dest_cb-= s->uvlinesize*7;
2473             dest_cr-= s->uvlinesize*7;
2474         }
2475         if(FRAME_MBAFF) {
2476             int list;
2477             for(list=0; list<h->list_count; list++){
2478                 if(!USES_LIST(mb_type, list))
2479                     continue;
2480                 if(IS_16X16(mb_type)){
2481                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2482                     fill_rectangle(ref, 4, 4, 8, (16+*ref)^(s->mb_y&1), 1);
2483                 }else{
2484                     for(i=0; i<16; i+=4){
2485                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2486                         int ref = h->ref_cache[list][scan8[i]];
2487                         if(ref >= 0)
2488                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, (16+ref)^(s->mb_y&1), 1);
2489                     }
2490                 }
2491             }
2492         }
2493     } else {
2494         linesize   = h->mb_linesize   = s->linesize;
2495         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2496 //        dct_offset = s->linesize * 16;
2497     }
2498
2499     if(transform_bypass){
2500         idct_dc_add =
2501         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2502     }else if(IS_8x8DCT(mb_type)){
2503         idct_dc_add = s->dsp.h264_idct8_dc_add;
2504         idct_add = s->dsp.h264_idct8_add;
2505     }else{
2506         idct_dc_add = s->dsp.h264_idct_dc_add;
2507         idct_add = s->dsp.h264_idct_add;
2508     }
2509
2510     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2511        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2512         int mbt_y = mb_y&~1;
2513         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2514         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2515         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2516         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2517     }
2518
2519     if (!simple && IS_INTRA_PCM(mb_type)) {
2520         unsigned int x, y;
2521
2522         // The pixels are stored in h->mb array in the same order as levels,
2523         // copy them in output in the correct order.
2524         for(i=0; i<16; i++) {
2525             for (y=0; y<4; y++) {
2526                 for (x=0; x<4; x++) {
2527                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2528                 }
2529             }
2530         }
2531         for(i=16; i<16+4; i++) {
2532             for (y=0; y<4; y++) {
2533                 for (x=0; x<4; x++) {
2534                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2535                 }
2536             }
2537         }
2538         for(i=20; i<20+4; i++) {
2539             for (y=0; y<4; y++) {
2540                 for (x=0; x<4; x++) {
2541                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2542                 }
2543             }
2544         }
2545     } else {
2546         if(IS_INTRA(mb_type)){
2547             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2548                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2549
2550             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2551                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2552                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2553             }
2554
2555             if(IS_INTRA4x4(mb_type)){
2556                 if(simple || !s->encoding){
2557                     if(IS_8x8DCT(mb_type)){
2558                         for(i=0; i<16; i+=4){
2559                             uint8_t * const ptr= dest_y + block_offset[i];
2560                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2561                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2562                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2563                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2564                             if(nnz){
2565                                 if(nnz == 1 && h->mb[i*16])
2566                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2567                                 else
2568                                     idct_add(ptr, h->mb + i*16, linesize);
2569                             }
2570                         }
2571                     }else
2572                     for(i=0; i<16; i++){
2573                         uint8_t * const ptr= dest_y + block_offset[i];
2574                         uint8_t *topright;
2575                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2576                         int nnz, tr;
2577
2578                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2579                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2580                             assert(mb_y || linesize <= block_offset[i]);
2581                             if(!topright_avail){
2582                                 tr= ptr[3 - linesize]*0x01010101;
2583                                 topright= (uint8_t*) &tr;
2584                             }else
2585                                 topright= ptr + 4 - linesize;
2586                         }else
2587                             topright= NULL;
2588
2589                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2590                         nnz = h->non_zero_count_cache[ scan8[i] ];
2591                         if(nnz){
2592                             if(is_h264){
2593                                 if(nnz == 1 && h->mb[i*16])
2594                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2595                                 else
2596                                     idct_add(ptr, h->mb + i*16, linesize);
2597                             }else
2598                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2599                         }
2600                     }
2601                 }
2602             }else{
2603                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2604                 if(is_h264){
2605                     if(!transform_bypass)
2606                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2607                 }else
2608                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2609             }
2610             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2611                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2612         }else if(is_h264){
2613             hl_motion(h, dest_y, dest_cb, dest_cr,
2614                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2615                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2616                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2617         }
2618
2619
2620         if(!IS_INTRA4x4(mb_type)){
2621             if(is_h264){
2622                 if(IS_INTRA16x16(mb_type)){
2623                     for(i=0; i<16; i++){
2624                         if(h->non_zero_count_cache[ scan8[i] ])
2625                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2626                         else if(h->mb[i*16])
2627                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2628                     }
2629                 }else{
2630                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2631                     for(i=0; i<16; i+=di){
2632                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2633                         if(nnz){
2634                             if(nnz==1 && h->mb[i*16])
2635                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2636                             else
2637                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2638                         }
2639                     }
2640                 }
2641             }else{
2642                 for(i=0; i<16; i++){
2643                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2644                         uint8_t * const ptr= dest_y + block_offset[i];
2645                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2646                     }
2647                 }
2648             }
2649         }
2650
2651         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2652             uint8_t *dest[2] = {dest_cb, dest_cr};
2653             if(transform_bypass){
2654                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2655             }else{
2656                 idct_add = s->dsp.h264_idct_add;
2657                 idct_dc_add = s->dsp.h264_idct_dc_add;
2658                 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]);
2659                 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]);
2660             }
2661             if(is_h264){
2662                 for(i=16; i<16+8; i++){
2663                     if(h->non_zero_count_cache[ scan8[i] ])
2664                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2665                     else if(h->mb[i*16])
2666                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2667                 }
2668             }else{
2669                 for(i=16; i<16+8; i++){
2670                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2671                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2672                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2673                     }
2674                 }
2675             }
2676         }
2677     }
2678     if(h->deblocking_filter) {
2679         if (!simple && FRAME_MBAFF) {
2680             //FIXME try deblocking one mb at a time?
2681             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2682             const int mb_y = s->mb_y - 1;
2683             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2684             const int mb_xy= mb_x + mb_y*s->mb_stride;
2685             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2686             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2687             if (!bottom) return;
2688             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2689             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2690             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2691
2692             if(IS_INTRA(mb_type_top | mb_type_bottom))
2693                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2694
2695             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2696             // deblock a pair
2697             // top
2698             s->mb_y--; h->mb_xy -= s->mb_stride;
2699             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);
2700             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2701             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2702             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2703             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2704             // bottom
2705             s->mb_y++; h->mb_xy += s->mb_stride;
2706             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2707             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2708             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2709             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2710             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2711         } else {
2712             tprintf(h->s.avctx, "call filter_mb\n");
2713             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2714             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2715             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2716             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2717             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2718         }
2719     }
2720 }
2721
2722 /**
2723  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2724  */
2725 static void hl_decode_mb_simple(H264Context *h){
2726     hl_decode_mb_internal(h, 1);
2727 }
2728
2729 /**
2730  * Process a macroblock; this handles edge cases, such as interlacing.
2731  */
2732 static void av_noinline hl_decode_mb_complex(H264Context *h){
2733     hl_decode_mb_internal(h, 0);
2734 }
2735
2736 static void hl_decode_mb(H264Context *h){
2737     MpegEncContext * const s = &h->s;
2738     const int mb_xy= h->mb_xy;
2739     const int mb_type= s->current_picture.mb_type[mb_xy];
2740     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 ||
2741                     (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || (ENABLE_H264_ENCODER && s->encoding) || ENABLE_SMALL;
2742
2743     if(ENABLE_H264_ENCODER && !s->decode)
2744         return;
2745
2746     if (is_complex)
2747         hl_decode_mb_complex(h);
2748     else hl_decode_mb_simple(h);
2749 }
2750
2751 static void pic_as_field(Picture *pic, const int parity){
2752     int i;
2753     for (i = 0; i < 4; ++i) {
2754         if (parity == PICT_BOTTOM_FIELD)
2755             pic->data[i] += pic->linesize[i];
2756         pic->reference = parity;
2757         pic->linesize[i] *= 2;
2758     }
2759 }
2760
2761 static int split_field_copy(Picture *dest, Picture *src,
2762                             int parity, int id_add){
2763     int match = !!(src->reference & parity);
2764
2765     if (match) {
2766         *dest = *src;
2767         pic_as_field(dest, parity);
2768         dest->pic_id *= 2;
2769         dest->pic_id += id_add;
2770     }
2771
2772     return match;
2773 }
2774
2775 /**
2776  * Split one reference list into field parts, interleaving by parity
2777  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2778  * set to look at the actual start of data for that field.
2779  *
2780  * @param dest output list
2781  * @param dest_len maximum number of fields to put in dest
2782  * @param src the source reference list containing fields and/or field pairs
2783  *            (aka short_ref/long_ref, or
2784  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2785  * @param src_len number of Picture's in source (pairs and unmatched fields)
2786  * @param parity the parity of the picture being decoded/needing
2787  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2788  * @return number of fields placed in dest
2789  */
2790 static int split_field_half_ref_list(Picture *dest, int dest_len,
2791                                      Picture *src,  int src_len,  int parity){
2792     int same_parity   = 1;
2793     int same_i        = 0;
2794     int opp_i         = 0;
2795     int out_i;
2796     int field_output;
2797
2798     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2799         if (same_parity && same_i < src_len) {
2800             field_output = split_field_copy(dest + out_i, src + same_i,
2801                                             parity, 1);
2802             same_parity = !field_output;
2803             same_i++;
2804
2805         } else if (opp_i < src_len) {
2806             field_output = split_field_copy(dest + out_i, src + opp_i,
2807                                             PICT_FRAME - parity, 0);
2808             same_parity = field_output;
2809             opp_i++;
2810
2811         } else {
2812             break;
2813         }
2814     }
2815
2816     return out_i;
2817 }
2818
2819 /**
2820  * Split the reference frame list into a reference field list.
2821  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2822  * The input list contains both reference field pairs and
2823  * unmatched reference fields; it is ordered as spec describes
2824  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2825  * unmatched field pairs are also present. Conceptually this is equivalent
2826  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2827  *
2828  * @param dest output reference list where ordered fields are to be placed
2829  * @param dest_len max number of fields to place at dest
2830  * @param src source reference list, as described above
2831  * @param src_len number of pictures (pairs and unmatched fields) in src
2832  * @param parity parity of field being currently decoded
2833  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2834  * @param long_i index into src array that holds first long reference picture,
2835  *        or src_len if no long refs present.
2836  */
2837 static int split_field_ref_list(Picture *dest, int dest_len,
2838                                 Picture *src,  int src_len,
2839                                 int parity,    int long_i){
2840
2841     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2842     dest += i;
2843     dest_len -= i;
2844
2845     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2846                                    src_len - long_i, parity);
2847     return i;
2848 }
2849
2850 /**
2851  * fills the default_ref_list.
2852  */
2853 static int fill_default_ref_list(H264Context *h){
2854     MpegEncContext * const s = &h->s;
2855     int i;
2856     int smallest_poc_greater_than_current = -1;
2857     int structure_sel;
2858     Picture sorted_short_ref[32];
2859     Picture field_entry_list[2][32];
2860     Picture *frame_list[2];
2861
2862     if (FIELD_PICTURE) {
2863         structure_sel = PICT_FRAME;
2864         frame_list[0] = field_entry_list[0];
2865         frame_list[1] = field_entry_list[1];
2866     } else {
2867         structure_sel = 0;
2868         frame_list[0] = h->default_ref_list[0];
2869         frame_list[1] = h->default_ref_list[1];
2870     }
2871
2872     if(h->slice_type_nos==FF_B_TYPE){
2873         int list;
2874         int len[2];
2875         int short_len[2];
2876         int out_i;
2877         int limit= INT_MIN;
2878
2879         /* sort frame according to POC in B slice */
2880         for(out_i=0; out_i<h->short_ref_count; out_i++){
2881             int best_i=INT_MIN;
2882             int best_poc=INT_MAX;
2883
2884             for(i=0; i<h->short_ref_count; i++){
2885                 const int poc= h->short_ref[i]->poc;
2886                 if(poc > limit && poc < best_poc){
2887                     best_poc= poc;
2888                     best_i= i;
2889                 }
2890             }
2891
2892             assert(best_i != INT_MIN);
2893
2894             limit= best_poc;
2895             sorted_short_ref[out_i]= *h->short_ref[best_i];
2896             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);
2897             if (-1 == smallest_poc_greater_than_current) {
2898                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2899                     smallest_poc_greater_than_current = out_i;
2900                 }
2901             }
2902         }
2903
2904         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2905
2906         // find the largest POC
2907         for(list=0; list<2; list++){
2908             int index = 0;
2909             int j= -99;
2910             int step= list ? -1 : 1;
2911
2912             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2913                 int sel;
2914                 while(j<0 || j>= h->short_ref_count){
2915                     if(j != -99 && step == (list ? -1 : 1))
2916                         return -1;
2917                     step = -step;
2918                     j= smallest_poc_greater_than_current + (step>>1);
2919                 }
2920                 sel = sorted_short_ref[j].reference | structure_sel;
2921                 if(sel != PICT_FRAME) continue;
2922                 frame_list[list][index  ]= sorted_short_ref[j];
2923                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2924             }
2925             short_len[list] = index;
2926
2927             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2928                 int sel;
2929                 if(h->long_ref[i] == NULL) continue;
2930                 sel = h->long_ref[i]->reference | structure_sel;
2931                 if(sel != PICT_FRAME) continue;
2932
2933                 frame_list[ list ][index  ]= *h->long_ref[i];
2934                 frame_list[ list ][index++].pic_id= i;
2935             }
2936             len[list] = index;
2937         }
2938
2939         for(list=0; list<2; list++){
2940             if (FIELD_PICTURE)
2941                 len[list] = split_field_ref_list(h->default_ref_list[list],
2942                                                  h->ref_count[list],
2943                                                  frame_list[list],
2944                                                  len[list],
2945                                                  s->picture_structure,
2946                                                  short_len[list]);
2947
2948             // swap the two first elements of L1 when L0 and L1 are identical
2949             if(list && len[0] > 1 && len[0] == len[1])
2950                 for(i=0; h->default_ref_list[0][i].data[0] == h->default_ref_list[1][i].data[0]; i++)
2951                     if(i == len[0]){
2952                         FFSWAP(Picture, h->default_ref_list[1][0], h->default_ref_list[1][1]);
2953                         break;
2954                     }
2955
2956             if(len[list] < h->ref_count[ list ])
2957                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2958         }
2959
2960
2961     }else{
2962         int index=0;
2963         int short_len;
2964         for(i=0; i<h->short_ref_count; i++){
2965             int sel;
2966             sel = h->short_ref[i]->reference | structure_sel;
2967             if(sel != PICT_FRAME) continue;
2968             frame_list[0][index  ]= *h->short_ref[i];
2969             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2970         }
2971         short_len = index;
2972         for(i = 0; i < 16; i++){
2973             int sel;
2974             if(h->long_ref[i] == NULL) continue;
2975             sel = h->long_ref[i]->reference | structure_sel;
2976             if(sel != PICT_FRAME) continue;
2977             frame_list[0][index  ]= *h->long_ref[i];
2978             frame_list[0][index++].pic_id= i;
2979         }
2980
2981         if (FIELD_PICTURE)
2982             index = split_field_ref_list(h->default_ref_list[0],
2983                                          h->ref_count[0], frame_list[0],
2984                                          index, s->picture_structure,
2985                                          short_len);
2986
2987         if(index < h->ref_count[0])
2988             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
2989     }
2990 #ifdef TRACE
2991     for (i=0; i<h->ref_count[0]; i++) {
2992         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]);
2993     }
2994     if(h->slice_type_nos==FF_B_TYPE){
2995         for (i=0; i<h->ref_count[1]; i++) {
2996             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]);
2997         }
2998     }
2999 #endif
3000     return 0;
3001 }
3002
3003 static void print_short_term(H264Context *h);
3004 static void print_long_term(H264Context *h);
3005
3006 /**
3007  * Extract structure information about the picture described by pic_num in
3008  * the current decoding context (frame or field). Note that pic_num is
3009  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
3010  * @param pic_num picture number for which to extract structure information
3011  * @param structure one of PICT_XXX describing structure of picture
3012  *                      with pic_num
3013  * @return frame number (short term) or long term index of picture
3014  *         described by pic_num
3015  */
3016 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3017     MpegEncContext * const s = &h->s;
3018
3019     *structure = s->picture_structure;
3020     if(FIELD_PICTURE){
3021         if (!(pic_num & 1))
3022             /* opposite field */
3023             *structure ^= PICT_FRAME;
3024         pic_num >>= 1;
3025     }
3026
3027     return pic_num;
3028 }
3029
3030 static int decode_ref_pic_list_reordering(H264Context *h){
3031     MpegEncContext * const s = &h->s;
3032     int list, index, pic_structure;
3033
3034     print_short_term(h);
3035     print_long_term(h);
3036     if(h->slice_type_nos==FF_I_TYPE) return 0; //FIXME move before function
3037
3038     for(list=0; list<h->list_count; list++){
3039         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3040
3041         if(get_bits1(&s->gb)){
3042             int pred= h->curr_pic_num;
3043
3044             for(index=0; ; index++){
3045                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3046                 unsigned int pic_id;
3047                 int i;
3048                 Picture *ref = NULL;
3049
3050                 if(reordering_of_pic_nums_idc==3)
3051                     break;
3052
3053                 if(index >= h->ref_count[list]){
3054                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3055                     return -1;
3056                 }
3057
3058                 if(reordering_of_pic_nums_idc<3){
3059                     if(reordering_of_pic_nums_idc<2){
3060                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3061                         int frame_num;
3062
3063                         if(abs_diff_pic_num > h->max_pic_num){
3064                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3065                             return -1;
3066                         }
3067
3068                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3069                         else                                pred+= abs_diff_pic_num;
3070                         pred &= h->max_pic_num - 1;
3071
3072                         frame_num = pic_num_extract(h, pred, &pic_structure);
3073
3074                         for(i= h->short_ref_count-1; i>=0; i--){
3075                             ref = h->short_ref[i];
3076                             assert(ref->reference);
3077                             assert(!ref->long_ref);
3078                             if(ref->data[0] != NULL &&
3079                                    ref->frame_num == frame_num &&
3080                                    (ref->reference & pic_structure) &&
3081                                    ref->long_ref == 0) // ignore non-existing pictures by testing data[0] pointer
3082                                 break;
3083                         }
3084                         if(i>=0)
3085                             ref->pic_id= pred;
3086                     }else{
3087                         int long_idx;
3088                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3089
3090                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3091
3092                         if(long_idx>31){
3093                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3094                             return -1;
3095                         }
3096                         ref = h->long_ref[long_idx];
3097                         assert(!(ref && !ref->reference));
3098                         if(ref && (ref->reference & pic_structure)){
3099                             ref->pic_id= pic_id;
3100                             assert(ref->long_ref);
3101                             i=0;
3102                         }else{
3103                             i=-1;
3104                         }
3105                     }
3106
3107                     if (i < 0) {
3108                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3109                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3110                     } else {
3111                         for(i=index; i+1<h->ref_count[list]; i++){
3112                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3113                                 break;
3114                         }
3115                         for(; i > index; i--){
3116                             h->ref_list[list][i]= h->ref_list[list][i-1];
3117                         }
3118                         h->ref_list[list][index]= *ref;
3119                         if (FIELD_PICTURE){
3120                             pic_as_field(&h->ref_list[list][index], pic_structure);
3121                         }
3122                     }
3123                 }else{
3124                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3125                     return -1;
3126                 }
3127             }
3128         }
3129     }
3130     for(list=0; list<h->list_count; list++){
3131         for(index= 0; index < h->ref_count[list]; index++){
3132             if(!h->ref_list[list][index].data[0])
3133                 h->ref_list[list][index]= s->current_picture;
3134         }
3135     }
3136
3137     if(h->slice_type_nos==FF_B_TYPE && !h->direct_spatial_mv_pred)
3138         direct_dist_scale_factor(h);
3139     direct_ref_list_init(h);
3140     return 0;
3141 }
3142
3143 static void fill_mbaff_ref_list(H264Context *h){
3144     int list, i, j;
3145     for(list=0; list<2; list++){ //FIXME try list_count
3146         for(i=0; i<h->ref_count[list]; i++){
3147             Picture *frame = &h->ref_list[list][i];
3148             Picture *field = &h->ref_list[list][16+2*i];
3149             field[0] = *frame;
3150             for(j=0; j<3; j++)
3151                 field[0].linesize[j] <<= 1;
3152             field[0].reference = PICT_TOP_FIELD;
3153             field[1] = field[0];
3154             for(j=0; j<3; j++)
3155                 field[1].data[j] += frame->linesize[j];
3156             field[1].reference = PICT_BOTTOM_FIELD;
3157
3158             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3159             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3160             for(j=0; j<2; j++){
3161                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3162                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3163             }
3164         }
3165     }
3166     for(j=0; j<h->ref_count[1]; j++){
3167         for(i=0; i<h->ref_count[0]; i++)
3168             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3169         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3170         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3171     }
3172 }
3173
3174 static int pred_weight_table(H264Context *h){
3175     MpegEncContext * const s = &h->s;
3176     int list, i;
3177     int luma_def, chroma_def;
3178
3179     h->use_weight= 0;
3180     h->use_weight_chroma= 0;
3181     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3182     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3183     luma_def = 1<<h->luma_log2_weight_denom;
3184     chroma_def = 1<<h->chroma_log2_weight_denom;
3185
3186     for(list=0; list<2; list++){
3187         for(i=0; i<h->ref_count[list]; i++){
3188             int luma_weight_flag, chroma_weight_flag;
3189
3190             luma_weight_flag= get_bits1(&s->gb);
3191             if(luma_weight_flag){
3192                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3193                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3194                 if(   h->luma_weight[list][i] != luma_def
3195                    || h->luma_offset[list][i] != 0)
3196                     h->use_weight= 1;
3197             }else{
3198                 h->luma_weight[list][i]= luma_def;
3199                 h->luma_offset[list][i]= 0;
3200             }
3201
3202             chroma_weight_flag= get_bits1(&s->gb);
3203             if(chroma_weight_flag){
3204                 int j;
3205                 for(j=0; j<2; j++){
3206                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3207                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3208                     if(   h->chroma_weight[list][i][j] != chroma_def
3209                        || h->chroma_offset[list][i][j] != 0)
3210                         h->use_weight_chroma= 1;
3211                 }
3212             }else{
3213                 int j;
3214                 for(j=0; j<2; j++){
3215                     h->chroma_weight[list][i][j]= chroma_def;
3216                     h->chroma_offset[list][i][j]= 0;
3217                 }
3218             }
3219         }
3220         if(h->slice_type_nos != FF_B_TYPE) break;
3221     }
3222     h->use_weight= h->use_weight || h->use_weight_chroma;
3223     return 0;
3224 }
3225
3226 static void implicit_weight_table(H264Context *h){
3227     MpegEncContext * const s = &h->s;
3228     int ref0, ref1;
3229     int cur_poc = s->current_picture_ptr->poc;
3230
3231     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3232        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3233         h->use_weight= 0;
3234         h->use_weight_chroma= 0;
3235         return;
3236     }
3237
3238     h->use_weight= 2;
3239     h->use_weight_chroma= 2;
3240     h->luma_log2_weight_denom= 5;
3241     h->chroma_log2_weight_denom= 5;
3242
3243     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3244         int poc0 = h->ref_list[0][ref0].poc;
3245         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3246             int poc1 = h->ref_list[1][ref1].poc;
3247             int td = av_clip(poc1 - poc0, -128, 127);
3248             if(td){
3249                 int tb = av_clip(cur_poc - poc0, -128, 127);
3250                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3251                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3252                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3253                     h->implicit_weight[ref0][ref1] = 32;
3254                 else
3255                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3256             }else
3257                 h->implicit_weight[ref0][ref1] = 32;
3258         }
3259     }
3260 }
3261
3262 /**
3263  * Mark a picture as no longer needed for reference. The refmask
3264  * argument allows unreferencing of individual fields or the whole frame.
3265  * If the picture becomes entirely unreferenced, but is being held for
3266  * display purposes, it is marked as such.
3267  * @param refmask mask of fields to unreference; the mask is bitwise
3268  *                anded with the reference marking of pic
3269  * @return non-zero if pic becomes entirely unreferenced (except possibly
3270  *         for display purposes) zero if one of the fields remains in
3271  *         reference
3272  */
3273 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3274     int i;
3275     if (pic->reference &= refmask) {
3276         return 0;
3277     } else {
3278         for(i = 0; h->delayed_pic[i]; i++)
3279             if(pic == h->delayed_pic[i]){
3280                 pic->reference=DELAYED_PIC_REF;
3281                 break;
3282             }
3283         return 1;
3284     }
3285 }
3286
3287 /**
3288  * instantaneous decoder refresh.
3289  */
3290 static void idr(H264Context *h){
3291     int i;
3292
3293     for(i=0; i<16; i++){
3294         if (h->long_ref[i] != NULL) {
3295             unreference_pic(h, h->long_ref[i], 0);
3296             h->long_ref[i]= NULL;
3297         }
3298     }
3299     h->long_ref_count=0;
3300
3301     for(i=0; i<h->short_ref_count; i++){
3302         unreference_pic(h, h->short_ref[i], 0);
3303         h->short_ref[i]= NULL;
3304     }
3305     h->short_ref_count=0;
3306     h->prev_frame_num= 0;
3307 }
3308
3309 /* forget old pics after a seek */
3310 static void flush_dpb(AVCodecContext *avctx){
3311     H264Context *h= avctx->priv_data;
3312     int i;
3313     for(i=0; i<MAX_DELAYED_PIC_COUNT; i++) {
3314         if(h->delayed_pic[i])
3315             h->delayed_pic[i]->reference= 0;
3316         h->delayed_pic[i]= NULL;
3317     }
3318     h->outputed_poc= INT_MIN;
3319     idr(h);
3320     if(h->s.current_picture_ptr)
3321         h->s.current_picture_ptr->reference= 0;
3322     h->s.first_field= 0;
3323     ff_mpeg_flush(avctx);
3324 }
3325
3326 /**
3327  * Find a Picture in the short term reference list by frame number.
3328  * @param frame_num frame number to search for
3329  * @param idx the index into h->short_ref where returned picture is found
3330  *            undefined if no picture found.
3331  * @return pointer to the found picture, or NULL if no pic with the provided
3332  *                 frame number is found
3333  */
3334 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3335     MpegEncContext * const s = &h->s;
3336     int i;
3337
3338     for(i=0; i<h->short_ref_count; i++){
3339         Picture *pic= h->short_ref[i];
3340         if(s->avctx->debug&FF_DEBUG_MMCO)
3341             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3342         if(pic->frame_num == frame_num) {
3343             *idx = i;
3344             return pic;
3345         }
3346     }
3347     return NULL;
3348 }
3349
3350 /**
3351  * Remove a picture from the short term reference list by its index in
3352  * that list.  This does no checking on the provided index; it is assumed
3353  * to be valid. Other list entries are shifted down.
3354  * @param i index into h->short_ref of picture to remove.
3355  */
3356 static void remove_short_at_index(H264Context *h, int i){
3357     assert(i >= 0 && i < h->short_ref_count);
3358     h->short_ref[i]= NULL;
3359     if (--h->short_ref_count)
3360         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3361 }
3362
3363 /**
3364  *
3365  * @return the removed picture or NULL if an error occurs
3366  */
3367 static Picture * remove_short(H264Context *h, int frame_num){
3368     MpegEncContext * const s = &h->s;
3369     Picture *pic;
3370     int i;
3371
3372     if(s->avctx->debug&FF_DEBUG_MMCO)
3373         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3374
3375     pic = find_short(h, frame_num, &i);
3376     if (pic)
3377         remove_short_at_index(h, i);
3378
3379     return pic;
3380 }
3381
3382 /**
3383  * Remove a picture from the long term reference list by its index in
3384  * that list.  This does no checking on the provided index; it is assumed
3385  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3386  * @param i index into h->long_ref of picture to remove.
3387  */
3388 static void remove_long_at_index(H264Context *h, int i){
3389     h->long_ref[i]= NULL;
3390     h->long_ref_count--;
3391 }
3392
3393 /**
3394  *
3395  * @return the removed picture or NULL if an error occurs
3396  */
3397 static Picture * remove_long(H264Context *h, int i){
3398     Picture *pic;
3399
3400     pic= h->long_ref[i];
3401     if (pic)
3402         remove_long_at_index(h, i);
3403
3404     return pic;
3405 }
3406
3407 /**
3408  * print short term list
3409  */
3410 static void print_short_term(H264Context *h) {
3411     uint32_t i;
3412     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3413         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3414         for(i=0; i<h->short_ref_count; i++){
3415             Picture *pic= h->short_ref[i];
3416             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3417         }
3418     }
3419 }
3420
3421 /**
3422  * print long term list
3423  */
3424 static void print_long_term(H264Context *h) {
3425     uint32_t i;