Modifies macroblock addressing and current macroblock y-position for field decoding.
[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
36 #include "cabac.h"
37
38 //#undef NDEBUG
39 #include <assert.h>
40
41 /**
42  * Value of Picture.reference when Picture is not a reference picture, but
43  * is held for delayed output.
44  */
45 #define DELAYED_PIC_REF 4
46
47 static VLC coeff_token_vlc[4];
48 static VLC chroma_dc_coeff_token_vlc;
49
50 static VLC total_zeros_vlc[15];
51 static VLC chroma_dc_total_zeros_vlc[3];
52
53 static VLC run_vlc[6];
54 static VLC run7_vlc;
55
56 static void svq3_luma_dc_dequant_idct_c(DCTELEM *block, int qp);
57 static void svq3_add_idct_c(uint8_t *dst, DCTELEM *block, int stride, int qp, int dc);
58 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);
59 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);
60
61 static av_always_inline uint32_t pack16to32(int a, int b){
62 #ifdef WORDS_BIGENDIAN
63    return (b&0xFFFF) + (a<<16);
64 #else
65    return (a&0xFFFF) + (b<<16);
66 #endif
67 }
68
69 const uint8_t ff_rem6[52]={
70 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,
71 };
72
73 const uint8_t ff_div6[52]={
74 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,
75 };
76
77
78 /**
79  * fill a rectangle.
80  * @param h height of the rectangle, should be a constant
81  * @param w width of the rectangle, should be a constant
82  * @param size the size of val (1 or 4), should be a constant
83  */
84 static av_always_inline void fill_rectangle(void *vp, int w, int h, int stride, uint32_t val, int size){
85     uint8_t *p= (uint8_t*)vp;
86     assert(size==1 || size==4);
87     assert(w<=4);
88
89     w      *= size;
90     stride *= size;
91
92     assert((((long)vp)&(FFMIN(w, STRIDE_ALIGN)-1)) == 0);
93     assert((stride&(w-1))==0);
94     if(w==2){
95         const uint16_t v= size==4 ? val : val*0x0101;
96         *(uint16_t*)(p + 0*stride)= v;
97         if(h==1) return;
98         *(uint16_t*)(p + 1*stride)= v;
99         if(h==2) return;
100         *(uint16_t*)(p + 2*stride)= v;
101         *(uint16_t*)(p + 3*stride)= v;
102     }else if(w==4){
103         const uint32_t v= size==4 ? val : val*0x01010101;
104         *(uint32_t*)(p + 0*stride)= v;
105         if(h==1) return;
106         *(uint32_t*)(p + 1*stride)= v;
107         if(h==2) return;
108         *(uint32_t*)(p + 2*stride)= v;
109         *(uint32_t*)(p + 3*stride)= v;
110     }else if(w==8){
111     //gcc can't optimize 64bit math on x86_32
112 #if defined(ARCH_X86_64) || (defined(MP_WORDSIZE) && MP_WORDSIZE >= 64)
113         const uint64_t v= val*0x0100000001ULL;
114         *(uint64_t*)(p + 0*stride)= v;
115         if(h==1) return;
116         *(uint64_t*)(p + 1*stride)= v;
117         if(h==2) return;
118         *(uint64_t*)(p + 2*stride)= v;
119         *(uint64_t*)(p + 3*stride)= v;
120     }else if(w==16){
121         const uint64_t v= val*0x0100000001ULL;
122         *(uint64_t*)(p + 0+0*stride)= v;
123         *(uint64_t*)(p + 8+0*stride)= v;
124         *(uint64_t*)(p + 0+1*stride)= v;
125         *(uint64_t*)(p + 8+1*stride)= v;
126         if(h==2) return;
127         *(uint64_t*)(p + 0+2*stride)= v;
128         *(uint64_t*)(p + 8+2*stride)= v;
129         *(uint64_t*)(p + 0+3*stride)= v;
130         *(uint64_t*)(p + 8+3*stride)= v;
131 #else
132         *(uint32_t*)(p + 0+0*stride)= val;
133         *(uint32_t*)(p + 4+0*stride)= val;
134         if(h==1) return;
135         *(uint32_t*)(p + 0+1*stride)= val;
136         *(uint32_t*)(p + 4+1*stride)= val;
137         if(h==2) return;
138         *(uint32_t*)(p + 0+2*stride)= val;
139         *(uint32_t*)(p + 4+2*stride)= val;
140         *(uint32_t*)(p + 0+3*stride)= val;
141         *(uint32_t*)(p + 4+3*stride)= val;
142     }else if(w==16){
143         *(uint32_t*)(p + 0+0*stride)= val;
144         *(uint32_t*)(p + 4+0*stride)= val;
145         *(uint32_t*)(p + 8+0*stride)= val;
146         *(uint32_t*)(p +12+0*stride)= val;
147         *(uint32_t*)(p + 0+1*stride)= val;
148         *(uint32_t*)(p + 4+1*stride)= val;
149         *(uint32_t*)(p + 8+1*stride)= val;
150         *(uint32_t*)(p +12+1*stride)= val;
151         if(h==2) return;
152         *(uint32_t*)(p + 0+2*stride)= val;
153         *(uint32_t*)(p + 4+2*stride)= val;
154         *(uint32_t*)(p + 8+2*stride)= val;
155         *(uint32_t*)(p +12+2*stride)= val;
156         *(uint32_t*)(p + 0+3*stride)= val;
157         *(uint32_t*)(p + 4+3*stride)= val;
158         *(uint32_t*)(p + 8+3*stride)= val;
159         *(uint32_t*)(p +12+3*stride)= val;
160 #endif
161     }else
162         assert(0);
163     assert(h==4);
164 }
165
166 static void fill_caches(H264Context *h, int mb_type, int for_deblock){
167     MpegEncContext * const s = &h->s;
168     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
169     int topleft_xy, top_xy, topright_xy, left_xy[2];
170     int topleft_type, top_type, topright_type, left_type[2];
171     int left_block[8];
172     int i;
173
174     //FIXME deblocking could skip the intra and nnz parts.
175     if(for_deblock && (h->slice_num == 1 || h->slice_table[mb_xy] == h->slice_table[mb_xy-s->mb_stride]) && !FRAME_MBAFF)
176         return;
177
178     //wow what a mess, why didn't they simplify the interlacing&intra stuff, i can't imagine that these complex rules are worth it
179
180     top_xy     = mb_xy  - (s->mb_stride << FIELD_PICTURE);
181     topleft_xy = top_xy - 1;
182     topright_xy= top_xy + 1;
183     left_xy[1] = left_xy[0] = mb_xy-1;
184     left_block[0]= 0;
185     left_block[1]= 1;
186     left_block[2]= 2;
187     left_block[3]= 3;
188     left_block[4]= 7;
189     left_block[5]= 10;
190     left_block[6]= 8;
191     left_block[7]= 11;
192     if(FRAME_MBAFF){
193         const int pair_xy          = s->mb_x     + (s->mb_y & ~1)*s->mb_stride;
194         const int top_pair_xy      = pair_xy     - s->mb_stride;
195         const int topleft_pair_xy  = top_pair_xy - 1;
196         const int topright_pair_xy = top_pair_xy + 1;
197         const int topleft_mb_frame_flag  = !IS_INTERLACED(s->current_picture.mb_type[topleft_pair_xy]);
198         const int top_mb_frame_flag      = !IS_INTERLACED(s->current_picture.mb_type[top_pair_xy]);
199         const int topright_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[topright_pair_xy]);
200         const int left_mb_frame_flag = !IS_INTERLACED(s->current_picture.mb_type[pair_xy-1]);
201         const int curr_mb_frame_flag = !IS_INTERLACED(mb_type);
202         const int bottom = (s->mb_y & 1);
203         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);
204         if (bottom
205                 ? !curr_mb_frame_flag // bottom macroblock
206                 : (!curr_mb_frame_flag && !top_mb_frame_flag) // top macroblock
207                 ) {
208             top_xy -= s->mb_stride;
209         }
210         if (bottom
211                 ? !curr_mb_frame_flag // bottom macroblock
212                 : (!curr_mb_frame_flag && !topleft_mb_frame_flag) // top macroblock
213                 ) {
214             topleft_xy -= s->mb_stride;
215         }
216         if (bottom
217                 ? !curr_mb_frame_flag // bottom macroblock
218                 : (!curr_mb_frame_flag && !topright_mb_frame_flag) // top macroblock
219                 ) {
220             topright_xy -= s->mb_stride;
221         }
222         if (left_mb_frame_flag != curr_mb_frame_flag) {
223             left_xy[1] = left_xy[0] = pair_xy - 1;
224             if (curr_mb_frame_flag) {
225                 if (bottom) {
226                     left_block[0]= 2;
227                     left_block[1]= 2;
228                     left_block[2]= 3;
229                     left_block[3]= 3;
230                     left_block[4]= 8;
231                     left_block[5]= 11;
232                     left_block[6]= 8;
233                     left_block[7]= 11;
234                 } else {
235                     left_block[0]= 0;
236                     left_block[1]= 0;
237                     left_block[2]= 1;
238                     left_block[3]= 1;
239                     left_block[4]= 7;
240                     left_block[5]= 10;
241                     left_block[6]= 7;
242                     left_block[7]= 10;
243                 }
244             } else {
245                 left_xy[1] += s->mb_stride;
246                 //left_block[0]= 0;
247                 left_block[1]= 2;
248                 left_block[2]= 0;
249                 left_block[3]= 2;
250                 //left_block[4]= 7;
251                 left_block[5]= 10;
252                 left_block[6]= 7;
253                 left_block[7]= 10;
254             }
255         }
256     }
257
258     h->top_mb_xy = top_xy;
259     h->left_mb_xy[0] = left_xy[0];
260     h->left_mb_xy[1] = left_xy[1];
261     if(for_deblock){
262         topleft_type = 0;
263         topright_type = 0;
264         top_type     = h->slice_table[top_xy     ] < 255 ? s->current_picture.mb_type[top_xy]     : 0;
265         left_type[0] = h->slice_table[left_xy[0] ] < 255 ? s->current_picture.mb_type[left_xy[0]] : 0;
266         left_type[1] = h->slice_table[left_xy[1] ] < 255 ? s->current_picture.mb_type[left_xy[1]] : 0;
267
268         if(FRAME_MBAFF && !IS_INTRA(mb_type)){
269             int list;
270             int v = *(uint16_t*)&h->non_zero_count[mb_xy][14];
271             for(i=0; i<16; i++)
272                 h->non_zero_count_cache[scan8[i]] = (v>>i)&1;
273             for(list=0; list<h->list_count; list++){
274                 if(USES_LIST(mb_type,list)){
275                     uint32_t *src = (uint32_t*)s->current_picture.motion_val[list][h->mb2b_xy[mb_xy]];
276                     uint32_t *dst = (uint32_t*)h->mv_cache[list][scan8[0]];
277                     int8_t *ref = &s->current_picture.ref_index[list][h->mb2b8_xy[mb_xy]];
278                     for(i=0; i<4; i++, dst+=8, src+=h->b_stride){
279                         dst[0] = src[0];
280                         dst[1] = src[1];
281                         dst[2] = src[2];
282                         dst[3] = src[3];
283                     }
284                     *(uint32_t*)&h->ref_cache[list][scan8[ 0]] =
285                     *(uint32_t*)&h->ref_cache[list][scan8[ 2]] = pack16to32(ref[0],ref[1])*0x0101;
286                     ref += h->b8_stride;
287                     *(uint32_t*)&h->ref_cache[list][scan8[ 8]] =
288                     *(uint32_t*)&h->ref_cache[list][scan8[10]] = pack16to32(ref[0],ref[1])*0x0101;
289                 }else{
290                     fill_rectangle(&h-> mv_cache[list][scan8[ 0]], 4, 4, 8, 0, 4);
291                     fill_rectangle(&h->ref_cache[list][scan8[ 0]], 4, 4, 8, (uint8_t)LIST_NOT_USED, 1);
292                 }
293             }
294         }
295     }else{
296         topleft_type = h->slice_table[topleft_xy ] == h->slice_num ? s->current_picture.mb_type[topleft_xy] : 0;
297         top_type     = h->slice_table[top_xy     ] == h->slice_num ? s->current_picture.mb_type[top_xy]     : 0;
298         topright_type= h->slice_table[topright_xy] == h->slice_num ? s->current_picture.mb_type[topright_xy]: 0;
299         left_type[0] = h->slice_table[left_xy[0] ] == h->slice_num ? s->current_picture.mb_type[left_xy[0]] : 0;
300         left_type[1] = h->slice_table[left_xy[1] ] == h->slice_num ? s->current_picture.mb_type[left_xy[1]] : 0;
301     }
302
303     if(IS_INTRA(mb_type)){
304         h->topleft_samples_available=
305         h->top_samples_available=
306         h->left_samples_available= 0xFFFF;
307         h->topright_samples_available= 0xEEEA;
308
309         if(!IS_INTRA(top_type) && (top_type==0 || h->pps.constrained_intra_pred)){
310             h->topleft_samples_available= 0xB3FF;
311             h->top_samples_available= 0x33FF;
312             h->topright_samples_available= 0x26EA;
313         }
314         for(i=0; i<2; i++){
315             if(!IS_INTRA(left_type[i]) && (left_type[i]==0 || h->pps.constrained_intra_pred)){
316                 h->topleft_samples_available&= 0xDF5F;
317                 h->left_samples_available&= 0x5F5F;
318             }
319         }
320
321         if(!IS_INTRA(topleft_type) && (topleft_type==0 || h->pps.constrained_intra_pred))
322             h->topleft_samples_available&= 0x7FFF;
323
324         if(!IS_INTRA(topright_type) && (topright_type==0 || h->pps.constrained_intra_pred))
325             h->topright_samples_available&= 0xFBFF;
326
327         if(IS_INTRA4x4(mb_type)){
328             if(IS_INTRA4x4(top_type)){
329                 h->intra4x4_pred_mode_cache[4+8*0]= h->intra4x4_pred_mode[top_xy][4];
330                 h->intra4x4_pred_mode_cache[5+8*0]= h->intra4x4_pred_mode[top_xy][5];
331                 h->intra4x4_pred_mode_cache[6+8*0]= h->intra4x4_pred_mode[top_xy][6];
332                 h->intra4x4_pred_mode_cache[7+8*0]= h->intra4x4_pred_mode[top_xy][3];
333             }else{
334                 int pred;
335                 if(!top_type || (IS_INTER(top_type) && h->pps.constrained_intra_pred))
336                     pred= -1;
337                 else{
338                     pred= 2;
339                 }
340                 h->intra4x4_pred_mode_cache[4+8*0]=
341                 h->intra4x4_pred_mode_cache[5+8*0]=
342                 h->intra4x4_pred_mode_cache[6+8*0]=
343                 h->intra4x4_pred_mode_cache[7+8*0]= pred;
344             }
345             for(i=0; i<2; i++){
346                 if(IS_INTRA4x4(left_type[i])){
347                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[0+2*i]];
348                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= h->intra4x4_pred_mode[left_xy[i]][left_block[1+2*i]];
349                 }else{
350                     int pred;
351                     if(!left_type[i] || (IS_INTER(left_type[i]) && h->pps.constrained_intra_pred))
352                         pred= -1;
353                     else{
354                         pred= 2;
355                     }
356                     h->intra4x4_pred_mode_cache[3+8*1 + 2*8*i]=
357                     h->intra4x4_pred_mode_cache[3+8*2 + 2*8*i]= pred;
358                 }
359             }
360         }
361     }
362
363
364 /*
365 0 . T T. T T T T
366 1 L . .L . . . .
367 2 L . .L . . . .
368 3 . T TL . . . .
369 4 L . .L . . . .
370 5 L . .. . . . .
371 */
372 //FIXME constraint_intra_pred & partitioning & nnz (lets hope this is just a typo in the spec)
373     if(top_type){
374         h->non_zero_count_cache[4+8*0]= h->non_zero_count[top_xy][4];
375         h->non_zero_count_cache[5+8*0]= h->non_zero_count[top_xy][5];
376         h->non_zero_count_cache[6+8*0]= h->non_zero_count[top_xy][6];
377         h->non_zero_count_cache[7+8*0]= h->non_zero_count[top_xy][3];
378
379         h->non_zero_count_cache[1+8*0]= h->non_zero_count[top_xy][9];
380         h->non_zero_count_cache[2+8*0]= h->non_zero_count[top_xy][8];
381
382         h->non_zero_count_cache[1+8*3]= h->non_zero_count[top_xy][12];
383         h->non_zero_count_cache[2+8*3]= h->non_zero_count[top_xy][11];
384
385     }else{
386         h->non_zero_count_cache[4+8*0]=
387         h->non_zero_count_cache[5+8*0]=
388         h->non_zero_count_cache[6+8*0]=
389         h->non_zero_count_cache[7+8*0]=
390
391         h->non_zero_count_cache[1+8*0]=
392         h->non_zero_count_cache[2+8*0]=
393
394         h->non_zero_count_cache[1+8*3]=
395         h->non_zero_count_cache[2+8*3]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
396
397     }
398
399     for (i=0; i<2; i++) {
400         if(left_type[i]){
401             h->non_zero_count_cache[3+8*1 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[0+2*i]];
402             h->non_zero_count_cache[3+8*2 + 2*8*i]= h->non_zero_count[left_xy[i]][left_block[1+2*i]];
403             h->non_zero_count_cache[0+8*1 +   8*i]= h->non_zero_count[left_xy[i]][left_block[4+2*i]];
404             h->non_zero_count_cache[0+8*4 +   8*i]= h->non_zero_count[left_xy[i]][left_block[5+2*i]];
405         }else{
406             h->non_zero_count_cache[3+8*1 + 2*8*i]=
407             h->non_zero_count_cache[3+8*2 + 2*8*i]=
408             h->non_zero_count_cache[0+8*1 +   8*i]=
409             h->non_zero_count_cache[0+8*4 +   8*i]= h->pps.cabac && !IS_INTRA(mb_type) ? 0 : 64;
410         }
411     }
412
413     if( h->pps.cabac ) {
414         // top_cbp
415         if(top_type) {
416             h->top_cbp = h->cbp_table[top_xy];
417         } else if(IS_INTRA(mb_type)) {
418             h->top_cbp = 0x1C0;
419         } else {
420             h->top_cbp = 0;
421         }
422         // left_cbp
423         if (left_type[0]) {
424             h->left_cbp = h->cbp_table[left_xy[0]] & 0x1f0;
425         } else if(IS_INTRA(mb_type)) {
426             h->left_cbp = 0x1C0;
427         } else {
428             h->left_cbp = 0;
429         }
430         if (left_type[0]) {
431             h->left_cbp |= ((h->cbp_table[left_xy[0]]>>((left_block[0]&(~1))+1))&0x1) << 1;
432         }
433         if (left_type[1]) {
434             h->left_cbp |= ((h->cbp_table[left_xy[1]]>>((left_block[2]&(~1))+1))&0x1) << 3;
435         }
436     }
437
438 #if 1
439     if(IS_INTER(mb_type) || IS_DIRECT(mb_type)){
440         int list;
441         for(list=0; list<h->list_count; list++){
442             if(!USES_LIST(mb_type, list) && !IS_DIRECT(mb_type) && !h->deblocking_filter){
443                 /*if(!h->mv_cache_clean[list]){
444                     memset(h->mv_cache [list],  0, 8*5*2*sizeof(int16_t)); //FIXME clean only input? clean at all?
445                     memset(h->ref_cache[list], PART_NOT_AVAILABLE, 8*5*sizeof(int8_t));
446                     h->mv_cache_clean[list]= 1;
447                 }*/
448                 continue;
449             }
450             h->mv_cache_clean[list]= 0;
451
452             if(USES_LIST(top_type, list)){
453                 const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
454                 const int b8_xy= h->mb2b8_xy[top_xy] + h->b8_stride;
455                 *(uint32_t*)h->mv_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 0];
456                 *(uint32_t*)h->mv_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 1];
457                 *(uint32_t*)h->mv_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 2];
458                 *(uint32_t*)h->mv_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy + 3];
459                 h->ref_cache[list][scan8[0] + 0 - 1*8]=
460                 h->ref_cache[list][scan8[0] + 1 - 1*8]= s->current_picture.ref_index[list][b8_xy + 0];
461                 h->ref_cache[list][scan8[0] + 2 - 1*8]=
462                 h->ref_cache[list][scan8[0] + 3 - 1*8]= s->current_picture.ref_index[list][b8_xy + 1];
463             }else{
464                 *(uint32_t*)h->mv_cache [list][scan8[0] + 0 - 1*8]=
465                 *(uint32_t*)h->mv_cache [list][scan8[0] + 1 - 1*8]=
466                 *(uint32_t*)h->mv_cache [list][scan8[0] + 2 - 1*8]=
467                 *(uint32_t*)h->mv_cache [list][scan8[0] + 3 - 1*8]= 0;
468                 *(uint32_t*)&h->ref_cache[list][scan8[0] + 0 - 1*8]= ((top_type ? LIST_NOT_USED : PART_NOT_AVAILABLE)&0xFF)*0x01010101;
469             }
470
471             for(i=0; i<2; i++){
472                 int cache_idx = scan8[0] - 1 + i*2*8;
473                 if(USES_LIST(left_type[i], list)){
474                     const int b_xy= h->mb2b_xy[left_xy[i]] + 3;
475                     const int b8_xy= h->mb2b8_xy[left_xy[i]] + 1;
476                     *(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]];
477                     *(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]];
478                     h->ref_cache[list][cache_idx  ]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[0+i*2]>>1)];
479                     h->ref_cache[list][cache_idx+8]= s->current_picture.ref_index[list][b8_xy + h->b8_stride*(left_block[1+i*2]>>1)];
480                 }else{
481                     *(uint32_t*)h->mv_cache [list][cache_idx  ]=
482                     *(uint32_t*)h->mv_cache [list][cache_idx+8]= 0;
483                     h->ref_cache[list][cache_idx  ]=
484                     h->ref_cache[list][cache_idx+8]= left_type[i] ? LIST_NOT_USED : PART_NOT_AVAILABLE;
485                 }
486             }
487
488             if((for_deblock || (IS_DIRECT(mb_type) && !h->direct_spatial_mv_pred)) && !FRAME_MBAFF)
489                 continue;
490
491             if(USES_LIST(topleft_type, list)){
492                 const int b_xy = h->mb2b_xy[topleft_xy] + 3 + 3*h->b_stride;
493                 const int b8_xy= h->mb2b8_xy[topleft_xy] + 1 + h->b8_stride;
494                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
495                 h->ref_cache[list][scan8[0] - 1 - 1*8]= s->current_picture.ref_index[list][b8_xy];
496             }else{
497                 *(uint32_t*)h->mv_cache[list][scan8[0] - 1 - 1*8]= 0;
498                 h->ref_cache[list][scan8[0] - 1 - 1*8]= topleft_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
499             }
500
501             if(USES_LIST(topright_type, list)){
502                 const int b_xy= h->mb2b_xy[topright_xy] + 3*h->b_stride;
503                 const int b8_xy= h->mb2b8_xy[topright_xy] + h->b8_stride;
504                 *(uint32_t*)h->mv_cache[list][scan8[0] + 4 - 1*8]= *(uint32_t*)s->current_picture.motion_val[list][b_xy];
505                 h->ref_cache[list][scan8[0] + 4 - 1*8]= s->current_picture.ref_index[list][b8_xy];
506             }else{
507                 *(uint32_t*)h->mv_cache [list][scan8[0] + 4 - 1*8]= 0;
508                 h->ref_cache[list][scan8[0] + 4 - 1*8]= topright_type ? LIST_NOT_USED : PART_NOT_AVAILABLE;
509             }
510
511             if((IS_SKIP(mb_type) || IS_DIRECT(mb_type)) && !FRAME_MBAFF)
512                 continue;
513
514             h->ref_cache[list][scan8[5 ]+1] =
515             h->ref_cache[list][scan8[7 ]+1] =
516             h->ref_cache[list][scan8[13]+1] =  //FIXME remove past 3 (init somewhere else)
517             h->ref_cache[list][scan8[4 ]] =
518             h->ref_cache[list][scan8[12]] = PART_NOT_AVAILABLE;
519             *(uint32_t*)h->mv_cache [list][scan8[5 ]+1]=
520             *(uint32_t*)h->mv_cache [list][scan8[7 ]+1]=
521             *(uint32_t*)h->mv_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
522             *(uint32_t*)h->mv_cache [list][scan8[4 ]]=
523             *(uint32_t*)h->mv_cache [list][scan8[12]]= 0;
524
525             if( h->pps.cabac ) {
526                 /* XXX beurk, Load mvd */
527                 if(USES_LIST(top_type, list)){
528                     const int b_xy= h->mb2b_xy[top_xy] + 3*h->b_stride;
529                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 0 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 0];
530                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 1 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 1];
531                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 2 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 2];
532                     *(uint32_t*)h->mvd_cache[list][scan8[0] + 3 - 1*8]= *(uint32_t*)h->mvd_table[list][b_xy + 3];
533                 }else{
534                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 0 - 1*8]=
535                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 1 - 1*8]=
536                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 2 - 1*8]=
537                     *(uint32_t*)h->mvd_cache [list][scan8[0] + 3 - 1*8]= 0;
538                 }
539                 if(USES_LIST(left_type[0], list)){
540                     const int b_xy= h->mb2b_xy[left_xy[0]] + 3;
541                     *(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]];
542                     *(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]];
543                 }else{
544                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 0*8]=
545                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 1*8]= 0;
546                 }
547                 if(USES_LIST(left_type[1], list)){
548                     const int b_xy= h->mb2b_xy[left_xy[1]] + 3;
549                     *(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]];
550                     *(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]];
551                 }else{
552                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 2*8]=
553                     *(uint32_t*)h->mvd_cache [list][scan8[0] - 1 + 3*8]= 0;
554                 }
555                 *(uint32_t*)h->mvd_cache [list][scan8[5 ]+1]=
556                 *(uint32_t*)h->mvd_cache [list][scan8[7 ]+1]=
557                 *(uint32_t*)h->mvd_cache [list][scan8[13]+1]= //FIXME remove past 3 (init somewhere else)
558                 *(uint32_t*)h->mvd_cache [list][scan8[4 ]]=
559                 *(uint32_t*)h->mvd_cache [list][scan8[12]]= 0;
560
561                 if(h->slice_type == B_TYPE){
562                     fill_rectangle(&h->direct_cache[scan8[0]], 4, 4, 8, 0, 1);
563
564                     if(IS_DIRECT(top_type)){
565                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0x01010101;
566                     }else if(IS_8X8(top_type)){
567                         int b8_xy = h->mb2b8_xy[top_xy] + h->b8_stride;
568                         h->direct_cache[scan8[0] + 0 - 1*8]= h->direct_table[b8_xy];
569                         h->direct_cache[scan8[0] + 2 - 1*8]= h->direct_table[b8_xy + 1];
570                     }else{
571                         *(uint32_t*)&h->direct_cache[scan8[0] - 1*8]= 0;
572                     }
573
574                     if(IS_DIRECT(left_type[0]))
575                         h->direct_cache[scan8[0] - 1 + 0*8]= 1;
576                     else if(IS_8X8(left_type[0]))
577                         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)];
578                     else
579                         h->direct_cache[scan8[0] - 1 + 0*8]= 0;
580
581                     if(IS_DIRECT(left_type[1]))
582                         h->direct_cache[scan8[0] - 1 + 2*8]= 1;
583                     else if(IS_8X8(left_type[1]))
584                         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)];
585                     else
586                         h->direct_cache[scan8[0] - 1 + 2*8]= 0;
587                 }
588             }
589
590             if(FRAME_MBAFF){
591 #define MAP_MVS\
592                     MAP_F2F(scan8[0] - 1 - 1*8, topleft_type)\
593                     MAP_F2F(scan8[0] + 0 - 1*8, top_type)\
594                     MAP_F2F(scan8[0] + 1 - 1*8, top_type)\
595                     MAP_F2F(scan8[0] + 2 - 1*8, top_type)\
596                     MAP_F2F(scan8[0] + 3 - 1*8, top_type)\
597                     MAP_F2F(scan8[0] + 4 - 1*8, topright_type)\
598                     MAP_F2F(scan8[0] - 1 + 0*8, left_type[0])\
599                     MAP_F2F(scan8[0] - 1 + 1*8, left_type[0])\
600                     MAP_F2F(scan8[0] - 1 + 2*8, left_type[1])\
601                     MAP_F2F(scan8[0] - 1 + 3*8, left_type[1])
602                 if(MB_FIELD){
603 #define MAP_F2F(idx, mb_type)\
604                     if(!IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
605                         h->ref_cache[list][idx] <<= 1;\
606                         h->mv_cache[list][idx][1] /= 2;\
607                         h->mvd_cache[list][idx][1] /= 2;\
608                     }
609                     MAP_MVS
610 #undef MAP_F2F
611                 }else{
612 #define MAP_F2F(idx, mb_type)\
613                     if(IS_INTERLACED(mb_type) && h->ref_cache[list][idx] >= 0){\
614                         h->ref_cache[list][idx] >>= 1;\
615                         h->mv_cache[list][idx][1] <<= 1;\
616                         h->mvd_cache[list][idx][1] <<= 1;\
617                     }
618                     MAP_MVS
619 #undef MAP_F2F
620                 }
621             }
622         }
623     }
624 #endif
625
626     h->neighbor_transform_size= !!IS_8x8DCT(top_type) + !!IS_8x8DCT(left_type[0]);
627 }
628
629 static inline void write_back_intra_pred_mode(H264Context *h){
630     MpegEncContext * const s = &h->s;
631     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
632
633     h->intra4x4_pred_mode[mb_xy][0]= h->intra4x4_pred_mode_cache[7+8*1];
634     h->intra4x4_pred_mode[mb_xy][1]= h->intra4x4_pred_mode_cache[7+8*2];
635     h->intra4x4_pred_mode[mb_xy][2]= h->intra4x4_pred_mode_cache[7+8*3];
636     h->intra4x4_pred_mode[mb_xy][3]= h->intra4x4_pred_mode_cache[7+8*4];
637     h->intra4x4_pred_mode[mb_xy][4]= h->intra4x4_pred_mode_cache[4+8*4];
638     h->intra4x4_pred_mode[mb_xy][5]= h->intra4x4_pred_mode_cache[5+8*4];
639     h->intra4x4_pred_mode[mb_xy][6]= h->intra4x4_pred_mode_cache[6+8*4];
640 }
641
642 /**
643  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
644  */
645 static inline int check_intra4x4_pred_mode(H264Context *h){
646     MpegEncContext * const s = &h->s;
647     static const int8_t top [12]= {-1, 0,LEFT_DC_PRED,-1,-1,-1,-1,-1, 0};
648     static const int8_t left[12]= { 0,-1, TOP_DC_PRED, 0,-1,-1,-1, 0,-1,DC_128_PRED};
649     int i;
650
651     if(!(h->top_samples_available&0x8000)){
652         for(i=0; i<4; i++){
653             int status= top[ h->intra4x4_pred_mode_cache[scan8[0] + i] ];
654             if(status<0){
655                 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);
656                 return -1;
657             } else if(status){
658                 h->intra4x4_pred_mode_cache[scan8[0] + i]= status;
659             }
660         }
661     }
662
663     if(!(h->left_samples_available&0x8000)){
664         for(i=0; i<4; i++){
665             int status= left[ h->intra4x4_pred_mode_cache[scan8[0] + 8*i] ];
666             if(status<0){
667                 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);
668                 return -1;
669             } else if(status){
670                 h->intra4x4_pred_mode_cache[scan8[0] + 8*i]= status;
671             }
672         }
673     }
674
675     return 0;
676 } //FIXME cleanup like next
677
678 /**
679  * checks if the top & left blocks are available if needed & changes the dc mode so it only uses the available blocks.
680  */
681 static inline int check_intra_pred_mode(H264Context *h, int mode){
682     MpegEncContext * const s = &h->s;
683     static const int8_t top [7]= {LEFT_DC_PRED8x8, 1,-1,-1};
684     static const int8_t left[7]= { TOP_DC_PRED8x8,-1, 2,-1,DC_128_PRED8x8};
685
686     if(mode > 6U) {
687         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);
688         return -1;
689     }
690
691     if(!(h->top_samples_available&0x8000)){
692         mode= top[ mode ];
693         if(mode<0){
694             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);
695             return -1;
696         }
697     }
698
699     if(!(h->left_samples_available&0x8000)){
700         mode= left[ mode ];
701         if(mode<0){
702             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);
703             return -1;
704         }
705     }
706
707     return mode;
708 }
709
710 /**
711  * gets the predicted intra4x4 prediction mode.
712  */
713 static inline int pred_intra_mode(H264Context *h, int n){
714     const int index8= scan8[n];
715     const int left= h->intra4x4_pred_mode_cache[index8 - 1];
716     const int top = h->intra4x4_pred_mode_cache[index8 - 8];
717     const int min= FFMIN(left, top);
718
719     tprintf(h->s.avctx, "mode:%d %d min:%d\n", left ,top, min);
720
721     if(min<0) return DC_PRED;
722     else      return min;
723 }
724
725 static inline void write_back_non_zero_count(H264Context *h){
726     MpegEncContext * const s = &h->s;
727     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
728
729     h->non_zero_count[mb_xy][0]= h->non_zero_count_cache[7+8*1];
730     h->non_zero_count[mb_xy][1]= h->non_zero_count_cache[7+8*2];
731     h->non_zero_count[mb_xy][2]= h->non_zero_count_cache[7+8*3];
732     h->non_zero_count[mb_xy][3]= h->non_zero_count_cache[7+8*4];
733     h->non_zero_count[mb_xy][4]= h->non_zero_count_cache[4+8*4];
734     h->non_zero_count[mb_xy][5]= h->non_zero_count_cache[5+8*4];
735     h->non_zero_count[mb_xy][6]= h->non_zero_count_cache[6+8*4];
736
737     h->non_zero_count[mb_xy][9]= h->non_zero_count_cache[1+8*2];
738     h->non_zero_count[mb_xy][8]= h->non_zero_count_cache[2+8*2];
739     h->non_zero_count[mb_xy][7]= h->non_zero_count_cache[2+8*1];
740
741     h->non_zero_count[mb_xy][12]=h->non_zero_count_cache[1+8*5];
742     h->non_zero_count[mb_xy][11]=h->non_zero_count_cache[2+8*5];
743     h->non_zero_count[mb_xy][10]=h->non_zero_count_cache[2+8*4];
744
745     if(FRAME_MBAFF){
746         // store all luma nnzs, for deblocking
747         int v = 0, i;
748         for(i=0; i<16; i++)
749             v += (!!h->non_zero_count_cache[scan8[i]]) << i;
750         *(uint16_t*)&h->non_zero_count[mb_xy][14] = v;
751     }
752 }
753
754 /**
755  * gets the predicted number of non zero coefficients.
756  * @param n block index
757  */
758 static inline int pred_non_zero_count(H264Context *h, int n){
759     const int index8= scan8[n];
760     const int left= h->non_zero_count_cache[index8 - 1];
761     const int top = h->non_zero_count_cache[index8 - 8];
762     int i= left + top;
763
764     if(i<64) i= (i+1)>>1;
765
766     tprintf(h->s.avctx, "pred_nnz L%X T%X n%d s%d P%X\n", left, top, n, scan8[n], i&31);
767
768     return i&31;
769 }
770
771 static inline int fetch_diagonal_mv(H264Context *h, const int16_t **C, int i, int list, int part_width){
772     const int topright_ref= h->ref_cache[list][ i - 8 + part_width ];
773     MpegEncContext *s = &h->s;
774
775     /* there is no consistent mapping of mvs to neighboring locations that will
776      * make mbaff happy, so we can't move all this logic to fill_caches */
777     if(FRAME_MBAFF){
778         const uint32_t *mb_types = s->current_picture_ptr->mb_type;
779         const int16_t *mv;
780         *(uint32_t*)h->mv_cache[list][scan8[0]-2] = 0;
781         *C = h->mv_cache[list][scan8[0]-2];
782
783         if(!MB_FIELD
784            && (s->mb_y&1) && i < scan8[0]+8 && topright_ref != PART_NOT_AVAILABLE){
785             int topright_xy = s->mb_x + (s->mb_y-1)*s->mb_stride + (i == scan8[0]+3);
786             if(IS_INTERLACED(mb_types[topright_xy])){
787 #define SET_DIAG_MV(MV_OP, REF_OP, X4, Y4)\
788                 const int x4 = X4, y4 = Y4;\
789                 const int mb_type = mb_types[(x4>>2)+(y4>>2)*s->mb_stride];\
790                 if(!USES_LIST(mb_type,list) && !IS_8X8(mb_type))\
791                     return LIST_NOT_USED;\
792                 mv = s->current_picture_ptr->motion_val[list][x4 + y4*h->b_stride];\
793                 h->mv_cache[list][scan8[0]-2][0] = mv[0];\
794                 h->mv_cache[list][scan8[0]-2][1] = mv[1] MV_OP;\
795                 return s->current_picture_ptr->ref_index[list][(x4>>1) + (y4>>1)*h->b8_stride] REF_OP;
796
797                 SET_DIAG_MV(*2, >>1, s->mb_x*4+(i&7)-4+part_width, s->mb_y*4-1);
798             }
799         }
800         if(topright_ref == PART_NOT_AVAILABLE
801            && ((s->mb_y&1) || i >= scan8[0]+8) && (i&7)==4
802            && h->ref_cache[list][scan8[0]-1] != PART_NOT_AVAILABLE){
803             if(!MB_FIELD
804                && IS_INTERLACED(mb_types[h->left_mb_xy[0]])){
805                 SET_DIAG_MV(*2, >>1, s->mb_x*4-1, (s->mb_y|1)*4+(s->mb_y&1)*2+(i>>4)-1);
806             }
807             if(MB_FIELD
808                && !IS_INTERLACED(mb_types[h->left_mb_xy[0]])
809                && i >= scan8[0]+8){
810                 // leftshift will turn LIST_NOT_USED into PART_NOT_AVAILABLE, but that's ok.
811                 SET_DIAG_MV(>>1, <<1, s->mb_x*4-1, (s->mb_y&~1)*4 - 1 + ((i-scan8[0])>>3)*2);
812             }
813         }
814 #undef SET_DIAG_MV
815     }
816
817     if(topright_ref != PART_NOT_AVAILABLE){
818         *C= h->mv_cache[list][ i - 8 + part_width ];
819         return topright_ref;
820     }else{
821         tprintf(s->avctx, "topright MV not available\n");
822
823         *C= h->mv_cache[list][ i - 8 - 1 ];
824         return h->ref_cache[list][ i - 8 - 1 ];
825     }
826 }
827
828 /**
829  * gets the predicted MV.
830  * @param n the block index
831  * @param part_width the width of the partition (4, 8,16) -> (1, 2, 4)
832  * @param mx the x component of the predicted motion vector
833  * @param my the y component of the predicted motion vector
834  */
835 static inline void pred_motion(H264Context * const h, int n, int part_width, int list, int ref, int * const mx, int * const my){
836     const int index8= scan8[n];
837     const int top_ref=      h->ref_cache[list][ index8 - 8 ];
838     const int left_ref=     h->ref_cache[list][ index8 - 1 ];
839     const int16_t * const A= h->mv_cache[list][ index8 - 1 ];
840     const int16_t * const B= h->mv_cache[list][ index8 - 8 ];
841     const int16_t * C;
842     int diagonal_ref, match_count;
843
844     assert(part_width==1 || part_width==2 || part_width==4);
845
846 /* mv_cache
847   B . . A T T T T
848   U . . L . . , .
849   U . . L . . . .
850   U . . L . . , .
851   . . . L . . . .
852 */
853
854     diagonal_ref= fetch_diagonal_mv(h, &C, index8, list, part_width);
855     match_count= (diagonal_ref==ref) + (top_ref==ref) + (left_ref==ref);
856     tprintf(h->s.avctx, "pred_motion match_count=%d\n", match_count);
857     if(match_count > 1){ //most common
858         *mx= mid_pred(A[0], B[0], C[0]);
859         *my= mid_pred(A[1], B[1], C[1]);
860     }else if(match_count==1){
861         if(left_ref==ref){
862             *mx= A[0];
863             *my= A[1];
864         }else if(top_ref==ref){
865             *mx= B[0];
866             *my= B[1];
867         }else{
868             *mx= C[0];
869             *my= C[1];
870         }
871     }else{
872         if(top_ref == PART_NOT_AVAILABLE && diagonal_ref == PART_NOT_AVAILABLE && left_ref != PART_NOT_AVAILABLE){
873             *mx= A[0];
874             *my= A[1];
875         }else{
876             *mx= mid_pred(A[0], B[0], C[0]);
877             *my= mid_pred(A[1], B[1], C[1]);
878         }
879     }
880
881     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);
882 }
883
884 /**
885  * gets the directionally predicted 16x8 MV.
886  * @param n the block index
887  * @param mx the x component of the predicted motion vector
888  * @param my the y component of the predicted motion vector
889  */
890 static inline void pred_16x8_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
891     if(n==0){
892         const int top_ref=      h->ref_cache[list][ scan8[0] - 8 ];
893         const int16_t * const B= h->mv_cache[list][ scan8[0] - 8 ];
894
895         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);
896
897         if(top_ref == ref){
898             *mx= B[0];
899             *my= B[1];
900             return;
901         }
902     }else{
903         const int left_ref=     h->ref_cache[list][ scan8[8] - 1 ];
904         const int16_t * const A= h->mv_cache[list][ scan8[8] - 1 ];
905
906         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);
907
908         if(left_ref == ref){
909             *mx= A[0];
910             *my= A[1];
911             return;
912         }
913     }
914
915     //RARE
916     pred_motion(h, n, 4, list, ref, mx, my);
917 }
918
919 /**
920  * gets the directionally predicted 8x16 MV.
921  * @param n the block index
922  * @param mx the x component of the predicted motion vector
923  * @param my the y component of the predicted motion vector
924  */
925 static inline void pred_8x16_motion(H264Context * const h, int n, int list, int ref, int * const mx, int * const my){
926     if(n==0){
927         const int left_ref=      h->ref_cache[list][ scan8[0] - 1 ];
928         const int16_t * const A=  h->mv_cache[list][ scan8[0] - 1 ];
929
930         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);
931
932         if(left_ref == ref){
933             *mx= A[0];
934             *my= A[1];
935             return;
936         }
937     }else{
938         const int16_t * C;
939         int diagonal_ref;
940
941         diagonal_ref= fetch_diagonal_mv(h, &C, scan8[4], list, 2);
942
943         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);
944
945         if(diagonal_ref == ref){
946             *mx= C[0];
947             *my= C[1];
948             return;
949         }
950     }
951
952     //RARE
953     pred_motion(h, n, 2, list, ref, mx, my);
954 }
955
956 static inline void pred_pskip_motion(H264Context * const h, int * const mx, int * const my){
957     const int top_ref = h->ref_cache[0][ scan8[0] - 8 ];
958     const int left_ref= h->ref_cache[0][ scan8[0] - 1 ];
959
960     tprintf(h->s.avctx, "pred_pskip: (%d) (%d) at %2d %2d\n", top_ref, left_ref, h->s.mb_x, h->s.mb_y);
961
962     if(top_ref == PART_NOT_AVAILABLE || left_ref == PART_NOT_AVAILABLE
963        || (top_ref == 0  && *(uint32_t*)h->mv_cache[0][ scan8[0] - 8 ] == 0)
964        || (left_ref == 0 && *(uint32_t*)h->mv_cache[0][ scan8[0] - 1 ] == 0)){
965
966         *mx = *my = 0;
967         return;
968     }
969
970     pred_motion(h, 0, 4, 0, 0, mx, my);
971
972     return;
973 }
974
975 static inline void direct_dist_scale_factor(H264Context * const h){
976     const int poc = h->s.current_picture_ptr->poc;
977     const int poc1 = h->ref_list[1][0].poc;
978     int i;
979     for(i=0; i<h->ref_count[0]; i++){
980         int poc0 = h->ref_list[0][i].poc;
981         int td = av_clip(poc1 - poc0, -128, 127);
982         if(td == 0 /* FIXME || pic0 is a long-term ref */){
983             h->dist_scale_factor[i] = 256;
984         }else{
985             int tb = av_clip(poc - poc0, -128, 127);
986             int tx = (16384 + (FFABS(td) >> 1)) / td;
987             h->dist_scale_factor[i] = av_clip((tb*tx + 32) >> 6, -1024, 1023);
988         }
989     }
990     if(FRAME_MBAFF){
991         for(i=0; i<h->ref_count[0]; i++){
992             h->dist_scale_factor_field[2*i] =
993             h->dist_scale_factor_field[2*i+1] = h->dist_scale_factor[i];
994         }
995     }
996 }
997 static inline void direct_ref_list_init(H264Context * const h){
998     MpegEncContext * const s = &h->s;
999     Picture * const ref1 = &h->ref_list[1][0];
1000     Picture * const cur = s->current_picture_ptr;
1001     int list, i, j;
1002     if(cur->pict_type == I_TYPE)
1003         cur->ref_count[0] = 0;
1004     if(cur->pict_type != B_TYPE)
1005         cur->ref_count[1] = 0;
1006     for(list=0; list<2; list++){
1007         cur->ref_count[list] = h->ref_count[list];
1008         for(j=0; j<h->ref_count[list]; j++)
1009             cur->ref_poc[list][j] = h->ref_list[list][j].poc;
1010     }
1011     if(cur->pict_type != B_TYPE || h->direct_spatial_mv_pred)
1012         return;
1013     for(list=0; list<2; list++){
1014         for(i=0; i<ref1->ref_count[list]; i++){
1015             const int poc = ref1->ref_poc[list][i];
1016             h->map_col_to_list0[list][i] = 0; /* bogus; fills in for missing frames */
1017             for(j=0; j<h->ref_count[list]; j++)
1018                 if(h->ref_list[list][j].poc == poc){
1019                     h->map_col_to_list0[list][i] = j;
1020                     break;
1021                 }
1022         }
1023     }
1024     if(FRAME_MBAFF){
1025         for(list=0; list<2; list++){
1026             for(i=0; i<ref1->ref_count[list]; i++){
1027                 j = h->map_col_to_list0[list][i];
1028                 h->map_col_to_list0_field[list][2*i] = 2*j;
1029                 h->map_col_to_list0_field[list][2*i+1] = 2*j+1;
1030             }
1031         }
1032     }
1033 }
1034
1035 static inline void pred_direct_motion(H264Context * const h, int *mb_type){
1036     MpegEncContext * const s = &h->s;
1037     const int mb_xy =   s->mb_x +   s->mb_y*s->mb_stride;
1038     const int b8_xy = 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1039     const int b4_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1040     const int mb_type_col = h->ref_list[1][0].mb_type[mb_xy];
1041     const int16_t (*l1mv0)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[0][b4_xy];
1042     const int16_t (*l1mv1)[2] = (const int16_t (*)[2]) &h->ref_list[1][0].motion_val[1][b4_xy];
1043     const int8_t *l1ref0 = &h->ref_list[1][0].ref_index[0][b8_xy];
1044     const int8_t *l1ref1 = &h->ref_list[1][0].ref_index[1][b8_xy];
1045     const int is_b8x8 = IS_8X8(*mb_type);
1046     unsigned int sub_mb_type;
1047     int i8, i4;
1048
1049 #define MB_TYPE_16x16_OR_INTRA (MB_TYPE_16x16|MB_TYPE_INTRA4x4|MB_TYPE_INTRA16x16|MB_TYPE_INTRA_PCM)
1050     if(IS_8X8(mb_type_col) && !h->sps.direct_8x8_inference_flag){
1051         /* FIXME save sub mb types from previous frames (or derive from MVs)
1052          * so we know exactly what block size to use */
1053         sub_mb_type = MB_TYPE_8x8|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_4x4 */
1054         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
1055     }else if(!is_b8x8 && (mb_type_col & MB_TYPE_16x16_OR_INTRA)){
1056         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1057         *mb_type =    MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_16x16 */
1058     }else{
1059         sub_mb_type = MB_TYPE_16x16|MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2; /* B_SUB_8x8 */
1060         *mb_type =    MB_TYPE_8x8|MB_TYPE_L0L1;
1061     }
1062     if(!is_b8x8)
1063         *mb_type |= MB_TYPE_DIRECT2;
1064     if(MB_FIELD)
1065         *mb_type |= MB_TYPE_INTERLACED;
1066
1067     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);
1068
1069     if(h->direct_spatial_mv_pred){
1070         int ref[2];
1071         int mv[2][2];
1072         int list;
1073
1074         /* FIXME interlacing + spatial direct uses wrong colocated block positions */
1075
1076         /* ref = min(neighbors) */
1077         for(list=0; list<2; list++){
1078             int refa = h->ref_cache[list][scan8[0] - 1];
1079             int refb = h->ref_cache[list][scan8[0] - 8];
1080             int refc = h->ref_cache[list][scan8[0] - 8 + 4];
1081             if(refc == -2)
1082                 refc = h->ref_cache[list][scan8[0] - 8 - 1];
1083             ref[list] = refa;
1084             if(ref[list] < 0 || (refb < ref[list] && refb >= 0))
1085                 ref[list] = refb;
1086             if(ref[list] < 0 || (refc < ref[list] && refc >= 0))
1087                 ref[list] = refc;
1088             if(ref[list] < 0)
1089                 ref[list] = -1;
1090         }
1091
1092         if(ref[0] < 0 && ref[1] < 0){
1093             ref[0] = ref[1] = 0;
1094             mv[0][0] = mv[0][1] =
1095             mv[1][0] = mv[1][1] = 0;
1096         }else{
1097             for(list=0; list<2; list++){
1098                 if(ref[list] >= 0)
1099                     pred_motion(h, 0, 4, list, ref[list], &mv[list][0], &mv[list][1]);
1100                 else
1101                     mv[list][0] = mv[list][1] = 0;
1102             }
1103         }
1104
1105         if(ref[1] < 0){
1106             *mb_type &= ~MB_TYPE_P0L1;
1107             sub_mb_type &= ~MB_TYPE_P0L1;
1108         }else if(ref[0] < 0){
1109             *mb_type &= ~MB_TYPE_P0L0;
1110             sub_mb_type &= ~MB_TYPE_P0L0;
1111         }
1112
1113         if(IS_16X16(*mb_type)){
1114             int a=0, b=0;
1115
1116             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, (uint8_t)ref[0], 1);
1117             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, (uint8_t)ref[1], 1);
1118             if(!IS_INTRA(mb_type_col)
1119                && (   (l1ref0[0] == 0 && FFABS(l1mv0[0][0]) <= 1 && FFABS(l1mv0[0][1]) <= 1)
1120                    || (l1ref0[0]  < 0 && l1ref1[0] == 0 && FFABS(l1mv1[0][0]) <= 1 && FFABS(l1mv1[0][1]) <= 1
1121                        && (h->x264_build>33 || !h->x264_build)))){
1122                 if(ref[0] > 0)
1123                     a= pack16to32(mv[0][0],mv[0][1]);
1124                 if(ref[1] > 0)
1125                     b= pack16to32(mv[1][0],mv[1][1]);
1126             }else{
1127                 a= pack16to32(mv[0][0],mv[0][1]);
1128                 b= pack16to32(mv[1][0],mv[1][1]);
1129             }
1130             fill_rectangle(&h->mv_cache[0][scan8[0]], 4, 4, 8, a, 4);
1131             fill_rectangle(&h->mv_cache[1][scan8[0]], 4, 4, 8, b, 4);
1132         }else{
1133             for(i8=0; i8<4; i8++){
1134                 const int x8 = i8&1;
1135                 const int y8 = i8>>1;
1136
1137                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1138                     continue;
1139                 h->sub_mb_type[i8] = sub_mb_type;
1140
1141                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mv[0][0],mv[0][1]), 4);
1142                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mv[1][0],mv[1][1]), 4);
1143                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[0], 1);
1144                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, (uint8_t)ref[1], 1);
1145
1146                 /* col_zero_flag */
1147                 if(!IS_INTRA(mb_type_col) && (   l1ref0[x8 + y8*h->b8_stride] == 0
1148                                               || (l1ref0[x8 + y8*h->b8_stride] < 0 && l1ref1[x8 + y8*h->b8_stride] == 0
1149                                                   && (h->x264_build>33 || !h->x264_build)))){
1150                     const int16_t (*l1mv)[2]= l1ref0[x8 + y8*h->b8_stride] == 0 ? l1mv0 : l1mv1;
1151                     if(IS_SUB_8X8(sub_mb_type)){
1152                         const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1153                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1154                             if(ref[0] == 0)
1155                                 fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1156                             if(ref[1] == 0)
1157                                 fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1158                         }
1159                     }else
1160                     for(i4=0; i4<4; i4++){
1161                         const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1162                         if(FFABS(mv_col[0]) <= 1 && FFABS(mv_col[1]) <= 1){
1163                             if(ref[0] == 0)
1164                                 *(uint32_t*)h->mv_cache[0][scan8[i8*4+i4]] = 0;
1165                             if(ref[1] == 0)
1166                                 *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] = 0;
1167                         }
1168                     }
1169                 }
1170             }
1171         }
1172     }else{ /* direct temporal mv pred */
1173         const int *map_col_to_list0[2] = {h->map_col_to_list0[0], h->map_col_to_list0[1]};
1174         const int *dist_scale_factor = h->dist_scale_factor;
1175
1176         if(FRAME_MBAFF){
1177             if(IS_INTERLACED(*mb_type)){
1178                 map_col_to_list0[0] = h->map_col_to_list0_field[0];
1179                 map_col_to_list0[1] = h->map_col_to_list0_field[1];
1180                 dist_scale_factor = h->dist_scale_factor_field;
1181             }
1182             if(IS_INTERLACED(*mb_type) != IS_INTERLACED(mb_type_col)){
1183                 /* FIXME assumes direct_8x8_inference == 1 */
1184                 const int pair_xy = s->mb_x + (s->mb_y&~1)*s->mb_stride;
1185                 int mb_types_col[2];
1186                 int y_shift;
1187
1188                 *mb_type = MB_TYPE_8x8|MB_TYPE_L0L1
1189                          | (is_b8x8 ? 0 : MB_TYPE_DIRECT2)
1190                          | (*mb_type & MB_TYPE_INTERLACED);
1191                 sub_mb_type = MB_TYPE_P0L0|MB_TYPE_P0L1|MB_TYPE_DIRECT2|MB_TYPE_16x16;
1192
1193                 if(IS_INTERLACED(*mb_type)){
1194                     /* frame to field scaling */
1195                     mb_types_col[0] = h->ref_list[1][0].mb_type[pair_xy];
1196                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1197                     if(s->mb_y&1){
1198                         l1ref0 -= 2*h->b8_stride;
1199                         l1ref1 -= 2*h->b8_stride;
1200                         l1mv0 -= 4*h->b_stride;
1201                         l1mv1 -= 4*h->b_stride;
1202                     }
1203                     y_shift = 0;
1204
1205                     if(   (mb_types_col[0] & MB_TYPE_16x16_OR_INTRA)
1206                        && (mb_types_col[1] & MB_TYPE_16x16_OR_INTRA)
1207                        && !is_b8x8)
1208                         *mb_type |= MB_TYPE_16x8;
1209                     else
1210                         *mb_type |= MB_TYPE_8x8;
1211                 }else{
1212                     /* field to frame scaling */
1213                     /* col_mb_y = (mb_y&~1) + (topAbsDiffPOC < bottomAbsDiffPOC ? 0 : 1)
1214                      * but in MBAFF, top and bottom POC are equal */
1215                     int dy = (s->mb_y&1) ? 1 : 2;
1216                     mb_types_col[0] =
1217                     mb_types_col[1] = h->ref_list[1][0].mb_type[pair_xy+s->mb_stride];
1218                     l1ref0 += dy*h->b8_stride;
1219                     l1ref1 += dy*h->b8_stride;
1220                     l1mv0 += 2*dy*h->b_stride;
1221                     l1mv1 += 2*dy*h->b_stride;
1222                     y_shift = 2;
1223
1224                     if((mb_types_col[0] & (MB_TYPE_16x16_OR_INTRA|MB_TYPE_16x8))
1225                        && !is_b8x8)
1226                         *mb_type |= MB_TYPE_16x16;
1227                     else
1228                         *mb_type |= MB_TYPE_8x8;
1229                 }
1230
1231                 for(i8=0; i8<4; i8++){
1232                     const int x8 = i8&1;
1233                     const int y8 = i8>>1;
1234                     int ref0, scale;
1235                     const int16_t (*l1mv)[2]= l1mv0;
1236
1237                     if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1238                         continue;
1239                     h->sub_mb_type[i8] = sub_mb_type;
1240
1241                     fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1242                     if(IS_INTRA(mb_types_col[y8])){
1243                         fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1244                         fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1245                         fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1246                         continue;
1247                     }
1248
1249                     ref0 = l1ref0[x8 + (y8*2>>y_shift)*h->b8_stride];
1250                     if(ref0 >= 0)
1251                         ref0 = map_col_to_list0[0][ref0*2>>y_shift];
1252                     else{
1253                         ref0 = map_col_to_list0[1][l1ref1[x8 + (y8*2>>y_shift)*h->b8_stride]*2>>y_shift];
1254                         l1mv= l1mv1;
1255                     }
1256                     scale = dist_scale_factor[ref0];
1257                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1258
1259                     {
1260                         const int16_t *mv_col = l1mv[x8*3 + (y8*6>>y_shift)*h->b_stride];
1261                         int my_col = (mv_col[1]<<y_shift)/2;
1262                         int mx = (scale * mv_col[0] + 128) >> 8;
1263                         int my = (scale * my_col + 128) >> 8;
1264                         fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1265                         fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-my_col), 4);
1266                     }
1267                 }
1268                 return;
1269             }
1270         }
1271
1272         /* one-to-one mv scaling */
1273
1274         if(IS_16X16(*mb_type)){
1275             int ref, mv0, mv1;
1276
1277             fill_rectangle(&h->ref_cache[1][scan8[0]], 4, 4, 8, 0, 1);
1278             if(IS_INTRA(mb_type_col)){
1279                 ref=mv0=mv1=0;
1280             }else{
1281                 const int ref0 = l1ref0[0] >= 0 ? map_col_to_list0[0][l1ref0[0]]
1282                                                 : map_col_to_list0[1][l1ref1[0]];
1283                 const int scale = dist_scale_factor[ref0];
1284                 const int16_t *mv_col = l1ref0[0] >= 0 ? l1mv0[0] : l1mv1[0];
1285                 int mv_l0[2];
1286                 mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1287                 mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1288                 ref= ref0;
1289                 mv0= pack16to32(mv_l0[0],mv_l0[1]);
1290                 mv1= pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1291             }
1292             fill_rectangle(&h->ref_cache[0][scan8[0]], 4, 4, 8, ref, 1);
1293             fill_rectangle(&h-> mv_cache[0][scan8[0]], 4, 4, 8, mv0, 4);
1294             fill_rectangle(&h-> mv_cache[1][scan8[0]], 4, 4, 8, mv1, 4);
1295         }else{
1296             for(i8=0; i8<4; i8++){
1297                 const int x8 = i8&1;
1298                 const int y8 = i8>>1;
1299                 int ref0, scale;
1300                 const int16_t (*l1mv)[2]= l1mv0;
1301
1302                 if(is_b8x8 && !IS_DIRECT(h->sub_mb_type[i8]))
1303                     continue;
1304                 h->sub_mb_type[i8] = sub_mb_type;
1305                 fill_rectangle(&h->ref_cache[1][scan8[i8*4]], 2, 2, 8, 0, 1);
1306                 if(IS_INTRA(mb_type_col)){
1307                     fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, 0, 1);
1308                     fill_rectangle(&h-> mv_cache[0][scan8[i8*4]], 2, 2, 8, 0, 4);
1309                     fill_rectangle(&h-> mv_cache[1][scan8[i8*4]], 2, 2, 8, 0, 4);
1310                     continue;
1311                 }
1312
1313                 ref0 = l1ref0[x8 + y8*h->b8_stride];
1314                 if(ref0 >= 0)
1315                     ref0 = map_col_to_list0[0][ref0];
1316                 else{
1317                     ref0 = map_col_to_list0[1][l1ref1[x8 + y8*h->b8_stride]];
1318                     l1mv= l1mv1;
1319                 }
1320                 scale = dist_scale_factor[ref0];
1321
1322                 fill_rectangle(&h->ref_cache[0][scan8[i8*4]], 2, 2, 8, ref0, 1);
1323                 if(IS_SUB_8X8(sub_mb_type)){
1324                     const int16_t *mv_col = l1mv[x8*3 + y8*3*h->b_stride];
1325                     int mx = (scale * mv_col[0] + 128) >> 8;
1326                     int my = (scale * mv_col[1] + 128) >> 8;
1327                     fill_rectangle(&h->mv_cache[0][scan8[i8*4]], 2, 2, 8, pack16to32(mx,my), 4);
1328                     fill_rectangle(&h->mv_cache[1][scan8[i8*4]], 2, 2, 8, pack16to32(mx-mv_col[0],my-mv_col[1]), 4);
1329                 }else
1330                 for(i4=0; i4<4; i4++){
1331                     const int16_t *mv_col = l1mv[x8*2 + (i4&1) + (y8*2 + (i4>>1))*h->b_stride];
1332                     int16_t *mv_l0 = h->mv_cache[0][scan8[i8*4+i4]];
1333                     mv_l0[0] = (scale * mv_col[0] + 128) >> 8;
1334                     mv_l0[1] = (scale * mv_col[1] + 128) >> 8;
1335                     *(uint32_t*)h->mv_cache[1][scan8[i8*4+i4]] =
1336                         pack16to32(mv_l0[0]-mv_col[0],mv_l0[1]-mv_col[1]);
1337                 }
1338             }
1339         }
1340     }
1341 }
1342
1343 static inline void write_back_motion(H264Context *h, int mb_type){
1344     MpegEncContext * const s = &h->s;
1345     const int b_xy = 4*s->mb_x + 4*s->mb_y*h->b_stride;
1346     const int b8_xy= 2*s->mb_x + 2*s->mb_y*h->b8_stride;
1347     int list;
1348
1349     if(!USES_LIST(mb_type, 0))
1350         fill_rectangle(&s->current_picture.ref_index[0][b8_xy], 2, 2, h->b8_stride, (uint8_t)LIST_NOT_USED, 1);
1351
1352     for(list=0; list<h->list_count; list++){
1353         int y;
1354         if(!USES_LIST(mb_type, list))
1355             continue;
1356
1357         for(y=0; y<4; y++){
1358             *(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];
1359             *(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];
1360         }
1361         if( h->pps.cabac ) {
1362             if(IS_SKIP(mb_type))
1363                 fill_rectangle(h->mvd_table[list][b_xy], 4, 4, h->b_stride, 0, 4);
1364             else
1365             for(y=0; y<4; y++){
1366                 *(uint64_t*)h->mvd_table[list][b_xy + 0 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+0 + 8*y];
1367                 *(uint64_t*)h->mvd_table[list][b_xy + 2 + y*h->b_stride]= *(uint64_t*)h->mvd_cache[list][scan8[0]+2 + 8*y];
1368             }
1369         }
1370
1371         {
1372             int8_t *ref_index = &s->current_picture.ref_index[list][b8_xy];
1373             ref_index[0+0*h->b8_stride]= h->ref_cache[list][scan8[0]];
1374             ref_index[1+0*h->b8_stride]= h->ref_cache[list][scan8[4]];
1375             ref_index[0+1*h->b8_stride]= h->ref_cache[list][scan8[8]];
1376             ref_index[1+1*h->b8_stride]= h->ref_cache[list][scan8[12]];
1377         }
1378     }
1379
1380     if(h->slice_type == B_TYPE && h->pps.cabac){
1381         if(IS_8X8(mb_type)){
1382             uint8_t *direct_table = &h->direct_table[b8_xy];
1383             direct_table[1+0*h->b8_stride] = IS_DIRECT(h->sub_mb_type[1]) ? 1 : 0;
1384             direct_table[0+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[2]) ? 1 : 0;
1385             direct_table[1+1*h->b8_stride] = IS_DIRECT(h->sub_mb_type[3]) ? 1 : 0;
1386         }
1387     }
1388 }
1389
1390 /**
1391  * Decodes a network abstraction layer unit.
1392  * @param consumed is the number of bytes used as input
1393  * @param length is the length of the array
1394  * @param dst_length is the number of decoded bytes FIXME here or a decode rbsp tailing?
1395  * @returns decoded bytes, might be src+1 if no escapes
1396  */
1397 static uint8_t *decode_nal(H264Context *h, uint8_t *src, int *dst_length, int *consumed, int length){
1398     int i, si, di;
1399     uint8_t *dst;
1400     int bufidx;
1401
1402 //    src[0]&0x80;                //forbidden bit
1403     h->nal_ref_idc= src[0]>>5;
1404     h->nal_unit_type= src[0]&0x1F;
1405
1406     src++; length--;
1407 #if 0
1408     for(i=0; i<length; i++)
1409         printf("%2X ", src[i]);
1410 #endif
1411     for(i=0; i+1<length; i+=2){
1412         if(src[i]) continue;
1413         if(i>0 && src[i-1]==0) i--;
1414         if(i+2<length && src[i+1]==0 && src[i+2]<=3){
1415             if(src[i+2]!=3){
1416                 /* startcode, so we must be past the end */
1417                 length=i;
1418             }
1419             break;
1420         }
1421     }
1422
1423     if(i>=length-1){ //no escaped 0
1424         *dst_length= length;
1425         *consumed= length+1; //+1 for the header
1426         return src;
1427     }
1428
1429     bufidx = h->nal_unit_type == NAL_DPC ? 1 : 0; // use second escape buffer for inter data
1430     h->rbsp_buffer[bufidx]= av_fast_realloc(h->rbsp_buffer[bufidx], &h->rbsp_buffer_size[bufidx], length);
1431     dst= h->rbsp_buffer[bufidx];
1432
1433     if (dst == NULL){
1434         return NULL;
1435     }
1436
1437 //printf("decoding esc\n");
1438     si=di=0;
1439     while(si<length){
1440         //remove escapes (very rare 1:2^22)
1441         if(si+2<length && src[si]==0 && src[si+1]==0 && src[si+2]<=3){
1442             if(src[si+2]==3){ //escape
1443                 dst[di++]= 0;
1444                 dst[di++]= 0;
1445                 si+=3;
1446                 continue;
1447             }else //next start code
1448                 break;
1449         }
1450
1451         dst[di++]= src[si++];
1452     }
1453
1454     *dst_length= di;
1455     *consumed= si + 1;//+1 for the header
1456 //FIXME store exact number of bits in the getbitcontext (it is needed for decoding)
1457     return dst;
1458 }
1459
1460 /**
1461  * identifies the exact end of the bitstream
1462  * @return the length of the trailing, or 0 if damaged
1463  */
1464 static int decode_rbsp_trailing(H264Context *h, uint8_t *src){
1465     int v= *src;
1466     int r;
1467
1468     tprintf(h->s.avctx, "rbsp trailing %X\n", v);
1469
1470     for(r=1; r<9; r++){
1471         if(v&1) return r;
1472         v>>=1;
1473     }
1474     return 0;
1475 }
1476
1477 /**
1478  * idct tranforms the 16 dc values and dequantize them.
1479  * @param qp quantization parameter
1480  */
1481 static void h264_luma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1482 #define stride 16
1483     int i;
1484     int temp[16]; //FIXME check if this is a good idea
1485     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1486     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1487
1488 //memset(block, 64, 2*256);
1489 //return;
1490     for(i=0; i<4; i++){
1491         const int offset= y_offset[i];
1492         const int z0= block[offset+stride*0] + block[offset+stride*4];
1493         const int z1= block[offset+stride*0] - block[offset+stride*4];
1494         const int z2= block[offset+stride*1] - block[offset+stride*5];
1495         const int z3= block[offset+stride*1] + block[offset+stride*5];
1496
1497         temp[4*i+0]= z0+z3;
1498         temp[4*i+1]= z1+z2;
1499         temp[4*i+2]= z1-z2;
1500         temp[4*i+3]= z0-z3;
1501     }
1502
1503     for(i=0; i<4; i++){
1504         const int offset= x_offset[i];
1505         const int z0= temp[4*0+i] + temp[4*2+i];
1506         const int z1= temp[4*0+i] - temp[4*2+i];
1507         const int z2= temp[4*1+i] - temp[4*3+i];
1508         const int z3= temp[4*1+i] + temp[4*3+i];
1509
1510         block[stride*0 +offset]= ((((z0 + z3)*qmul + 128 ) >> 8)); //FIXME think about merging this into decode_resdual
1511         block[stride*2 +offset]= ((((z1 + z2)*qmul + 128 ) >> 8));
1512         block[stride*8 +offset]= ((((z1 - z2)*qmul + 128 ) >> 8));
1513         block[stride*10+offset]= ((((z0 - z3)*qmul + 128 ) >> 8));
1514     }
1515 }
1516
1517 #if 0
1518 /**
1519  * dct tranforms the 16 dc values.
1520  * @param qp quantization parameter ??? FIXME
1521  */
1522 static void h264_luma_dc_dct_c(DCTELEM *block/*, int qp*/){
1523 //    const int qmul= dequant_coeff[qp][0];
1524     int i;
1525     int temp[16]; //FIXME check if this is a good idea
1526     static const int x_offset[4]={0, 1*stride, 4* stride,  5*stride};
1527     static const int y_offset[4]={0, 2*stride, 8* stride, 10*stride};
1528
1529     for(i=0; i<4; i++){
1530         const int offset= y_offset[i];
1531         const int z0= block[offset+stride*0] + block[offset+stride*4];
1532         const int z1= block[offset+stride*0] - block[offset+stride*4];
1533         const int z2= block[offset+stride*1] - block[offset+stride*5];
1534         const int z3= block[offset+stride*1] + block[offset+stride*5];
1535
1536         temp[4*i+0]= z0+z3;
1537         temp[4*i+1]= z1+z2;
1538         temp[4*i+2]= z1-z2;
1539         temp[4*i+3]= z0-z3;
1540     }
1541
1542     for(i=0; i<4; i++){
1543         const int offset= x_offset[i];
1544         const int z0= temp[4*0+i] + temp[4*2+i];
1545         const int z1= temp[4*0+i] - temp[4*2+i];
1546         const int z2= temp[4*1+i] - temp[4*3+i];
1547         const int z3= temp[4*1+i] + temp[4*3+i];
1548
1549         block[stride*0 +offset]= (z0 + z3)>>1;
1550         block[stride*2 +offset]= (z1 + z2)>>1;
1551         block[stride*8 +offset]= (z1 - z2)>>1;
1552         block[stride*10+offset]= (z0 - z3)>>1;
1553     }
1554 }
1555 #endif
1556
1557 #undef xStride
1558 #undef stride
1559
1560 static void chroma_dc_dequant_idct_c(DCTELEM *block, int qp, int qmul){
1561     const int stride= 16*2;
1562     const int xStride= 16;
1563     int a,b,c,d,e;
1564
1565     a= block[stride*0 + xStride*0];
1566     b= block[stride*0 + xStride*1];
1567     c= block[stride*1 + xStride*0];
1568     d= block[stride*1 + xStride*1];
1569
1570     e= a-b;
1571     a= a+b;
1572     b= c-d;
1573     c= c+d;
1574
1575     block[stride*0 + xStride*0]= ((a+c)*qmul) >> 7;
1576     block[stride*0 + xStride*1]= ((e+b)*qmul) >> 7;
1577     block[stride*1 + xStride*0]= ((a-c)*qmul) >> 7;
1578     block[stride*1 + xStride*1]= ((e-b)*qmul) >> 7;
1579 }
1580
1581 #if 0
1582 static void chroma_dc_dct_c(DCTELEM *block){
1583     const int stride= 16*2;
1584     const int xStride= 16;
1585     int a,b,c,d,e;
1586
1587     a= block[stride*0 + xStride*0];
1588     b= block[stride*0 + xStride*1];
1589     c= block[stride*1 + xStride*0];
1590     d= block[stride*1 + xStride*1];
1591
1592     e= a-b;
1593     a= a+b;
1594     b= c-d;
1595     c= c+d;
1596
1597     block[stride*0 + xStride*0]= (a+c);
1598     block[stride*0 + xStride*1]= (e+b);
1599     block[stride*1 + xStride*0]= (a-c);
1600     block[stride*1 + xStride*1]= (e-b);
1601 }
1602 #endif
1603
1604 /**
1605  * gets the chroma qp.
1606  */
1607 static inline int get_chroma_qp(H264Context *h, int t, int qscale){
1608     return h->pps.chroma_qp_table[t][qscale & 0xff];
1609 }
1610
1611 //FIXME need to check that this does not overflow signed 32 bit for low qp, i am not sure, it's very close
1612 //FIXME check that gcc inlines this (and optimizes intra & separate_dc stuff away)
1613 static inline int quantize_c(DCTELEM *block, uint8_t *scantable, int qscale, int intra, int separate_dc){
1614     int i;
1615     const int * const quant_table= quant_coeff[qscale];
1616     const int bias= intra ? (1<<QUANT_SHIFT)/3 : (1<<QUANT_SHIFT)/6;
1617     const unsigned int threshold1= (1<<QUANT_SHIFT) - bias - 1;
1618     const unsigned int threshold2= (threshold1<<1);
1619     int last_non_zero;
1620
1621     if(separate_dc){
1622         if(qscale<=18){
1623             //avoid overflows
1624             const int dc_bias= intra ? (1<<(QUANT_SHIFT-2))/3 : (1<<(QUANT_SHIFT-2))/6;
1625             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT-2)) - dc_bias - 1;
1626             const unsigned int dc_threshold2= (dc_threshold1<<1);
1627
1628             int level= block[0]*quant_coeff[qscale+18][0];
1629             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1630                 if(level>0){
1631                     level= (dc_bias + level)>>(QUANT_SHIFT-2);
1632                     block[0]= level;
1633                 }else{
1634                     level= (dc_bias - level)>>(QUANT_SHIFT-2);
1635                     block[0]= -level;
1636                 }
1637 //                last_non_zero = i;
1638             }else{
1639                 block[0]=0;
1640             }
1641         }else{
1642             const int dc_bias= intra ? (1<<(QUANT_SHIFT+1))/3 : (1<<(QUANT_SHIFT+1))/6;
1643             const unsigned int dc_threshold1= (1<<(QUANT_SHIFT+1)) - dc_bias - 1;
1644             const unsigned int dc_threshold2= (dc_threshold1<<1);
1645
1646             int level= block[0]*quant_table[0];
1647             if(((unsigned)(level+dc_threshold1))>dc_threshold2){
1648                 if(level>0){
1649                     level= (dc_bias + level)>>(QUANT_SHIFT+1);
1650                     block[0]= level;
1651                 }else{
1652                     level= (dc_bias - level)>>(QUANT_SHIFT+1);
1653                     block[0]= -level;
1654                 }
1655 //                last_non_zero = i;
1656             }else{
1657                 block[0]=0;
1658             }
1659         }
1660         last_non_zero= 0;
1661         i=1;
1662     }else{
1663         last_non_zero= -1;
1664         i=0;
1665     }
1666
1667     for(; i<16; i++){
1668         const int j= scantable[i];
1669         int level= block[j]*quant_table[j];
1670
1671 //        if(   bias+level >= (1<<(QMAT_SHIFT - 3))
1672 //           || bias-level >= (1<<(QMAT_SHIFT - 3))){
1673         if(((unsigned)(level+threshold1))>threshold2){
1674             if(level>0){
1675                 level= (bias + level)>>QUANT_SHIFT;
1676                 block[j]= level;
1677             }else{
1678                 level= (bias - level)>>QUANT_SHIFT;
1679                 block[j]= -level;
1680             }
1681             last_non_zero = i;
1682         }else{
1683             block[j]=0;
1684         }
1685     }
1686
1687     return last_non_zero;
1688 }
1689
1690 static inline void mc_dir_part(H264Context *h, Picture *pic, int n, int square, int chroma_height, int delta, int list,
1691                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1692                            int src_x_offset, int src_y_offset,
1693                            qpel_mc_func *qpix_op, h264_chroma_mc_func chroma_op){
1694     MpegEncContext * const s = &h->s;
1695     const int mx= h->mv_cache[list][ scan8[n] ][0] + src_x_offset*8;
1696     int my=       h->mv_cache[list][ scan8[n] ][1] + src_y_offset*8;
1697     const int luma_xy= (mx&3) + ((my&3)<<2);
1698     uint8_t * src_y = pic->data[0] + (mx>>2) + (my>>2)*h->mb_linesize;
1699     uint8_t * src_cb, * src_cr;
1700     int extra_width= h->emu_edge_width;
1701     int extra_height= h->emu_edge_height;
1702     int emu=0;
1703     const int full_mx= mx>>2;
1704     const int full_my= my>>2;
1705     const int pic_width  = 16*s->mb_width;
1706     const int pic_height = 16*s->mb_height >> (MB_MBAFF || FIELD_PICTURE);
1707
1708     if(!pic->data[0]) //FIXME this is unacceptable, some senseable error concealment must be done for missing reference frames
1709         return;
1710
1711     if(mx&7) extra_width -= 3;
1712     if(my&7) extra_height -= 3;
1713
1714     if(   full_mx < 0-extra_width
1715        || full_my < 0-extra_height
1716        || full_mx + 16/*FIXME*/ > pic_width + extra_width
1717        || full_my + 16/*FIXME*/ > pic_height + extra_height){
1718         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);
1719             src_y= s->edge_emu_buffer + 2 + 2*h->mb_linesize;
1720         emu=1;
1721     }
1722
1723     qpix_op[luma_xy](dest_y, src_y, h->mb_linesize); //FIXME try variable height perhaps?
1724     if(!square){
1725         qpix_op[luma_xy](dest_y + delta, src_y + delta, h->mb_linesize);
1726     }
1727
1728     if(ENABLE_GRAY && s->flags&CODEC_FLAG_GRAY) return;
1729
1730     if(MB_MBAFF || FIELD_PICTURE){
1731         // chroma offset when predicting from a field of opposite parity
1732         my += 2 * ((s->mb_y & 1) - (h->ref_cache[list][scan8[n]] & 1));
1733         emu |= (my>>3) < 0 || (my>>3) + 8 >= (pic_height>>1);
1734     }
1735     src_cb= pic->data[1] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1736     src_cr= pic->data[2] + (mx>>3) + (my>>3)*h->mb_uvlinesize;
1737
1738     if(emu){
1739         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);
1740             src_cb= s->edge_emu_buffer;
1741     }
1742     chroma_op(dest_cb, src_cb, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1743
1744     if(emu){
1745         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);
1746             src_cr= s->edge_emu_buffer;
1747     }
1748     chroma_op(dest_cr, src_cr, h->mb_uvlinesize, chroma_height, mx&7, my&7);
1749 }
1750
1751 static inline void mc_part_std(H264Context *h, int n, int square, int chroma_height, int delta,
1752                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1753                            int x_offset, int y_offset,
1754                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1755                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1756                            int list0, int list1){
1757     MpegEncContext * const s = &h->s;
1758     qpel_mc_func *qpix_op=  qpix_put;
1759     h264_chroma_mc_func chroma_op= chroma_put;
1760
1761     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1762     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1763     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1764     x_offset += 8*s->mb_x;
1765     y_offset += 8*(s->mb_y >> (MB_MBAFF || FIELD_PICTURE));
1766
1767     if(list0){
1768         Picture *ref= &h->ref_list[0][ h->ref_cache[0][ scan8[n] ] ];
1769         mc_dir_part(h, ref, n, square, chroma_height, delta, 0,
1770                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1771                            qpix_op, chroma_op);
1772
1773         qpix_op=  qpix_avg;
1774         chroma_op= chroma_avg;
1775     }
1776
1777     if(list1){
1778         Picture *ref= &h->ref_list[1][ h->ref_cache[1][ scan8[n] ] ];
1779         mc_dir_part(h, ref, n, square, chroma_height, delta, 1,
1780                            dest_y, dest_cb, dest_cr, x_offset, y_offset,
1781                            qpix_op, chroma_op);
1782     }
1783 }
1784
1785 static inline void mc_part_weighted(H264Context *h, int n, int square, int chroma_height, int delta,
1786                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1787                            int x_offset, int y_offset,
1788                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1789                            h264_weight_func luma_weight_op, h264_weight_func chroma_weight_op,
1790                            h264_biweight_func luma_weight_avg, h264_biweight_func chroma_weight_avg,
1791                            int list0, int list1){
1792     MpegEncContext * const s = &h->s;
1793
1794     dest_y  += 2*x_offset + 2*y_offset*h->  mb_linesize;
1795     dest_cb +=   x_offset +   y_offset*h->mb_uvlinesize;
1796     dest_cr +=   x_offset +   y_offset*h->mb_uvlinesize;
1797     x_offset += 8*s->mb_x;
1798     y_offset += 8*(s->mb_y >> (MB_MBAFF || FIELD_PICTURE));
1799
1800     if(list0 && list1){
1801         /* don't optimize for luma-only case, since B-frames usually
1802          * use implicit weights => chroma too. */
1803         uint8_t *tmp_cb = s->obmc_scratchpad;
1804         uint8_t *tmp_cr = s->obmc_scratchpad + 8;
1805         uint8_t *tmp_y  = s->obmc_scratchpad + 8*h->mb_uvlinesize;
1806         int refn0 = h->ref_cache[0][ scan8[n] ];
1807         int refn1 = h->ref_cache[1][ scan8[n] ];
1808
1809         mc_dir_part(h, &h->ref_list[0][refn0], n, square, chroma_height, delta, 0,
1810                     dest_y, dest_cb, dest_cr,
1811                     x_offset, y_offset, qpix_put, chroma_put);
1812         mc_dir_part(h, &h->ref_list[1][refn1], n, square, chroma_height, delta, 1,
1813                     tmp_y, tmp_cb, tmp_cr,
1814                     x_offset, y_offset, qpix_put, chroma_put);
1815
1816         if(h->use_weight == 2){
1817             int weight0 = h->implicit_weight[refn0][refn1];
1818             int weight1 = 64 - weight0;
1819             luma_weight_avg(  dest_y,  tmp_y,  h->  mb_linesize, 5, weight0, weight1, 0);
1820             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, 5, weight0, weight1, 0);
1821             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, 5, weight0, weight1, 0);
1822         }else{
1823             luma_weight_avg(dest_y, tmp_y, h->mb_linesize, h->luma_log2_weight_denom,
1824                             h->luma_weight[0][refn0], h->luma_weight[1][refn1],
1825                             h->luma_offset[0][refn0] + h->luma_offset[1][refn1]);
1826             chroma_weight_avg(dest_cb, tmp_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1827                             h->chroma_weight[0][refn0][0], h->chroma_weight[1][refn1][0],
1828                             h->chroma_offset[0][refn0][0] + h->chroma_offset[1][refn1][0]);
1829             chroma_weight_avg(dest_cr, tmp_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1830                             h->chroma_weight[0][refn0][1], h->chroma_weight[1][refn1][1],
1831                             h->chroma_offset[0][refn0][1] + h->chroma_offset[1][refn1][1]);
1832         }
1833     }else{
1834         int list = list1 ? 1 : 0;
1835         int refn = h->ref_cache[list][ scan8[n] ];
1836         Picture *ref= &h->ref_list[list][refn];
1837         mc_dir_part(h, ref, n, square, chroma_height, delta, list,
1838                     dest_y, dest_cb, dest_cr, x_offset, y_offset,
1839                     qpix_put, chroma_put);
1840
1841         luma_weight_op(dest_y, h->mb_linesize, h->luma_log2_weight_denom,
1842                        h->luma_weight[list][refn], h->luma_offset[list][refn]);
1843         if(h->use_weight_chroma){
1844             chroma_weight_op(dest_cb, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1845                              h->chroma_weight[list][refn][0], h->chroma_offset[list][refn][0]);
1846             chroma_weight_op(dest_cr, h->mb_uvlinesize, h->chroma_log2_weight_denom,
1847                              h->chroma_weight[list][refn][1], h->chroma_offset[list][refn][1]);
1848         }
1849     }
1850 }
1851
1852 static inline void mc_part(H264Context *h, int n, int square, int chroma_height, int delta,
1853                            uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1854                            int x_offset, int y_offset,
1855                            qpel_mc_func *qpix_put, h264_chroma_mc_func chroma_put,
1856                            qpel_mc_func *qpix_avg, h264_chroma_mc_func chroma_avg,
1857                            h264_weight_func *weight_op, h264_biweight_func *weight_avg,
1858                            int list0, int list1){
1859     if((h->use_weight==2 && list0 && list1
1860         && (h->implicit_weight[ h->ref_cache[0][scan8[n]] ][ h->ref_cache[1][scan8[n]] ] != 32))
1861        || h->use_weight==1)
1862         mc_part_weighted(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1863                          x_offset, y_offset, qpix_put, chroma_put,
1864                          weight_op[0], weight_op[3], weight_avg[0], weight_avg[3], list0, list1);
1865     else
1866         mc_part_std(h, n, square, chroma_height, delta, dest_y, dest_cb, dest_cr,
1867                     x_offset, y_offset, qpix_put, chroma_put, qpix_avg, chroma_avg, list0, list1);
1868 }
1869
1870 static inline void prefetch_motion(H264Context *h, int list){
1871     /* fetch pixels for estimated mv 4 macroblocks ahead
1872      * optimized for 64byte cache lines */
1873     MpegEncContext * const s = &h->s;
1874     const int refn = h->ref_cache[list][scan8[0]];
1875     if(refn >= 0){
1876         const int mx= (h->mv_cache[list][scan8[0]][0]>>2) + 16*s->mb_x + 8;
1877         const int my= (h->mv_cache[list][scan8[0]][1]>>2) + 16*s->mb_y;
1878         uint8_t **src= h->ref_list[list][refn].data;
1879         int off= mx + (my + (s->mb_x&3)*4)*h->mb_linesize + 64;
1880         s->dsp.prefetch(src[0]+off, s->linesize, 4);
1881         off= (mx>>1) + ((my>>1) + (s->mb_x&7))*s->uvlinesize + 64;
1882         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);
1883     }
1884 }
1885
1886 static void hl_motion(H264Context *h, uint8_t *dest_y, uint8_t *dest_cb, uint8_t *dest_cr,
1887                       qpel_mc_func (*qpix_put)[16], h264_chroma_mc_func (*chroma_put),
1888                       qpel_mc_func (*qpix_avg)[16], h264_chroma_mc_func (*chroma_avg),
1889                       h264_weight_func *weight_op, h264_biweight_func *weight_avg){
1890     MpegEncContext * const s = &h->s;
1891     const int mb_xy= s->mb_x + s->mb_y*s->mb_stride;
1892     const int mb_type= s->current_picture.mb_type[mb_xy];
1893
1894     assert(IS_INTER(mb_type));
1895
1896     prefetch_motion(h, 0);
1897
1898     if(IS_16X16(mb_type)){
1899         mc_part(h, 0, 1, 8, 0, dest_y, dest_cb, dest_cr, 0, 0,
1900                 qpix_put[0], chroma_put[0], qpix_avg[0], chroma_avg[0],
1901                 &weight_op[0], &weight_avg[0],
1902                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1903     }else if(IS_16X8(mb_type)){
1904         mc_part(h, 0, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 0,
1905                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1906                 &weight_op[1], &weight_avg[1],
1907                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1908         mc_part(h, 8, 0, 4, 8, dest_y, dest_cb, dest_cr, 0, 4,
1909                 qpix_put[1], chroma_put[0], qpix_avg[1], chroma_avg[0],
1910                 &weight_op[1], &weight_avg[1],
1911                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1912     }else if(IS_8X16(mb_type)){
1913         mc_part(h, 0, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 0, 0,
1914                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1915                 &weight_op[2], &weight_avg[2],
1916                 IS_DIR(mb_type, 0, 0), IS_DIR(mb_type, 0, 1));
1917         mc_part(h, 4, 0, 8, 8*h->mb_linesize, dest_y, dest_cb, dest_cr, 4, 0,
1918                 qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1919                 &weight_op[2], &weight_avg[2],
1920                 IS_DIR(mb_type, 1, 0), IS_DIR(mb_type, 1, 1));
1921     }else{
1922         int i;
1923
1924         assert(IS_8X8(mb_type));
1925
1926         for(i=0; i<4; i++){
1927             const int sub_mb_type= h->sub_mb_type[i];
1928             const int n= 4*i;
1929             int x_offset= (i&1)<<2;
1930             int y_offset= (i&2)<<1;
1931
1932             if(IS_SUB_8X8(sub_mb_type)){
1933                 mc_part(h, n, 1, 4, 0, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1934                     qpix_put[1], chroma_put[1], qpix_avg[1], chroma_avg[1],
1935                     &weight_op[3], &weight_avg[3],
1936                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1937             }else if(IS_SUB_8X4(sub_mb_type)){
1938                 mc_part(h, n  , 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1939                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1940                     &weight_op[4], &weight_avg[4],
1941                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1942                 mc_part(h, n+2, 0, 2, 4, dest_y, dest_cb, dest_cr, x_offset, y_offset+2,
1943                     qpix_put[2], chroma_put[1], qpix_avg[2], chroma_avg[1],
1944                     &weight_op[4], &weight_avg[4],
1945                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1946             }else if(IS_SUB_4X8(sub_mb_type)){
1947                 mc_part(h, n  , 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset, y_offset,
1948                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1949                     &weight_op[5], &weight_avg[5],
1950                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1951                 mc_part(h, n+1, 0, 4, 4*h->mb_linesize, dest_y, dest_cb, dest_cr, x_offset+2, y_offset,
1952                     qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1953                     &weight_op[5], &weight_avg[5],
1954                     IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1955             }else{
1956                 int j;
1957                 assert(IS_SUB_4X4(sub_mb_type));
1958                 for(j=0; j<4; j++){
1959                     int sub_x_offset= x_offset + 2*(j&1);
1960                     int sub_y_offset= y_offset +   (j&2);
1961                     mc_part(h, n+j, 1, 2, 0, dest_y, dest_cb, dest_cr, sub_x_offset, sub_y_offset,
1962                         qpix_put[2], chroma_put[2], qpix_avg[2], chroma_avg[2],
1963                         &weight_op[6], &weight_avg[6],
1964                         IS_DIR(sub_mb_type, 0, 0), IS_DIR(sub_mb_type, 0, 1));
1965                 }
1966             }
1967         }
1968     }
1969
1970     prefetch_motion(h, 1);
1971 }
1972
1973 static void decode_init_vlc(void){
1974     static int done = 0;
1975
1976     if (!done) {
1977         int i;
1978         done = 1;
1979
1980         init_vlc(&chroma_dc_coeff_token_vlc, CHROMA_DC_COEFF_TOKEN_VLC_BITS, 4*5,
1981                  &chroma_dc_coeff_token_len [0], 1, 1,
1982                  &chroma_dc_coeff_token_bits[0], 1, 1, 1);
1983
1984         for(i=0; i<4; i++){
1985             init_vlc(&coeff_token_vlc[i], COEFF_TOKEN_VLC_BITS, 4*17,
1986                      &coeff_token_len [i][0], 1, 1,
1987                      &coeff_token_bits[i][0], 1, 1, 1);
1988         }
1989
1990         for(i=0; i<3; i++){
1991             init_vlc(&chroma_dc_total_zeros_vlc[i], CHROMA_DC_TOTAL_ZEROS_VLC_BITS, 4,
1992                      &chroma_dc_total_zeros_len [i][0], 1, 1,
1993                      &chroma_dc_total_zeros_bits[i][0], 1, 1, 1);
1994         }
1995         for(i=0; i<15; i++){
1996             init_vlc(&total_zeros_vlc[i], TOTAL_ZEROS_VLC_BITS, 16,
1997                      &total_zeros_len [i][0], 1, 1,
1998                      &total_zeros_bits[i][0], 1, 1, 1);
1999         }
2000
2001         for(i=0; i<6; i++){
2002             init_vlc(&run_vlc[i], RUN_VLC_BITS, 7,
2003                      &run_len [i][0], 1, 1,
2004                      &run_bits[i][0], 1, 1, 1);
2005         }
2006         init_vlc(&run7_vlc, RUN7_VLC_BITS, 16,
2007                  &run_len [6][0], 1, 1,
2008                  &run_bits[6][0], 1, 1, 1);
2009     }
2010 }
2011
2012 static void free_tables(H264Context *h){
2013     int i;
2014     H264Context *hx;
2015     av_freep(&h->intra4x4_pred_mode);
2016     av_freep(&h->chroma_pred_mode_table);
2017     av_freep(&h->cbp_table);
2018     av_freep(&h->mvd_table[0]);
2019     av_freep(&h->mvd_table[1]);
2020     av_freep(&h->direct_table);
2021     av_freep(&h->non_zero_count);
2022     av_freep(&h->slice_table_base);
2023     h->slice_table= NULL;
2024
2025     av_freep(&h->mb2b_xy);
2026     av_freep(&h->mb2b8_xy);
2027
2028     for(i = 0; i < MAX_SPS_COUNT; i++)
2029         av_freep(h->sps_buffers + i);
2030
2031     for(i = 0; i < MAX_PPS_COUNT; i++)
2032         av_freep(h->pps_buffers + i);
2033
2034     for(i = 0; i < h->s.avctx->thread_count; i++) {
2035         hx = h->thread_context[i];
2036         if(!hx) continue;
2037         av_freep(&hx->top_borders[1]);
2038         av_freep(&hx->top_borders[0]);
2039         av_freep(&hx->s.obmc_scratchpad);
2040         av_freep(&hx->s.allocated_edge_emu_buffer);
2041     }
2042 }
2043
2044 static void init_dequant8_coeff_table(H264Context *h){
2045     int i,q,x;
2046     const int transpose = (h->s.dsp.h264_idct8_add != ff_h264_idct8_add_c); //FIXME ugly
2047     h->dequant8_coeff[0] = h->dequant8_buffer[0];
2048     h->dequant8_coeff[1] = h->dequant8_buffer[1];
2049
2050     for(i=0; i<2; i++ ){
2051         if(i && !memcmp(h->pps.scaling_matrix8[0], h->pps.scaling_matrix8[1], 64*sizeof(uint8_t))){
2052             h->dequant8_coeff[1] = h->dequant8_buffer[0];
2053             break;
2054         }
2055
2056         for(q=0; q<52; q++){
2057             int shift = ff_div6[q];
2058             int idx = ff_rem6[q];
2059             for(x=0; x<64; x++)
2060                 h->dequant8_coeff[i][q][transpose ? (x>>3)|((x&7)<<3) : x] =
2061                     ((uint32_t)dequant8_coeff_init[idx][ dequant8_coeff_init_scan[((x>>1)&12) | (x&3)] ] *
2062                     h->pps.scaling_matrix8[i][x]) << shift;
2063         }
2064     }
2065 }
2066
2067 static void init_dequant4_coeff_table(H264Context *h){
2068     int i,j,q,x;
2069     const int transpose = (h->s.dsp.h264_idct_add != ff_h264_idct_add_c); //FIXME ugly
2070     for(i=0; i<6; i++ ){
2071         h->dequant4_coeff[i] = h->dequant4_buffer[i];
2072         for(j=0; j<i; j++){
2073             if(!memcmp(h->pps.scaling_matrix4[j], h->pps.scaling_matrix4[i], 16*sizeof(uint8_t))){
2074                 h->dequant4_coeff[i] = h->dequant4_buffer[j];
2075                 break;
2076             }
2077         }
2078         if(j<i)
2079             continue;
2080
2081         for(q=0; q<52; q++){
2082             int shift = ff_div6[q] + 2;
2083             int idx = ff_rem6[q];
2084             for(x=0; x<16; x++)
2085                 h->dequant4_coeff[i][q][transpose ? (x>>2)|((x<<2)&0xF) : x] =
2086                     ((uint32_t)dequant4_coeff_init[idx][(x&1) + ((x>>2)&1)] *
2087                     h->pps.scaling_matrix4[i][x]) << shift;
2088         }
2089     }
2090 }
2091
2092 static void init_dequant_tables(H264Context *h){
2093     int i,x;
2094     init_dequant4_coeff_table(h);
2095     if(h->pps.transform_8x8_mode)
2096         init_dequant8_coeff_table(h);
2097     if(h->sps.transform_bypass){
2098         for(i=0; i<6; i++)
2099             for(x=0; x<16; x++)
2100                 h->dequant4_coeff[i][0][x] = 1<<6;
2101         if(h->pps.transform_8x8_mode)
2102             for(i=0; i<2; i++)
2103                 for(x=0; x<64; x++)
2104                     h->dequant8_coeff[i][0][x] = 1<<6;
2105     }
2106 }
2107
2108
2109 /**
2110  * allocates tables.
2111  * needs width/height
2112  */
2113 static int alloc_tables(H264Context *h){
2114     MpegEncContext * const s = &h->s;
2115     const int big_mb_num= s->mb_stride * (s->mb_height+1);
2116     int x,y;
2117
2118     CHECKED_ALLOCZ(h->intra4x4_pred_mode, big_mb_num * 8  * sizeof(uint8_t))
2119
2120     CHECKED_ALLOCZ(h->non_zero_count    , big_mb_num * 16 * sizeof(uint8_t))
2121     CHECKED_ALLOCZ(h->slice_table_base  , (big_mb_num+s->mb_stride) * sizeof(uint8_t))
2122     CHECKED_ALLOCZ(h->cbp_table, big_mb_num * sizeof(uint16_t))
2123
2124     if( h->pps.cabac ) {
2125         CHECKED_ALLOCZ(h->chroma_pred_mode_table, big_mb_num * sizeof(uint8_t))
2126         CHECKED_ALLOCZ(h->mvd_table[0], 32*big_mb_num * sizeof(uint16_t));
2127         CHECKED_ALLOCZ(h->mvd_table[1], 32*big_mb_num * sizeof(uint16_t));
2128         CHECKED_ALLOCZ(h->direct_table, 32*big_mb_num * sizeof(uint8_t));
2129     }
2130
2131     memset(h->slice_table_base, -1, (big_mb_num+s->mb_stride)  * sizeof(uint8_t));
2132     h->slice_table= h->slice_table_base + s->mb_stride*2 + 1;
2133
2134     CHECKED_ALLOCZ(h->mb2b_xy  , big_mb_num * sizeof(uint32_t));
2135     CHECKED_ALLOCZ(h->mb2b8_xy , big_mb_num * sizeof(uint32_t));
2136     for(y=0; y<s->mb_height; y++){
2137         for(x=0; x<s->mb_width; x++){
2138             const int mb_xy= x + y*s->mb_stride;
2139             const int b_xy = 4*x + 4*y*h->b_stride;
2140             const int b8_xy= 2*x + 2*y*h->b8_stride;
2141
2142             h->mb2b_xy [mb_xy]= b_xy;
2143             h->mb2b8_xy[mb_xy]= b8_xy;
2144         }
2145     }
2146
2147     s->obmc_scratchpad = NULL;
2148
2149     if(!h->dequant4_coeff[0])
2150         init_dequant_tables(h);
2151
2152     return 0;
2153 fail:
2154     free_tables(h);
2155     return -1;
2156 }
2157
2158 /**
2159  * Mimic alloc_tables(), but for every context thread.
2160  */
2161 static void clone_tables(H264Context *dst, H264Context *src){
2162     dst->intra4x4_pred_mode       = src->intra4x4_pred_mode;
2163     dst->non_zero_count           = src->non_zero_count;
2164     dst->slice_table              = src->slice_table;
2165     dst->cbp_table                = src->cbp_table;
2166     dst->mb2b_xy                  = src->mb2b_xy;
2167     dst->mb2b8_xy                 = src->mb2b8_xy;
2168     dst->chroma_pred_mode_table   = src->chroma_pred_mode_table;
2169     dst->mvd_table[0]             = src->mvd_table[0];
2170     dst->mvd_table[1]             = src->mvd_table[1];
2171     dst->direct_table             = src->direct_table;
2172
2173     dst->s.obmc_scratchpad = NULL;
2174     ff_h264_pred_init(&dst->hpc, src->s.codec_id);
2175 }
2176
2177 /**
2178  * Init context
2179  * Allocate buffers which are not shared amongst multiple threads.
2180  */
2181 static int context_init(H264Context *h){
2182     MpegEncContext * const s = &h->s;
2183
2184     CHECKED_ALLOCZ(h->top_borders[0], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2185     CHECKED_ALLOCZ(h->top_borders[1], h->s.mb_width * (16+8+8) * sizeof(uint8_t))
2186
2187     // edge emu needs blocksize + filter length - 1 (=17x17 for halfpel / 21x21 for h264)
2188     CHECKED_ALLOCZ(s->allocated_edge_emu_buffer,
2189                    (s->width+64)*2*21*2); //(width + edge + align)*interlaced*MBsize*tolerance
2190     s->edge_emu_buffer= s->allocated_edge_emu_buffer + (s->width+64)*2*21;
2191     return 0;
2192 fail:
2193     return -1; // free_tables will clean up for us
2194 }
2195
2196 static void common_init(H264Context *h){
2197     MpegEncContext * const s = &h->s;
2198
2199     s->width = s->avctx->width;
2200     s->height = s->avctx->height;
2201     s->codec_id= s->avctx->codec->id;
2202
2203     ff_h264_pred_init(&h->hpc, s->codec_id);
2204
2205     h->dequant_coeff_pps= -1;
2206     s->unrestricted_mv=1;
2207     s->decode=1; //FIXME
2208
2209     memset(h->pps.scaling_matrix4, 16, 6*16*sizeof(uint8_t));
2210     memset(h->pps.scaling_matrix8, 16, 2*64*sizeof(uint8_t));
2211 }
2212
2213 static int decode_init(AVCodecContext *avctx){
2214     H264Context *h= avctx->priv_data;
2215     MpegEncContext * const s = &h->s;
2216
2217     MPV_decode_defaults(s);
2218
2219     s->avctx = avctx;
2220     common_init(h);
2221
2222     s->out_format = FMT_H264;
2223     s->workaround_bugs= avctx->workaround_bugs;
2224
2225     // set defaults
2226 //    s->decode_mb= ff_h263_decode_mb;
2227     s->quarter_sample = 1;
2228     s->low_delay= 1;
2229     avctx->pix_fmt= PIX_FMT_YUV420P;
2230
2231     decode_init_vlc();
2232
2233     if(avctx->extradata_size > 0 && avctx->extradata &&
2234        *(char *)avctx->extradata == 1){
2235         h->is_avc = 1;
2236         h->got_avcC = 0;
2237     } else {
2238         h->is_avc = 0;
2239     }
2240
2241     h->thread_context[0] = h;
2242     return 0;
2243 }
2244
2245 static int frame_start(H264Context *h){
2246     MpegEncContext * const s = &h->s;
2247     int i;
2248
2249     if(MPV_frame_start(s, s->avctx) < 0)
2250         return -1;
2251     ff_er_frame_start(s);
2252     /*
2253      * MPV_frame_start uses pict_type to derive key_frame.
2254      * This is incorrect for H.264; IDR markings must be used.
2255      * Zero here; IDR markings per slice in frame or fields are OR'd in later.
2256      * See decode_nal_units().
2257      */
2258     s->current_picture_ptr->key_frame= 0;
2259
2260     assert(s->linesize && s->uvlinesize);
2261
2262     for(i=0; i<16; i++){
2263         h->block_offset[i]= 4*((scan8[i] - scan8[0])&7) + 4*s->linesize*((scan8[i] - scan8[0])>>3);
2264         h->block_offset[24+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->linesize*((scan8[i] - scan8[0])>>3);
2265     }
2266     for(i=0; i<4; i++){
2267         h->block_offset[16+i]=
2268         h->block_offset[20+i]= 4*((scan8[i] - scan8[0])&7) + 4*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2269         h->block_offset[24+16+i]=
2270         h->block_offset[24+20+i]= 4*((scan8[i] - scan8[0])&7) + 8*s->uvlinesize*((scan8[i] - scan8[0])>>3);
2271     }
2272
2273     /* can't be in alloc_tables because linesize isn't known there.
2274      * FIXME: redo bipred weight to not require extra buffer? */
2275     for(i = 0; i < s->avctx->thread_count; i++)
2276         if(!h->thread_context[i]->s.obmc_scratchpad)
2277             h->thread_context[i]->s.obmc_scratchpad = av_malloc(16*2*s->linesize + 8*2*s->uvlinesize);
2278
2279     /* some macroblocks will be accessed before they're available */
2280     if(FRAME_MBAFF || s->avctx->thread_count > 1)
2281         memset(h->slice_table, -1, (s->mb_height*s->mb_stride-1) * sizeof(uint8_t));
2282
2283 //    s->decode= (s->flags&CODEC_FLAG_PSNR) || !s->encoding || s->current_picture.reference /*|| h->contains_intra*/ || 1;
2284     return 0;
2285 }
2286
2287 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){
2288     MpegEncContext * const s = &h->s;
2289     int i;
2290
2291     src_y  -=   linesize;
2292     src_cb -= uvlinesize;
2293     src_cr -= uvlinesize;
2294
2295     // There are two lines saved, the line above the the top macroblock of a pair,
2296     // and the line above the bottom macroblock
2297     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2298     for(i=1; i<17; i++){
2299         h->left_border[i]= src_y[15+i*  linesize];
2300     }
2301
2302     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  16*linesize);
2303     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+16*linesize);
2304
2305     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2306         h->left_border[17  ]= h->top_borders[0][s->mb_x][16+7];
2307         h->left_border[17+9]= h->top_borders[0][s->mb_x][24+7];
2308         for(i=1; i<9; i++){
2309             h->left_border[i+17  ]= src_cb[7+i*uvlinesize];
2310             h->left_border[i+17+9]= src_cr[7+i*uvlinesize];
2311         }
2312         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+8*uvlinesize);
2313         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+8*uvlinesize);
2314     }
2315 }
2316
2317 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){
2318     MpegEncContext * const s = &h->s;
2319     int temp8, i;
2320     uint64_t temp64;
2321     int deblock_left;
2322     int deblock_top;
2323     int mb_xy;
2324
2325     if(h->deblocking_filter == 2) {
2326         mb_xy = s->mb_x + s->mb_y*s->mb_stride;
2327         deblock_left = h->slice_table[mb_xy] == h->slice_table[mb_xy - 1];
2328         deblock_top  = h->slice_table[mb_xy] == h->slice_table[h->top_mb_xy];
2329     } else {
2330         deblock_left = (s->mb_x > 0);
2331         deblock_top =  (s->mb_y > 0);
2332     }
2333
2334     src_y  -=   linesize + 1;
2335     src_cb -= uvlinesize + 1;
2336     src_cr -= uvlinesize + 1;
2337
2338 #define XCHG(a,b,t,xchg)\
2339 t= a;\
2340 if(xchg)\
2341     a= b;\
2342 b= t;
2343
2344     if(deblock_left){
2345         for(i = !deblock_top; i<17; i++){
2346             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2347         }
2348     }
2349
2350     if(deblock_top){
2351         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2352         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2353         if(s->mb_x+1 < s->mb_width){
2354             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2355         }
2356     }
2357
2358     if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2359         if(deblock_left){
2360             for(i = !deblock_top; i<9; i++){
2361                 XCHG(h->left_border[i+17  ], src_cb[i*uvlinesize], temp8, xchg);
2362                 XCHG(h->left_border[i+17+9], src_cr[i*uvlinesize], temp8, xchg);
2363             }
2364         }
2365         if(deblock_top){
2366             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2367             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2368         }
2369     }
2370 }
2371
2372 static inline void backup_pair_border(H264Context *h, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr, int linesize, int uvlinesize){
2373     MpegEncContext * const s = &h->s;
2374     int i;
2375
2376     src_y  -= 2 *   linesize;
2377     src_cb -= 2 * uvlinesize;
2378     src_cr -= 2 * uvlinesize;
2379
2380     // There are two lines saved, the line above the the top macroblock of a pair,
2381     // and the line above the bottom macroblock
2382     h->left_border[0]= h->top_borders[0][s->mb_x][15];
2383     h->left_border[1]= h->top_borders[1][s->mb_x][15];
2384     for(i=2; i<34; i++){
2385         h->left_border[i]= src_y[15+i*  linesize];
2386     }
2387
2388     *(uint64_t*)(h->top_borders[0][s->mb_x]+0)= *(uint64_t*)(src_y +  32*linesize);
2389     *(uint64_t*)(h->top_borders[0][s->mb_x]+8)= *(uint64_t*)(src_y +8+32*linesize);
2390     *(uint64_t*)(h->top_borders[1][s->mb_x]+0)= *(uint64_t*)(src_y +  33*linesize);
2391     *(uint64_t*)(h->top_borders[1][s->mb_x]+8)= *(uint64_t*)(src_y +8+33*linesize);
2392
2393     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2394         h->left_border[34     ]= h->top_borders[0][s->mb_x][16+7];
2395         h->left_border[34+   1]= h->top_borders[1][s->mb_x][16+7];
2396         h->left_border[34+18  ]= h->top_borders[0][s->mb_x][24+7];
2397         h->left_border[34+18+1]= h->top_borders[1][s->mb_x][24+7];
2398         for(i=2; i<18; i++){
2399             h->left_border[i+34   ]= src_cb[7+i*uvlinesize];
2400             h->left_border[i+34+18]= src_cr[7+i*uvlinesize];
2401         }
2402         *(uint64_t*)(h->top_borders[0][s->mb_x]+16)= *(uint64_t*)(src_cb+16*uvlinesize);
2403         *(uint64_t*)(h->top_borders[0][s->mb_x]+24)= *(uint64_t*)(src_cr+16*uvlinesize);
2404         *(uint64_t*)(h->top_borders[1][s->mb_x]+16)= *(uint64_t*)(src_cb+17*uvlinesize);
2405         *(uint64_t*)(h->top_borders[1][s->mb_x]+24)= *(uint64_t*)(src_cr+17*uvlinesize);
2406     }
2407 }
2408
2409 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){
2410     MpegEncContext * const s = &h->s;
2411     int temp8, i;
2412     uint64_t temp64;
2413     int deblock_left = (s->mb_x > 0);
2414     int deblock_top  = (s->mb_y > 1);
2415
2416     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);
2417
2418     src_y  -= 2 *   linesize + 1;
2419     src_cb -= 2 * uvlinesize + 1;
2420     src_cr -= 2 * uvlinesize + 1;
2421
2422 #define XCHG(a,b,t,xchg)\
2423 t= a;\
2424 if(xchg)\
2425     a= b;\
2426 b= t;
2427
2428     if(deblock_left){
2429         for(i = (!deblock_top)<<1; i<34; i++){
2430             XCHG(h->left_border[i     ], src_y [i*  linesize], temp8, xchg);
2431         }
2432     }
2433
2434     if(deblock_top){
2435         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+0), *(uint64_t*)(src_y +1), temp64, xchg);
2436         XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+8), *(uint64_t*)(src_y +9), temp64, 1);
2437         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+0), *(uint64_t*)(src_y +1 +linesize), temp64, xchg);
2438         XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+8), *(uint64_t*)(src_y +9 +linesize), temp64, 1);
2439         if(s->mb_x+1 < s->mb_width){
2440             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x+1]), *(uint64_t*)(src_y +17), temp64, 1);
2441             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x+1]), *(uint64_t*)(src_y +17 +linesize), temp64, 1);
2442         }
2443     }
2444
2445     if(!ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2446         if(deblock_left){
2447             for(i = (!deblock_top) << 1; i<18; i++){
2448                 XCHG(h->left_border[i+34   ], src_cb[i*uvlinesize], temp8, xchg);
2449                 XCHG(h->left_border[i+34+18], src_cr[i*uvlinesize], temp8, xchg);
2450             }
2451         }
2452         if(deblock_top){
2453             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+16), *(uint64_t*)(src_cb+1), temp64, 1);
2454             XCHG(*(uint64_t*)(h->top_borders[0][s->mb_x]+24), *(uint64_t*)(src_cr+1), temp64, 1);
2455             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+16), *(uint64_t*)(src_cb+1 +uvlinesize), temp64, 1);
2456             XCHG(*(uint64_t*)(h->top_borders[1][s->mb_x]+24), *(uint64_t*)(src_cr+1 +uvlinesize), temp64, 1);
2457         }
2458     }
2459 }
2460
2461 static av_always_inline void hl_decode_mb_internal(H264Context *h, int simple){
2462     MpegEncContext * const s = &h->s;
2463     const int mb_x= s->mb_x;
2464     const int mb_y= s->mb_y;
2465     const int mb_xy= mb_x + mb_y*s->mb_stride;
2466     const int mb_type= s->current_picture.mb_type[mb_xy];
2467     uint8_t  *dest_y, *dest_cb, *dest_cr;
2468     int linesize, uvlinesize /*dct_offset*/;
2469     int i;
2470     int *block_offset = &h->block_offset[0];
2471     const unsigned int bottom = mb_y & 1;
2472     const int transform_bypass = (s->qscale == 0 && h->sps.transform_bypass), is_h264 = (simple || s->codec_id == CODEC_ID_H264);
2473     void (*idct_add)(uint8_t *dst, DCTELEM *block, int stride);
2474     void (*idct_dc_add)(uint8_t *dst, DCTELEM *block, int stride);
2475
2476     dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2477     dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2478     dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2479
2480     s->dsp.prefetch(dest_y + (s->mb_x&3)*4*s->linesize + 64, s->linesize, 4);
2481     s->dsp.prefetch(dest_cb + (s->mb_x&7)*s->uvlinesize + 64, dest_cr - dest_cb, 2);
2482
2483     if (!simple && MB_FIELD) {
2484         linesize   = h->mb_linesize   = s->linesize * 2;
2485         uvlinesize = h->mb_uvlinesize = s->uvlinesize * 2;
2486         block_offset = &h->block_offset[24];
2487         if(mb_y&1){ //FIXME move out of this func?
2488             dest_y -= s->linesize*15;
2489             dest_cb-= s->uvlinesize*7;
2490             dest_cr-= s->uvlinesize*7;
2491         }
2492         if(FRAME_MBAFF) {
2493             int list;
2494             for(list=0; list<h->list_count; list++){
2495                 if(!USES_LIST(mb_type, list))
2496                     continue;
2497                 if(IS_16X16(mb_type)){
2498                     int8_t *ref = &h->ref_cache[list][scan8[0]];
2499                     fill_rectangle(ref, 4, 4, 8, 16+*ref^(s->mb_y&1), 1);
2500                 }else{
2501                     for(i=0; i<16; i+=4){
2502                         //FIXME can refs be smaller than 8x8 when !direct_8x8_inference ?
2503                         int ref = h->ref_cache[list][scan8[i]];
2504                         if(ref >= 0)
2505                             fill_rectangle(&h->ref_cache[list][scan8[i]], 2, 2, 8, 16+ref^(s->mb_y&1), 1);
2506                     }
2507                 }
2508             }
2509         }
2510     } else {
2511         linesize   = h->mb_linesize   = s->linesize;
2512         uvlinesize = h->mb_uvlinesize = s->uvlinesize;
2513 //        dct_offset = s->linesize * 16;
2514     }
2515
2516     if(transform_bypass){
2517         idct_dc_add =
2518         idct_add = IS_8x8DCT(mb_type) ? s->dsp.add_pixels8 : s->dsp.add_pixels4;
2519     }else if(IS_8x8DCT(mb_type)){
2520         idct_dc_add = s->dsp.h264_idct8_dc_add;
2521         idct_add = s->dsp.h264_idct8_add;
2522     }else{
2523         idct_dc_add = s->dsp.h264_idct_dc_add;
2524         idct_add = s->dsp.h264_idct_add;
2525     }
2526
2527     if(!simple && FRAME_MBAFF && h->deblocking_filter && IS_INTRA(mb_type)
2528        && (!bottom || !IS_INTRA(s->current_picture.mb_type[mb_xy-s->mb_stride]))){
2529         int mbt_y = mb_y&~1;
2530         uint8_t *top_y  = s->current_picture.data[0] + (mbt_y * 16* s->linesize  ) + mb_x * 16;
2531         uint8_t *top_cb = s->current_picture.data[1] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2532         uint8_t *top_cr = s->current_picture.data[2] + (mbt_y * 8 * s->uvlinesize) + mb_x * 8;
2533         xchg_pair_border(h, top_y, top_cb, top_cr, s->linesize, s->uvlinesize, 1);
2534     }
2535
2536     if (!simple && IS_INTRA_PCM(mb_type)) {
2537         unsigned int x, y;
2538
2539         // The pixels are stored in h->mb array in the same order as levels,
2540         // copy them in output in the correct order.
2541         for(i=0; i<16; i++) {
2542             for (y=0; y<4; y++) {
2543                 for (x=0; x<4; x++) {
2544                     *(dest_y + block_offset[i] + y*linesize + x) = h->mb[i*16+y*4+x];
2545                 }
2546             }
2547         }
2548         for(i=16; i<16+4; i++) {
2549             for (y=0; y<4; y++) {
2550                 for (x=0; x<4; x++) {
2551                     *(dest_cb + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2552                 }
2553             }
2554         }
2555         for(i=20; i<20+4; i++) {
2556             for (y=0; y<4; y++) {
2557                 for (x=0; x<4; x++) {
2558                     *(dest_cr + block_offset[i] + y*uvlinesize + x) = h->mb[i*16+y*4+x];
2559                 }
2560             }
2561         }
2562     } else {
2563         if(IS_INTRA(mb_type)){
2564             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2565                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 1, simple);
2566
2567             if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2568                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cb, uvlinesize);
2569                 h->hpc.pred8x8[ h->chroma_pred_mode ](dest_cr, uvlinesize);
2570             }
2571
2572             if(IS_INTRA4x4(mb_type)){
2573                 if(simple || !s->encoding){
2574                     if(IS_8x8DCT(mb_type)){
2575                         for(i=0; i<16; i+=4){
2576                             uint8_t * const ptr= dest_y + block_offset[i];
2577                             const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2578                             const int nnz = h->non_zero_count_cache[ scan8[i] ];
2579                             h->hpc.pred8x8l[ dir ](ptr, (h->topleft_samples_available<<i)&0x8000,
2580                                                    (h->topright_samples_available<<i)&0x4000, linesize);
2581                             if(nnz){
2582                                 if(nnz == 1 && h->mb[i*16])
2583                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2584                                 else
2585                                     idct_add(ptr, h->mb + i*16, linesize);
2586                             }
2587                         }
2588                     }else
2589                     for(i=0; i<16; i++){
2590                         uint8_t * const ptr= dest_y + block_offset[i];
2591                         uint8_t *topright;
2592                         const int dir= h->intra4x4_pred_mode_cache[ scan8[i] ];
2593                         int nnz, tr;
2594
2595                         if(dir == DIAG_DOWN_LEFT_PRED || dir == VERT_LEFT_PRED){
2596                             const int topright_avail= (h->topright_samples_available<<i)&0x8000;
2597                             assert(mb_y || linesize <= block_offset[i]);
2598                             if(!topright_avail){
2599                                 tr= ptr[3 - linesize]*0x01010101;
2600                                 topright= (uint8_t*) &tr;
2601                             }else
2602                                 topright= ptr + 4 - linesize;
2603                         }else
2604                             topright= NULL;
2605
2606                         h->hpc.pred4x4[ dir ](ptr, topright, linesize);
2607                         nnz = h->non_zero_count_cache[ scan8[i] ];
2608                         if(nnz){
2609                             if(is_h264){
2610                                 if(nnz == 1 && h->mb[i*16])
2611                                     idct_dc_add(ptr, h->mb + i*16, linesize);
2612                                 else
2613                                     idct_add(ptr, h->mb + i*16, linesize);
2614                             }else
2615                                 svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, 0);
2616                         }
2617                     }
2618                 }
2619             }else{
2620                 h->hpc.pred16x16[ h->intra16x16_pred_mode ](dest_y , linesize);
2621                 if(is_h264){
2622                     if(!transform_bypass)
2623                         h264_luma_dc_dequant_idct_c(h->mb, s->qscale, h->dequant4_coeff[0][s->qscale][0]);
2624                 }else
2625                     svq3_luma_dc_dequant_idct_c(h->mb, s->qscale);
2626             }
2627             if(h->deblocking_filter && (simple || !FRAME_MBAFF))
2628                 xchg_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, 0, simple);
2629         }else if(is_h264){
2630             hl_motion(h, dest_y, dest_cb, dest_cr,
2631                       s->me.qpel_put, s->dsp.put_h264_chroma_pixels_tab,
2632                       s->me.qpel_avg, s->dsp.avg_h264_chroma_pixels_tab,
2633                       s->dsp.weight_h264_pixels_tab, s->dsp.biweight_h264_pixels_tab);
2634         }
2635
2636
2637         if(!IS_INTRA4x4(mb_type)){
2638             if(is_h264){
2639                 if(IS_INTRA16x16(mb_type)){
2640                     for(i=0; i<16; i++){
2641                         if(h->non_zero_count_cache[ scan8[i] ])
2642                             idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2643                         else if(h->mb[i*16])
2644                             idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2645                     }
2646                 }else{
2647                     const int di = IS_8x8DCT(mb_type) ? 4 : 1;
2648                     for(i=0; i<16; i+=di){
2649                         int nnz = h->non_zero_count_cache[ scan8[i] ];
2650                         if(nnz){
2651                             if(nnz==1 && h->mb[i*16])
2652                                 idct_dc_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2653                             else
2654                                 idct_add(dest_y + block_offset[i], h->mb + i*16, linesize);
2655                         }
2656                     }
2657                 }
2658             }else{
2659                 for(i=0; i<16; i++){
2660                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){ //FIXME benchmark weird rule, & below
2661                         uint8_t * const ptr= dest_y + block_offset[i];
2662                         svq3_add_idct_c(ptr, h->mb + i*16, linesize, s->qscale, IS_INTRA(mb_type) ? 1 : 0);
2663                     }
2664                 }
2665             }
2666         }
2667
2668         if(simple || !ENABLE_GRAY || !(s->flags&CODEC_FLAG_GRAY)){
2669             uint8_t *dest[2] = {dest_cb, dest_cr};
2670             if(transform_bypass){
2671                 idct_add = idct_dc_add = s->dsp.add_pixels4;
2672             }else{
2673                 idct_add = s->dsp.h264_idct_add;
2674                 idct_dc_add = s->dsp.h264_idct_dc_add;
2675                 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]);
2676                 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]);
2677             }
2678             if(is_h264){
2679                 for(i=16; i<16+8; i++){
2680                     if(h->non_zero_count_cache[ scan8[i] ])
2681                         idct_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2682                     else if(h->mb[i*16])
2683                         idct_dc_add(dest[(i&4)>>2] + block_offset[i], h->mb + i*16, uvlinesize);
2684                 }
2685             }else{
2686                 for(i=16; i<16+8; i++){
2687                     if(h->non_zero_count_cache[ scan8[i] ] || h->mb[i*16]){
2688                         uint8_t * const ptr= dest[(i&4)>>2] + block_offset[i];
2689                         svq3_add_idct_c(ptr, h->mb + i*16, uvlinesize, chroma_qp[s->qscale + 12] - 12, 2);
2690                     }
2691                 }
2692             }
2693         }
2694     }
2695     if(h->deblocking_filter) {
2696         if (!simple && FRAME_MBAFF) {
2697             //FIXME try deblocking one mb at a time?
2698             // the reduction in load/storing mvs and such might outweigh the extra backup/xchg_border
2699             const int mb_y = s->mb_y - 1;
2700             uint8_t  *pair_dest_y, *pair_dest_cb, *pair_dest_cr;
2701             const int mb_xy= mb_x + mb_y*s->mb_stride;
2702             const int mb_type_top   = s->current_picture.mb_type[mb_xy];
2703             const int mb_type_bottom= s->current_picture.mb_type[mb_xy+s->mb_stride];
2704             if (!bottom) return;
2705             pair_dest_y  = s->current_picture.data[0] + (mb_y * 16* s->linesize  ) + mb_x * 16;
2706             pair_dest_cb = s->current_picture.data[1] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2707             pair_dest_cr = s->current_picture.data[2] + (mb_y * 8 * s->uvlinesize) + mb_x * 8;
2708
2709             if(IS_INTRA(mb_type_top | mb_type_bottom))
2710                 xchg_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize, 0);
2711
2712             backup_pair_border(h, pair_dest_y, pair_dest_cb, pair_dest_cr, s->linesize, s->uvlinesize);
2713             // deblock a pair
2714             // top
2715             s->mb_y--;
2716             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);
2717             fill_caches(h, mb_type_top, 1); //FIXME don't fill stuff which isn't used by filter_mb
2718             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy]);
2719             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy]);
2720             filter_mb(h, mb_x, mb_y, pair_dest_y, pair_dest_cb, pair_dest_cr, linesize, uvlinesize);
2721             // bottom
2722             s->mb_y++;
2723             tprintf(h->s.avctx, "call mbaff filter_mb\n");
2724             fill_caches(h, mb_type_bottom, 1); //FIXME don't fill stuff which isn't used by filter_mb
2725             h->chroma_qp[0] = get_chroma_qp(h, 0, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2726             h->chroma_qp[1] = get_chroma_qp(h, 1, s->current_picture.qscale_table[mb_xy+s->mb_stride]);
2727             filter_mb(h, mb_x, mb_y+1, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2728         } else {
2729             tprintf(h->s.avctx, "call filter_mb\n");
2730             backup_mb_border(h, dest_y, dest_cb, dest_cr, linesize, uvlinesize, simple);
2731             fill_caches(h, mb_type, 1); //FIXME don't fill stuff which isn't used by filter_mb
2732             filter_mb_fast(h, mb_x, mb_y, dest_y, dest_cb, dest_cr, linesize, uvlinesize);
2733         }
2734     }
2735 }
2736
2737 /**
2738  * Process a macroblock; this case avoids checks for expensive uncommon cases.
2739  */
2740 static void hl_decode_mb_simple(H264Context *h){
2741     hl_decode_mb_internal(h, 1);
2742 }
2743
2744 /**
2745  * Process a macroblock; this handles edge cases, such as interlacing.
2746  */
2747 static void av_noinline hl_decode_mb_complex(H264Context *h){
2748     hl_decode_mb_internal(h, 0);
2749 }
2750
2751 static void hl_decode_mb(H264Context *h){
2752     MpegEncContext * const s = &h->s;
2753     const int mb_x= s->mb_x;
2754     const int mb_y= s->mb_y;
2755     const int mb_xy= mb_x + mb_y*s->mb_stride;
2756     const int mb_type= s->current_picture.mb_type[mb_xy];
2757     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;
2758
2759     if(!s->decode)
2760         return;
2761
2762     if (is_complex)
2763         hl_decode_mb_complex(h);
2764     else hl_decode_mb_simple(h);
2765 }
2766
2767 static void pic_as_field(Picture *pic, const int bottom){
2768     int i;
2769     for (i = 0; i < 4; ++i) {
2770         if (bottom)
2771             pic->data[i] += pic->linesize[i];
2772         pic->linesize[i] *= 2;
2773     }
2774 }
2775
2776 static int split_field_copy(Picture *dest, Picture *src,
2777                             int parity, int id_add){
2778     int match = !!(src->reference & parity);
2779
2780     if (match) {
2781         *dest = *src;
2782         pic_as_field(dest, parity == PICT_BOTTOM_FIELD);
2783         dest->pic_id *= 2;
2784         dest->pic_id += id_add;
2785     }
2786
2787     return match;
2788 }
2789
2790 /**
2791  * Split one reference list into field parts, interleaving by parity
2792  * as per H.264 spec section 8.2.4.2.5. Output fields have their data pointers
2793  * set to look at the actual start of data for that field.
2794  *
2795  * @param dest output list
2796  * @param dest_len maximum number of fields to put in dest
2797  * @param src the source reference list containing fields and/or field pairs
2798  *            (aka short_ref/long_ref, or
2799  *             refFrameListXShortTerm/refFrameListLongTerm in spec-speak)
2800  * @param src_len number of Picture's in source (pairs and unmatched fields)
2801  * @param parity the parity of the picture being decoded/needing
2802  *        these ref pics (PICT_{TOP,BOTTOM}_FIELD)
2803  * @return number of fields placed in dest
2804  */
2805 static int split_field_half_ref_list(Picture *dest, int dest_len,
2806                                      Picture *src,  int src_len,  int parity){
2807     int same_parity   = 1;
2808     int same_i        = 0;
2809     int opp_i         = 0;
2810     int out_i;
2811     int field_output;
2812
2813     for (out_i = 0; out_i < dest_len; out_i += field_output) {
2814         if (same_parity && same_i < src_len) {
2815             field_output = split_field_copy(dest + out_i, src + same_i,
2816                                             parity, 1);
2817             same_parity = !field_output;
2818             same_i++;
2819
2820         } else if (opp_i < src_len) {
2821             field_output = split_field_copy(dest + out_i, src + opp_i,
2822                                             PICT_FRAME - parity, 0);
2823             same_parity = field_output;
2824             opp_i++;
2825
2826         } else {
2827             break;
2828         }
2829     }
2830
2831     return out_i;
2832 }
2833
2834 /**
2835  * Split the reference frame list into a reference field list.
2836  * This implements H.264 spec 8.2.4.2.5 for a combined input list.
2837  * The input list contains both reference field pairs and
2838  * unmatched reference fields; it is ordered as spec describes
2839  * RefPicListX for frames in 8.2.4.2.1 and 8.2.4.2.3, except that
2840  * unmatched field pairs are also present. Conceptually this is equivalent
2841  * to concatenation of refFrameListXShortTerm with refFrameListLongTerm.
2842  *
2843  * @param dest output reference list where ordered fields are to be placed
2844  * @param dest_len max number of fields to place at dest
2845  * @param src source reference list, as described above
2846  * @param src_len number of pictures (pairs and unmatched fields) in src
2847  * @param parity parity of field being currently decoded
2848  *        (one of PICT_{TOP,BOTTOM}_FIELD)
2849  * @param long_i index into src array that holds first long reference picture,
2850  *        or src_len if no long refs present.
2851  */
2852 static int split_field_ref_list(Picture *dest, int dest_len,
2853                                 Picture *src,  int src_len,
2854                                 int parity,    int long_i){
2855
2856     int i = split_field_half_ref_list(dest, dest_len, src, long_i, parity);
2857     dest += i;
2858     dest_len -= i;
2859
2860     i += split_field_half_ref_list(dest, dest_len, src + long_i,
2861                                    src_len - long_i, parity);
2862     return i;
2863 }
2864
2865 /**
2866  * fills the default_ref_list.
2867  */
2868 static int fill_default_ref_list(H264Context *h){
2869     MpegEncContext * const s = &h->s;
2870     int i;
2871     int smallest_poc_greater_than_current = -1;
2872     int structure_sel;
2873     Picture sorted_short_ref[32];
2874     Picture field_entry_list[2][32];
2875     Picture *frame_list[2];
2876
2877     if (FIELD_PICTURE) {
2878         structure_sel = PICT_FRAME;
2879         frame_list[0] = field_entry_list[0];
2880         frame_list[1] = field_entry_list[1];
2881     } else {
2882         structure_sel = 0;
2883         frame_list[0] = h->default_ref_list[0];
2884         frame_list[1] = h->default_ref_list[1];
2885     }
2886
2887     if(h->slice_type==B_TYPE){
2888         int list;
2889         int len[2];
2890         int short_len[2];
2891         int out_i;
2892         int limit= INT_MIN;
2893
2894         /* sort frame according to poc in B slice */
2895         for(out_i=0; out_i<h->short_ref_count; out_i++){
2896             int best_i=INT_MIN;
2897             int best_poc=INT_MAX;
2898
2899             for(i=0; i<h->short_ref_count; i++){
2900                 const int poc= h->short_ref[i]->poc;
2901                 if(poc > limit && poc < best_poc){
2902                     best_poc= poc;
2903                     best_i= i;
2904                 }
2905             }
2906
2907             assert(best_i != INT_MIN);
2908
2909             limit= best_poc;
2910             sorted_short_ref[out_i]= *h->short_ref[best_i];
2911             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);
2912             if (-1 == smallest_poc_greater_than_current) {
2913                 if (h->short_ref[best_i]->poc >= s->current_picture_ptr->poc) {
2914                     smallest_poc_greater_than_current = out_i;
2915                 }
2916             }
2917         }
2918
2919         tprintf(h->s.avctx, "current poc: %d, smallest_poc_greater_than_current: %d\n", s->current_picture_ptr->poc, smallest_poc_greater_than_current);
2920
2921         // find the largest poc
2922         for(list=0; list<2; list++){
2923             int index = 0;
2924             int j= -99;
2925             int step= list ? -1 : 1;
2926
2927             for(i=0; i<h->short_ref_count && index < h->ref_count[list]; i++, j+=step) {
2928                 int sel;
2929                 while(j<0 || j>= h->short_ref_count){
2930                     if(j != -99 && step == (list ? -1 : 1))
2931                         return -1;
2932                     step = -step;
2933                     j= smallest_poc_greater_than_current + (step>>1);
2934                 }
2935                 sel = sorted_short_ref[j].reference | structure_sel;
2936                 if(sel != PICT_FRAME) continue;
2937                 frame_list[list][index  ]= sorted_short_ref[j];
2938                 frame_list[list][index++].pic_id= sorted_short_ref[j].frame_num;
2939             }
2940             short_len[list] = index;
2941
2942             for(i = 0; i < 16 && index < h->ref_count[ list ]; i++){
2943                 int sel;
2944                 if(h->long_ref[i] == NULL) continue;
2945                 sel = h->long_ref[i]->reference | structure_sel;
2946                 if(sel != PICT_FRAME) continue;
2947
2948                 frame_list[ list ][index  ]= *h->long_ref[i];
2949                 frame_list[ list ][index++].pic_id= i;;
2950             }
2951             len[list] = index;
2952
2953             if(list && (smallest_poc_greater_than_current<=0 || smallest_poc_greater_than_current>=h->short_ref_count) && (1 < index)){
2954                 // swap the two first elements of L1 when
2955                 // L0 and L1 are identical
2956                 Picture temp= frame_list[1][0];
2957                 frame_list[1][0] = frame_list[1][1];
2958                 frame_list[1][1] = temp;
2959             }
2960
2961         }
2962
2963         for(list=0; list<2; list++){
2964             if (FIELD_PICTURE)
2965                 len[list] = split_field_ref_list(h->default_ref_list[list],
2966                                                  h->ref_count[list],
2967                                                  frame_list[list],
2968                                                  len[list],
2969                                                  s->picture_structure,
2970                                                  short_len[list]);
2971
2972             if(len[list] < h->ref_count[ list ])
2973                 memset(&h->default_ref_list[list][len[list]], 0, sizeof(Picture)*(h->ref_count[ list ] - len[list]));
2974         }
2975
2976
2977     }else{
2978         int index=0;
2979         int short_len;
2980         for(i=0; i<h->short_ref_count; i++){
2981             int sel;
2982             sel = h->short_ref[i]->reference | structure_sel;
2983             if(sel != PICT_FRAME) continue;
2984             frame_list[0][index  ]= *h->short_ref[i];
2985             frame_list[0][index++].pic_id= h->short_ref[i]->frame_num;
2986         }
2987         short_len = index;
2988         for(i = 0; i < 16; i++){
2989             int sel;
2990             if(h->long_ref[i] == NULL) continue;
2991             sel = h->long_ref[i]->reference | structure_sel;
2992             if(sel != PICT_FRAME) continue;
2993             frame_list[0][index  ]= *h->long_ref[i];
2994             frame_list[0][index++].pic_id= i;;
2995         }
2996
2997         if (FIELD_PICTURE)
2998             index = split_field_ref_list(h->default_ref_list[0],
2999                                          h->ref_count[0], frame_list[0],
3000                                          index, s->picture_structure,
3001                                          short_len);
3002
3003         if(index < h->ref_count[0])
3004             memset(&h->default_ref_list[0][index], 0, sizeof(Picture)*(h->ref_count[0] - index));
3005     }
3006 #ifdef TRACE
3007     for (i=0; i<h->ref_count[0]; i++) {
3008         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]);
3009     }
3010     if(h->slice_type==B_TYPE){
3011         for (i=0; i<h->ref_count[1]; i++) {
3012             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]);
3013         }
3014     }
3015 #endif
3016     return 0;
3017 }
3018
3019 static void print_short_term(H264Context *h);
3020 static void print_long_term(H264Context *h);
3021
3022 /**
3023  * Extract structure information about the picture described by pic_num in
3024  * the current decoding context (frame or field). Note that pic_num is
3025  * picture number without wrapping (so, 0<=pic_num<max_pic_num).
3026  * @param pic_num picture number for which to extract structure information
3027  * @param structure one of PICT_XXX describing structure of picture
3028  *                      with pic_num
3029  * @return frame number (short term) or long term index of picture
3030  *         described by pic_num
3031  */
3032 static int pic_num_extract(H264Context *h, int pic_num, int *structure){
3033     MpegEncContext * const s = &h->s;
3034
3035     *structure = s->picture_structure;
3036     if(FIELD_PICTURE){
3037         if (!(pic_num & 1))
3038             /* opposite field */
3039             *structure ^= PICT_FRAME;
3040         pic_num >>= 1;
3041     }
3042
3043     return pic_num;
3044 }
3045
3046 static int decode_ref_pic_list_reordering(H264Context *h){
3047     MpegEncContext * const s = &h->s;
3048     int list, index, pic_structure;
3049
3050     print_short_term(h);
3051     print_long_term(h);
3052     if(h->slice_type==I_TYPE || h->slice_type==SI_TYPE) return 0; //FIXME move before func
3053
3054     for(list=0; list<h->list_count; list++){
3055         memcpy(h->ref_list[list], h->default_ref_list[list], sizeof(Picture)*h->ref_count[list]);
3056
3057         if(get_bits1(&s->gb)){
3058             int pred= h->curr_pic_num;
3059
3060             for(index=0; ; index++){
3061                 unsigned int reordering_of_pic_nums_idc= get_ue_golomb(&s->gb);
3062                 unsigned int pic_id;
3063                 int i;
3064                 Picture *ref = NULL;
3065
3066                 if(reordering_of_pic_nums_idc==3)
3067                     break;
3068
3069                 if(index >= h->ref_count[list]){
3070                     av_log(h->s.avctx, AV_LOG_ERROR, "reference count overflow\n");
3071                     return -1;
3072                 }
3073
3074                 if(reordering_of_pic_nums_idc<3){
3075                     if(reordering_of_pic_nums_idc<2){
3076                         const unsigned int abs_diff_pic_num= get_ue_golomb(&s->gb) + 1;
3077                         int frame_num;
3078
3079                         if(abs_diff_pic_num >= h->max_pic_num){
3080                             av_log(h->s.avctx, AV_LOG_ERROR, "abs_diff_pic_num overflow\n");
3081                             return -1;
3082                         }
3083
3084                         if(reordering_of_pic_nums_idc == 0) pred-= abs_diff_pic_num;
3085                         else                                pred+= abs_diff_pic_num;
3086                         pred &= h->max_pic_num - 1;
3087
3088                         frame_num = pic_num_extract(h, pred, &pic_structure);
3089
3090                         for(i= h->short_ref_count-1; i>=0; i--){
3091                             ref = h->short_ref[i];
3092                             assert(ref->reference);
3093                             assert(!ref->long_ref);
3094                             if(ref->data[0] != NULL &&
3095                                    ref->frame_num == frame_num &&
3096                                    (ref->reference & pic_structure) &&
3097                                    ref->long_ref == 0) // ignore non existing pictures by testing data[0] pointer
3098                                 break;
3099                         }
3100                         if(i>=0)
3101                             ref->pic_id= pred;
3102                     }else{
3103                         int long_idx;
3104                         pic_id= get_ue_golomb(&s->gb); //long_term_pic_idx
3105
3106                         long_idx= pic_num_extract(h, pic_id, &pic_structure);
3107
3108                         if(long_idx>31){
3109                             av_log(h->s.avctx, AV_LOG_ERROR, "long_term_pic_idx overflow\n");
3110                             return -1;
3111                         }
3112                         ref = h->long_ref[long_idx];
3113                         assert(!(ref && !ref->reference));
3114                         if(ref && (ref->reference & pic_structure)){
3115                             ref->pic_id= pic_id;
3116                             assert(ref->long_ref);
3117                             i=0;
3118                         }else{
3119                             i=-1;
3120                         }
3121                     }
3122
3123                     if (i < 0) {
3124                         av_log(h->s.avctx, AV_LOG_ERROR, "reference picture missing during reorder\n");
3125                         memset(&h->ref_list[list][index], 0, sizeof(Picture)); //FIXME
3126                     } else {
3127                         for(i=index; i+1<h->ref_count[list]; i++){
3128                             if(ref->long_ref == h->ref_list[list][i].long_ref && ref->pic_id == h->ref_list[list][i].pic_id)
3129                                 break;
3130                         }
3131                         for(; i > index; i--){
3132                             h->ref_list[list][i]= h->ref_list[list][i-1];
3133                         }
3134                         h->ref_list[list][index]= *ref;
3135                         if (FIELD_PICTURE){
3136                             int bot = pic_structure == PICT_BOTTOM_FIELD;
3137                             pic_as_field(&h->ref_list[list][index], bot);
3138                         }
3139                     }
3140                 }else{
3141                     av_log(h->s.avctx, AV_LOG_ERROR, "illegal reordering_of_pic_nums_idc\n");
3142                     return -1;
3143                 }
3144             }
3145         }
3146     }
3147     for(list=0; list<h->list_count; list++){
3148         for(index= 0; index < h->ref_count[list]; index++){
3149             if(!h->ref_list[list][index].data[0])
3150                 h->ref_list[list][index]= s->current_picture;
3151         }
3152     }
3153
3154     if(h->slice_type==B_TYPE && !h->direct_spatial_mv_pred)
3155         direct_dist_scale_factor(h);
3156     direct_ref_list_init(h);
3157     return 0;
3158 }
3159
3160 static void fill_mbaff_ref_list(H264Context *h){
3161     int list, i, j;
3162     for(list=0; list<2; list++){ //FIXME try list_count
3163         for(i=0; i<h->ref_count[list]; i++){
3164             Picture *frame = &h->ref_list[list][i];
3165             Picture *field = &h->ref_list[list][16+2*i];
3166             field[0] = *frame;
3167             for(j=0; j<3; j++)
3168                 field[0].linesize[j] <<= 1;
3169             field[1] = field[0];
3170             for(j=0; j<3; j++)
3171                 field[1].data[j] += frame->linesize[j];
3172
3173             h->luma_weight[list][16+2*i] = h->luma_weight[list][16+2*i+1] = h->luma_weight[list][i];
3174             h->luma_offset[list][16+2*i] = h->luma_offset[list][16+2*i+1] = h->luma_offset[list][i];
3175             for(j=0; j<2; j++){
3176                 h->chroma_weight[list][16+2*i][j] = h->chroma_weight[list][16+2*i+1][j] = h->chroma_weight[list][i][j];
3177                 h->chroma_offset[list][16+2*i][j] = h->chroma_offset[list][16+2*i+1][j] = h->chroma_offset[list][i][j];
3178             }
3179         }
3180     }
3181     for(j=0; j<h->ref_count[1]; j++){
3182         for(i=0; i<h->ref_count[0]; i++)
3183             h->implicit_weight[j][16+2*i] = h->implicit_weight[j][16+2*i+1] = h->implicit_weight[j][i];
3184         memcpy(h->implicit_weight[16+2*j],   h->implicit_weight[j], sizeof(*h->implicit_weight));
3185         memcpy(h->implicit_weight[16+2*j+1], h->implicit_weight[j], sizeof(*h->implicit_weight));
3186     }
3187 }
3188
3189 static int pred_weight_table(H264Context *h){
3190     MpegEncContext * const s = &h->s;
3191     int list, i;
3192     int luma_def, chroma_def;
3193
3194     h->use_weight= 0;
3195     h->use_weight_chroma= 0;
3196     h->luma_log2_weight_denom= get_ue_golomb(&s->gb);
3197     h->chroma_log2_weight_denom= get_ue_golomb(&s->gb);
3198     luma_def = 1<<h->luma_log2_weight_denom;
3199     chroma_def = 1<<h->chroma_log2_weight_denom;
3200
3201     for(list=0; list<2; list++){
3202         for(i=0; i<h->ref_count[list]; i++){
3203             int luma_weight_flag, chroma_weight_flag;
3204
3205             luma_weight_flag= get_bits1(&s->gb);
3206             if(luma_weight_flag){
3207                 h->luma_weight[list][i]= get_se_golomb(&s->gb);
3208                 h->luma_offset[list][i]= get_se_golomb(&s->gb);
3209                 if(   h->luma_weight[list][i] != luma_def
3210                    || h->luma_offset[list][i] != 0)
3211                     h->use_weight= 1;
3212             }else{
3213                 h->luma_weight[list][i]= luma_def;
3214                 h->luma_offset[list][i]= 0;
3215             }
3216
3217             chroma_weight_flag= get_bits1(&s->gb);
3218             if(chroma_weight_flag){
3219                 int j;
3220                 for(j=0; j<2; j++){
3221                     h->chroma_weight[list][i][j]= get_se_golomb(&s->gb);
3222                     h->chroma_offset[list][i][j]= get_se_golomb(&s->gb);
3223                     if(   h->chroma_weight[list][i][j] != chroma_def
3224                        || h->chroma_offset[list][i][j] != 0)
3225                         h->use_weight_chroma= 1;
3226                 }
3227             }else{
3228                 int j;
3229                 for(j=0; j<2; j++){
3230                     h->chroma_weight[list][i][j]= chroma_def;
3231                     h->chroma_offset[list][i][j]= 0;
3232                 }
3233             }
3234         }
3235         if(h->slice_type != B_TYPE) break;
3236     }
3237     h->use_weight= h->use_weight || h->use_weight_chroma;
3238     return 0;
3239 }
3240
3241 static void implicit_weight_table(H264Context *h){
3242     MpegEncContext * const s = &h->s;
3243     int ref0, ref1;
3244     int cur_poc = s->current_picture_ptr->poc;
3245
3246     if(   h->ref_count[0] == 1 && h->ref_count[1] == 1
3247        && h->ref_list[0][0].poc + h->ref_list[1][0].poc == 2*cur_poc){
3248         h->use_weight= 0;
3249         h->use_weight_chroma= 0;
3250         return;
3251     }
3252
3253     h->use_weight= 2;
3254     h->use_weight_chroma= 2;
3255     h->luma_log2_weight_denom= 5;
3256     h->chroma_log2_weight_denom= 5;
3257
3258     for(ref0=0; ref0 < h->ref_count[0]; ref0++){
3259         int poc0 = h->ref_list[0][ref0].poc;
3260         for(ref1=0; ref1 < h->ref_count[1]; ref1++){
3261             int poc1 = h->ref_list[1][ref1].poc;
3262             int td = av_clip(poc1 - poc0, -128, 127);
3263             if(td){
3264                 int tb = av_clip(cur_poc - poc0, -128, 127);
3265                 int tx = (16384 + (FFABS(td) >> 1)) / td;
3266                 int dist_scale_factor = av_clip((tb*tx + 32) >> 6, -1024, 1023) >> 2;
3267                 if(dist_scale_factor < -64 || dist_scale_factor > 128)
3268                     h->implicit_weight[ref0][ref1] = 32;
3269                 else
3270                     h->implicit_weight[ref0][ref1] = 64 - dist_scale_factor;
3271             }else
3272                 h->implicit_weight[ref0][ref1] = 32;
3273         }
3274     }
3275 }
3276
3277 /**
3278  * Mark a picture as no longer needed for reference. The refmask
3279  * argument allows unreferencing of individual fields or the whole frame.
3280  * If the picture becomes entirely unreferenced, but is being held for
3281  * display purposes, it is marked as such.
3282  * @param refmask mask of fields to unreference; the mask is bitwise
3283  *                anded with the reference marking of pic
3284  * @return non-zero if pic becomes entirely unreferenced (except possibly
3285  *         for display purposes) zero if one of the fields remains in
3286  *         reference
3287  */
3288 static inline int unreference_pic(H264Context *h, Picture *pic, int refmask){
3289     int i;
3290     if (pic->reference &= refmask) {
3291         return 0;
3292     } else {
3293         if(pic == h->delayed_output_pic)
3294             pic->reference=DELAYED_PIC_REF;
3295         else{
3296             for(i = 0; h->delayed_pic[i]; i++)
3297                 if(pic == h->delayed_pic[i]){
3298                     pic->reference=DELAYED_PIC_REF;
3299                     break;
3300                 }
3301         }
3302         return 1;
3303     }
3304 }
3305
3306 /**
3307  * instantaneous decoder refresh.
3308  */
3309 static void idr(H264Context *h){
3310     int i;
3311
3312     for(i=0; i<16; i++){
3313         if (h->long_ref[i] != NULL) {
3314             unreference_pic(h, h->long_ref[i], 0);
3315             h->long_ref[i]= NULL;
3316         }
3317     }
3318     h->long_ref_count=0;
3319
3320     for(i=0; i<h->short_ref_count; i++){
3321         unreference_pic(h, h->short_ref[i], 0);
3322         h->short_ref[i]= NULL;
3323     }
3324     h->short_ref_count=0;
3325 }
3326
3327 /* forget old pics after a seek */
3328 static void flush_dpb(AVCodecContext *avctx){
3329     H264Context *h= avctx->priv_data;
3330     int i;
3331     for(i=0; i<16; i++) {
3332         if(h->delayed_pic[i])
3333             h->delayed_pic[i]->reference= 0;
3334         h->delayed_pic[i]= NULL;
3335     }
3336     if(h->delayed_output_pic)
3337         h->delayed_output_pic->reference= 0;
3338     h->delayed_output_pic= NULL;
3339     idr(h);
3340     if(h->s.current_picture_ptr)
3341         h->s.current_picture_ptr->reference= 0;
3342 }
3343
3344 /**
3345  * Find a Picture in the short term reference list by frame number.
3346  * @param frame_num frame number to search for
3347  * @param idx the index into h->short_ref where returned picture is found
3348  *            undefined if no picture found.
3349  * @return pointer to the found picture, or NULL if no pic with the provided
3350  *                 frame number is found
3351  */
3352 static Picture * find_short(H264Context *h, int frame_num, int *idx){
3353     MpegEncContext * const s = &h->s;
3354     int i;
3355
3356     for(i=0; i<h->short_ref_count; i++){
3357         Picture *pic= h->short_ref[i];
3358         if(s->avctx->debug&FF_DEBUG_MMCO)
3359             av_log(h->s.avctx, AV_LOG_DEBUG, "%d %d %p\n", i, pic->frame_num, pic);
3360         if(pic->frame_num == frame_num) {
3361             *idx = i;
3362             return pic;
3363         }
3364     }
3365     return NULL;
3366 }
3367
3368 /**
3369  * Remove a picture from the short term reference list by its index in
3370  * that list.  This does no checking on the provided index; it is assumed
3371  * to be valid. Other list entries are shifted down.
3372  * @param i index into h->short_ref of picture to remove.
3373  */
3374 static void remove_short_at_index(H264Context *h, int i){
3375     assert(i > 0 && i < h->short_ref_count);
3376     h->short_ref[i]= NULL;
3377     if (--h->short_ref_count)
3378         memmove(&h->short_ref[i], &h->short_ref[i+1], (h->short_ref_count - i)*sizeof(Picture*));
3379 }
3380
3381 /**
3382  *
3383  * @return the removed picture or NULL if an error occurs
3384  */
3385 static Picture * remove_short(H264Context *h, int frame_num){
3386     MpegEncContext * const s = &h->s;
3387     Picture *pic;
3388     int i;
3389
3390     if(s->avctx->debug&FF_DEBUG_MMCO)
3391         av_log(h->s.avctx, AV_LOG_DEBUG, "remove short %d count %d\n", frame_num, h->short_ref_count);
3392
3393     pic = find_short(h, frame_num, &i);
3394     if (pic)
3395         remove_short_at_index(h, i);
3396
3397     return pic;
3398 }
3399
3400 /**
3401  * Remove a picture from the long term reference list by its index in
3402  * that list.  This does no checking on the provided index; it is assumed
3403  * to be valid. The removed entry is set to NULL. Other entries are unaffected.
3404  * @param i index into h->long_ref of picture to remove.
3405  */
3406 static void remove_long_at_index(H264Context *h, int i){
3407     h->long_ref[i]= NULL;
3408     h->long_ref_count--;
3409 }
3410
3411 /**
3412  *
3413  * @return the removed picture or NULL if an error occurs
3414  */
3415 static Picture * remove_long(H264Context *h, int i){
3416     Picture *pic;
3417
3418     pic= h->long_ref[i];
3419     if (pic)
3420         remove_long_at_index(h, i);
3421
3422     return pic;
3423 }
3424
3425 /**
3426  * print short term list
3427  */
3428 static void print_short_term(H264Context *h) {
3429     uint32_t i;
3430     if(h->s.avctx->debug&FF_DEBUG_MMCO) {
3431         av_log(h->s.avctx, AV_LOG_DEBUG, "short term list:\n");
3432         for(i=0; i<h->short_ref_count; i++){
3433             Picture *pic= h->short_ref[i];
3434             av_log(h->s.avctx, AV_LOG_DEBUG, "%d fn:%d poc:%d %p\n", i, pic->frame_num, pic->poc, pic->data[0]);
3435