841d85479b0afb0a87d7dad967f79ef1bcb9d848
[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
39 //#undef NDEBUG
40 #include <assert.h>
41
42 /**
43  * Value of Picture.reference when Picture is not a reference picture, but
44  * is held for delayed output.
45  */
46 #define DELAYED_PIC_REF 4
47
48 static VLC coeff_token_vlc[4];
49 static VLC chroma_dc_coeff_token_vlc;
50
51 static VLC total_zeros_vlc[15];
52 static VLC chroma_dc_total_zeros_vlc[3];
53
54 static VLC run_vlc[6];
55 static VLC run7_vlc;
56
57 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
58 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
59 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);
60 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);
61
62 static av_always_inline uint32_t pack16to32(int a, int b){
63 #ifdef WORDS_BIGENDIAN
64    return (b&0xFFFF) + (a<<16);
65 #else
66    return (a&0xFFFF) + (b<<16);
67 #endif
68 }
69
70 const uint8_t ff_rem6[52]={
71 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,
72 };
73
74 const uint8_t ff_div6[52]={
75 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,
76 };
77
78
79 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
80     MpegEncContext * const s = &h->s;
81     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
82     int topleft_xy, top_xy, topright_xy, left_xy[2];
83     int topleft_type, top_type, topright_type, left_type[2];
84     int left_block[8];
85     int i;
86
87     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
88
89     //FIXME deblocking could skip the intra and nnz parts.
90     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[top_xy]) && !FRAME_MBAFF)
91         return;
92
93     //wow what a mess, why didn't they simplify the interlacing&intra stuff, i can't imagine that these complex rules are worth it
94
95     topleft_xy = top_xy - 1;
96     topright_xy= top_xy + 1;
97     left_xy[1] = left_xy[0] = mb_xy-1;
98     left_block[0]= 0;
99     left_block[1]= 1;
100     left_block[2]= 2;
101     left_block[3]= 3;
102     left_block[4]= 7;
103     left_block[5]= 10;
104     left_block[6]= 8;
105     left_block[7]= 11;
106     if(FRAME_MBAFF){
107         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
108         const int top_pair_xy      = pair_xy     - s->mb_stride;
109         const int topleft_pair_xy  = top_pair_xy - 1;
110         const int topright_pair_xy = top_pair_xy + 1;
111         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
112         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
113         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
114         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
115         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
116         const int bottom = (s->mb_y & 1);
117         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);
118         if (bottom
119                 ? !curr_mb_frame_flag // bottom macroblock
120                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
121                 ) {
122             top_xy -= s->mb_stride;
123         }
124         if (bottom
125                 ? !curr_mb_frame_flag // bottom macroblock
126                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
127                 ) {
128             topleft_xy -= s->mb_stride;
129         }
130         if (bottom
131                 ? !curr_mb_frame_flag // bottom macroblock
132                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
133                 ) {
134             topright_xy -= s->mb_stride;
135         }
136         if (left_mb_frame_flag != curr_mb_frame_flag) {
137             left_xy[1] = left_xy[0] = pair_xy - 1;
138             if (curr_mb_frame_flag) {
139                 if (bottom) {
140                     left_block[0]= 2;
141                     left_block[1]= 2;
142                     left_block[2]= 3;
143                     left_block[3]= 3;
144                     left_block[4]= 8;
145                     left_block[5]= 11;
146                     left_block[6]= 8;
147                     left_block[7]= 11;
148                 } else {
149                     left_block[0]= 0;
150                     left_block[1]= 0;
151                     left_block[2]= 1;
152                     left_block[3]= 1;
153                     left_block[4]= 7;
154                     left_block[5]= 10;
155                     left_block[6]= 7;
156                     left_block[7]= 10;
157                 }
158             } else {
159                 left_xy[1] += s->mb_stride;
160                 //left_block[0]= 0;
161                 left_block[1]= 2;
162                 left_block[2]= 0;
163                 left_block[3]= 2;
164                 //left_block[4]= 7;
165                 left_block[5]= 10;
166                 left_block[6]= 7;
167                 left_block[7]= 10;
168             }
169         }
170     }
171
172     h->top_mb_xy = top_xy;
173     h->left_mb_xy[0] = left_xy[0];
174     h->left_mb_xy[1] = left_xy[1];
175     if(for_deblock){
176         topleft_type = 0;
177         topright_type = 0;
178         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
179         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
180         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
181
182         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
183             int list;
184             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
185             for(i=0; i<16; i++)
186                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
187             for(list=0; list<h->list_count; list++){
188                 if(USES_LIST(mb_type,list)){
189                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
190                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
191                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
192                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
193                         dst[0] = src[0];
194                         dst[1] = src[1];
195                         dst[2] = src[2];
196                         dst[3] = src[3];
197                     }
198                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
199                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
200                     ref += h->b8_stride;
201                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
202                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
203                 }else{
204                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
205                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
206                 }
207             }
208         }
209     }else{
210         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
211         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
212         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
213         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
214         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
215     }
216
217     if(IS_INTRA(mb_type)){
218         h->topleft_samples_available=
219         h->top_samples_available=
220         h->left_samples_available= 0xFFFF;
221         h->topright_samples_available= 0xEEEA;
222
223         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
224             h->topleft_samples_available= 0xB3FF;
225             h->top_samples_available= 0x33FF;
226             h->topright_samples_available= 0x26EA;
227         }
228         for(i=0; i<2; i++){
229             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
230                 h->topleft_samples_available&= 0xDF5F;
231                 h->left_samples_available&= 0x5F5F;
232             }
233         }
234
235         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
236             h->topleft_samples_available&= 0x7FFF;
237
238         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
239             h->topright_samples_available&= 0xFBFF;
240
241         if(IS_INTRA4x4(mb_type)){
242             if(IS_INTRA4x4(top_type)){
243                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
244                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
245                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
246                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
247             }else{
248                 int pred;
249                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
250                     pred= -1;
251                 else{
252                     pred= 2;
253                 }
254                 h->intra4x4_pred_mode_cache[4+8*0]=
255                 h->intra4x4_pred_mode_cache[5+8*0]=
256                 h->intra4x4_pred_mode_cache[6+8*0]=
257                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
258             }
259             for(i=0; i<2; i++){
260                 if(IS_INTRA4x4(left_type[i])){
261                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
262                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
263                 }else{
264                     int pred;
265                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
266                         pred= -1;
267                     else{
268                         pred= 2;
269                     }
270                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
271                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
272                 }
273             }
274         }
275     }
276
277
278 /*
279 0 . T T. T T T T
280 1 L . .L . . . .
281 2 L . .L . . . .
282 3 . T TL . . . .
283 4 L . .L . . . .
284 5 L . .. . . . .
285 */
286 //FIXME constraint_intra_pred & partitioning & nnz (lets hope this is just a typo in the spec)
287     if(top_type){
288         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
289         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
290         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
291         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
292
293         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
294         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
295
296         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
297         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
298
299     }else{
300         h->non_zero_count_cache[4+8*0]=
301         h->non_zero_count_cache[5+8*0]=
302         h->non_zero_count_cache[6+8*0]=
303         h->non_zero_count_cache[7+8*0]=
304
305         h->non_zero_count_cache[1+8*0]=
306         h->non_zero_count_cache[2+8*0]=
307
308         h->non_zero_count_cache[1+8*3]=
309         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
310
311     }
312
313     for (i=0; i<2; i++) {
314         if(left_type[i]){
315             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
316             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
317             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
318             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
319         }else{
320             h->non_zero_count_cache[3+8*1 + 2*8*i]=
321             h->non_zero_count_cache[3+8*2 + 2*8*i]=
322             h->non_zero_count_cache[0+8*1 +   8*i]=
323             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
324         }
325     }
326
327     if( h->pps.cabac ) {
328         // top_cbp
329         if(top_type) {
330             h->top_cbp = h->cbp_table[top_xy];
331         } else if(IS_INTRA(mb_type)) {
332             h->top_cbp = 0x1C0;
333         } else {
334             h->top_cbp = 0;
335         }
336         // left_cbp
337         if (left_type[0]) {
338             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
339         } else if(IS_INTRA(mb_type)) {
340             h->left_cbp = 0x1C0;
341         } else {
342             h->left_cbp = 0;
343         }
344         if (left_type[0]) {
345             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
346         }
347         if (left_type[1]) {
348             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
349         }
350     }
351
352 #if 1
353     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
354         int list;
355         for(list=0; list<h->list_count; list++){
356             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
357                 /*if(!h->mv_cache_clean[list]){
358                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
359                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
360                     h->mv_cache_clean[list]= 1;
361                 }*/
362                 continue;
363             }
364             h->mv_cache_clean[list]= 0;
365
366             if(USES_LIST(top_type, list)){
367                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
368                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
369                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
370                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
371                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
372                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
373                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
374                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
375                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
376                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
377             }else{
378                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
379                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
380                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
381                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
382                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
383             }
384
385             for(i=0; i<2; i++){
386                 int cache_idx = scan8[0] - 1 + i*2*8;
387                 if(USES_LIST(left_type[i], list)){
388                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
389                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
390                     *(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]];
391                     *(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]];
392                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
393                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
394                 }else{
395                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
396                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
397                     h->ref_cache[list][cache_idx  ]=
398                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
399                 }
400             }
401
402             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
403                 continue;
404
405             if(USES_LIST(topleft_type, list)){
406                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + 3*h->b_stride;
407                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + h->b8_stride;
408                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
409                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
410             }else{
411                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
412                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
413             }
414
415             if(USES_LIST(topright_type, list)){
416                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
417                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
418                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
419                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
420             }else{
421                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
422                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
423             }
424
425             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
426                 continue;
427
428             h->ref_cache[list][scan8[5 ]+1] =
429             h->ref_cache[list][scan8[7 ]+1] =
430             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
431             h->ref_cache[list][scan8[4 ]] =
432             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
433             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
434             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
435             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
436             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
437             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
438
439             if( h->pps.cabac ) {
440                 /* XXX beurk, Load mvd */
441                 if(USES_LIST(top_type, list)){
442                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
443                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
444                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
445                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
446                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
447                 }else{
448                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
449                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
450                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
451                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
452                 }
453                 if(USES_LIST(left_type[0], list)){
454                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
455                     *(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]];
456                     *(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]];
457                 }else{
458                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
459                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
460                 }
461                 if(USES_LIST(left_type[1], list)){
462                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
463                     *(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]];
464                     *(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]];
465                 }else{
466                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
467                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
468                 }
469                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
470                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
471                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
472                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
473                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
474
475                 if(h->slice_type == B_TYPE){
476                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
477
478                     if(IS_DIRECT(top_type)){
479                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
480                     }else if(IS_8X8(top_type)){
481                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
482                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
483                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
484                     }else{
485                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
486                     }
487
488                     if(IS_DIRECT(left_type[0]))
489                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
490                     else if(IS_8X8(left_type[0]))
491                         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)];
492                     else
493                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
494
495                     if(IS_DIRECT(left_type[1]))
496                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
497                     else if(IS_8X8(left_type[1]))
498                         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)];
499                     else
500                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
501                 }
502             }
503
504             if(FRAME_MBAFF){
505 #define MAP_MVS\
506                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
507                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
508                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
509                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
510                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
511                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
512                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
513                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
514                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
515                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
516                 if(MB_FIELD){
517 #define MAP_F2F(idx, mb_type)\
518                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
519                         h->ref_cache[list][idx] <<= 1;\
520                         h->mv_cache[list][idx][1] /= 2;\
521                         h->mvd_cache[list][idx][1] /= 2;\
522                     }
523                     MAP_MVS
524 #undef MAP_F2F
525                 }else{
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] <<= 1;\
530                         h->mvd_cache[list][idx][1] <<= 1;\
531                     }
532                     MAP_MVS
533 #undef MAP_F2F
534                 }
535             }
536         }
537     }
538 #endif
539
540     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
541 }
542
543 static inline void write_back_intra_pred_mode(H264Context *h){
544     MpegEncContext * const s = &h->s;
545     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
546
547     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
548     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
549     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
550     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
551     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
552     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
553     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
554 }
555
556 /**
557  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
558  */
559 static inline int check_intra4x4_pred_mode(H264Context *h){
560     MpegEncContext * const s = &h->s;
561     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
562     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
563     int i;
564
565     if(!(h->top_samples_available&0x8000)){
566         for(i=0; i<4; i++){
567             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
568             if(status<0){
569                 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);
570                 return -1;
571             } else if(status){
572                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
573             }
574         }
575     }
576
577     if(!(h->left_samples_available&0x8000)){
578         for(i=0; i<4; i++){
579             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
580             if(status<0){
581                 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);
582                 return -1;
583             } else if(status){
584                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
585             }
586         }
587     }
588
589     return 0;
590 } //FIXME cleanup like next
591
592 /**
593  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
594  */
595 static inline int check_intra_pred_mode(H264Context *h, int mode){
596     MpegEncContext * const s = &h->s;
597     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
598     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
599
600     if(mode > 6U) {
601         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);
602         return -1;
603     }
604
605     if(!(h->top_samples_available&0x8000)){
606         mode= top[ mode ];
607         if(mode<0){
608             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);
609             return -1;
610         }
611     }
612
613     if(!(h->left_samples_available&0x8000)){
614         mode= left[ mode ];
615         if(mode<0){
616             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);
617             return -1;
618         }
619     }
620
621     return mode;
622 }
623
624 /**
625  * gets the predicted intra4x4 prediction mode.
626  */
627 static inline int pred_intra_mode(H264Context *h, int n){
628     const int index8= scan8[n];
629     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
630     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
631     const int min= FFMIN(left, top);
632
633     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
634
635     if(min<0) return DC_PRED;
636     else      return min;
637 }
638
639 static inline void write_back_non_zero_count(H264Context *h){
640     MpegEncContext * const s = &h->s;
641     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
642
643     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
644     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
645     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
646     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
647     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
648     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
649     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
650
651     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
652     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
653     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
654
655     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
656     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
657     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
658
659     if(FRAME_MBAFF){
660         // store all luma nnzs, for deblocking
661         int v = 0, i;
662         for(i=0; i<16; i++)
663             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
664         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
665     }
666 }
667
668 /**
669  * gets the predicted number of non zero coefficients.
670  * @param n block index
671  */
672 static inline int pred_non_zero_count(H264Context *h, int n){
673     const int index8= scan8[n];
674     const int left= h->non_zero_count_cache[index8 - 1];
675     const int top = h->non_zero_count_cache[index8 - 8];
676     int i= left + top;
677
678     if(i<64) i= (i+1)>>1;
679
680     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
681
682     return i&31;
683 }
684
685 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
686     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
687     MpegEncContext *s = &h->s;
688
689     /* there is no consistent mapping of mvs to neighboring locations that will
690      * make mbaff happy, so we can't move all this logic to fill_caches */
691     if(FRAME_MBAFF){
692         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
693         const int16_t *mv;
694         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
695         *C = h->mv_cache[list][scan8[0]-2];
696
697         if(!MB_FIELD
698            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
699             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
700             if(IS_INTERLACED(mb_types[topright_xy])){
701 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
702                 const int x4 = X4, y4 = Y4;\
703                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
704                 if(!USES_LIST(mb_type,list) && !IS_8X8(mb_type))\
705                     return LIST_NOT_USED;\
706                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
707                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
708                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
709                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
710
711                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
712             }
713         }
714         if(topright_ref == PART_NOT_AVAILABLE
715            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
716            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
717             if(!MB_FIELD
718                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
719                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
720             }
721             if(MB_FIELD
722                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
723                && i >= scan8[0]+8){
724                 // leftshift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's ok.
725                 SET_DIAG_MV(>>1, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
726             }
727         }
728 #undef SET_DIAG_MV
729     }
730
731     if(topright_ref != PART_NOT_AVAILABLE){
732         *C= h->mv_cache[list][ i - 8 + part_width ];
733         return topright_ref;
734     }else{
735         tprintf(s->avctx, "topright MV not available\n");
736
737         *C= h->mv_cache[list][ i - 8 - 1 ];
738         return h->ref_cache[list][ i - 8 - 1 ];
739     }
740 }
741
742 /**
743  * gets the predicted MV.
744  * @param n the block index
745  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
746  * @param mx the x component of the predicted motion vector
747  * @param my the y component of the predicted motion vector
748  */
749 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
750     const int index8= scan8[n];
751     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
752     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
753     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
754     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
755     const int16_t * C;
756     int diagonal_ref, match_count;
757
758     assert(part_width==1 || part_width==2 || part_width==4);
759
760 /* mv_cache
761   B . . A T T T T
762   U . . L . . , .
763   U . . L . . . .
764   U . . L . . , .
765   . . . L . . . .
766 */
767
768     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
769     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
770     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
771     if(match_count > 1){ //most common
772         *mx= mid_pred(A[0], B[0], C[0]);
773         *my= mid_pred(A[1], B[1], C[1]);
774     }else if(match_count==1){
775         if(left_ref==ref){
776             *mx= A[0];
777             *my= A[1];
778         }else if(top_ref==ref){
779             *mx= B[0];
780             *my= B[1];
781         }else{
782             *mx= C[0];
783             *my= C[1];
784         }
785     }else{
786         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
787             *mx= A[0];
788             *my= A[1];
789         }else{
790             *mx= mid_pred(A[0], B[0], C[0]);
791             *my= mid_pred(A[1], B[1], C[1]);
792         }
793     }
794
795     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);
796 }
797
798 /**
799  * gets the directionally predicted 16x8 MV.
800  * @param n the block index
801  * @param mx the x component of the predicted motion vector
802  * @param my the y component of the predicted motion vector
803  */
804 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
805     if(n==0){
806         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
807         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
808
809         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);
810
811         if(top_ref == ref){
812             *mx= B[0];
813             *my= B[1];
814             return;
815         }
816     }else{
817         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
818         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
819
820         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);
821
822         if(left_ref == ref){
823             *mx= A[0];
824             *my= A[1];
825             return;
826         }
827     }
828
829     //RARE
830     pred_motion(h, n, 4, list, ref, mx, my);
831 }
832
833 /**
834  * gets the directionally predicted 8x16 MV.
835  * @param n the block index
836  * @param mx the x component of the predicted motion vector
837  * @param my the y component of the predicted motion vector
838  */
839 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
840     if(n==0){
841         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
842         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
843
844         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);
845
846         if(left_ref == ref){
847             *mx= A[0];
848             *my= A[1];
849             return;
850         }
851     }else{
852         const int16_t * C;
853         int diagonal_ref;
854
855         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
856
857         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);
858
859         if(diagonal_ref == ref){
860             *mx= C[0];
861             *my= C[1];
862             return;
863         }
864     }
865
866     //RARE
867     pred_motion(h, n, 2, list, ref, mx, my);
868 }
869
870 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
871     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
872     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
873
874     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
875
876     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
877        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
878        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
879
880         *mx = *my = 0;
881         return;
882     }
883
884     pred_motion(h, 0, 4, 0, 0, mx, my);
885
886     return;
887 }
888
889 static inline void direct_dist_scale_factor(H264Context * const h){
890     const int poc = h->s.current_picture_ptr->poc;
891     const int poc1 = h->ref_list[1][0].poc;
892     int i;
893     for(i=0; i<h->ref_count[0]; i++){
894         int poc0 = h->ref_list[0][i].poc;
895         int td = av_clip(poc1 - poc0, -128, 127);
896         if(td == 0 /* FIXME || pic0 is a long-term ref */){
897             h->dist_scale_factor[i] = 256;
898         }else{
899             int tb = av_clip(poc - poc0, -128, 127);
900             int tx = (16384 + (FFABS(td) >> 1)) / td;
901             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
902         }
903     }
904     if(FRAME_MBAFF){
905         for(i=0; i<h->ref_count[0]; i++){
906             h->dist_scale_factor_field[2*i] =
907             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
908         }
909     }
910 }
911 static inline void direct_ref_list_init(H264Context * const h){
912     MpegEncContext * const s = &h->s;
913     Picture * const ref1 = &h->ref_list[1][0];
914     Picture * const cur = s->current_picture_ptr;
915     int list, i, j;
916     if(cur->pict_type == I_TYPE)
917         cur->ref_count[0] = 0;
918     if(cur->pict_type != B_TYPE)
919         cur->ref_count[1] = 0;
920     for(list=0; list<2; list++){
921         cur->ref_count[list] = h->ref_count[list];
922         for(j=0; j<h->ref_count[list]; j++)
923             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
924     }
925     if(cur->pict_type != B_TYPE || h->direct_spatial_mv_pred)
926         return;
927     for(list=0; list<2; list++){
928         for(i=0; i<ref1->ref_count[list]; i++){
929             const int poc = ref1->ref_poc[list][i];
930             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
931             for(j=0; j<h->ref_count[list]; j++)
932                 if(h->ref_list[list][j].poc == poc){
933                     h->map_col_to_list0[list][i] = j;
934                     break;
935                 }
936         }
937     }
938     if(FRAME_MBAFF){
939         for(list=0; list<2; list++){
940             for(i=0; i<ref1->ref_count[list]; i++){
941                 j = h->map_col_to_list0[list][i];
942                 h->map_col_to_list0_field[list][2*i] = 2*j;
943                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
944             }
945         }
946     }
947 }
948
949 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
950     MpegEncContext * const s = &h->s;
951     const int mb_xy =   s->mb_x +   s->mb_y*s->mb_stride;
952     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
953     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
954     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
955     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
956     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
957     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
958     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
959     const int is_b8x8 = IS_8X8(*mb_type);
960     unsigned int sub_mb_type;
961     int i8, i4;
962
963 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
964     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
965         /* FIXME save sub mb types from previous frames (or derive from MVs)
966          * so we know exactly what block size to use */
967         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
968         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
969     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
970         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
971         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
972     }else{
973         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
974         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
975     }
976     if(!is_b8x8)
977         *mb_type |= MB_TYPE_DIRECT2;
978     if(MB_FIELD)
979         *mb_type |= MB_TYPE_INTERLACED;
980
981     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);
982
983     if(h->direct_spatial_mv_pred){
984         int ref[2];
985         int mv[2][2];
986         int list;
987
988         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
989
990         /* ref = min(neighbors) */
991         for(list=0; list<2; list++){
992             int refa = h->ref_cache[list][scan8[0] - 1];
993             int refb = h->ref_cache[list][scan8[0] - 8];
994             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
995             if(refc == -2)
996                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
997             ref[list] = refa;
998             if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
999                 ref[list] = refb;
1000             if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
1001                 ref[list] = refc;
1002             if(ref[list] < 0)
1003                 ref[list] = -1;
1004         }
1005
1006         if(ref[0] < 0 && ref[1] < 0){
1007             ref[0] = ref[1] = 0;
1008             mv[0][0] = mv[0][1] =
1009             mv[1][0] = mv[1][1] = 0;
1010         }else{
1011             for(list=0; list<2; list++){
1012                 if(ref[list] >= 0)
1013                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1014                 else
1015                     mv[list][0] = mv[list][1] = 0;
1016             }
1017         }
1018
1019         if(ref[1] < 0){
1020             *mb_type &= ~MB_TYPE_P0L1;
1021             sub_mb_type &= ~MB_TYPE_P0L1;
1022         }else if(ref[0] < 0){
1023             *mb_type &= ~MB_TYPE_P0L0;
1024             sub_mb_type &= ~MB_TYPE_P0L0;
1025         }
1026
1027         if(IS_16X16(*mb_type)){
1028             int a=0, b=0;
1029
1030             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1031             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1032             if(!IS_INTRA(mb_type_col)
1033                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1034                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1035                        && (h->x264_build>33 || !h->x264_build)))){
1036                 if(ref[0] > 0)
1037                     a= pack16to32(mv[0][0],mv[0][1]);
1038                 if(ref[1] > 0)
1039                     b= pack16to32(mv[1][0],mv[1][1]);
1040             }else{
1041                 a= pack16to32(mv[0][0],mv[0][1]);
1042                 b= pack16to32(mv[1][0],mv[1][1]);
1043             }
1044             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1045             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1046         }else{
1047             for(i8=0; i8<4; i8++){
1048                 const int x8 = i8&1;
1049                 const int y8 = i8>>1;
1050
1051                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1052                     continue;
1053                 h->sub_mb_type[i8] = sub_mb_type;
1054
1055                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1056                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1057                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1058                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1059
1060                 /* col_zero_flag */
1061                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1062                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1063                                                   && (h->x264_build>33 || !h->x264_build)))){
1064                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1065                     if(IS_SUB_8X8(sub_mb_type)){
1066                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1067                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1068                             if(ref[0] == 0)
1069                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1070                             if(ref[1] == 0)
1071                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1072                         }
1073                     }else
1074                     for(i4=0; i4<4; i4++){
1075                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1076                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1077                             if(ref[0] == 0)
1078                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1079                             if(ref[1] == 0)
1080                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1081                         }
1082                     }
1083                 }
1084             }
1085         }
1086     }else{ /* direct temporal mv pred */
1087         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1088         const int *dist_scale_factor = h->dist_scale_factor;
1089
1090         if(FRAME_MBAFF){
1091             if(IS_INTERLACED(*mb_type)){
1092                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1093                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1094                 dist_scale_factor = h->dist_scale_factor_field;
1095             }
1096             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1097                 /* FIXME assumes direct_8x8_inference == 1 */
1098                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1099                 int mb_types_col[2];
1100                 int y_shift;
1101
1102                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1103                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1104                          | (*mb_type & MB_TYPE_INTERLACED);
1105                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1106
1107                 if(IS_INTERLACED(*mb_type)){
1108                     /* frame to field scaling */
1109                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1110                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1111                     if(s->mb_y&1){
1112                         l1ref0 -= 2*h->b8_stride;
1113                         l1ref1 -= 2*h->b8_stride;
1114                         l1mv0 -= 4*h->b_stride;
1115                         l1mv1 -= 4*h->b_stride;
1116                     }
1117                     y_shift = 0;
1118
1119                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1120                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1121                        && !is_b8x8)
1122                         *mb_type |= MB_TYPE_16x8;
1123                     else
1124                         *mb_type |= MB_TYPE_8x8;
1125                 }else{
1126                     /* field to frame scaling */
1127                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1128                      * but in MBAFF, top and bottom POC are equal */
1129                     int dy = (s->mb_y&1) ? 1 : 2;
1130                     mb_types_col[0] =
1131                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1132                     l1ref0 += dy*h->b8_stride;
1133                     l1ref1 += dy*h->b8_stride;
1134                     l1mv0 += 2*dy*h->b_stride;
1135                     l1mv1 += 2*dy*h->b_stride;
1136                     y_shift = 2;
1137
1138                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1139                        && !is_b8x8)
1140                         *mb_type |= MB_TYPE_16x16;
1141                     else
1142                         *mb_type |= MB_TYPE_8x8;
1143                 }
1144
1145                 for(i8=0; i8<4; i8++){
1146                     const int x8 = i8&1;
1147                     const int y8 = i8>>1;
1148                     int ref0, scale;
1149                     const int16_t (*l1mv)[2]= l1mv0;
1150
1151                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1152                         continue;
1153                     h->sub_mb_type[i8] = sub_mb_type;
1154
1155                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1156                     if(IS_INTRA(mb_types_col[y8])){
1157                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1158                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1159                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1160                         continue;
1161                     }
1162
1163                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1164                     if(ref0 >= 0)
1165                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1166                     else{
1167                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1168                         l1mv= l1mv1;
1169                     }
1170                     scale = dist_scale_factor[ref0];
1171                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1172
1173                     {
1174                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1175                         int my_col = (mv_col[1]<<y_shift)/2;
1176                         int mx = (scale * mv_col[0] + 128) >> 8;
1177                         int my = (scale * my_col + 128) >> 8;
1178                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1179                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1180                     }
1181                 }
1182                 return;
1183             }
1184         }
1185
1186         /* one-to-one mv scaling */
1187
1188         if(IS_16X16(*mb_type)){
1189             int ref, mv0, mv1;
1190
1191             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1192             if(IS_INTRA(mb_type_col)){
1193                 ref=mv0=mv1=0;
1194             }else{
1195                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1196                                                 : map_col_to_list0[1][l1ref1[0]];
1197                 const int scale = dist_scale_factor[ref0];
1198                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1199                 int mv_l0[2];
1200                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1201                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1202                 ref= ref0;
1203                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1204                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1205             }
1206             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1207             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1208             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1209         }else{
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                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1220                 if(IS_INTRA(mb_type_col)){
1221                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1222                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1223                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1224                     continue;
1225                 }
1226
1227                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1228                 if(ref0 >= 0)
1229                     ref0 = map_col_to_list0[0][ref0];
1230                 else{
1231                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1232                     l1mv= l1mv1;
1233                 }
1234                 scale = dist_scale_factor[ref0];
1235
1236                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1237                 if(IS_SUB_8X8(sub_mb_type)){
1238                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1239                     int mx = (scale * mv_col[0] + 128) >> 8;
1240                     int my = (scale * mv_col[1] + 128) >> 8;
1241                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1242                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1243                 }else
1244                 for(i4=0; i4<4; i4++){
1245                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1246                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1247                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1248                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1249                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1250                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1251                 }
1252             }
1253         }
1254     }
1255 }
1256
1257 static inline void write_back_motion(H264Context *h, int mb_type){
1258     MpegEncContext * const s = &h->s;
1259     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1260     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1261     int list;
1262
1263     if(!USES_LIST(mb_type, 0))
1264         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1265
1266     for(list=0; list<h->list_count; list++){
1267         int y;
1268         if(!USES_LIST(mb_type, list))
1269             continue;
1270
1271         for(y=0; y<4; y++){
1272             *(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];
1273             *(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];
1274         }
1275         if( h->pps.cabac ) {
1276             if(IS_SKIP(mb_type))
1277                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1278             else
1279             for(y=0; y<4; y++){
1280                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1281                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1282             }
1283         }
1284
1285         {
1286             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1287             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1288             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1289             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1290             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1291         }
1292     }
1293
1294     if(h->slice_type == B_TYPE && h->pps.cabac){
1295         if(IS_8X8(mb_type)){
1296             uint8_t *direct_table = &h->direct_table[b8_xy];
1297             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1298             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1299             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1300         }
1301     }
1302 }
1303
1304 /**
1305  * Decodes a network abstraction layer unit.
1306  * @param consumed is the number of bytes used as input
1307  * @param length is the length of the array
1308  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1309  * @returns decoded bytes, might be src+1 if no escapes
1310  */
1311 static uint8_t *decode_nal(H264Context *h, uint8_t *src, int *dst_length, int *consumed, int length){
1312     int i, si, di;
1313     uint8_t *dst;
1314     int bufidx;
1315
1316 //    src[0]&0x80;                //forbidden bit
1317     h->nal_ref_idc= src[0]>>5;
1318     h->nal_unit_type= src[0]&0x1F;
1319
1320     src++; length--;
1321 #if 0
1322     for(i=0; i<length; i++)
1323         printf("%2X ", src[i]);
1324 #endif
1325     for(i=0; i+1<length; i+=2){
1326         if(src[i]) continue;
1327         if(i>0 && src[i-1]==0) i--;
1328         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1329             if(src[i+2]!=3){
1330                 /* startcode, so we must be past the end */
1331                 length=i;
1332             }
1333             break;
1334         }
1335     }
1336
1337     if(i>=length-1){ //no escaped 0
1338         *dst_length= length;
1339         *consumed= length+1; //+1 for the header
1340         return src;
1341     }
1342
1343     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1344     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1345     dst= h->rbsp_buffer[bufidx];
1346
1347     if (dst == NULL){
1348         return NULL;
1349     }
1350
1351 //printf("decoding esc\n");
1352     si=di=0;
1353     while(si<length){
1354         //remove escapes (very rare 1:2^22)
1355         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1356             if(src[si+2]==3){ //escape
1357                 dst[di++]= 0;
1358                 dst[di++]= 0;
1359                 si+=3;
1360                 continue;
1361             }else //next start code
1362                 break;
1363         }
1364
1365         dst[di++]= src[si++];
1366     }
1367
1368     *dst_length= di;
1369     *consumed= si + 1;//+1 for the header
1370 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1371     return dst;
1372 }
1373
1374 /**
1375  * identifies the exact end of the bitstream
1376  * @return the length of the trailing, or 0 if damaged
1377  */
1378 static int decode_rbsp_trailing(H264Context *h, uint8_t *src){
1379     int v= *src;
1380     int r;
1381
1382     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1383
1384     for(r=1; r<9; r++){
1385         if(v&1) return r;
1386         v>>=1;
1387     }
1388     return 0;
1389 }
1390
1391 /**
1392  * idct tranforms the 16 dc values and dequantize them.
1393  * @param qp quantization parameter
1394  */
1395 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1396 #define stride 16
1397     int i;
1398     int temp[16]; //FIXME check if this is a good idea
1399     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1400     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1401
1402 //memset(block, 64, 2*256);
1403 //return;
1404     for(i=0; i<4; i++){
1405         const int offset= y_offset[i];
1406         const int z0= block[offset+stride*0] + block[offset+stride*4];
1407         const int z1= block[offset+stride*0] - block[offset+stride*4];
1408         const int z2= block[offset+stride*1] - block[offset+stride*5];
1409         const int z3= block[offset+stride*1] + block[offset+stride*5];
1410
1411         temp[4*i+0]= z0+z3;
1412         temp[4*i+1]= z1+z2;
1413         temp[4*i+2]= z1-z2;
1414         temp[4*i+3]= z0-z3;
1415     }
1416
1417     for(i=0; i<4; i++){
1418         const int offset= x_offset[i];
1419         const int z0= temp[4*0+i] + temp[4*2+i];
1420         const int z1= temp[4*0+i] - temp[4*2+i];
1421         const int z2= temp[4*1+i] - temp[4*3+i];
1422         const int z3= temp[4*1+i] + temp[4*3+i];
1423
1424         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_resdual
1425         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1426         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1427         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1428     }
1429 }
1430
1431 #if 0
1432 /**
1433  * dct tranforms the 16 dc values.
1434  * @param qp quantization parameter ??? FIXME
1435  */
1436 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1437 //    const int qmul= dequant_coeff[qp][0];
1438     int i;
1439     int temp[16]; //FIXME check if this is a good idea
1440     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1441     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1442
1443     for(i=0; i<4; i++){
1444         const int offset= y_offset[i];
1445         const int z0= block[offset+stride*0] + block[offset+stride*4];
1446         const int z1= block[offset+stride*0] - block[offset+stride*4];
1447         const int z2= block[offset+stride*1] - block[offset+stride*5];
1448         const int z3= block[offset+stride*1] + block[offset+stride*5];
1449
1450         temp[4*i+0]= z0+z3;
1451         temp[4*i+1]= z1+z2;
1452         temp[4*i+2]= z1-z2;
1453         temp[4*i+3]= z0-z3;
1454     }
1455
1456     for(i=0; i<4; i++){
1457         const int offset= x_offset[i];
1458         const int z0= temp[4*0+i] + temp[4*2+i];
1459         const int z1= temp[4*0+i] - temp[4*2+i];
1460         const int z2= temp[4*1+i] - temp[4*3+i];
1461         const int z3= temp[4*1+i] + temp[4*3+i];
1462
1463         block[stride*0 +offset]= (z0 + z3)>>1;
1464         block[stride*2 +offset]= (z1 + z2)>>1;
1465         block[stride*8 +offset]= (z1 - z2)>>1;
1466         block[stride*10+offset]= (z0 - z3)>>1;
1467     }
1468 }
1469 #endif
1470
1471 #undef xStride
1472 #undef stride
1473
1474 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1475     const int stride= 16*2;
1476     const int xStride= 16;
1477     int a,b,c,d,e;
1478
1479     a= block[stride*0 + xStride*0];
1480     b= block[stride*0 + xStride*1];
1481     c= block[stride*1 + xStride*0];
1482     d= block[stride*1 + xStride*1];
1483
1484     e= a-b;
1485     a= a+b;
1486     b= c-d;
1487     c= c+d;
1488
1489     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1490     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1491     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1492     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1493 }
1494
1495 #if 0
1496 static void chroma_dc_dct_c(DCTELEM *block){
1497     const int stride= 16*2;
1498     const int xStride= 16;
1499     int a,b,c,d,e;
1500
1501     a= block[stride*0 + xStride*0];
1502     b= block[stride*0 + xStride*1];
1503     c= block[stride*1 + xStride*0];
1504     d= block[stride*1 + xStride*1];
1505
1506     e= a-b;
1507     a= a+b;
1508     b= c-d;
1509     c= c+d;
1510
1511     block[stride*0 + xStride*0]= (a+c);
1512     block[stride*0 + xStride*1]= (e+b);
1513     block[stride*1 + xStride*0]= (a-c);
1514     block[stride*1 + xStride*1]= (e-b);
1515 }
1516 #endif
1517
1518 /**
1519  * gets the chroma qp.
1520  */
1521 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1522     return h->pps.chroma_qp_table[t][qscale & 0xff];
1523 }
1524
1525 //FIXME need to check that this does not overflow signed 32 bit for low qp, i am not sure, it's very close
1526 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1527 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1528     int i;
1529     const int * const quant_table= quant_coeff[qscale];
1530     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1531     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1532     const unsigned int threshold2= (threshold1<<1);
1533     int last_non_zero;
1534
1535     if(separate_dc){
1536         if(qscale<=18){
1537             //avoid overflows
1538             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1539             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1540             const unsigned int dc_threshold2= (dc_threshold1<<1);
1541
1542             int level= block[0]*quant_coeff[qscale+18][0];
1543             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1544                 if(level>0){
1545                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1546                     block[0]= level;
1547                 }else{
1548                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1549                     block[0]= -level;
1550                 }
1551 //                last_non_zero = i;
1552             }else{
1553                 block[0]=0;
1554             }
1555         }else{
1556             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1557             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1558             const unsigned int dc_threshold2= (dc_threshold1<<1);
1559
1560             int level= block[0]*quant_table[0];
1561             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1562                 if(level>0){
1563                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1564                     block[0]= level;
1565                 }else{
1566                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1567                     block[0]= -level;
1568                 }
1569 //                last_non_zero = i;
1570             }else{
1571                 block[0]=0;
1572             }
1573         }
1574         last_non_zero= 0;
1575         i=1;
1576     }else{
1577         last_non_zero= -1;
1578         i=0;
1579     }
1580
1581     for(; i<16; i++){
1582         const int j= scantable[i];
1583         int level= block[j]*quant_table[j];
1584
1585 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1586 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1587         if(((unsigned)(level+threshold1))>threshold2){
1588             if(level>0){
1589                 level= (bias + level)>>QUANT_SHIFT;
1590                 block[j]= level;
1591             }else{
1592                 level= (bias - level)>>QUANT_SHIFT;
1593                 block[j]= -level;
1594             }
1595             last_non_zero = i;
1596         }else{
1597             block[j]=0;
1598         }
1599     }
1600
1601     return last_non_zero;
1602 }
1603
1604 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1605                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1606                            int src_x_offset, int src_y_offset,
1607                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1608     MpegEncContext * const s = &h->s;
1609     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1610     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1611     const int luma_xy= (mx&3) + ((my&3)<<2);
1612     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1613     uint8_t * src_cb, * src_cr;
1614     int extra_width= h->emu_edge_width;
1615     int extra_height= h->emu_edge_height;
1616     int emu=0;
1617     const int full_mx= mx>>2;
1618     const int full_my= my>>2;
1619     const int pic_width  = 16*s->mb_width;
1620     const int pic_height = 16*s->mb_height >> MB_FIELD;
1621
1622     if(!pic->data[0]) //FIXME this is unacceptable, some senseable error concealment must be done for missing reference frames
1623         return;
1624
1625     if(mx&7) extra_width -= 3;
1626     if(my&7) extra_height -= 3;
1627
1628     if(   full_mx < 0-extra_width
1629        || full_my < 0-extra_height
1630        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1631        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1632         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);
1633             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1634         emu=1;
1635     }
1636
1637     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1638     if(!square){
1639         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1640     }
1641
1642     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1643
1644     if(MB_FIELD){
1645         // chroma offset when predicting from a field of opposite parity
1646         my += 2 * ((s->mb_y & 1) - (pic->reference - 1));
1647         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1648     }
1649     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1650     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1651
1652     if(emu){
1653         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);
1654             src_cb= s->edge_emu_buffer;
1655     }
1656     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1657
1658     if(emu){
1659         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);
1660             src_cr= s->edge_emu_buffer;
1661     }
1662     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1663 }
1664
1665 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1666                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1667                            int x_offset, int y_offset,
1668                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1669                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1670                            int list0, int list1){
1671     MpegEncContext * const s = &h->s;
1672     qpel_mc_func *qpix_op=  qpix_put;
1673     h264_chroma_mc_func chroma_op= chroma_put;
1674
1675     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1676     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1677     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1678     x_offset += 8*s->mb_x;
1679     y_offset += 8*(s->mb_y >> MB_FIELD);
1680
1681     if(list0){
1682         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1683         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1684                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1685                            qpix_op, chroma_op);
1686
1687         qpix_op=  qpix_avg;
1688         chroma_op= chroma_avg;
1689     }
1690
1691     if(list1){
1692         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1693         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1694                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1695                            qpix_op, chroma_op);
1696     }
1697 }
1698
1699 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1700                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1701                            int x_offset, int y_offset,
1702                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1703                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1704                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1705                            int list0, int list1){
1706     MpegEncContext * const s = &h->s;
1707
1708     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1709     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1710     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1711     x_offset += 8*s->mb_x;
1712     y_offset += 8*(s->mb_y >> MB_FIELD);
1713
1714     if(list0 && list1){
1715         /* don't optimize for luma-only case, since B-frames usually
1716          * use implicit weights => chroma too. */
1717         uint8_t *tmp_cb = s->obmc_scratchpad;
1718         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1719         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1720         int refn0 = h->ref_cache[0][ scan8[n] ];
1721         int refn1 = h->ref_cache[1][ scan8[n] ];
1722
1723         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1724                     dest_y, dest_cb, dest_cr,
1725                     x_offset, y_offset, qpix_put, chroma_put);
1726         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1727                     tmp_y, tmp_cb, tmp_cr,
1728                     x_offset, y_offset, qpix_put, chroma_put);
1729
1730         if(h->use_weight == 2){
1731             int weight0 = h->implicit_weight[refn0][refn1];
1732             int weight1 = 64 - weight0;
1733             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1734             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1735             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1736         }else{
1737             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1738                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1739                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1740             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1741                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1742                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1743             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1744                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1745                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1746         }
1747     }else{
1748         int list = list1 ? 1 : 0;
1749         int refn = h->ref_cache[list][ scan8[n] ];
1750         Picture *ref= &h->ref_list[list][refn];
1751         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1752                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1753                     qpix_put, chroma_put);
1754
1755         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1756                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1757         if(h->use_weight_chroma){
1758             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1759                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1760             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1761                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1762         }
1763     }
1764 }
1765
1766 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1767                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1768                            int x_offset, int y_offset,
1769                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1770                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1771                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1772                            int list0, int list1){
1773     if((h->use_weight==2 && list0 && list1
1774         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1775        || h->use_weight==1)
1776         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1777                          x_offset, y_offset, qpix_put, chroma_put,
1778                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1779     else
1780         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1781                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1782 }
1783
1784 static inline void prefetch_motion(H264Context *h, int list){
1785     /* fetch pixels for estimated mv 4 macroblocks ahead
1786      * optimized for 64byte cache lines */
1787     MpegEncContext * const s = &h->s;
1788     const int refn = h->ref_cache[list][scan8[0]];
1789     if(refn >= 0){
1790         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1791         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1792         uint8_t **src= h->ref_list[list][refn].data;
1793         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1794         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1795         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1796         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1797     }
1798 }
1799
1800 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1801                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1802                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1803                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1804     MpegEncContext * const s = &h->s;
1805     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
1806     const int mb_type= s->current_picture.mb_type[mb_xy];
1807
1808     assert(IS_INTER(mb_type));
1809
1810     prefetch_motion(h, 0);
1811
1812     if(IS_16X16(mb_type)){
1813         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1814                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1815                 &weight_op[0], &weight_avg[0],
1816                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1817     }else if(IS_16X8(mb_type)){
1818         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1819                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1820                 &weight_op[1], &weight_avg[1],
1821                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1822         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1823                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1824                 &weight_op[1], &weight_avg[1],
1825                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1826     }else if(IS_8X16(mb_type)){
1827         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1828                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1829                 &weight_op[2], &weight_avg[2],
1830                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1831         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1832                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1833                 &weight_op[2], &weight_avg[2],
1834                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1835     }else{
1836         int i;
1837
1838         assert(IS_8X8(mb_type));
1839
1840         for(i=0; i<4; i++){
1841             const int sub_mb_type= h->sub_mb_type[i];
1842             const int n= 4*i;
1843             int x_offset= (i&1)<<2;
1844             int y_offset= (i&2)<<1;
1845
1846             if(IS_SUB_8X8(sub_mb_type)){
1847                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1848                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1849                     &weight_op[3], &weight_avg[3],
1850                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1851             }else if(IS_SUB_8X4(sub_mb_type)){
1852                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1853                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1854                     &weight_op[4], &weight_avg[4],
1855                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1856                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1857                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1858                     &weight_op[4], &weight_avg[4],
1859                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1860             }else if(IS_SUB_4X8(sub_mb_type)){
1861                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1862                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1863                     &weight_op[5], &weight_avg[5],
1864                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1865                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1866                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1867                     &weight_op[5], &weight_avg[5],
1868                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1869             }else{
1870                 int j;
1871                 assert(IS_SUB_4X4(sub_mb_type));
1872                 for(j=0; j<4; j++){
1873                     int sub_x_offset= x_offset + 2*(j&1);
1874                     int sub_y_offset= y_offset +   (j&2);
1875                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1876                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1877                         &weight_op[6], &weight_avg[6],
1878                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1879                 }
1880             }
1881         }
1882     }
1883
1884     prefetch_motion(h, 1);
1885 }
1886
1887 static void decode_init_vlc(void){
1888     static int done = 0;
1889
1890     if (!done) {
1891         int i;
1892         done = 1;
1893
1894         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1895                  &chroma_dc_coeff_token_len [0], 1, 1,
1896                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1897
1898         for(i=0; i<4; i++){
1899             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1900                      &coeff_token_len [i][0], 1, 1,
1901                      &coeff_token_bits[i][0], 1, 1, 1);
1902         }
1903
1904         for(i=0; i<3; i++){
1905             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1906                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1907                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1908         }
1909         for(i=0; i<15; i++){
1910             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1911                      &total_zeros_len [i][0], 1, 1,
1912                      &total_zeros_bits[i][0], 1, 1, 1);
1913         }
1914
1915         for(i=0; i<6; i++){
1916             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
1917                      &run_len [i][0], 1, 1,
1918                      &run_bits[i][0], 1, 1, 1);
1919         }
1920         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
1921                  &run_len [6][0], 1, 1,
1922                  &run_bits[6][0], 1, 1, 1);
1923     }
1924 }
1925
1926 static void free_tables(H264Context *h){
1927     int i;
1928     H264Context *hx;
1929     av_freep(&h->intra4x4_pred_mode);
1930     av_freep(&h->chroma_pred_mode_table);
1931     av_freep(&h->cbp_table);
1932     av_freep(&h->mvd_table[0]);
1933     av_freep(&h->mvd_table[1]);
1934     av_freep(&h->direct_table);
1935     av_freep(&h->non_zero_count);
1936     av_freep(&h->slice_table_base);
1937     h->slice_table= NULL;
1938
1939     av_freep(&h->mb2b_xy);
1940     av_freep(&h->mb2b8_xy);
1941
1942     for(i = 0; i < MAX_SPS_COUNT; i++)
1943         av_freep(h->sps_buffers + i);
1944
1945     for(i = 0; i < MAX_PPS_COUNT; i++)
1946         av_freep(h->pps_buffers + i);
1947
1948     for(i = 0; i < h->s.avctx->thread_count; i++) {
1949         hx = h->thread_context[i];
1950         if(!hx) continue;
1951         av_freep(&hx->top_borders[1]);
1952         av_freep(&hx->top_borders[0]);
1953         av_freep(&hx->s.obmc_scratchpad);
1954         av_freep(&hx->s.allocated_edge_emu_buffer);
1955     }
1956 }
1957
1958 static void init_dequant8_coeff_table(H264Context *h){
1959     int i,q,x;
1960     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
1961     h->dequant8_coeff[0] = h->dequant8_buffer[0];
1962     h->dequant8_coeff[1] = h->dequant8_buffer[1];
1963
1964     for(i=0; i<2; i++ ){
1965         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
1966             h->dequant8_coeff[1] = h->dequant8_buffer[0];
1967             break;
1968         }
1969
1970         for(q=0; q<52; q++){
1971             int shift = ff_div6[q];
1972             int idx = ff_rem6[q];
1973             for(x=0; x<64; x++)
1974                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
1975                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
1976                     h->pps.scaling_matrix8[i][x]) << shift;
1977         }
1978     }
1979 }
1980
1981 static void init_dequant4_coeff_table(H264Context *h){
1982     int i,j,q,x;
1983     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
1984     for(i=0; i<6; i++ ){
1985         h->dequant4_coeff[i] = h->dequant4_buffer[i];
1986         for(j=0; j<i; j++){
1987             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
1988                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
1989                 break;
1990             }
1991         }
1992         if(j<i)
1993             continue;
1994
1995         for(q=0; q<52; q++){
1996             int shift = ff_div6[q] + 2;
1997             int idx = ff_rem6[q];
1998             for(x=0; x<16; x++)
1999                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2000                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2001                     h->pps.scaling_matrix4[i][x]) << shift;
2002         }
2003     }
2004 }
2005
2006 static void init_dequant_tables(H264Context *h){
2007     int i,x;
2008     init_dequant4_coeff_table(h);
2009     if(h->pps.transform_8x8_mode)
2010         init_dequant8_coeff_table(h);
2011     if(h->sps.transform_bypass){
2012         for(i=0; i<6; i++)
2013             for(x=0; x<16; x++)
2014                 h->dequant4_coeff[i][0][x] = 1<<6;
2015         if(h->pps.transform_8x8_mode)
2016             for(i=0; i<2; i++)
2017                 for(x=0; x<64; x++)
2018                     h->dequant8_coeff[i][0][x] = 1<<6;
2019     }
2020 }
2021
2022
2023 /**
2024  * allocates tables.
2025  * needs width/height
2026  */
2027 static int alloc_tables(H264Context *h){
2028     MpegEncContext * const s = &h->s;
2029     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2030     int x,y;
2031
2032     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2033
2034     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2035     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2036     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2037
2038         CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2039         CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2040         CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2041         CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2042
2043     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2044     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2045
2046     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2047     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2048     for(y=0; y<s->mb_height; y++){
2049         for(x=0; x<s->mb_width; x++){
2050             const int mb_xy= x + y*s->mb_stride;
2051             const int b_xy = 4*x + 4*y*h->b_stride;
2052             const int b8_xy= 2*x + 2*y*h->b8_stride;
2053
2054             h->mb2b_xy [mb_xy]= b_xy;
2055             h->mb2b8_xy[mb_xy]= b8_xy;
2056         }
2057     }
2058
2059     s->obmc_scratchpad = NULL;
2060
2061     if(!h->dequant4_coeff[0])
2062         init_dequant_tables(h);
2063
2064     return 0;
2065 fail:
2066     free_tables(h);
2067     return -1;
2068 }
2069
2070 /**
2071  * Mimic alloc_tables(), but for every context thread.
2072  */
2073 static void clone_tables(H264Context *dst, H264Context *src){
2074     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2075     dst->non_zero_count           = src->non_zero_count;
2076     dst->slice_table              = src->slice_table;
2077     dst->cbp_table                = src->cbp_table;
2078     dst->mb2b_xy                  = src->mb2b_xy;
2079     dst->mb2b8_xy                 = src->mb2b8_xy;
2080     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2081     dst->mvd_table[0]             = src->mvd_table[0];
2082     dst->mvd_table[1]             = src->mvd_table[1];
2083     dst->direct_table             = src->direct_table;
2084
2085     dst->s.obmc_scratchpad = NULL;
2086     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2087 }
2088
2089 /**
2090  * Init context
2091  * Allocate buffers which are not shared amongst multiple threads.
2092  */
2093 static int context_init(H264Context *h){
2094     MpegEncContext * const s = &h->s;
2095
2096     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2097     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2098
2099     // edge emu needs blocksize + filter length - 1 (=17x17 for halfpel / 21x21 for h264)
2100     CHECKED_ALLOCZ(s->allocated_edge_emu_buffer,
2101                    (s->width+64)*2*21*2); //(width + edge + align)*interlaced*MBsize*tolerance
2102     s->edge_emu_buffer= s->allocated_edge_emu_buffer + (s->width+64)*2*21;
2103     return 0;
2104 fail:
2105     return -1; // free_tables will clean up for us
2106 }
2107
2108 static void common_init(H264Context *h){
2109     MpegEncContext * const s = &h->s;
2110
2111     s->width = s->avctx->width;
2112     s->height = s->avctx->height;
2113     s->codec_id= s->avctx->codec->id;
2114
2115     ff_h264_pred_init(&h->hpc, s->codec_id);
2116
2117     h->dequant_coeff_pps= -1;
2118     s->unrestricted_mv=1;
2119     s->decode=1; //FIXME
2120
2121     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2122     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2123 }
2124
2125 static int decode_init(AVCodecContext *avctx){
2126     H264Context *h= avctx->priv_data;
2127     MpegEncContext * const s = &h->s;
2128
2129     MPV_decode_defaults(s);
2130
2131     s->avctx = avctx;
2132     common_init(h);
2133
2134     s->out_format = FMT_H264;
2135     s->workaround_bugs= avctx->workaround_bugs;
2136
2137     // set defaults
2138 //    s->decode_mb= ff_h263_decode_mb;
2139     s->quarter_sample = 1;
2140     s->low_delay= 1;
2141     avctx->pix_fmt= PIX_FMT_YUV420P;
2142
2143     decode_init_vlc();
2144
2145     if(avctx->extradata_size > 0 && avctx->extradata &&
2146        *(char *)avctx->extradata == 1){
2147         h->is_avc = 1;
2148         h->got_avcC = 0;
2149     } else {
2150         h->is_avc = 0;
2151     }
2152
2153     h->thread_context[0] = h;
2154     return 0;
2155 }
2156
2157 static int frame_start(H264Context *h){
2158     MpegEncContext * const s = &h->s;
2159     int i;
2160
2161     if(MPV_frame_start(s, s->avctx) < 0)
2162         return -1;
2163     ff_er_frame_start(s);
2164     /*
2165      * MPV_frame_start uses pict_type to derive key_frame.
2166      * This is incorrect for H.264; IDR markings must be used.
2167      * Zero here; IDR markings per slice in frame or fields are OR'd in later.
2168      * See decode_nal_units().
2169      */
2170     s->current_picture_ptr->key_frame= 0;
2171
2172     assert(s->linesize && s->uvlinesize);
2173
2174     for(i=0; i<16; i++){
2175         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2176         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2177     }
2178     for(i=0; i<4; i++){
2179         h->block_offset[16+i]=
2180         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2181         h->block_offset[24+16+i]=
2182         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2183     }
2184
2185     /* can't be in alloc_tables because linesize isn't known there.
2186      * FIXME: redo bipred weight to not require extra buffer? */
2187     for(i = 0; i < s->avctx->thread_count; i++)
2188         if(!h->thread_context[i]->s.obmc_scratchpad)
2189             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2190
2191     /* some macroblocks will be accessed before they're available */
2192     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2193         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2194
2195 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2196     return 0;
2197 }
2198
2199 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){
2200     MpegEncContext * const s = &h->s;
2201     int i;
2202
2203     src_y  -=   linesize;
2204     src_cb -= uvlinesize;
2205     src_cr -= uvlinesize;
2206
2207     // There are two lines saved, the line above the the top macroblock of a pair,
2208     // and the line above the bottom macroblock
2209     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2210     for(i=1; i<17; i++){
2211         h->left_border[i]= src_y[15+i*  linesize];
2212     }
2213
2214     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2215     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2216
2217     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2218         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2219         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2220         for(i=1; i<9; i++){
2221             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2222             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2223         }
2224         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2225         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2226     }
2227 }
2228
2229 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){
2230     MpegEncContext * const s = &h->s;
2231     int temp8, i;
2232     uint64_t temp64;
2233     int deblock_left;
2234     int deblock_top;
2235     int mb_xy;
2236
2237     if(h->deblocking_filter == 2) {
2238         mb_xy = s->mb_x + s->mb_y*s->mb_stride;
2239         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2240         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2241     } else {
2242         deblock_left = (s->mb_x > 0);
2243         deblock_top =  (s->mb_y > 0);
2244     }
2245
2246     src_y  -=   linesize + 1;
2247     src_cb -= uvlinesize + 1;
2248     src_cr -= uvlinesize + 1;
2249
2250 #define XCHG(a,b,t,xchg)\
2251 t= a;\
2252 if(xchg)\
2253     a= b;\
2254 b= t;
2255
2256     if(deblock_left){
2257         for(i = !deblock_top; i<17; i++){
2258             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2259         }
2260     }
2261
2262     if(deblock_top){
2263         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2264         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2265         if(s->mb_x+1 < s->mb_width){
2266             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2267         }
2268     }
2269
2270     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2271         if(deblock_left){
2272             for(i = !deblock_top; i<9; i++){
2273                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2274                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2275             }
2276         }
2277         if(deblock_top){
2278             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2279             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2280         }
2281     }
2282 }
2283
2284 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2285     MpegEncContext * const s = &h->s;
2286     int i;
2287
2288     src_y  -= 2 *   linesize;
2289     src_cb -= 2 * uvlinesize;
2290     src_cr -= 2 * uvlinesize;
2291
2292     // There are two lines saved, the line above the the top macroblock of a pair,
2293     // and the line above the bottom macroblock
2294     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2295     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2296     for(i=2; i<34; i++){
2297         h->left_border[i]= src_y[15+i*  linesize];
2298     }
2299
2300     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2301     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2302     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2303     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2304
2305     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2306         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2307         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2308         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2309         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2310         for(i=2; i<18; i++){
2311             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2312             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2313         }
2314         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2315         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2316         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2317         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2318     }
2319 }
2320
2321 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){
2322     MpegEncContext * const s = &h->s;
2323     int temp8, i;
2324     uint64_t temp64;
2325     int deblock_left = (s->mb_x > 0);
2326     int deblock_top  = (s->mb_y > 1);
2327
2328     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);
2329
2330     src_y  -= 2 *   linesize + 1;
2331     src_cb -= 2 * uvlinesize + 1;
2332     src_cr -= 2 * uvlinesize + 1;
2333
2334 #define XCHG(a,b,t,xchg)\
2335 t= a;\
2336 if(xchg)\
2337     a= b;\
2338 b= t;
2339
2340     if(deblock_left){
2341         for(i = (!deblock_top)<<1; i<34; i++){
2342             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2343         }
2344     }
2345
2346     if(deblock_top){
2347         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2348         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2349         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2350         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2351         if(s->mb_x+1 < s->mb_width){
2352             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2353             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2354         }
2355     }
2356
2357     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2358         if(deblock_left){
2359             for(i = (!deblock_top) << 1; i<18; i++){
2360                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2361                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2362             }
2363         }
2364         if(deblock_top){
2365             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2366             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2367             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2368             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2369         }
2370     }
2371 }
2372
2373 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2374     MpegEncContext * const s = &h->s;
2375     const int mb_x= s->mb_x;
2376     const int mb_y= s->mb_y;
2377     const int mb_xy= mb_x + mb_y*s->mb_stride;
2378     const int mb_type= s->current_picture.mb_type[mb_xy];
2379     uint8_t  *dest_y, *dest_cb, *dest_cr;
2380     int linesize, uvlinesize /*dct_offset*/;
2381     int i;
2382     int *block_offset = &h->block_offset[0];
2383     const unsigned int bottom = mb_y & 1;
2384     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2385     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2386     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2387
2388     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2389     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2390     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2391
2392     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2393     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2394
2395     if (!simple && MB_FIELD) {
2396         linesize   = h->mb_linesize   = s->linesize * 2;
2397         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2398         block_offset = &h->block_offset[24];
2399         if(mb_y&1){ //FIXME move out of this func?
2400             dest_y -= s->linesize*15;
2401             dest_cb-= s->uvlinesize*7;
2402             dest_cr-= s->uvlinesize*7;
2403         }
2404         if(FRAME_MBAFF) {
2405             int list;
2406             for(list=0; list<h->list_count; list++){
2407                 if(!USES_LIST(mb_type, list))
2408                     continue;
2409                 if(IS_16X16(mb_type)){
2410                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2411                     fill_rectangle(ref, 4, 4, 8, 16+*ref^(s->mb_y&1), 1);
2412                 }else{
2413                     for(i=0; i<16; i+=4){
2414                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2415                         int ref = h->ref_cache[list][scan8[i]];
2416                         if(ref >= 0)
2417                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, 16+ref^(s->mb_y&1), 1);
2418                     }
2419                 }
2420             }
2421         }
2422     } else {
2423         linesize   = h->mb_linesize   = s->linesize;
2424         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2425 //        dct_offset = s->linesize * 16;
2426     }
2427
2428     if(transform_bypass){
2429         idct_dc_add =
2430         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2431     }else if(IS_8x8DCT(mb_type)){
2432         idct_dc_add = s->dsp.h264_idct8_dc_add;
2433         idct_add = s->dsp.h264_idct8_add;
2434     }else{
2435         idct_dc_add = s->dsp.h264_idct_dc_add;
2436         idct_add = s->dsp.h264_idct_add;
2437     }
2438
2439     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2440        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2441         int mbt_y = mb_y&~1;
2442         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2443         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2444         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2445         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2446     }
2447
2448     if (!simple && IS_INTRA_PCM(mb_type)) {
2449         unsigned int x, y;
2450
2451         // The pixels are stored in h->mb array in the same order as levels,
2452         // copy them in output in the correct order.
2453         for(i=0; i<16; i++) {
2454             for (y=0; y<4; y++) {
2455                 for (x=0; x<4; x++) {
2456                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2457                 }
2458             }
2459         }
2460         for(i=16; i<16+4; i++) {
2461             for (y=0; y<4; y++) {
2462                 for (x=0; x<4; x++) {
2463                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2464                 }
2465             }
2466         }
2467         for(i=20; i<20+4; i++) {
2468             for (y=0; y<4; y++) {
2469                 for (x=0; x<4; x++) {
2470                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2471                 }
2472             }
2473         }
2474     } else {
2475         if(IS_INTRA(mb_type)){
2476             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2477                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2478
2479             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2480                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2481                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2482             }
2483
2484             if(IS_INTRA4x4(mb_type)){
2485                 if(simple || !s->encoding){
2486                     if(IS_8x8DCT(mb_type)){
2487                         for(i=0; i<16; i+=4){
2488                             uint8_t * const ptr= dest_y + block_offset[i];
2489                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2490                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2491                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2492                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2493                             if(nnz){
2494                                 if(nnz == 1 && h->mb[i*16])
2495                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2496                                 else
2497                                     idct_add(ptr, h->mb + i*16, linesize);
2498                             }
2499                         }
2500                     }else
2501                     for(i=0; i<16; i++){
2502                         uint8_t * const ptr= dest_y + block_offset[i];
2503                         uint8_t *topright;
2504                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2505                         int nnz, tr;
2506
2507                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2508                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2509                             assert(mb_y || linesize <= block_offset[i]);
2510                             if(!topright_avail){
2511                                 tr= ptr[3 - linesize]*0x01010101;
2512                                 topright= (uint8_t*) &tr;
2513                             }else
2514                                 topright= ptr + 4 - linesize;
2515                         }else
2516                             topright= NULL;
2517
2518                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2519                         nnz = h->non_zero_count_cache[ scan8[i] ];
2520                         if(nnz){
2521                             if(is_h264){
2522                                 if(nnz == 1 && h->mb[i*16])
2523                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2524                                 else
2525                                     idct_add(ptr, h->mb + i*16, linesize);
2526                             }else
2527                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2528                         }
2529                     }
2530                 }
2531             }else{
2532                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2533                 if(is_h264){
2534                     if(!transform_bypass)
2535                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2536                 }else
2537                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2538             }
2539             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2540                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2541         }else if(is_h264){
2542             hl_motion(h, dest_y, dest_cb, dest_cr,
2543                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2544                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2545                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2546         }
2547
2548
2549         if(!IS_INTRA4x4(mb_type)){
2550             if(is_h264){
2551                 if(IS_INTRA16x16(mb_type)){
2552                     for(i=0; i<16; i++){
2553                         if(h->non_zero_count_cache[ scan8[i] ])
2554                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2555                         else if(h->mb[i*16])
2556                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2557                     }
2558                 }else{
2559                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2560                     for(i=0; i<16; i+=di){
2561                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2562                         if(nnz){
2563                             if(nnz==1 && h->mb[i*16])
2564                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2565                             else
2566                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2567                         }
2568                     }
2569                 }
2570             }else{
2571                 for(i=0; i<16; i++){
2572                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2573                         uint8_t * const ptr= dest_y + block_offset[i];
2574                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2575                     }
2576                 }
2577             }
2578         }
2579
2580         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2581             uint8_t *dest[2] = {dest_cb, dest_cr};
2582             if(transform_bypass){
2583                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2584             }else{
2585                 idct_add = s->dsp.h264_idct_add;
2586                 idct_dc_add = s->dsp.h264_idct_dc_add;
2587                 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]);
2588                 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]);
2589             }
2590             if(is_h264){
2591                 for(i=16; i<16+8; i++){
2592                     if(h->non_zero_count_cache[ scan8[i] ])
2593                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2594                     else if(h->mb[i*16])
2595                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2596                 }
2597             }else{
2598                 for(i=16; i<16+8; i++){
2599                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2600                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2601                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2602                     }
2603                 }
2604             }
2605         }
2606     }
2607     if(h->deblocking_filter) {
2608         if (!simple && FRAME_MBAFF) {
2609             //FIXME try deblocking one mb at a time?
2610             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2611             const int mb_y = s->mb_y - 1;
2612             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2613             const int mb_xy= mb_x + mb_y*s->mb_stride;
2614             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2615             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2616             if (!bottom) return;
2617             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2618             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2619             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2620
2621             if(IS_INTRA(mb_type_top | mb_type_bottom))
2622                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2623
2624             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2625             // deblock a pair
2626             // top
2627             s->mb_y--;
2628             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);
2629             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2630             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2631             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2632             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2633             // bottom
2634             s->mb_y++;
2635             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2636             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2637             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2638             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2639             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2640         } else {
2641             tprintf(h->s.avctx, "call filter_mb\n");
2642             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2643             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2644             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2645         }
2646     }
2647 }
2648
2649 /**
2650  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2651  */
2652 static void hl_decode_mb_simple(H264Context *h){
2653     hl_decode_mb_internal(h, 1);
2654 }
2655
2656 /**
2657  * Process a macroblock; this handles edge cases, such as interlacing.
2658  */
2659 static void av_noinline hl_decode_mb_complex(H264Context *h){
2660     hl_decode_mb_internal(h, 0);
2661 }
2662
2663 static void hl_decode_mb(H264Context *h){
2664     MpegEncContext * const s = &h->s;
2665     const int mb_x= s->mb_x;
2666     const int mb_y= s->mb_y;
2667     const int mb_xy= mb_x + mb_y*s->mb_stride;
2668     const int mb_type= s->current_picture.mb_type[mb_xy];
2669     int is_complex = FRAME_MBAFF || MB_FIELD || IS_INTRA_PCM(mb_type) || s->codec_id != CODEC_ID_H264 || (ENABLE_GRAY && (s->flags&CODEC_FLAG_GRAY)) || s->encoding;
2670
2671     if(!s->decode)
2672         return;
2673
2674     if (is_complex)
2675         hl_decode_mb_complex(h);
2676     else hl_decode_mb_simple(h);
2677 }
2678
2679 static void pic_as_field(Picture *pic, const int parity){
2680     int i;
2681     for (i = 0; i < 4; ++i) {
2682         if (parity == PICT_BOTTOM_FIELD)
2683             pic->data[i] += pic->linesize[i];
2684         pic->reference = parity;
2685         pic->linesize[i] *= 2;
2686     }
2687 }
2688
2689 static int split_field_copy(Picture *dest, Picture *src,
2690                             int parity, int id_add){
2691     int match = !!(src->reference & parity);
2692
2693     if (match) {
2694         *dest = *src;
2695         pic_as_field(dest, parity);
2696         dest->pic_id *= 2;
2697         dest->pic_id += id_add;
2698     }
2699
2700     return match;
2701 }
2702
2703 /**
2704  * Split one reference list into field parts, interleaving by parity
2705  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2706  * set to look at the actual start of data for that field.
2707  *
2708  * @param dest output list
2709  * @param dest_len maximum number of fields to put in dest
2710  * @param src the source reference list containing fields and/or field pairs
2711  *            (aka short_ref/long_ref, or
2712  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2713  * @param src_len number of Picture's in source (pairs and unmatched fields)
2714  * @param parity the parity of the picture being decoded/needing
2715  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2716  * @return number of fields placed in dest
2717  */
2718 static int split_field_half_ref_list(Picture *dest, int dest_len,
2719                                      Picture *src,  int src_len,  int parity){
2720     int same_parity   = 1;
2721     int same_i        = 0;
2722     int opp_i         = 0;
2723     int out_i;
2724     int field_output;
2725
2726     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2727         if (same_parity && same_i < src_len) {
2728             field_output = split_field_copy(dest + out_i, src + same_i,
2729                                             parity, 1);
2730             same_parity = !field_output;
2731             same_i++;
2732
2733         } else if (opp_i < src_len) {
2734             field_output = split_field_copy(dest + out_i, src + opp_i,
2735                                             PICT_FRAME - parity, 0);
2736             same_parity = field_output;
2737             opp_i++;
2738
2739         } else {
2740             break;
2741         }
2742     }
2743
2744     return out_i;
2745 }
2746
2747 /**
2748  * Split the reference frame list into a reference field list.
2749  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2750  * The input list contains both reference field pairs and
2751  * unmatched reference fields; it is ordered as spec describes
2752  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2753  * unmatched field pairs are also present. Conceptually this is equivalent
2754  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2755  *
2756  * @param dest output reference list where ordered fields are to be placed
2757  * @param dest_len max number of fields to place at dest
2758  * @param src source reference list, as described above
2759  * @param src_len number of pictures (pairs and unmatched fields) in src
2760  * @param parity parity of field being currently decoded
2761  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2762  * @param long_i index into src array that holds first long reference picture,
2763  *        or src_len if no long refs present.
2764  */
2765 static int split_field_ref_list(Picture *dest, int dest_len,
2766                                 Picture *src,  int src_len,
2767                                 int parity,    int long_i){
2768
2769     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2770     dest += i;
2771     dest_len -= i;
2772
2773     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2774                                    src_len - long_i, parity);
2775     return i;
2776 }
2777
2778 /**
2779  * fills the default_ref_list.
2780  */
2781 static int fill_default_ref_list(H264Context *h){
2782     MpegEncContext * const s = &h->s;
2783     int i;
2784     int smallest_poc_greater_than_current = -1;
2785     int structure_sel;
2786     Picture sorted_short_ref[32];
2787     Picture field_entry_list[2][32];
2788     Picture *frame_list[2];
2789
2790     if (FIELD_PICTURE) {
2791         structure_sel = PICT_FRAME;
2792         frame_list[0] = field_entry_list[0];
2793         frame_list[1] = field_entry_list[1];
2794     } else {
2795         structure_sel = 0;
2796         frame_list[0] = h->default_ref_list[0];
2797         frame_list[1] = h->default_ref_list[1];
2798     }
2799
2800     if(h->slice_type==B_TYPE){
2801         int list;
2802         int len[2];
2803         int short_len[2];
2804         int out_i;
2805         int limit= INT_MIN;
2806
2807         /* sort frame according to poc in B slice */
2808         for(out_i=0; out_i<h->short_ref_count; out_i++){
2809             int best_i=INT_MIN;
2810             int best_poc=INT_MAX;
2811
2812             for(i=0; i<h->short_ref_count; i++){
2813                 const int poc= h->short_ref[i]->poc;
2814                 if(poc > limit && poc < best_poc){
2815                     best_poc= poc;
2816                     best_i= i;
2817                 }
2818             }
2819
2820             assert(best_i != INT_MIN);
2821
2822             limit= best_poc;
2823             sorted_short_ref[out_i]= *h->short_ref[best_i];
2824             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);
2825             if (-1 == smallest_poc_greater_than_current) {
2826                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2827                     smallest_poc_greater_than_current = out_i;
2828                 }
2829             }
2830         }
2831
2832         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2833
2834         // find the largest poc
2835         for(list=0; list<2; list++){
2836             int index = 0;
2837             int j= -99;
2838             int step= list ? -1 : 1;
2839
2840             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2841                 int sel;
2842                 while(j<0 || j>= h->short_ref_count){
2843                     if(j != -99 && step == (list ? -1 : 1))
2844                         return -1;
2845                     step = -step;
2846                     j= smallest_poc_greater_than_current + (step>>1);
2847                 }
2848                 sel = sorted_short_ref[j].reference | structure_sel;
2849                 if(sel != PICT_FRAME) continue;
2850                 frame_list[list][index  ]= sorted_short_ref[j];
2851                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2852             }
2853             short_len[list] = index;
2854
2855             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2856                 int sel;
2857                 if(h->long_ref[i] == NULL) continue;
2858                 sel = h->long_ref[i]->reference | structure_sel;
2859                 if(sel != PICT_FRAME) continue;
2860
2861                 frame_list[ list ][index  ]= *h->long_ref[i];
2862                 frame_list[ list ][index++].pic_id= i;;
2863             }
2864             len[list] = index;
2865
2866             if(list && (smallest_poc_greater_than_current<=0 || smallest_poc_greater_than_current>=h->short_ref_count) && (1 < index)){
2867                 // swap the two first elements of L1 when
2868                 // L0 and L1 are identical
2869                 Picture temp= frame_list[1][0];
2870                 frame_list[1][0] = frame_list[1][1];
2871                 frame_list[1][1] = temp;
2872             }
2873
2874         }
2875
2876         for(list=0; list<2; list++){
2877             if (FIELD_PICTURE)
2878                 len[list] = split_field_ref_list(h->default_ref_list[list],
2879                                                  h->ref_count[list],
2880                                                  frame_list[list],
2881                                                  len[list],
2882                                                  s->picture_structure,
2883                                                  short_len[list]);
2884
2885             if(len[list] < h->ref_count[ list ])
2886                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2887         }
2888
2889
2890     }else{
2891         int index=0;
2892         int short_len;
2893         for(i=0; i<h->short_ref_count; i++){
2894             int sel;
2895             sel = h->short_ref[i]->reference | structure_sel;
2896             if(sel != PICT_FRAME) continue;
2897             frame_list[0][index  ]= *h->short_ref[i];
2898             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2899         }
2900         short_len = index;
2901         for(i = 0; i < 16; i++){
2902             int sel;
2903             if(h->long_ref[i] == NULL) continue;
2904             sel = h->long_ref[i]->reference | structure_sel;
2905             if(sel != PICT_FRAME) continue;
2906             frame_list[0][index  ]= *h->long_ref[i];
2907             frame_list[0][index++].pic_id= i;;
2908         }
2909
2910         if (FIELD_PICTURE)
2911             index = split_field_ref_list(h->default_ref_list[0],
2912                                          h->ref_count[0], frame_list[0],
2913                                          index, s->picture_structure,
2914                                          short_len);
2915
2916         if(index < h->ref_count[0])
2917             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
2918     }
2919 #ifdef TRACE
2920     for (i=0; i<h->ref_count[0]; i++) {
2921         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]);
2922     }
2923     if(h->slice_type==B_TYPE){
2924         for (i=0; i<h->ref_count[1]; i++) {
2925             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[0][i].data[0]);
2926         }
2927     }
2928 #endif
2929     return 0;
2930 }
2931
2932 static void print_short_term(H264Context *h);
2933 static void print_long_term(H264Context *h);
2934
2935 /**
2936  * Extract structure information about the picture described by pic_num in
2937  * the current decoding context (frame or field). Note that pic_num is
2938  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
2939  * @param pic_num picture number for which to extract structure information
2940  * @param structure one of PICT_XXX describing structure of picture
2941  *                      with pic_num
2942  * @return frame number (short term) or long term index of picture
2943  *         described by pic_num
2944  */
2945 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
2946     MpegEncContext * const s = &h->s;
2947
2948     *structure = s->picture_structure;
2949     if(FIELD_PICTURE){
2950         if (!(pic_num & 1))
2951             /* opposite field */
2952             *structure ^= PICT_FRAME;
2953         pic_num >>= 1;
2954     }
2955
2956     return pic_num;
2957 }
2958
2959 static int decode_ref_pic_list_reordering(H264Context *h){
2960     MpegEncContext * const s = &h->s;
2961     int list, index, pic_structure;
2962
2963     print_short_term(h);
2964     print_long_term(h);
2965     if(h->slice_type==I_TYPE || h->slice_type==SI_TYPE) return 0; //FIXME move before func
2966
2967     for(list=0; list<h->list_count; list++){
2968         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
2969
2970         if(get_bits1(&s->gb)){
2971             int pred= h->curr_pic_num;
2972
2973             for(index=0; ; index++){
2974                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
2975                 unsigned int pic_id;
2976                 int i;
2977                 Picture *ref = NULL;
2978
2979                 if(reordering_of_pic_nums_idc==3)
2980                     break;
2981
2982                 if(index >= h->ref_count[list]){
2983                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
2984                     return -1;
2985                 }
2986
2987                 if(reordering_of_pic_nums_idc<3){
2988                     if(reordering_of_pic_nums_idc<2){
2989                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
2990                         int frame_num;
2991
2992                         if(abs_diff_pic_num > h->max_pic_num){
2993                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
2994                             return -1;
2995                         }
2996
2997                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
2998                         else                                pred+= abs_diff_pic_num;
2999                         pred &= h->max_pic_num - 1;
3000
3001                         frame_num = pic_num_extract(h, pred, &pic_structure);
3002
3003                         for(i= h->short_ref_count-1; i>=0; i--){
3004                             ref = h->short_ref[i];
3005                             assert(ref->reference);
3006                             assert(!ref->long_ref);
3007                             if(ref->data[0] != NULL &&
3008                                    ref->frame_num == frame_num &&
3009                                    (ref->reference & pic_structure) &&
3010                                    ref->long_ref == 0) // ignore non existing pictures by testing data[0] pointer
3011                                 break;
3012                         }
3013                         if(i>=0)
3014                             ref->pic_id= pred;
3015                     }else{
3016                         int long_idx;
3017                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3018
3019                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3020
3021                         if(long_idx>31){
3022                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3023                             return -1;
3024                         }
3025                         ref = h->long_ref[long_idx];
3026                         assert(!(ref && !ref->reference));
3027                         if(ref && (ref->reference & pic_structure)){
3028                             ref->pic_id= pic_id;
3029                             assert(ref->long_ref);
3030                             i=0;
3031                         }else{
3032                             i=-1;
3033                         }
3034                     }
3035
3036                     if (i < 0) {
3037                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3038                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3039                     } else {
3040                         for(i=index; i+1<h->ref_count[list]; i++){
3041                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3042                                 break;
3043                         }
3044                         for(; i > index; i--){
3045                             h->ref_list[list][i]= h->ref_list[list][i-1];
3046                         }
3047                         h->ref_list[list][index]= *ref;
3048                         if (FIELD_PICTURE){
3049                             pic_as_field(&h->ref_list[list][index], pic_structure);
3050                         }
3051                     }
3052                 }else{
3053                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3054                     return -1;
3055                 }
3056             }
3057         }
3058     }
3059     for(list=0; list<h->list_count; list++){
3060         for(index= 0; index < h->ref_count[list]; index++){
3061             if(!h->ref_list[list][index].data[0])
3062                 h->ref_list[list][index]= s->current_picture;
3063         }
3064     }
3065
3066     if(h->slice_type==B_TYPE && !h->direct_spatial_mv_pred)
3067         direct_dist_scale_factor(h);
3068     direct_ref_list_init(h);
3069     return 0;
3070 }
3071
3072 static void fill_mbaff_ref_list(H264Context *h){
3073     int list, i, j;
3074     for(list=0; list<2; list++){ //FIXME try list_count
3075         for(i=0; i<h->ref_count[list]; i++){
3076             Picture *frame = &h->ref_list[list][i];
3077             Picture *field = &h->ref_list[list][16+2*i];
3078             field[0] = *frame;
3079             for(j=0; j<3; j++)
3080                 field[0].linesize[j] <<= 1;
3081             field[0].reference = PICT_TOP_FIELD;
3082             field[1] = field[0];
3083             for(j=0; j<3; j++)
3084                 field[1].data[j] += frame->linesize[j];
3085             field[1].reference = PICT_BOTTOM_FIELD;
3086
3087             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3088             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3089             for(j=0; j<2; j++){
3090                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3091                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3092             }
3093         }
3094     }
3095     for(j=0; j<h->ref_count[1]; j++){
3096         for(i=0; i<h->ref_count[0]; i++)
3097             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3098         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3099         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3100     }
3101 }
3102
3103 static int pred_weight_table(H264Context *h){
3104     MpegEncContext * const s = &h->s;
3105     int list, i;
3106     int luma_def, chroma_def;
3107
3108     h->use_weight= 0;
3109     h->use_weight_chroma= 0;
3110     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3111     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3112     luma_def = 1<<h->luma_log2_weight_denom;
3113     chroma_def = 1<<h->chroma_log2_weight_denom;
3114
3115     for(list=0; list<2; list++){
3116         for(i=0; i<h->ref_count[list]; i++){
3117             int luma_weight_flag, chroma_weight_flag;
3118
3119             luma_weight_flag= get_bits1(&s->gb);
3120             if(luma_weight_flag){
3121                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3122                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3123                 if(   h->luma_weight[list][i] != luma_def
3124                    || h->luma_offset[list][i] != 0)
3125                     h->use_weight= 1;
3126             }else{
3127                 h->luma_weight[list][i]= luma_def;
3128                 h->luma_offset[list][i]= 0;
3129             }
3130
3131             chroma_weight_flag= get_bits1(&s->gb);
3132             if(chroma_weight_flag){
3133                 int j;
3134                 for(j=0; j<2; j++){
3135                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3136                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3137                     if(   h->chroma_weight[list][i][j] != chroma_def
3138                        || h->chroma_offset[list][i][j] != 0)
3139                         h->use_weight_chroma= 1;
3140                 }
3141             }else{
3142                 int j;
3143                 for(j=0; j<2; j++){
3144                     h->chroma_weight[list][i][j]= chroma_def;
3145                     h->chroma_offset[list][i][j]= 0;
3146                 }
3147             }
3148         }
3149         if(h->slice_type != B_TYPE) break;
3150     }
3151     h->use_weight= h->use_weight || h->use_weight_chroma;
3152     return 0;
3153 }
3154
3155 static void implicit_weight_table(H264Context *h){
3156     MpegEncContext * const s = &h->s;
3157     int ref0, ref1;
3158     int cur_poc = s->current_picture_ptr->poc;
3159
3160     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3161        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3162         h->use_weight= 0;
3163         h->use_weight_chroma= 0;
3164         return;
3165     }
3166
3167     h->use_weight= 2;
3168     h->use_weight_chroma= 2;
3169     h->luma_log2_weight_denom= 5;
3170     h->chroma_log2_weight_denom= 5;
3171
3172     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3173         int poc0 = h->ref_list[0][ref0].poc;
3174         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3175             int poc1 = h->ref_list[1][ref1].poc;
3176             int td = av_clip(poc1 - poc0, -128, 127);
3177             if(td){
3178                 int tb = av_clip(cur_poc - poc0, -128, 127);
3179                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3180                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3181                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3182                     h->implicit_weight[ref0][ref1] = 32;
3183                 else
3184                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3185             }else
3186                 h->implicit_weight[ref0][ref1] = 32;
3187         }
3188     }
3189 }
3190
3191 /**
3192  * Mark a picture as no longer needed for reference. The refmask
3193  * argument allows unreferencing of individual fields or the whole frame.
3194  * If the picture becomes entirely unreferenced, but is being held for
3195  * display purposes, it is marked as such.
3196  * @param refmask mask of fields to unreference; the mask is bitwise
3197  *                anded with the reference marking of pic
3198  * @return non-zero if pic becomes entirely unreferenced (except possibly
3199  *         for display purposes) zero if one of the fields remains in
3200  *         reference
3201  */
3202 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3203     int i;
3204     if (pic->reference &= refmask) {
3205         return 0;
3206     } else {
3207         if(pic == h->delayed_output_pic)
3208             pic->reference=DELAYED_PIC_REF;
3209         else{
3210             for(i = 0; h->delayed_pic[i]; i++)
3211                 if(pic == h->delayed_pic[i]){
3212                     pic->reference=DELAYED_PIC_REF;
3213                     break;
3214                 }
3215         }
3216         return 1;
3217     }
3218 }
3219
3220 /**
3221  * instantaneous decoder refresh.
3222  */
3223 static void idr(H264Context *h){
3224     int i;
3225
3226     for(i=0; i<16; i++){
3227         if (h->long_ref[i] != NULL) {
3228             unreference_pic(h, h->long_ref[i], 0);
3229             h->long_ref[i]= NULL;
3230         }
3231     }
3232     h->long_ref_count=0;
3233
3234     for(i=0; i<h->short_ref_count; i++){
3235         unreference_pic(h, h->short_ref[i], 0);
3236         h->short_ref[i]= NULL;
3237     }
3238     h->short_ref_count=0;
3239 }
3240
3241 /* forget old pics after a seek */
3242 static void flush_dpb(AVCodecContext *avctx){
3243     H264Context *h= avctx->priv_data;
3244     int i;
3245     for(i=0; i<16; i++) {
3246         if(h->delayed_pic[i])
3247             h->delayed_pic[i]->reference= 0;
3248         h->delayed_pic[i]= NULL;
3249     }
3250     if(h->delayed_output_pic)
3251         h->delayed_output_pic->reference= 0;
3252     h->delayed_output_pic= NULL;
3253     idr(h);
3254     if(h->s.current_picture_ptr)
3255         h->s.current_picture_ptr->reference= 0;
3256     h->s.first_field= 0;
3257     ff_mpeg_flush(avctx);
3258 }
3259
3260 /**
3261  * Find a Picture in the short term reference list by frame number.
3262  * @param frame_num frame number to search for
3263  * @param idx the index into h->short_ref where returned picture is found
3264  *            undefined if no picture found.
3265  * @return pointer to the found picture, or NULL if no pic with the provided
3266  *                 frame number is found
3267  */
3268 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3269     MpegEncContext * const s = &h->s;
3270     int i;
3271
3272     for(i=0; i<h->short_ref_count; i++){
3273         Picture *pic= h->short_ref[i];
3274         if(s->avctx->debug&FF_DEBUG_MMCO)
3275             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3276         if(pic->frame_num == frame_num) {
3277             *idx = i;
3278             return pic;
3279         }
3280     }
3281     return NULL;
3282 }
3283
3284 /**
3285  * Remove a picture from the short term reference list by its index in
3286  * that list.  This does no checking on the provided index; it is assumed
3287  * to be valid. Other list entries are shifted down.
3288  * @param i index into h->short_ref of picture to remove.
3289  */
3290 static void remove_short_at_index(H264Context *h, int i){
3291     assert(i > 0 && i < h->short_ref_count);
3292     h->short_ref[i]= NULL;
3293     if (--h->short_ref_count)
3294         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3295 }
3296
3297 /**
3298  *
3299  * @return the removed picture or NULL if an error occurs
3300  */
3301 static Picture * remove_short(H264Context *h, int frame_num){
3302     MpegEncContext * const s = &h->s;
3303     Picture *pic;
3304     int i;
3305
3306     if(s->avctx->debug&FF_DEBUG_MMCO)
3307         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3308
3309     pic = find_short(h, frame_num, &i);
3310     if (pic)
3311         remove_short_at_index(h, i);
3312
3313     return pic;
3314 }
3315
3316 /**
3317  * Remove a picture from the long term reference list by its index in
3318  * that list.  This does no checking on the provided index; it is assumed
3319  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3320  * @param i index into h->long_ref of picture to remove.
3321  */
3322 static void remove_long_at_index(H264Context *h, int i){
3323     h->long_ref[i]= NULL;
3324     h->long_ref_count--;
3325 }
3326
3327 /**
3328  *
3329  * @return the removed picture or NULL if an error occurs
3330  */
3331 static Picture * remove_long(H264Context *h, int i){
3332     Picture *pic;
3333
3334     pic= h->long_ref[i];
3335     if (pic)
3336         remove_long_at_index(h, i);
3337
3338     return pic;
3339 }
3340
3341 /**
3342  * print short term list
3343  */
3344 static void print_short_term(H264Context *h) {
3345     uint32_t i;
3346     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3347         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3348         for(i=0; i<h->short_ref_count; i++){
3349             Picture *pic= h->short_ref[i];
3350             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3351         }
3352     }
3353 }
3354
3355 /**
3356  * print long term list
3357  */
3358 static void print_long_term(H264Context *h) {
3359     uint32_t i;
3360     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3361         av_log(h->s.avctx, AV_LOG_DEBUG, "long term list:\n");
3362         for(i = 0; i < 16; i++){
3363             Picture *pic= h->long_ref[i];
3364             if (pic) {
3365                 av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3366             }
3367         }
3368     }
3369 }
3370
3371 /**
3372  * Executes the reference picture marking (memory management control operations).
3373  */
3374 static int execute_ref_pic_marking(H264Context *h, MMCO *mmco, int mmco_count){
3375     MpegEncContext * const s = &h->s;
3376     int i, j;
3377     int current_ref_assigned=0;
3378     Picture *pic;
3379
3380     if((s->avctx->debug&FF_DEBUG_MMCO) && mmco_count==0)
3381         av_log(h->s.avctx, AV_LOG_DEBUG, "no mmco here\n");
3382
3383     for(i=0; i<mmco_count; i++){
3384         int structure, frame_num, unref_pic;
3385         if(s->avctx->debug&FF_DEBUG_MMCO)
3386             av_log(h->s.avctx, AV_LOG_DEBUG, "mmco:%d %d %d\n", h->mmco[i].opcode, h->mmco[i].short_pic_num, h->mmco[i].long_arg);
3387
3388         switch(mmco[i].opcode){
3389         case MMCO_SHORT2UNUSED:
3390             if(s->avctx->debug&FF_DEBUG_MMCO)
3391                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref short %d count %d\n", h->mmco[i].short_pic_num, h->short_ref_count);
3392             frame_num = pic_num_extract(h, mmco[i].short_pic_num, &structure);
3393             pic = find_short(h, frame_num, &j);
3394             if (pic) {
3395                 if (unreference_pic(h, pic, structure ^ PICT_FRAME))
3396                     remove_short_at_index(h, j);
3397             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3398                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref short failure\n");
3399             break;
3400         case MMCO_SHORT2LONG:
3401             if (FIELD_PICTURE && mmco[i].long_arg < h->long_ref_count &&
3402                     h->long_ref[mmco[i].long_arg]->frame_num ==
3403                                               mmco[i].short_pic_num / 2) {
3404                 /* do nothing, we've already moved this field pair. */
3405             } else {
3406                 int frame_num = mmco[i].short_pic_num >> FIELD_PICTURE;
3407
3408                 pic= remove_long(h, mmco[i].long_arg);
3409                 if(pic) unreference_pic(h, pic, 0);
3410
3411                 h->long_ref[ mmco[i].long_arg ]= remove_short(h, frame_num);
3412                 if (h->long_ref[ mmco[i].long_arg ]){
3413                     h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3414                     h->long_ref_count++;
3415                 }
3416             }
3417             break;
3418         case MMCO_LONG2UNUSED:
3419             j = pic_num_extract(h, mmco[i].long_arg, &structure);
3420             pic = h->long_ref[j];
3421             if (pic) {
3422                 if (unreference_pic(h, pic, structure ^ PICT_FRAME))
3423                     remove_long_at_index(h, j);
3424             } else if(s->avctx->debug&FF_DEBUG_MMCO)
3425                 av_log(h->s.avctx, AV_LOG_DEBUG, "mmco: unref long failure\n");
3426             break;
3427         case MMCO_LONG:
3428             unref_pic = 1;
3429             if (FIELD_PICTURE && !s->first_field) {
3430                 if (h->long_ref[mmco[i].long_arg] == s->current_picture_ptr) {
3431                     /* Just mark second field as referenced */
3432                     unref_pic = 0;
3433                 } else if (s->current_picture_ptr->reference) {
3434                     /* First field in pair is in short term list or
3435                      * at a different long term index.
3436                      * This is not allowed; see 7.4.3, notes 2 and 3.
3437                      * Report the problem and keep the pair where it is,
3438                      * and mark this field valid.
3439                      */
3440                     av_log(h->s.avctx, AV_LOG_ERROR,
3441                         "illegal long term reference assignment for second "
3442                         "field in complementary field pair (first field is "
3443                         "short term or has non-matching long index)\n");
3444                     unref_pic = 0;
3445                 }
3446             }
3447
3448             if (unref_pic) {
3449                 pic= remove_long(h, mmco[i].long_arg);
3450                 if(pic) unreference_pic(h, pic, 0);
3451
3452                 h->long_ref[ mmco[i].long_arg ]= s->current_picture_ptr;
3453                 h->long_ref[ mmco[i].long_arg ]->long_ref=1;
3454                 h->long_ref_count++;
3455             }
3456
3457             s->current_picture_ptr->reference |= s->picture_structure;
3458             current_ref_assigned=1;
3459             break;
3460         case MMCO_SET_MAX_LONG:
3461             assert(mmco[i].long_arg <= 16);
3462             // just remove the long term which index is greater than new max
3463             for(j = mmco[i].long_arg; j<16; j++){
3464                 pic = remove_long(h, j);
3465                 if (pic) unreference_pic(h, pic, 0);
3466             }
3467             break;
3468         case MMCO_RESET:
3469             while(h->short_ref_count){
3470                 pic= remove_short(h, h->short_ref[0]->frame_num);
3471                 if(pic) unreference_pic(h, pic, 0);
3472             }
3473             for(j = 0; j < 16; j++) {
3474                 pic= remove_long(h, j);
3475                 if(pic) unreference_pic(h, pic, 0);
3476             }
3477             break;
3478         default: assert(0);
3479         }
3480     }
3481
3482     if (!current_ref_assigned && FIELD_PICTURE &&
3483             !s->first_field && s->current_picture_ptr->reference) {
3484
3485         /* Second field of complementary field pair; the first field of
3486          * which is already referenced. If short referenced, it
3487          * should be first entry in short_ref. If not, it must exist
3488          * in long_ref; trying to put it on the short list here is an
3489          * error in the encoded bit stream (ref: 7.4.3, NOTE 2 and 3).
3490          */
3491         if (h->short_ref_count && h->short_ref[0] == s->current_picture_ptr) {
3492             /* Just mark the second field valid */
3493             s->current_picture_ptr->reference = PICT_FRAME;
3494         } else if (s->current_picture_ptr->long_ref) {
3495             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term reference "
3496                                              "assignment for second field "
3497                                              "in complementary field pair "
3498                                              "(first field is long term)\n");
3499         } else {
3500             /*
3501              * First field in reference, but not in any sensible place on our
3502              * reference lists. This shouldn't happen unless reference
3503              * handling somewhere else is wrong.
3504              */
3505             assert(0);
3506         }
3507         current_ref_assigned = 1;
3508     }
3509
3510     if(!current_ref_assigned){
3511         pic= remove_short(h, s->current_picture_ptr->frame_num);
3512         if(pic){
3513             unreference_pic(h, pic, 0);
3514             av_log(h->s.avctx, AV_LOG_ERROR, "illegal short term buffer state detected\n");
3515         }
3516
3517         if(h->short_ref_count)
3518             memmove(&h->short_ref[1], &h->short_ref[0], h->short_ref_count*sizeof(Picture*));
3519
3520         h->short_ref[0]= s->current_picture_ptr;
3521         h->short_ref[0]->long_ref=0;
3522         h->short_ref_count++;
3523         s->current_picture_ptr->reference |= s->picture_structure;
3524     }
3525
3526     if (h->long_ref_count + h->short_ref_count > h->sps.ref_frame_count){
3527
3528         /* We have too many reference frames, probably due to corrupted
3529          * stream. Need to discard one frame. Prevents overrun of the
3530          * short_ref and long_ref buffers.
3531          */
3532         av_log(h->s.avctx, AV_LOG_ERROR,
3533                "number of reference frames exceeds max (probably "
3534                "corrupt input), discarding one\n");
3535
3536         if (h->long_ref_count) {
3537             for (i = 0; i < 16; ++i)
3538                 if (h->long_ref[i])
3539                     break;
3540
3541             assert(i < 16);
3542             pic = h->long_ref[i];
3543             remove_long_at_index(h, i);
3544         } else {
3545             pic = h->short_ref[h->short_ref_count - 1];
3546             remove_short_at_index(h, h->short_ref_count - 1);
3547         }
3548         unreference_pic(h, pic, 0);
3549     }
3550
3551     print_short_term(h);
3552     print_long_term(h);
3553     return 0;
3554 }
3555
3556 static int decode_ref_pic_marking(H264Context *h, GetBitContext *gb){
3557     MpegEncContext * const s = &h->s;
3558     int i;
3559
3560     if(h->nal_unit_type == NAL_IDR_SLICE){ //FIXME fields
3561         s->broken_link= get_bits1(gb) -1;
3562         h->mmco[0].long_arg= get_bits1(gb) - 1; // current_long_term_idx
3563         if(h->mmco[0].long_arg == -1)
3564             h->mmco_index= 0;
3565         else{
3566             h->mmco[0].opcode= MMCO_LONG;
3567             h->mmco_index= 1;
3568         }
3569     }else{
3570         if(get_bits1(gb)){ // adaptive_ref_pic_marking_mode_flag
3571             for(i= 0; i<MAX_MMCO_COUNT; i++) {
3572                 MMCOOpcode opcode= get_ue_golomb(gb);
3573
3574                 h->mmco[i].opcode= opcode;
3575                 if(opcode==MMCO_SHORT2UNUSED || opcode==MMCO_SHORT2LONG){
3576                     h->mmco[i].short_pic_num= (h->curr_pic_num - get_ue_golomb(gb) - 1) & (h->max_pic_num - 1);
3577 /*                    if(h->mmco[i].short_pic_num >= h->short_ref_count || h->short_ref[ h->mmco[i].short_pic_num ] == NULL){
3578                         av_log(s->avctx, AV_LOG_ERROR, "illegal short ref in memory management control operation %d\n", mmco);
3579                         return -1;
3580                     }*/
3581                 }
3582                 if(opcode==MMCO_SHORT2LONG || opcode==MMCO_LONG2UNUSED || opcode==MMCO_LONG || opcode==MMCO_SET_MAX_LONG){
3583                     unsigned int long_arg= get_ue_golomb(gb);
3584                     if(long_arg >= 32 || (long_arg >= 16 && !(opcode == MMCO_LONG2UNUSED && FIELD_PICTURE))){
3585                         av_log(h->s.avctx, AV_LOG_ERROR, "illegal long ref in memory management control operation %d\n", opcode);
3586                         return -1;
3587                     }
3588                     h->mmco[i].long_arg= long_arg;
3589                 }
3590
3591                 if(opcode > (unsigned)MMCO_LONG){
3592                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal memory management control operation %d\n", opcode);
3593                     return -1;
3594                 }
3595                 if(opcode == MMCO_END)
3596                     break;
3597             }
3598             h->mmco_index= i;
3599         }else{
3600             assert(h->long_ref_count + h->short_ref_count <= h->sps.ref_frame_count);
3601
3602             if(h->short_ref_count && h->long_ref_count + h->short_ref_count == h->sps.ref_frame_count &&
3603                     !(FIELD_PICTURE && !s->first_field && s->current_picture_ptr->reference)) {
3604                 h->mmco[0].opcode= MMCO_SHORT2UNUSED;
3605                 h->mmco[0].short_pic_num= h->short_ref[ h->short_ref_count - 1 ]->frame_num;
3606                 h->mmco_index= 1;
3607                 if (FIELD_PICTURE) {
3608                     h->mmco[0].short_pic_num *= 2;
3609                     h->mmco[1].opcode= MMCO_SHORT2UNUSED;
3610                     h->mmco[1].short_pic_num= h->mmco[0].short_pic_num + 1;
3611                     h->mmco_index= 2;
3612                 }
3613             }else
3614                 h->mmco_index= 0;
3615         }
3616     }
3617
3618     return 0;
3619 }
3620
3621 static int init_poc(H264Context *h){
3622     MpegEncContext * const s = &h->s;
3623     const int max_frame_num= 1<<h->sps.log2_max_frame_num;
3624     int field_poc[2];
3625
3626     if(h->nal_unit_type == NAL_IDR_SLICE){
3627         h->frame_num_offset= 0;
3628     }else{
3629         if(h->frame_num < h->prev_frame_num)
3630             h->frame_num_offset= h->prev_frame_num_offset + max_frame_num;
3631         else
3632             h->frame_num_offset= h->prev_frame_num_offset;
3633     }
3634
3635     if(h->sps.poc_type==0){
3636         const int max_poc_lsb= 1<<h->sps.log2_max_poc_lsb;
3637
3638         if(h->nal_unit_type == NAL_IDR_SLICE){
3639              h->prev_poc_msb=
3640              h->prev_poc_lsb= 0;
3641         }
3642
3643         if     (h->poc_lsb < h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb >= max_poc_lsb/2)
3644             h->poc_msb = h->prev_poc_msb + max_poc_lsb;
3645         else if(h->poc_lsb > h->prev_poc_lsb && h->prev_poc_lsb - h->poc_lsb < -max_poc_lsb/2)
3646             h->poc_msb = h->prev_poc_msb - max_poc_lsb;
3647         else
3648             h->poc_msb = h->prev_poc_msb;
3649 //printf("poc: %d %d\n", h->poc_msb, h->poc_lsb);
3650         field_poc[0] =
3651         field_poc[1] = h->poc_msb + h->poc_lsb;
3652         if(s->picture_structure == PICT_FRAME)
3653             field_poc[1] += h->delta_poc_bottom;
3654     }else if(h->sps.poc_type==1){
3655         int abs_frame_num, expected_delta_per_poc_cycle, expectedpoc;
3656         int i;
3657
3658         if(h->sps.poc_cycle_length != 0)
3659             abs_frame_num = h->frame_num_offset + h->frame_num;
3660         else
3661             abs_frame_num = 0;
3662
3663         if(h->nal_ref_idc==0 && abs_frame_num > 0)
3664             abs_frame_num--;
3665
3666         expected_delta_per_poc_cycle = 0;
3667         for(i=0; i < h->sps.poc_cycle_length; i++)
3668             expected_delta_per_poc_cycle += h->sps.offset_for_ref_frame[ i ]; //FIXME integrate during sps parse
3669
3670         if(abs_frame_num > 0){
3671             int poc_cycle_cnt          = (abs_frame_num - 1) / h->sps.poc_cycle_length;
3672             int frame_num_in_poc_cycle = (abs_frame_num - 1) % h->sps.poc_cycle_length;
3673
3674             expectedpoc = poc_cycle_cnt * expected_delta_per_poc_cycle;
3675             for(i = 0; i <= frame_num_in_poc_cycle; i++)
3676                 expectedpoc = expectedpoc + h->sps.offset_for_ref_frame[ i ];
3677         } else
3678             expectedpoc = 0;
3679
3680         if(h->nal_ref_idc == 0)
3681             expectedpoc = expectedpoc + h->sps.offset_for_non_ref_pic;
3682
3683         field_poc[0] = expectedpoc + h->delta_poc[0];
3684         field_poc[1] = field_poc[0] + h->sps.offset_for_top_to_bottom_field;
3685
3686         if(s->picture_structure == PICT_FRAME)
3687             field_poc[1] += h->delta_poc[1];
3688     }else{
3689         int poc;
3690         if(h->nal_unit_type == NAL_IDR_SLICE){
3691             poc= 0;
3692         }else{
3693             if(h->nal_ref_idc) poc= 2*(h->frame_num_offset + h->frame_num);
3694             else               poc= 2*(h->frame_num_offset + h->frame_num) - 1;
3695         }
3696         field_poc[0]= poc;
3697         field_poc[1]= poc;
3698     }
3699
3700     if(s->picture_structure != PICT_BOTTOM_FIELD) {
3701         s->current_picture_ptr->field_poc[0]= field_poc[0];
3702         s->current_picture_ptr->poc = field_poc[0];
3703     }
3704     if(s->picture_structure != PICT_TOP_FIELD) {
3705         s->current_picture_ptr->field_poc[1]= field_poc[1];
3706         s->current_picture_ptr->poc = field_poc[1];
3707     }
3708     if(!FIELD_PICTURE || !s->first_field) {
3709         Picture *cur = s->current_picture_ptr;
3710         cur->poc= FFMIN(cur->field_poc[0], cur->field_poc[1]);
3711     }
3712
3713     return 0;
3714 }
3715
3716
3717 /**
3718  * initialize scan tables
3719  */
3720 static void init_scan_tables(H264Context *h){
3721     MpegEncContext * const s = &h->s;
3722     int i;
3723     if(s->dsp.h264_idct_add == ff_h264_idct_add_c){ //FIXME little ugly
3724         memcpy(h->zigzag_scan, zigzag_scan, 16*sizeof(uint8_t));
3725         memcpy(h-> field_scan,  field_scan, 16*sizeof(uint8_t));
3726     }else{
3727         for(i=0; i<16; i++){
3728 #define T(x) (x>>2) | ((x<<2) & 0xF)
3729             h->zigzag_scan[i] = T(zigzag_scan[i]);
3730             h-> field_scan[i] = T( field_scan[i]);
3731 #undef T
3732         }
3733     }
3734     if(s->dsp.h264_idct8_add == ff_h264_idct8_add_c){
3735         memcpy(h->zigzag_scan8x8,       zigzag_scan8x8,       64*sizeof(uint8_t));
3736         memcpy(h->zigzag_scan8x8_cavlc, zigzag_scan8x8_cavlc, 64*sizeof(uint8_t));
3737         memcpy(h->field_scan8x8,        field_scan8x8,        64*sizeof(uint8_t));
3738         memcpy(h->field_scan8x8_cavlc,  field_scan8x8_cavlc,  64*sizeof(uint8_t));
3739     }else{
3740         for(i=0; i<64; i++){
3741 #define T(x) (x>>3) | ((x&7)<<3)
3742             h->zigzag_scan8x8[i]       = T(zigzag_scan8x8[i]);
3743             h->zigzag_scan8x8_cavlc[i] = T(zigzag_scan8x8_cavlc[i]);
3744             h->field_scan8x8[i]        = T(field_scan8x8[i]);
3745             h->field_scan8x8_cavlc[i]  = T(field_scan8x8_cavlc[i]);
3746 #undef T
3747         }
3748     }
3749     if(h->sps.transform_bypass){ //FIXME same ugly
3750         h->zigzag_scan_q0          = zigzag_scan;
3751         h->zigzag_scan8x8_q0       = zigzag_scan8x8;
3752         h->zigzag_scan8x8_cavlc_q0 = zigzag_scan8x8_cavlc;
3753         h->field_scan_q0           = field_scan;
3754         h->field_scan8x8_q0        = field_scan8x8;
3755         h->field_scan8x8_cavlc_q0  = field_scan8x8_cavlc;
3756     }else{
3757         h->zigzag_scan_q0          = h->zigzag_scan;
3758         h->zigzag_scan8x8_q0       = h->zigzag_scan8x8;
3759         h->zigzag_scan8x8_cavlc_q0 = h->zigzag_scan8x8_cavlc;
3760         h->field_scan_q0           = h->field_scan;
3761         h->field_scan8x8_q0        = h->field_scan8x8;
3762         h->field_scan8x8_cavlc_q0  = h->field_scan8x8_cavlc;
3763     }
3764 }
3765
3766 /**
3767  * Replicates H264 "master" context to thread contexts.
3768  */
3769 static void clone_slice(H264Context *dst, H264Context *src)
3770 {
3771     memcpy(dst->block_offset,     src->block_offset, sizeof(dst->block_offset));
3772     dst->s.current_picture_ptr  = src->s.current_picture_ptr;
3773     dst->s.current_picture      = src->s.current_picture;
3774     dst->s.linesize             = src->s.linesize;
3775     dst->s.uvlinesize           = src->s.uvlinesize;
3776     dst->s.first_field          = src->s.first_field;
3777
3778     dst->prev_poc_msb           = src->prev_poc_msb;
3779     dst->prev_poc_lsb           = src->prev_poc_lsb;
3780     dst->prev_frame_num_offset  = src->prev_frame_num_offset;
3781     dst->prev_frame_num         = src->prev_frame_num;
3782     dst->short_ref_count        = src->short_ref_count;
3783
3784     memcpy(dst->short_ref,        src->short_ref,        sizeof(dst->short_ref));
3785     memcpy(dst->long_ref,         src->long_ref,         sizeof(dst->long_ref));
3786     memcpy(dst->default_ref_list, src->default_ref_list, sizeof(dst->default_ref_list));
3787     memcpy(dst->ref_list,         src->ref_list,         sizeof(dst->ref_list));
3788
3789     memcpy(dst->dequant4_coeff,   src->dequant4_coeff,   sizeof(src->dequant4_coeff));
3790     memcpy(dst->dequant8_coeff,   src->dequant8_coeff,   sizeof(src->dequant8_coeff));
3791 }
3792
3793 /**
3794  * decodes a slice header.
3795  * this will allso call MPV_common_init() and frame_start() as needed
3796  *
3797  * @param h h264context
3798  * @param h0 h264 master context (differs from 'h' when doing sliced based parallel decoding)
3799  *
3800  * @return 0 if okay, <0 if an error occured, 1 if decoding must not be multithreaded
3801  */
3802 static int decode_slice_header(H264Context *h, H264Context *h0){
3803     MpegEncContext * const s = &h->s;
3804     MpegEncContext * const s0 = &h0->s;
3805     unsigned int first_mb_in_slice;
3806     unsigned int pps_id;
3807     int num_ref_idx_active_override_flag;
3808     static const uint8_t slice_type_map[5]= {P_TYPE, B_TYPE, I_TYPE, SP_TYPE, SI_TYPE};
3809     unsigned int slice_type, tmp, i;
3810     int default_ref_list_done = 0;
3811     int last_pic_structure;
3812
3813     s->dropable= h->nal_ref_idc == 0;
3814
3815     if((s->avctx->flags2 & CODEC_FLAG2_FAST) && !h->nal_ref_idc){
3816         s->me.qpel_put= s->dsp.put_2tap_qpel_pixels_tab;
3817         s->me.qpel_avg= s->dsp.avg_2tap_qpel_pixels_tab;
3818     }else{
3819         s->me.qpel_put= s->dsp.put_h264_qpel_pixels_tab;
3820         s->me.qpel_avg= s->dsp.avg_h264_qpel_pixels_tab;
3821     }
3822
3823     first_mb_in_slice= get_ue_golomb(&s->gb);
3824
3825     if((s->flags2 & CODEC_FLAG2_CHUNKS) && first_mb_in_slice == 0){
3826         h0->current_slice = 0;
3827         if (!s0->first_field)
3828             s->current_picture_ptr= NULL;
3829     }
3830
3831     slice_type= get_ue_golomb(&s->gb);
3832     if(slice_type > 9){
3833         av_log(h->s.avctx, AV_LOG_ERROR, "slice type too large (%d) at %d %d\n", h->slice_type, s->mb_x, s->mb_y);
3834         return -1;
3835     }
3836     if(slice_type > 4){
3837         slice_type -= 5;
3838         h->slice_type_fixed=1;
3839     }else
3840         h->slice_type_fixed=0;
3841
3842     slice_type= slice_type_map[ slice_type ];
3843     if (slice_type == I_TYPE
3844         || (h0->current_slice != 0 && slice_type == h0->last_slice_type) ) {
3845         default_ref_list_done = 1;
3846     }
3847     h->slice_type= slice_type;
3848
3849     s->pict_type= h->slice_type; // to make a few old func happy, it's wrong though
3850     if (s->pict_type == B_TYPE && s0->last_picture_ptr == NULL) {
3851         av_log(h->s.avctx, AV_LOG_ERROR,
3852                "B picture before any references, skipping\n");
3853         return -1;
3854     }
3855
3856     pps_id= get_ue_golomb(&s->gb);
3857     if(pps_id>=MAX_PPS_COUNT){
3858         av_log(h->s.avctx, AV_LOG_ERROR, "pps_id out of range\n");
3859         return -1;
3860     }
3861     if(!h0->pps_buffers[pps_id]) {
3862         av_log(h->s.avctx, AV_LOG_ERROR, "non existing PPS referenced\n");
3863         return -1;
3864     }
3865     h->pps= *h0->pps_buffers[pps_id];
3866
3867     if(!h0->sps_buffers[h->pps.sps_id]) {
3868         av_log(h->s.avctx, AV_LOG_ERROR, "non existing SPS referenced\n");
3869         return -1;
3870     }
3871     h->sps = *h0->sps_buffers[h->pps.sps_id];
3872
3873     if(h == h0 && h->dequant_coeff_pps != pps_id){
3874         h->dequant_coeff_pps = pps_id;
3875         init_dequant_tables(h);
3876     }
3877
3878     s->mb_width= h->sps.mb_width;
3879     s->mb_height= h->sps.mb_height * (2 - h->sps.frame_mbs_only_flag);
3880
3881     h->b_stride=  s->mb_width*4;
3882     h->b8_stride= s->mb_width*2;
3883
3884     s->width = 16*s->mb_width - 2*(h->sps.crop_left + h->sps.crop_right );
3885     if(h->sps.frame_mbs_only_flag)
3886         s->height= 16*s->mb_height - 2*(h->sps.crop_top  + h->sps.crop_bottom);
3887     else
3888         s->height= 16*s->mb_height - 4*(h->sps.crop_top  + h->sps.crop_bottom); //FIXME recheck
3889
3890     if (s->context_initialized
3891         && (   s->width != s->avctx->width || s->height != s->avctx->height)) {
3892         if(h != h0)
3893             return -1;   // width / height changed during parallelized decoding
3894         free_tables(h);
3895         MPV_common_end(s);
3896     }
3897     if (!s->context_initialized) {
3898         if(h != h0)
3899             return -1;  // we cant (re-)initialize context during parallel decoding
3900         if (MPV_common_init(s) < 0)
3901             return -1;
3902         s->first_field = 0;
3903
3904         init_scan_tables(h);
3905         alloc_tables(h);
3906
3907         for(i = 1; i < s->avctx->thread_count; i++) {
3908             H264Context *c;
3909             c = h->thread_context[i] = av_malloc(sizeof(H264Context));
3910             memcpy(c, h, sizeof(MpegEncContext));
3911             memset(&c->s + 1, 0, sizeof(H264Context) - sizeof(MpegEncContext));
3912             c->sps = h->sps;
3913             c->pps = h->pps;
3914             init_scan_tables(c);
3915             clone_tables(c, h);
3916         }
3917
3918         for(i = 0; i < s->avctx->thread_count; i++)
3919             if(context_init(h->thread_context[i]) < 0)
3920                 return -1;
3921
3922         s->avctx->width = s->width;
3923         s->avctx->height = s->height;
3924         s->avctx->sample_aspect_ratio= h->sps.sar;
3925         if(!s->avctx->sample_aspect_ratio.den)
3926             s->avctx->sample_aspect_ratio.den = 1;
3927
3928         if(h->sps.timing_info_present_flag){
3929             s->avctx->time_base= (AVRational){h->sps.num_units_in_tick * 2, h->sps.time_scale};
3930             if(h->x264_build > 0 && h->x264_build < 44)
3931                 s->avctx->time_base.den *= 2;
3932             av_reduce(&s->avctx->time_base.num, &s->avctx->time_base.den,
3933                       s->avctx->time_base.num, s->avctx->time_base.den, 1<<30);
3934         }
3935     }
3936
3937     h->frame_num= get_bits(&s->gb, h->sps.log2_max_frame_num);
3938
3939     h->mb_mbaff = 0;
3940     h->mb_aff_frame = 0;
3941     last_pic_structure = s0->picture_structure;
3942     if(h->sps.frame_mbs_only_flag){
3943         s->picture_structure= PICT_FRAME;
3944     }else{
3945         if(get_bits1(&s->gb)) { //field_pic_flag
3946             s->picture_structure= PICT_TOP_FIELD + get_bits1(&s->gb); //bottom_field_flag
3947         } else {
3948             s->picture_structure= PICT_FRAME;
3949             h->mb_aff_frame = h->sps.mb_aff;
3950         }
3951     }
3952
3953     if(h0->current_slice == 0){
3954         /* See if we have a decoded first field looking for a pair... */
3955         if (s0->first_field) {
3956             assert(s0->current_picture_ptr);
3957             assert(s0->current_picture_ptr->data[0]);
3958             assert(s0->current_picture_ptr->reference != DELAYED_PIC_REF);
3959
3960             /* figure out if we have a complementary field pair */
3961             if (!FIELD_PICTURE || s->picture_structure == last_pic_structure) {
3962                 /*
3963                  * Previous field is unmatched. Don't display it, but let it
3964                  * remain for reference if marked as such.
3965                  */
3966                 s0->current_picture_ptr = NULL;
3967                 s0->first_field = FIELD_PICTURE;
3968
3969             } else {
3970                 if (h->nal_ref_idc &&
3971                         s0->current_picture_ptr->reference &&
3972                         s0->current_picture_ptr->frame_num != h->frame_num) {
3973                     /*
3974                      * This and previous field were reference, but had
3975                      * different frame_nums. Consider this field first in
3976                      * pair. Throw away previous field except for reference
3977                      * purposes.
3978                      */
3979                     s0->first_field = 1;
3980                     s0->current_picture_ptr = NULL;
3981
3982                 } else {
3983                     /* Second field in complementary pair */
3984                     s0->first_field = 0;
3985                 }
3986             }
3987
3988         } else {
3989             /* Frame or first field in a potentially complementary pair */
3990             assert(!s0->current_picture_ptr);
3991             s0->first_field = FIELD_PICTURE;
3992         }
3993
3994         if((!FIELD_PICTURE || s0->first_field) && frame_start(h) < 0) {
3995             s0->first_field = 0;
3996             return -1;
3997         }
3998     }
3999     if(h != h0)
4000         clone_slice(h, h0);
4001
4002     s->current_picture_ptr->frame_num= h->frame_num; //FIXME frame_num cleanup
4003
4004     assert(s->mb_num == s->mb_width * s->mb_height);
4005     if(first_mb_in_slice << FIELD_OR_MBAFF_PICTURE >= s->mb_num ||
4006        first_mb_in_slice                    >= s->mb_num){
4007         av_log(h->s.avctx, AV_LOG_ERROR, "first_mb_in_slice overflow\n");
4008         return -1;
4009     }
4010     s->resync_mb_x = s->mb_x = first_mb_in_slice % s->mb_width;
4011     s->resync_mb_y = s->mb_y = (first_mb_in_slice / s->mb_width) << FIELD_OR_MBAFF_PICTURE;
4012     if (s->picture_structure == PICT_BOTTOM_FIELD)
4013         s->resync_mb_y = s->mb_y = s->mb_y + 1;
4014     assert(s->mb_y < s->mb_height);
4015
4016     if(s->picture_structure==PICT_FRAME){
4017         h->curr_pic_num=   h->frame_num;
4018         h->max_pic_num= 1<< h->sps.log2_max_frame_num;
4019     }else{
4020         h->curr_pic_num= 2*h->frame_num + 1;
4021         h->max_pic_num= 1<<(h->sps.log2_max_frame_num + 1);
4022     }
4023